Vectors, Matrices, and Least Squares

(Working Title)

S. Boyd and L. Vandenberghe

DRAFT December 27, 2016

Contents

Preface vii

I Vectors 1
1 Vectors 3
1.1 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Vector addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Scalar-vector multiplication . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Inner product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Complexity of vector computations . . . . . . . . . . . . . . . . . . . . 21

2 Linear functions 25
2.1 Linear functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Taylor approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Regression model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3 Norm and distance 37

3.1 Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Standard deviation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4 Clustering 55
4.1 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 A clustering objective . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3 The k-means algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

5 Linear independence 73
5.1 Linear dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.3 Orthonormal vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.4 Gram-Schmidt algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 80
iv Contents

II Matrices 87
6 Matrices 89
6.1 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2 Zero and identity matrices . . . . . . . . . . . . . . . . . . . . . . . . 95
6.3 Transpose and addition . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.4 Matrix-vector multiplication . . . . . . . . . . . . . . . . . . . . . . . . 99
6.5 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7 Matrix examples 105

7.1 Geometric transformations . . . . . . . . . . . . . . . . . . . . . . . . 105
7.2 Selectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.3 Incidence matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.4 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

8 Linear equations 119

8.1 Linear and affine functions . . . . . . . . . . . . . . . . . . . . . . . . 119
8.2 Linear function models . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.3 Systems of linear equations . . . . . . . . . . . . . . . . . . . . . . . . 124

9 Linear dynamical systems 131

9.1 Linear dynamical systems . . . . . . . . . . . . . . . . . . . . . . . . . 131
9.2 Population dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.3 Epidemic dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
9.4 Motion of a mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
9.5 Supply chain dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 140

10 Matrix multiplication 143

10.1 Matrix-matrix multiplication . . . . . . . . . . . . . . . . . . . . . . . 143
10.2 Composition of linear functions . . . . . . . . . . . . . . . . . . . . . . 149
10.3 Matrix power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
10.4 QR factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

11 Matrix inverses 157

11.1 Left and right inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11.2 Inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
11.3 Solving linear equations . . . . . . . . . . . . . . . . . . . . . . . . . . 165
11.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.5 Pseudo-inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

III Least squares 175

12 Least squares 177
12.1 Least squares problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
12.2 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
12.3 Solving least squares problems . . . . . . . . . . . . . . . . . . . . . . 183
12.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Contents v

13 Least squares data fitting 189

13.1 Least squares data fitting . . . . . . . . . . . . . . . . . . . . . . . . . 189
13.2 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
13.3 Feature engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

14 Least squares classification 221

14.1 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
14.2 Least squares classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
14.3 Multi-class classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

15 Multi-objective least squares 241

15.1 Multi-objective least squares . . . . . . . . . . . . . . . . . . . . . . . 241
15.2 Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
15.3 Estimation and inversion . . . . . . . . . . . . . . . . . . . . . . . . . 248
15.4 Regularized data fitting . . . . . . . . . . . . . . . . . . . . . . . . . . 259
15.5 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

16 Constrained least squares 265

16.1 Constrained least squares problem . . . . . . . . . . . . . . . . . . . . 265
16.2 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
16.3 Solving constrained least squares problems . . . . . . . . . . . . . . . . 273

17 Constrained least squares applications 279

17.1 Portfolio optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
17.2 Linear quadratic control . . . . . . . . . . . . . . . . . . . . . . . . . . 288
17.3 Linear quadratic state estimation . . . . . . . . . . . . . . . . . . . . . 294

18 Nonlinear least squares 301

18.1 Nonlinear equations and least squares . . . . . . . . . . . . . . . . . . 301
18.2 Gauss-Newton algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 306
18.3 Levenberg-Marquardt algorithm . . . . . . . . . . . . . . . . . . . . . . 311
18.4 Nonlinear model fitting . . . . . . . . . . . . . . . . . . . . . . . . . . 318
18.5 Nonlinear least squares classification . . . . . . . . . . . . . . . . . . . 320

19 Constrained nonlinear least squares 331

19.1 Constrained nonlinear least squares . . . . . . . . . . . . . . . . . . . . 331
19.2 Penalty algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
19.3 Augmented Lagrangian algorithm . . . . . . . . . . . . . . . . . . . . . 334
19.4 Nonlinear control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337

Appendices 347
A Notation 349

B Complexity 351
vi Contents

C Derivatives and optimization 353

C.1 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
C.2 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
C.3 Lagrange multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358

D Whats next 361

Index 363
Preface

This book is meant to provide a basic introduction to vectors, matrices, and least
squares methods, with a focus on applications. Vectors, matrices, and least squares
are basic topics in applied linear algebra. They are now used in a very wide range
of applications, including data fitting, machine learning and artificial intelligence,
tomography, navigation, image processing, finance, and control, to name just a few.
Our goal is to give the beginning student, with little or no prior exposure to linear
algebra, a good grounding in the basic ideas, as well as an appreciation for how
they are used in many applications.
The background required of the reader is familiarity with basic mathematical
notation. We use calculus in just a few places, but it does not play a critical
role and is not a strict prerequisite. Even though the book covers many topics
that are traditionally taught as part of probability and statistics, such as fitting
mathematical models to data, no knowledge of or background in probability and
statistics is needed.
The book covers less mathematics than a typical one on applied linear algebra.
We use only one theoretical concept from linear algeba, linear independence, and
only one computational too, the QR factorization; our approach to most applica-
tions relies on only one method, least squares (or some extension). In this sense
the book aims for intellectual economy: We cover many applications with just a
few basic ideas, concepts, and methods. The mathematics we do present, however,
is complete, in that we carefully justify every mathematical statement. In contrast
to most other applied linear algebra books, however, we describe many applica-
tions, including some that are typically considered advanced topics, like document
classification, control, state estimation, and portfolio optimization.
The book does not require any knowledge of computer programming, and can
be used as a conventional textbook, by reading the chapters and working the ex-
ercises. This approach misses out on one of the most compelling reasons to learn
the material: You can actually use the ideas and methods described in this book
to do practical things like build a prediction model from data, enhance images, or
optimize an investment portfolio. The growing power of computers, together with
the development of high level computer languages and packages that support vector
and matrix computation, have made it easy to use the methods described in this
book for real applications. We hope that every student of this book will comple-
ment their reading with computer programming exercises and projects, including
some that involve real data and problems.
If you read the whole book, work some of the exercises, and carry out computer
viii Preface

exercises to implement or use the ideas and methods, you will learn quite a lot.
While there will still be much for you to learn, you will know many of the basic
ideas behind modern data science and many other application areas, and you will
be empowered to use the methods for your own applications.
The book is structured into three parts. Part I introduces the reader to vectors,
and various vector operations and functions like addition, inner product, distance,
and angle. We also describe how vectors are used to represent word counts in a
document, time series, attributes of a patient, sales of a product, an audio track,
or an image. Part II does the same for matrices, culminating with matrix inverses.
Part III, on least squares, is the payoff, at least in terms of the applications. We
show how the simple and natural idea of approximately solving a set of over-
determined equations, and a few extensions of this basic idea, can be used to solve
a wide range of practical problems.
The whole book can be covered in a 15 week (semester) course; a 10 week
(quarter) course can cover most of the material, by skipping a few applications and
perhaps the last two chapters on nonlinear least squares.

Stephen Boyd Stanford, California

Lieven Vandenberghe Los Angeles, California
 Part I

Vectors
 Chapter 1

Vectors

In this chapter we introduce vectors and some common operations on them. We

describe some settings in which vectors are used.

1.1 Vectors
A vector is an ordered finite list of numbers. Vectors are usually written as vertical
arrays, surrounded by square or curved brackets, as in

1.1 1.1
0.0 0.0

3.6
or 3.6 .

7.2 7.2

They can also be written as numbers separated by commas and surrounded by

parentheses. In this notation style, the vector above is written as

(1.1, 0, 3.6, 7.2).

The elements (or entries, coefficients, components) of a vector are the values in the
array. The size (also called dimension or length) of the vector is the number of
elements it contains. The vector above, for example, has size four; its third entry
is 3.6. A vector of size n is called an n-vector. A 1-vector is considered to be the
same as a number, i.e., we do not distinguish between the 1-vector [ 1.3 ] and the
number 1.3.
We often use symbols to denote vectors. If we denote an n-vector using the
symbol a, the ith element of the vector a is denoted ai , where the subscript i is an
integer index that runs from 1 to n, the size of the vector.
Two vectors a and b are equal, which we denote a = b, if they have the same
size, and each of the corresponding entries is the same. If a and b are n-vectors,
then a = b means a1 = b1 , . . . , an = bn .
4 1 Vectors

The numbers or values of the elements in a vector are called scalars. We will
focus on the case that arises in most applications, where the scalars are real num-
bers. In this case we refer to vectors as real vectors. (Occasionally other types of
scalars arise, for example, complex numbers, in which case we refer to the vector
as a complex vector.) The set of all real numbers is written as R, and the set of all
real n-vectors is denoted Rn , so a Rn is another way to say that a is an n-vector
with real entries. Here we use set notation: a Rn means that a is an element of
the set Rn ; see appendix A.

Block or stacked vectors. It is sometimes useful to define vectors by concatenat-

ing or stacking two or more vectors, as in

b
a = c ,
d

where a, b, c, and d are vectors. If b is an m-vector, c is an n-vector, and d is a

p-vector, this defines the (m + n + p)-vector

a = (b1 , b2 , . . . , bm , c1 , c2 , . . . , cn , d1 , d2 , . . . , dp ).

The stacked vector a is also written as a = (b, c, d).

Stacked vectors can include scalars (numbers). For example if a is a 3-vector,
(1, a) is the 4-vector (1, a1 , a2 , a3 ).

Subvectors. In the equation above, we say that b, c, and d are subvectors or slices
of a, with sizes m, n, and p, respectively. One notation used to denote subvectors
uses colon notation. If a is a vector, then ar:s is the vector of size s r + 1, with
entries ar , . . . , as :
ar:s = (ar , . . . , as ).
The subscript r : s is called the index range. Thus, in our example above, we have

b = a1:m , c = am+1:m+n , d = am+n+1:m+n+p .

Colon notation is not completely standard, but it is growing in popularity.

Notational conventions. Some authors try to use notation that helps the reader
distinguish between vectors and scalars (numbers). For example, Greek letters
(, , . . . ) might be used for numbers, and lower-case letters (a, x, f , . . . ) for
vectors. Other notational conventions include vectors given in bold font (g), or
vectors written with arrows above them (~a). These notational conventions are not
standardized, so you should be prepared to figure out what things are (i.e., scalars
or vectors) despite the authors notational scheme (if any exists).

Indexing. We should also give a couple of warnings concerning the subscripted

index notation ai . The first warning concerns the range of the index. In many
computer languages, arrays of length n are indexed from i = 0 to n 1. But in
1.1 Vectors 5

standard mathematical notation, n-vectors are indexed from i = 1 to i = n, so in

this book, vectors will be indexed from i = 1 to i = n.
The next warning concerns an ambiguity in the notation ai , used for the ith
element of a vector a. The same notation will occasionally refer to the ith vector
in a collection or list of k vectors a1 , . . . , ak . Whether a3 means the third element
of a vector a (in which case a3 is a number), or the third vector in some list of
vectors (in which case a3 is a vector) should be clear from the context. When we
need to refer to an element of a vector that is in an indexed collection of vectors,
we can write (ai )j to refer to the jth entry of ai , the ith vector in our list.

Zero vectors. A zero vector is a vector with all elements equal to zero. Sometimes
the zero vector of size n is written as 0n , where the subscript denotes the size.
But usually a zero vector is denoted just 0, the same symbol used to denote the
number 0. In this case you have to figure out the size of the zero vector from the
context. (We will see how this is done later.)
Even though zero vectors of different sizes are different vectors, we use the same
symbol 0 to denote them. In computer programming this is called overloading: the
symbol 0 is overloaded because it can mean different things depending on the
context (e.g., the equation it appears in).

Unit vectors. A (standard) unit vector is a vector with all elements equal to zero,
except one element which is equal to one. The ith unit vector (of size n) is the
unit vector with ith element one, and is denoted ei . For example, the vectors

1 0 0
e1 = 0 , e2 = 1 , e3 = 0
0 0 1

are the three unit vectors of size 3. The notation for unit vectors is an example of
the ambiguity in notation noted above. Here, ei denotes the ith unit vector, and
not the ith element of a vector e. Thus we can describe the ith unit n-vector ei as

1 j=i
(ei )j =
0 j 6= i,

for i, j = 1, . . . , n. On the left-hand side ei is an n-vector; (ei )j is a number, its jth

entry. As with zero vectors, the size of ei is usually determined from the context.

Ones vector. We use the notation 1n for the n-vector with all its elements equal
to one. We also write 1 if the size of the vector can be determined from the
context. (Some authors use e to denote a vector of all ones, but we will not use
this notation.) The vector 1 is sometimes called the ones vector.

Sparsity. A vector is said to be sparse if many of its entries are zero; its sparsity
pattern is the set of indices of nonzero entries. The number of the nonzero entries
of an n-vector x is denoted nnz(x). Unit vectors are sparse, since they have only
one nonzero entry. The zero vector is the sparsest possible vector, since it has no
nonzero entries. Sparse vectors arise in many applications.
6 1 Vectors

x2 x

x1

Figure 1.1 The 2-vector x specifies the position (shown as a dot) with coor-
dinates x1 and x2 in a plane.

x
x2

x1

Figure 1.2 The 2-vector x represents a displacement in the plane (shown as

an arrow) by x1 in the first axis and x2 in the second.

Examples
An n-vector can be used to represent n quantities or values in an application. In
some cases the values are similar in nature (for example, they are given in the same
physical units); in others, the quantities represented by the entries of the vector are
quite different from each other. We briefly describe below some typical examples,
many of which we will see throughout the book.

Location and displacement. A 2-vector can be used to represent a position or

location in a 2-dimensional (2-D) space, i.e., a plane, as shown in figure 1.1. A
3-vector is used to represent a location or position of some point in 3-dimensional
(3-D) space. The entries of the vector give the coordinates of the position or
location.
A vector can also be used to represent a displacement in a plane or 3-D space,
in which case it is typically drawn as an arrow, as shown in figure 1.2. A vector can
also be used to represent the velocity or acceleration, at a given time, of a point
that moves in a plane or 3-D space.

Color. A 3-vector can represent a color, with its entries giving the Red, Green,
and Blue (RGB) intensity values (often between 0 and 1). The vector (0, 0, 0)
represents black, the vector (0, 1, 0) represents a bright pure green color, and the
vector (1, 0.5, 0.5) represents a shade of pink. This is illustrated in figure 1.3.
1.1 Vectors 7

(1, 0, 0) (0, 1, 0) (0, 0, 1)

(1, 1, 0) (1, 0.5, 0.5) (0.5, 0.5, 0.5)

Figure 1.3 Six colors and their RGB vectors.

Quantities. An n-vector q can represent the amounts or quantities of n different

resources or products held (or produced, or required) by an entity such as a com-
pany. Negative entries mean an amount of the resource owed to another party (or
consumed, or to be disposed of). For example, a bill of materials is a vector that
gives the amounts of n resources required to create a product or carry out a task.

Portfolio. An n-vector s can represent a stock portfolio or investment in n dif-

ferent assets, with si giving the number of shares of asset i held. The vector
(100, 50, 20) represents a portfolio consisting of 100 shares of asset 1, 50 shares of
asset 2, and 20 shares of asset 3. Short positions (i.e., shares that you owe another
party) are represented by negative entries in a portfolio vector. The entries of the
portfolio vector can also be given in dollar values, or fractions of the total dollar
amount invested.

Values across a population. An n-vector can give the values of some quantity
across a population of individuals or entities. For example, an n-vector b can give
the blood pressure of a collection of n patients, with bi the blood pressure of patient
i, for i = 1, . . . , n.

Proportions. A vector w can be used to give fractions or proportions out of n

choices, outcomes, or options, with wi the fraction with choice or outcome i. In
this case the entries are nonnegative and add up to one. Such vectors can also be
interpreted as the recipes for a mixture of n items, an allocation across n entities,
or as probability values in a probability space with n outcomes. For example, a
uniform mixture of 4 outcomes is represented as the 4-vector (1/4, 1/4, 1/4, 1/4).

Time series. An n-vector can represent a time series or signal, that is, the value
of some quantity at different times. (The entries in a vector that represents a time
series are sometimes called samples, especially when the quantity is something
measured.) An audio (sound) signal can be represented as a vector whose entries
8 1 Vectors

90

85

80

xi ( F) 75

70

65
0 10 20 30 40 50
i

Figure 1.4 Hourly temperature in downtown Los Angeles on August 5 and

6, 2015 (starting at 12:47AM, ending at 11:47PM).

give the value of acoustic pressure at equally spaced times (typically 48000 or 44100
per second). A vector might give the hourly rainfall (or temperature, or barometric
pressure) at some location, over some time period. When a vector represents a time
series, it is natural to plot xi versus i with lines connecting consecutive time series
values. (These lines carry no information; they are added only to make the plot
easier to understand visually.) An example is shown in figure 1.4, where the 48-
vector x gives the hourly temperature in downtown Los Angeles.

Daily return. A vector can represent the daily return of a stock, i.e., its fractional
increase (or decrease if negative) in value over the day. For example the return time
series vector (0.022, +0.014, +0.004) means the stock price went down 2.2% on
the first day, then up 1.4% the next day, and up again 0.4% on the third day. In
this example, the samples are not uniformly spaced in time; the index refers to
trading days, and does not include weekends or market holidays. A vector can
represent the daily (or quarterly, hourly, or minute-by-minute) value of any other
quantity of interest for an asset, such as price or volume.

Cash flow. A cash flow into and out of an entity (say, a company) can be repre-
sented by a vector, with positive representing payments to the entity, and negative
representing payment by the entity. For example, with entries giving cash flow
each quarter, the vector (1000, 10, 10, 10, 1010) represents a one year loan of
$1000, with 1% interest only payments made each quarter, and the principal and
last interest payment at the end.
1.1 Vectors 9

0.65 0.05 0.20

0.28 0.00 0.90

Figure 1.5 8 8 image and the grayscale levels at six pixels.

Images. A monochrome (black and white) image is an array of M N pixels

(square patches with uniform grayscale level) with M rows and N columns. Each
of the M N pixels has a grayscale or intensity value, with 0 corresponding to black
and 1 corresponding to bright white. (Other ranges are also used.) An image can
be represented by a vector of length M N , with the elements giving grayscale levels
at the pixel locations, typically ordered column-wise or row-wise.
Figure 1.5 shows a simple example, an 88 image. (This is a very low resolution;
typical values of M and N are in the hundreds or thousands.) With the vector
entries arranged row-wise, the associated 64-vector is

x = (0.65, 0.05, 0.20, . . . , 0.28, 0.00, 0.90).

A color M N pixel image is described by a vector of length 3M N , with the

entries giving the R, G, and B values for each pixel, in some agreed-upon order.

Video. A monochrome video, i.e., a time series of length K images of M N

pixels, can be represented by a vector of length KM N (again, in some particular
order).

Word count and histogram. A vector of length n can represent the number of
times each word in a dictionary of n words appears in a document. For example,
(25, 2, 0) means that the first dictionary word appears 25 times, the second one
twice, and the third one not at all. (Typical dictionaries used for document word
counts have many more than 3 elements.) A small example is shown in figure 1.6. A
variation is to have the entries of the vector give the histogram of word frequencies
in the document, so that, e.g., x5 = 0.003 means that 0.3% of all the words in the
document are the fifth word in the dictionary.
It is common practice to count variations of a word (say, the same word stem
with different endings) as the same word; for example, rain, rains, raining, and
10 1 Vectors

Word count vectors are used in computer based document analysis.

Each entry of the word count vector is the number of times the as-
sociated dictionary word appears in the document.

word 3
in
2

number
1

horse
0

the 4
document 1

Figure 1.6 A snippet of text (top), the dictionary (bottom left), and word
count vector (bottom right).

rained might all be counted as rain. Reducing each word to its stem is called
stemming. It is also common practice to exclude words that are too common (such
as a or the) or extremely rare. These are referred to as stop words.

Customer purchases. An n-vector p can be used to record a particular customers

purchases from some business over some period of time, with pi the quantity of
item i the customer has purchased, for i = 1, . . . , n. (Unless n is small, we would
expect many of these entries to be zero, meaning the customer has not purchased
those items.) In one variation, pi represents the total dollar value of item i the
customer has purchased.

Occurrence or subsets. An n-vector o can be used to record whether or not each

of n different events has occurred, with oi = 0 meaning that event i did not occur,
and oi = 1 meaning that it did occur. Such a vector encodes a subset of a collection
of n objects, with oi = 1 meaning that object i is contained in the subset, and oi = 0
meaning that object i is not in the subset.

Features or attributes. In many applications a vector collects together n different

quantities that pertain to a single thing or object. The quantities can be measure-
ments, or quantities that can be measured or derived from the object. Such a
vector is sometimes called a feature vector, and its entries are called the features
or attributes. For example, a 6-vector f could give the age, height, weight, blood
pressure, temperature, and gender of a patient admitted to a hospital. (The last
entry of the vector could be encoded as f6 = 0 for Male, f6 = 1 for Female.) In this
example, the quantities represented by the entries of the vector are quite different,
with different physical units.

Vector entry labels. In applications such as the ones described above each entry
of a vector has a meaning, such as the count of a specific word in a document, the
number of shares of a specific stock held in a portfolio, or the rainfall in a specific
 1.2 Vector addition 11

hour. It is common to keep a separate list of labels or tags that explain or annotate
the meaning of the vector entries. As an example, we might associate the portfolio
vector (100, 50, 20) with the list of ticker symbols (AAPL, INTC, AMZN), so we
know that assets 1, 2, and 3 are Apple, Intel, and Amazon. In some applications,
such as an image, the meaning or ordering of the entries follow known conventions
or standards.

1.2 Vector addition

Two vectors of the same size can be added together by adding the corresponding
elements, to form another vector of the same size, called the sum of the vectors.
Vector addition is denoted by the symbol +. (Thus the symbol + is overloaded
to mean scalar addition when scalars appear on its left- and right-hand sides, and
vector addition when vectors appear on its left- and right-hand sides.) For example,

0 1 1
7 + 2 = 9 .
3 0 3

Vector subtraction is similar. As an example,

     
1 1 0
= .
9 1 8

The result of vector subtraction is called the difference of the two vectors.

Properties. Several properties of vector addition are easily verified. For any vec-
tors a, b, and c of the same size we have the following.
Vector addition is commutative: a + b = b + a.
Vector addition is associative: (a + b) + c = a + (b + c). We can therefore
write both as a + b + c.
a + 0 = 0 + a = a. Adding the zero vector to a vector has no effect. (This
is an example where the size of the zero vector follows from the context: It
must be the same as the size of a.)
a a = 0. Subtracting a vector from itself yields the zero vector. (Here too
the size of 0 is the size of a.)
To show that these properties hold, we argue using the definition of vector
addition and vector equality. As an example, let us show that for any n-vectors a
and b, we have a + b = b + a. The ith entry of a + b is, by the definition of vector
addition, ai + bi . The ith entry of b + a is bi + ai . For any two numbers we have
ai + bi = bi + ai , so the ith entries of the vectors a + b and b + a are the same.
This is true for all of the entries, so by the definition of vector equality, we have
a + b = b + a.
12 1 Vectors

a
a+b
b
b b+a

Figure 1.7 Left. The lowest dark arrow shows the displacement a; the dis-
placement b, shown as a dashed arrow, starts from the head of the displace-
ment a and ends at the sum displacement a + b, shown as the longer dark
arrow. Right. The displacement b + a.

p+a

Figure 1.8 The vector p + a is the position of the point represented by p

displaced by the displacement represented by a.

Verifying identities like the ones above, and many others we will encounter
later, can be tedious. But it is important to understand that the various properties
we will list can be derived using elementary arguments like the one above. We
recommend that the reader select a few of the properties we will see, and attempt
to derive them, just to see that it can be done. (Deriving all of them is overkill.)

Examples.

Displacements. When vectors a and b represent displacements, the sum a + b

is the net displacement found by first displacing by a, then displacing by b,
as shown in figure 1.7. Note that we arrive at the same vector if we first
displace by b and then a. If the vector p represents a position and the vector
a represents a displacement, then p+a is the position of the point p, displaced
by a, as shown in figure 1.8.

Displacements between two points. If the vectors p and q represent the posi-
tions of two points in 2-D or 3-D space, then p q is the displacement vector
from q to p, as illustrated in figure 1.9.

Word counts. If a and b are word count vectors (using the same dictionary)
1.2 Vector addition 13

pq p

Figure 1.9 The vector p q represents the displacement from the point
represented by q to the point represented by p.

for two documents, the sum a + b is the word count vector of a new document
created by combining the original two (in either order). The word count
difference vector a b gives the number of times more each word appears in
the first document than the second.
Bill of materials. Suppose q1 , . . . , qN are n-vectors that give the quantities of
n different resources required to accomplish N tasks. Then the sum n-vector
q1 + + qN gives the bill of materials for completing all N tasks.
Market clearing. Suppose the n-vector qi represents the amounts of n goods
or resources produced (when positive) or consumed (when negative) by agent
i, for i = 1, . . . , N , so (q5 )4 = 3.2 means that agent 5 consumes 3.2 units of
resource 4. The sum s = q1 + + qN is the n-vector of total net surplus of
the resources (or shortfall, when the entries are negative). When s = 0, we
have a closed market, which means that the total amount of each resource
produced by the agents balances the total amount consumed. In other words,
the n resources are exchanged among the agents. In this case we say that the
market clears (with the resource vectors q1 , . . . , qN ).
Audio addition. When a and b are vectors representing audio signals over
the same period of time, the sum a + b is an audio signal that is perceived as
containing both audio signals combined into one. If a represents a recording of
a voice, and b a recording of music (of the same length), the audio signal a + b
will be perceived as containing both the voice recording and, simultaneously,
the music.
Feature differences. If f and g are n-vectors that give n feature values for two
items, the difference vector d = f g gives the difference in feature values for
the two objects. For example, d7 = 0 means that the two objects have the
same value for feature 7; d3 = 1.67 means that the first objects third feature
value exceeds the second objects third feature value by 1.67.
Time series. If a and b represent time series of the same quantity, such as
daily profit at two different stores, then a + b represents a time series which is
the total daily profit at the two stores. An example (with monthly rainfall)
is shown in figure 1.10.
14 1 Vectors

8
Los Angeles
San Francisco
Sum
6

Rainfall (inches)
4

1 2 3 4 5 6 7 8 9 10 11 12
k

Figure 1.10 Average monthly rainfall in inches measured in downtown Los

Angeles and San Francisco International Airport, and their sum. Averages
are 30-year averages (1981-2010).

Portfolio trading. Suppose s is an n-vector giving the number of shares of n

assets in a portfolio, and b is an n-vector giving the number of shares of the
assets that we buy (when bi is positive) or sell (when bi is negative). After
the asset purchases and sales, our portfolio is given by s + b, the sum of the
original portfolio vector and the purchase vector b, which is also called the
trade vector or trade list. (The same interpretation works when the portfolio
and trade vectors are given in dollar value.)

Addition notation in computer languages. Some computer languages for manip-

ulating vectors define the sum of a vector and a scalar, as the vector obtained by
adding the scalar to each element of the vector. This is not standard mathematical
notation, however, so we will not use it. Even more confusing, in some computer
languages the plus symbol is used to denote concatenation of arrays, which means
putting one array after another, as in (1, 2) + (3, 4, 5) = (1, 2, 3, 4, 5). While this
notation might give a valid expression in some computer languages, it is not stan-
dard mathematical notation, and we will not use it in this book. In general, it is
very important to distinguish between mathematical notation for vectors (which
we use) and the syntax of specific computer languages or software packages for
manipulating vectors.
 1.3 Scalar-vector multiplication 15

1.3 Scalar-vector multiplication

Another operation is scalar multiplication or scalar-vector multiplication, in which
a vector is multiplied by a scalar (i.e., number), which is done by multiplying
every element of the vector by the scalar. Scalar multiplication is denoted by
juxtaposition, typically with the scalar on the left, as in

1 2
(2) 9 = 18 .
6 12

Scalar-vector multiplication can also be written with the scalar on the right, as in

1 1.5
9 (1.5) = 13.5 .
6 9

The meaning is the same: It is the vector obtained by multiplying each element
by the scalar. A similar notation is a/2, where a is a vector, meaning (1/2)a. The
scalar-vector product (1)a is written simply as a. Note that 0 a = 0 (where the
left-hand zero is the scalar zero, and the right-hand zero is a vector zero of the
same size as a).

Properties. By definition, we have a = a, for any scalar and any vector a.

This is called the commutative property of scalar-vector multiplication; it means
that scalar-vector multiplication can be written in either order.
Scalar multiplication obeys several other laws that are easy to figure out from
the definition. For example, it satisfies the associative property: If a is a vector
and and are scalars, we have

()a = (a).

On the left-hand side we see scalar-scalar multiplication () and scalar-vector

multiplication; on the right we see two scalar-vector products. As a consequence,
we can write the vector above as a, since it does not matter whether we interpret
this as (a) or ()a.
The associative property holds also when we denote scalar-vector multiplication
with the scalar on the right. For example, we have (a) = (a), and consequently
we can write both as a. As a convention, however, this vector is normally written
as a or as ()a.
If a is a vector and , are scalars, then

( + )a = a + a.

(This is the left-distributive property of scalar-vector multiplication.) Scalar mul-

tiplication, like ordinary multiplication, has higher precedence in equations than
vector addition, so the right-hand side here, a+a, means (a)+(a). It is useful
to identify the symbols appearing in this formula above. The + symbol on the left
is addition of scalars, while the + symbol on the right denotes vector addition.
16 1 Vectors

When scalar multiplication is written with the scalar on the right, we have the
right-distributive property:

a( + ) = a + a.

Scalar-vector multiplication also satisfies another version of the right-distributive

property:
(a + b) = a + b

for any scalar and any n-vectors a and b. In this equation, both of the + symbols
refer to the addition of n-vectors.

Linear combinations. If a1 , . . . , am are n-vectors, and 1 , . . . , m are scalars,

the n-vector
1 a1 + + m am

is called a linear combination of the vectors a1 , . . . , an . The scalars 1 , . . . , m

are called the coefficients of the linear combination.
As a simple but important application, we can write any n-vector b as a linear
combination of the standard unit vectors, as

b = b1 e 1 + + bn e n . (1.1)

In this equation bi are the entries in b (i.e., scalars), and ei is the ith unit vector.
A specific example is

1 1 0 0
3 = (1) 0 + 3 1 + 5 0 .
5 0 0 1

Special linear combinations. Some linear combinations of the vectors a1 , . . . , am

have special names. For example, the linear combination with 1 = = m = 1,
given by a1 + + am , is the sum of the vectors, and the linear combination with
1 = = m = 1/m, given by (1/m)(a1 + + am ), is the average of the vectors.
When the coefficients sum to one, i.e., 1 + + m = 1, the linear combination
is called an affine combination. When the coefficients in an affine combination are
nonnegative, it is called a convex combination or a mixture. The coefficients in an
affine or convex combination are sometimes given as percentages, which add up to
100%.

Linear combinations of linear combinations. If vectors b1 , . . . , bk are each a

linear combination of vectors a1 , . . . , am , and c is a linear combination of b1 , . . . ,
bk , then c is a linear combination of a1 , . . . , am . We will encounter this important
idea in several later chapters, and will introduce notation to describe it concisely
then.
1.3 Scalar-vector multiplication 17

1.5a
a

(1.5)a

Figure 1.11 1.5a represents the displacement in the direction of the displace-
ment a, with magnitude scaled by 1.5; (1.5)a represents the displacement
in the opposite direction, also with magnitude scaled by 1.5.

Examples.
Displacements. When a vector a represents a displacement, and > 0, a
is a displacement in the same direction of a, with its magnitude scaled by
. When < 0, a represents a displacement in the opposite direction of a,
with magnitude scaled by ||. This is illustrated in figure 1.11.
Materials requirements. Suppose the n-vector q is the bill of materials for
producing one unit of some product, i.e., qi is the amount of raw material
required to produce one unit of product. To produce units of the product
will then require raw materials given by q. (Here we assume that 0.)
Audio scaling. If a is a vector representing an audio signal, the scalar-vector
product a is perceived as the same audio signal, but changed in volume
(loudness) by the factor ||. For example, when = 1/2 (or = 1/2), a
is perceived as the same audio signal, but quieter.
Audio mixing. When a1 , . . . , am are vectors representing audio signals (over
the same period of time, for example, simultaneously recorded), they are
called tracks. The linear combination 1 a1 + + m am is perceived as a
mixture (also called a mix ) of the audio tracks, with relative loudness given
by |1 |, . . . , |m |. A producer in a studio, or a sound engineer at a live show,
chooses values of 1 , . . . , m to give a good balance between the different
instruments, vocals, and drums.
Cash flow replication. Suppose that c1 , . . . , cm are vectors that represent cash
flows, such as particular types of loans or investments. The linear combination
f = 1 c1 + + m cm represents another cash flow. We say that the cash
flow f has been replicated by the (linear combination of the) original cash
18 1 Vectors

b
= 1.2

a = 0.4

= 0.4

Figure 1.12 The affine combination (1 )a + b for different values of .

These points are on the line passing through a and b; for between 0 and
1, the points are on the line segment between a and b.

flows c1 , . . . , cm . As an example, c1 = (1, 1.1, 0) represents a $1 loan from

period 1 to period 2 with 10% interest, and c2 = (0, 1, 1.1) represents a $1
loan from period 2 to period 3 with 10% interest. The linear combination
d = c1 + 1.1c2 = (1, 0, 1.21)
represents a two period loan of $1 in period 1, with compounded 10% interest.
Here we have replicated a two period loan from two one period loans.
Line and segment. When a and b are different n-vectors, the affine combi-
nation c = (1 )a + b, where is a scalar, describes a point on the line
passing through a and b. When 0 1, c is a convex combination of a
and b, and is said to lie on the segment between a and b. For n = 2 and
n = 3, with the vectors representing coordinates of 2-D or 3-D points, this
agrees with the usual geometric notion of line and segment. But we can also
talk about the line passing through two vectors of dimension 100. This is
illustrated in figure 1.12.

1.4 Inner product

The (standard) inner product (also called dot product) of two n-vectors is defined
as the scalar
aT b = a1 b1 + a2 b2 + + an bn ,
the sum of the products of corresponding entries. (The origin of the superscript
in the inner product notation aT b will be explained in chapter 6.) Some other
notations for the inner product (that we will not use in this book) are ha, bi, ha|bi,
(a, b), and a b. (In the notation used in this book, (a, b) denotes a stacked vector
of length 2n.) As you might guess, there is also a vector outer product, which we
will encounter later, in 10.1. As a specific example of the inner product, we have
T
1 1
2 0 = 7.
2 3
1.4 Inner product 19

When n = 1, the inner product reduces to the usual product of two numbers.

Properties. The inner product satisfies some simple properties that are easily
verified from the definition. If a, b, and c are vectors of the same size, and is a
scalar, we have the following.
Commutativity. aT b = bT a. The order of the two vector arguments in the
inner product does not matter.
Associativity with scalar multiplication. (a)T b = (aT b), so we can write
both as aT b.
Distributivity with vector addition. (a + b)T c = aT c + bT c. The inner product
can be distributed across vector addition.
These can be combined to obtain other identities, such as aT (b) = (aT b), or
aT (b + c) = aT b + aT c. As another useful example, we have, for any vectors
a, b, c, d of the same size,
(a + b)T (c + d) = aT c + aT d + bT c + bT d.
This formula expresses an inner product on the left-hand side as a sum of four
inner products on the right-hand side, and is analogous to expanding a product of
sums in algebra. Note that on the left-hand side, the two addition symbols refer to
vector addition, whereas on the right-hand side, the three addition symbols refer
to scalar (number) addition.

General examples.
Unit vector. eTi a = ai . The inner product of a vector with the ith standard
unit vector gives (or picks out) the ith element a.
Sum. 1T a = a1 + + an . The inner product of a vector with the vector of
ones gives the sum of the elements of the vector.
Average. (1/n)T a = (a1 + + an )/n. The inner product of an n-vector with
the vector 1/n gives the average of the elements of the vector.
Sum of squares. aT a = a21 + + a2n . The inner product of a vector with
itself gives the sum of the squares of the elements of the vector.
Selective sum. Let b be a vector all of whose entries are either 0 or 1. Then
bT a is the sum of the elements in a for which bi = 1.

Block vectors. If the vectors a and b are block vectors, and the corresponding
blocks have the same sizes (in which case we say they conform), then we have
T
a1 b1
aT b = ... ... = aT1 b1 + + aTk bk .

ak bk
The inner product of block vectors is the sum of the inner products of the blocks.
20 1 Vectors

A B
6
2
3
1
7 5
4

Figure 1.13 Two sets A and B, containing seven objects.

Applications. The inner product is useful in many applications, a few of which

we list here.
Co-occurrence. If a and b are n-vectors that describe occurrence, i.e., each
of their elements is either 0 or 1, then aT b gives the total number of indices
for which ai and bi are both one, that is, the total number of co-occurrences.
If we interpret the vectors a and b as describing subsets of n objects, then
aT b gives the number of objects in the intersection of the two subsets. This
is illustrated in figure 1.13, for two subsets A and B of 7 objects, labeled
1, . . . , 7, with corresponding occurrence vectors
a = (0, 1, 1, 1, 1, 1, 1), b = (1, 0, 1, 0, 1, 0, 0).
Here we have aT b = 2, which is the number of objects in both A and B (i.e.,
objects 3 and 5).
Weights, features, and score. When the vector f represents a set of features of
an object, and w is a vector of the same size (often called a weight vector ), the
inner product wT f is the sum of the feature values, scaled (or weighted) by
the weights, and is sometimes called a score. For example, if the features are
associated with a loan applicant (e.g., age, income, . . . ) we might interpret
s = wT f as a credit score. In this example we can interpret wi as the weight
given to feature i in forming the score.
Price-quantity. If p represents a vector of prices of n goods, and q is a vector
of quantities of the n goods (say, the bill of materials for a product), then
their inner product pT q is the total cost of the goods given by the vector q.
Speed-time. Suppose the vector s gives the speed of a vehicle traveling over
n segments, with t a vector of the time taken to traverse the segments. Then
sT t is the total distance travelled.
Probability and expected values. Suppose the n-vector p has nonnegative
entries that sum to one, so it describes a set of proportions among n items,
or a set of probabilities of n outcomes, one of which must occur. Suppose
f is another n-vector, where we interpret fi as the value of some quantity
if outcome i occurs. Then f T p gives the expected value or average of the
quantity, under the probabilities (or fractions) given by p.
 1.5 Complexity of vector computations 21

Polynomial evaluation. Suppose the n-vector c represents the coefficients of

a polynomial p of degree n 1 or less:

p(x) = c1 + c2 x + + cn1 xn2 + cn xn1 .

Let t be a number, and let z = (1, t, t2 , . . . , tn1 ) be the n-vector of powers

of t. Then cT z = p(t), the value of the polynomial p at the point t. So the
inner product of a polynomial coefficient vector and vector of powers of a
number evaluates the polynomial at the number.

Discounted total. Let c be an n-vector representing a cash flow, with ci the

cash received (when ci > 0) in period i. Let d be the n-vector defined as

d = (1, 1/(1 + r), . . . , 1/(1 + r)n1 ),

where r 0 is an interest rate. Then

dT c = c1 + c2 /(1 + r) + + cn /(1 + r)n1

is the discounted total of the cash flow, i.e., its net present value (NPV), with
interest rate r.

Portfolio value. Suppose s is an n-vector representing the holdings in a port-

folio of n different assets (in shares, with negative meaning short positions).
If p is an n-vector giving the prices of the assets, then pT s is the total value
of the portfolio.

Portfolio return. Suppose r is the vector of (fractional) returns of n assets

over some time period, i.e., the asset relative price changes

pfinal
i pinitial
i
ri = initial
, i = 1, . . . , n,
pi

where pinitial
i and pfinal
i are the prices of asset i at the beginning and end
of the investment period. If h is an n-vector giving our portfolio, with hi
denoting the dollar value of asset i held, then the inner product rT h is the
total return of the portfolio, in dollars, over the period. If w represents the
fractional (dollar) holdings of our portfolio, then rT w gives the total return
of the portfolio. For example, if rT w = 0.09, then our portfolio return is 9%.
If we had invested $10000 initially, we would have earned $900.

1.5 Complexity of vector computations

Computer representation of vectors. Computers store real numbers using float-
ing point format, which represents a real number using a block of 64 bits (0s and
1s), or 8 bytes (groups of 8 bits). A very wide range of numbers can be represented
in floating point format, to an accuracy that is good enough for almost all practical
22 1 Vectors

applications. Integers are stored in a more compact format, and are represented
exactly. Vectors are stored as arrays of floating point numbers (or integers, when
the entries are all integers). Storing an n-vector requires 8n bytes to store. Cur-
rent memory and storage devices, with capacities measured in many gigabytes (109
bytes), can easily store vectors with dimensions in the millions or billions. Sparse
vectors are stored in a more efficient way that keeps track of indexes and values of
the nonzero entries.

Flop counts and complexity. So far we have seen only a few vector operations,
like scalar multiplication, vector addition, and the inner product. How quickly
these operations can be carried out depends very much on the computer hardware
and software, and the size of the vector.
A very rough estimate of the time required to carry out some computation, such
as an inner product, can be found by counting the total number of basic arithmetic
operations (addition, subtraction, multiplication, and division of two numbers).
Since numbers are stored in floating point format on computers, these operations
are called floating point operations, or FLOPs. This term is in such common use
that the acronym is now written in lower case letters, as flops, and the speed with
which a computer can carry out flops is expressed in Gflop/s (gigaflops per second,
i.e., billions of flops per second). Typical current values are in the range of 1
10 Gflop/s, but this can vary by several orders of magnitude. The actual time it
takes a computer to carry out some computation depends on many other factors
beyond the total number of flops required, so time estimates based on counting
flops are very crude, and are not meant to be more accurate than a factor of ten
or so. For this reason, gross approximations (such as ignoring a factor of 2) can be
used when counting the flops required in a computation.
The complexity of an operation is the number of flops required to carry it out, as
a function of the size or sizes of the input to the operation. Usually the complexity
is highly simplified, dropping terms that are small or negligible (compared to other
terms) when the sizes of the inputs are large.

Complexity of vector operations. Scalar-vector multiplication ax, where x is an

n-vector, requires n multiplications, i.e., axi for i = 1, . . . , n. Vector addition x + y
of two n-vectors takes n additions, i.e., xi + yi for i = 1, . . . , n. Computing the
inner product xT y = x1 y1 + + xn yn of two n-vectors takes 2n 1 flops, n scalar
multiplications and n1 scalar additions. So scalar multiplication, vector addition,
and the inner product of n-vectors require n, n, and 2n 1 flops, respectively.
We only need an estimate, so we simplify the last to 2n flops, and say that the
complexity of scalar multiplication, vector addition, and the inner product of n-
vectors is n, n, and 2n flops, respectively. We can guess that a 1 Gflop/s computer
can compute the inner product of two vectors of size one million in around one
thousandth of a second, but we should not be surprised if the actual time differs
by a factor of 10 from this value.
The order of the computation is obtained by ignoring any constant that multi-
plies a power of the dimension. So we say that the three vector operations scalar
multiplication, vector addition, and inner product have order n. Ignoring the fac-
tor of 2 dropped in the actual complexity of the inner product is reasonable, since
1.5 Complexity of vector computations 23

we do not expect flop counts to predict the running time with an accuracy better
than a factor of 2. The order is useful in understanding how the time to execute
the computation will scale when the size of the operands changes. An order n
computation should take around 10 times longer to carry out its computation on
an input that is 10 times bigger.

Complexity of sparse vector operations. If x is sparse, then computing ax re-

quires nnz(x) flops. If x and y are sparse, computing x + y requires no more than
min{nnz(x), nnz(y)} flops (since no arithmetic operations are required to compute
(x + y)i when either xi or yi is zero). If the sparsity patterns of x and y do not
overlap (intersect), then zero flops are needed to compute x + y. The inner product
calculation is similar: computing xT y requires no more than 2 min{nnz(x), nnz(y)}
flops. When the sparsity patterns of x and y do not overlap, computing xT y re-
quires zero flops, since xT y = 0 in this case.
24 1 Vectors
 Chapter 2

Linear functions

In this chapter we introduce linear and affine functions, and describe some common
settings where they arise, including regression models.

2.1 Linear functions

Function notation. The notation f : Rn R means that f is a function that
maps real n-vectors to real numbers, i.e., it is a scalar-valued function of n-vectors.
If x is an n-vector, then f (x), which is a scalar, denotes the value of the function f
at x. (In the notation f (x), x is referred to as the argument of the function.) We
can also interpret f as a function of n scalar arguments, the entries of the vector
argument, in which case we write f (x) as
f (x) = f (x1 , x2 , . . . , xn ).
Here we refer to x1 , . . . , xn as the arguments of f . We sometimes say that f is
real-valued, or scalar-valued, to emphasize that f (x) is a real number or scalar.
To describe a function f : Rn R, we have to specify what its value is for any
possible argument x Rn . For example, we can define a function f : R4 R by
f (x) = x1 + x2 x24
for any 4-vector x. In words, we might describe f as the sum of the first two
elements of its argument, minus the square of the last entry of the argument.
(This particular function does not depend on the third element of its argument.)
Sometimes we introduce a function without formally assigning a symbol for it,
by directly giving a formula for its value in terms of its arguments, or describing
how to find its value from its arguments. An example is the sum function, whose
value is x1 + + xn . We can give a name to the value of the function, as in
y = x1 + + xn , and say that y is a function of x, in this case, the sum of its
entries.
Many functions are not given by formulas or equations. As an example, suppose
f : R4 R is the function that gives the lift on a particular airplane, as a function
26 2 Linear functions

of the 4-vector x, where x1 is the angle of attack of the airplane (i.e., the angle
between the airplane body and its direction of motion), x2 is its air speed, x3 is
the air density, and x4 is the angle of the airplane elevator control surface.

The inner product function. Suppose a is an n-vector. We can define a scalar-

valued function f of n-vectors, given by

f (x) = aT x = a1 x1 + a2 x2 + + an xn (2.1)

for any n-vector x. This function gives the inner product of its n-dimensional
argument x with some (fixed) n-vector a. We can also think of f as forming a
weighted sum of the elements of x; the elements of a give the weights used in
forming the weighted sum.

Superposition and linearity. The inner product function f defined in (2.1) satisfies
the property

f (x + y) = aT (x + y)
= aT (x) + aT (y)
= (aT x) + (aT y)
= f (x) + f (y)

for all n-vectors x, y, and all scalars , . This property is called superposition.
A function that satisfies the superposition property is called linear. We have just
shown that the inner product with a fixed vector is a linear function.
The superposition equality

f (x + y) = f (x) + f (y) (2.2)

looks deceptively simple; it is easy to read it as just a re-arrangement of the paren-

theses and the order of a few terms. But in fact it says a lot. On the left-hand
side, the term x + y involves vector addition and scalar-vector multiplication.
On the right-hand side, f (x) + f (y) involves ordinary scalar multiplication and
scalar addition.
If a function f is linear, superposition extends to linear combinations of any
number of vectors, and not just linear combinations of two vectors: We have

f (1 x1 + + k xk ) = 1 f (x1 ) + + k f (xk ),

for any n vectors x1 , . . . , xk , and any scalars 1 , . . . , k . (This more general k-

term form of superposition reduces to the two-term form given above when k = 2.)
To see this, we note that

f (1 x1 + + k xk ) = 1 f (x1 ) + f (2 x2 + + k xk )
= 1 f (x1 ) + 2 f (x2 ) + f (3 x3 + + k xk )
..
.
= 1 f (x1 ) + + k f (xk ).
2.1 Linear functions 27

In the first line here, we apply (two-term) superposition to the argument

1 x1 + (1)(2 x2 + + k xk ),

and in the other lines we apply this recursively.

Inner product representation of a linear function. We saw above that a function

defined as the inner product of its argument with some fixed vector is linear. The
converse is also true: If a function is linear, then it can be expressed as the inner
product of its argument with some fixed vector.
Suppose f is a scalar-valued function of n-vectors, and is linear, i.e., (2.2) holds
for all n-vectors x, y, and all scalars , . Then there is an n-vector a such that
f (x) = aT x for all x. We call aT x the inner product representation of f .
To see this, we use the identity (1.1) to express an arbitrary n-vector x as
x = x1 e1 + + xn en . If f is linear, then by multi-term superposition we have

f (x) = f (x1 e1 + + xn en )
= x1 f (e1 ) + + xn f (en )
= aT x,

with a = (f (e1 ), f (e2 ), . . . , f (en )). The formula just derived,

f (x) = x1 f (e1 ) + x2 f (e2 ) + + xn f (en ) (2.3)

which holds for any linear scalar-valued function f , has several interesting impli-
cations. Suppose, for example, that the linear function f is given as a subroutine
(or a physical system) that computes (or results in the output) f (x) when we give
the argument (or input) x. Once we have found f (e1 ), . . . , f (en ), by n calls to the
subroutine (or n experiments), we can predict (or simulate) what f (x) will be, for
any vector x, using the formula (2.3).
The representation of a linear function f as f (x) = aT x is unique, which means
that there is only one vector a for which f (x) = aT x holds for all x. To see this,
suppose that we have f (x) = aT x for all x, and also f (x) = bT x for all x. Taking
x = ei , we have f (ei ) = aT ei = ai , using the formula f (x) = aT x. Using the
formula f (x) = bT x, we have f (ei ) = bT ei = bi . These two numbers must be the
same, so we have ai = bi . Repeating this argument for i = 1, . . . , n, we conclude
that the corresponding elements in a and b are the same, so a = b.

Examples.

Average. The mean or average value of an n-vector is defined as

f (x) = (x1 + x2 + + xn )/n,

and is denoted avg(x) (and sometimes x). The average of a vector is a linear
function. It can be expressed as avg(x) = aT x with

a = (1/n, . . . , 1/n) = 1/n.

28 2 Linear functions

Maximum. The maximum element of an n-vector x, f (x) = max{x1 , . . . , xn },

is not a linear function (except when n = 1). We can show this by a coun-
terexample for n = 2. Take x = (1, 1), y = (1, 1), = 1/2, = 1/2.
Then
f (x + y) = 0 6= f (x) + f (y) = 1.

Affine functions. A linear function plus a constant is called an affine function. A

function f : Rn R is affine if and only if it can be expressed as f (x) = aT x + b
for some n-vector a and scalar b, which is sometimes called the offset. For example,
the function on 3-vectors defined by

f (x) = 2.3 2x1 + 1.3x2 x3 ,

is affine, with b = 2.3, a = (2, 1.3, 1).

Any affine scalar-valued function satisfies the following variation on the super-
position property:
f (x + y) = f (x) + f (y),
for all n-vectors x, y, and all scalars , that satisfy + = 1. For linear functions,
superposition holds for any coefficients and ; for affine functions, it holds when
the coefficients sum to one (i.e., when the argument is an affine combination).
This restricted superposition property for affine functions is useful in showing
that a function f is not affine: We find vectors x, y, and numbers and with
+ = 1, and verify that f (x + y) 6= f (x) + f (y). This shows that f cannot
be affine. As an example, we verified above that superposition does not hold for
the maximum function (with n > 1); the coefficients in our counterexample are
= = 1/2, which sum to one, which allows us to conclude that the maximum
function is not affine.
The converse is also true: Any scalar-valued function that satisfies the restricted
superposition property is affine. An analog of the formula (2.3) is

f (x) = f (0) + x1 (f (e1 ) f (0)) + + xn (f (en ) f (0)) ,

which holds when f is affine, and x is any n-vector. This formula shows that for
an affine function, once we know the n + 1 numbers f (0), f (e1 ), . . . , f (en ), we can
predict (or reconstruct or evaluate) f (x) for any n-vector x.
In some contexts affine functions are called linear. For example, when x is a
scalar, the function f defined as f (x) = x + is sometimes referred to as a linear
function of x, perhaps because its graph is a line. But when 6= 0, f is not a linear
function of x, in the standard mathematical sense; it is an affine function of x.
In this book we will distinguish between linear and affine functions. Two simple
examples are shown in figure 2.1.

A civil engineering example. Many scalar-valued functions that arise in science

and engineering are well approximated by linear or affine functions. As a typical
example, consider a steel structure like a bridge, and let w be an n-vector that
gives the weight of the load on the bridge in n specific locations, in metric tons.
These loads will cause the bridge to deform (move and change shape) slightly. Let s
2.1 Linear functions 29

f (x) g(x)

x x

Figure 2.1 Left. The function f is linear. Right. The function g is affine,
but not linear.

w1 w2 w3

Figure 2.2 A bridge with weights w1 , w2 , w3 applied in 3 locations. These

weights cause the bridge to sag in the middle, by an amount s. (The sag is
exaggerated in this diagram.)
30 2 Linear functions

w1 w2 w3 Measured sag Predicted sag

1 0 0 0.12
0 1 0 0.31
0 0 1 0.26
0.5 1.1 0.3 0.481 0.479
1.5 0.8 1.2 0.736 0.740

Table 2.1 Loadings on a bridge (first three columns), and the associated
measured sag at a certain point (fourth column) and the predicted sag using
the affine model constructed from the first three experiments (fifth column).

denote the distance that a specific point on the bridge sags, in centimeters, due
to the load w. This is shown in figure 2.2. For weights the bridge is designed
to handle, the sag is very well approximated as a linear function s = f (x). This
function can be expressed as an inner product, s = cT w, for some n-vector c. From
the equation s = c1 w1 + + cn wn , we see that c1 w1 is the amount of the sag that
is due to the weight w1 , and similarly for the other weights. The coefficients ci ,
which have units of cm/ton, are called compliances, and give the sensitivity of the
sag with respect to loads applied at the n locations.
The vector c can be computed by (numerically) solving a partial differential
equation, given the detailed design of the bridge and the mechanical properties of
the steel used to construct it. This is always done during the design of a bridge.
The vector c can also be measured once the bridge is built, using the formula (2.3).
We apply the load w = e1 , which means that we place a one ton load at the first
load position on the bridge, with no load at the other positions. We can then
measure the sag, which is c1 . We repeat this experiment, moving the one ton load
to positions 2, 3, . . . , n, which gives us the coefficients c2 , . . . , cn . At this point
we have the vector c, so we can now predict what the sag will be with any other
loading. To check our measurements (and linearity of the sag function) we might
measure the sag under other more complicated loadings, and in each case compare
our prediction (i.e., cT w) with the actual measured sag.
Table 2.1 shows what the results of these experiments might look like, with each
row representing an experiment (i.e., placing the loads and measuring the sag). In
the last two rows we compare the measured sag and the predicted sag, using the
linear function with coefficients found in the first three experiments.

2.2 Taylor approximation

In many applications, scalar-valued functions of n variables, or relations between
n variables and a scalar one, can be approximated as linear or affine functions. In
these cases we sometimes refer to the linear or affine function relating the vari-
ables and the scalar variable as a model, to remind us that the relation is only an
2.2 Taylor approximation 31

approximation, and not exact.

Differential calculus gives us an organized way to find an approximate affine
model. Suppose that f : Rn R is differentiable, which means that its par-
tial derivatives exist (see C.1). Let z be an n-vector. The (first-order) Taylor
approximation of f near (or at) the point z is the function f(x) of x defined as

f f
f(x) = f (z) + (z)(x1 z1 ) + + (z)(xn zn ),
x1 xn
f
where x i
(z) denotes the partial derivative of f with respect to its ith argument,
evaluated at the n-vector z. The hat appearing over f on the left-hand side is a
common notational hint that it is an approximation of the function f .
The first-order Taylor approximation f(x) is a very good approximation of f (x)
when all xi are near the associated zi . Sometimes f is written with a second vector
argument, as f(x; z), to show the point z at which the approximation is developed.
The first term in the Taylor approximation is a constant; the other terms can be
interpreted as the contribution to the (approximate) change in the function value
(from f (z)) due to the change in xi (from zi ).
Evidently f is an affine function of x. (It is sometimes called the linear approx-
imation of f near z, even though it is in general affine, and not linear.) It can be
written compactly using inner product notation as

f(x) = f (z) + f (z)T (x z), (2.4)

where f (z) is an n-vector, the gradient of f (at the point z),

f
x1 (z)

f (z) = ..
. (2.5)

.
f
xn (z)

The first term in the Taylor approximation (2.4) is the constant f (z), the value of
the function when x = z. The second term is the inner product of the gradient of
f at z and the deviation or perturbation of x from z, i.e., x z.
We can express the first-order Taylor approximation as a linear function plus a
constant,
f(x) = f (z)T x + (f (z) f (z)T z),
but the form (2.4) is perhaps easier to interpret.
The first-order Taylor approximation gives us an organized way to construct
an affine approximation of a function f : Rn R, near a given point z, when
there is a formula or equation that describes f , and it is differentiable. A simple
example, for n = 1, is shown in figure 2.3. Over the full x-axis scale shown, the
Taylor approximation f does not give a good approximation of the function f . But
for x near z, the Taylor approximation is very good.

Example. Consider the function f : R2 R given by f (x) = x1 + exp(x2 x1 ),

which is not linear or affine. To find the Taylor approximation f near the point
32 2 Linear functions

f (x)

f(x)

Figure 2.3 A function f of one variable, and the first-order Taylor approxi-
mation f(x) = f (z) + f 0 (z)(x z) at z.

x f (x) f(x) |f(x) f (x)|

(1.00, 2.00) 3.7183 3.7183 0.0000
(0.96, 1.98) 3.7332 3.7326 0.0005
(1.10, 2.11) 3.8456 3.8455 0.0001
(0.85, 2.05) 4.1701 4.1119 0.0582
(1.25, 2.41) 4.4399 4.4032 0.0367

Table 2.2 Some values of x (first column), the function value f (x) (sec-
ond column), the Taylor approximation f(x) (third column), and the error
(fourth column).

z = (1, 2), we take partial derivatives to obtain

 
1 exp(z2 z1 )
f (z) = ,
exp(z2 z1 )

which evaluates to (1.7183, 2.7183) at z = (1, 2). The Taylor approximation at

z = (1, 2) is then

f(x) = 3.7183 + (1.7183, 2.7183)T (x (1, 2))

= 3.7183 1.7183(x1 1) + 2.7183(x2 2).

Table 2.2 shows f (x) and f(x), and the approximation error |f(x) f (x)|, for some
values of x relatively near z. We can see that f is indeed a very good approximation
of f , especially when x is near z.
 2.3 Regression model 33

2.3 Regression model

In this section we describe a very commonly used affine function, especially when
the n-vector x represents a feature vector. The affine function of x given by

y = xT + v, (2.6)

where is an n-vector and v is a scalar, is called a regression model. In this context,

the entries of x are called the regressors, and y is called the prediction, since the
regression model is typically an approximation or prediction of some true value y,
which is called the dependent variable, outcome, or label.
The vector is called the weight vector, and the scalar v is called the offset or
intercept in the regression model. Together, and v are called the parameters in
the regression model. (We will see in chapter 13 how the parameters in a regression
model can be estimated or guessed, based on some past or known observations of
the feature vector x and the associated outcome y.) The symbol y is used in
the regression model to emphasize that it is an estimate or prediction of some
outcome y.
The entries in the weight vector have a simple interpretation: i is the amount
by which y increases (if i > 0) when feature i increases by one (with all other
features the same). If i is small, the prediction y doesnt depend too strongly on
feature i. The offset v is the value of y when all features have the value 0.
The regression model is very interpretable when all of the features have values
that are either 0 or 1, which occurs when the features represent which of two
outcomes holds. As a simple example consider a regression model for lifespan of a
person in some group, with x1 = 0 if the person is female (x1 = 1 if male), x2 = 1
if the person has type II diabetes, and x3 = 1 if the person smokes cigarettes. In
this case, v is the regression model estimate for lifespan of a female nondiabetic
nonsmoker; 1 is the increase in estimated lifespan if the person is male, 2 is the
increase in estimated lifespan if the person is diabetic, and 3 is the increase in
estimated lifespan if the person smokes cigarettes. (In a model that fits real data,
all three of these coefficients would be negative, meaning that they decrease the
regression model estimate of lifespan.)
Vector stacking can be used to lump the weights and offset in the regression
model (2.6) into a single parameter vector, which simplifies the notation a bit. We
create a new regression vector x, with n + 1 entries, as x = (1, x). We can think
of x as a new feature vector, consisting of all n original features, and one new
feature added (x1 ) at the beginning, which always has the value one. We define
the parameter vector = (v, ), so the regression model (2.6) has the simple inner
product form
 T  
1 v
y = xT + v = = xT . (2.7)
x

Often we omit the tildes, and simply write this as y = xT , where we assume that
the first feature in x is the constant 1. A feature that always has the value 1 is
not particularly informative or interesting, but it does simplify the notation in a
regression model.
34 2 Linear functions

House x1 (area) x2 (beds) y (price) y (prediction)

1 0.846 1 115.00 161.37
2 1.324 2 234.50 213.61
3 1.150 3 198.00 168.88
4 3.037 4 528.00 430.67
5 3.984 5 572.50 552.66

Table 2.3 Five houses with associated feature vectors shown in the second
and third columns. The fourth and fifth column give the actual price, and
the price predicted by the regression model.

House price regression model. As a simple example of a regression model, sup-

pose that y is the selling price of a house in some neighborhood, over some time
period, and the 2-vector x contains attributes of the house:

x1 is the house area (in 1000 square feet),

x2 is the number of bedrooms.

If y represents the selling price of the house, in thousands of dollars, the regression
model y = xT + v predicts the price in terms of the attributes or features. This
regression model is not meant to describe an exact relationship between the house
attributes and its selling price; it is a model or approximation. Indeed, we would
expect such a model to give, at best, only a crude approximation of selling price.
As a specific numerical example, consider the regression model parameters

= (148.73, 18.85), v = 54.40. (2.8)

These parameter values were found using the methods we will see in chapter 13,
based on records of sales for 774 houses in the Sacramento area. Table 2.3 shows
the feature vectors x for five houses that sold during the period, the actual sale
price y, and the predicted price y from the regression model above. Figure 2.4
shows the predicted and actual sale prices for 774 houses, including the five houses
in the table, on a scatter plot, with actual price on the horizontal axis and predicted
price on the vertical axis.
We can see that this particular regression model gives reasonable, but not very
accurate, predictions of the actual sale price. (Regression models for house prices
that are used in practice use many more than two regressors, and are much more
accurate.)
The model parameters in (2.8) are readily interpreted. The parameter 1 =
148.73 is the amount the regression model price prediction increases (in thousands
of dollars) when the house area increases by 1000 square feet (with the same number
of bedrooms). The parameter 2 = 18.85 is the price prediction increase with
the addition of one bedroom, with the total house area held constant, in units of
thousands of dollars per bedroom. It might seem strange that 2 is negative, since
one imagines that adding a bedroom to a house would increase its sale price, not
2.3 Regression model 35

800
Predicted price y (thousand dollars)

600
House 5

House 4
400

House 1

200

House 2
House 3

0
0 200 400 600 800
Actual price y (thousand dollars)

Figure 2.4 Scatter plot of actual and predicted sale prices for 774 houses
sold in Sacramento during a five-day period.
36 2 Linear functions

decrease it. The regression model (2.8) does predict that adding a bedroom to
a house will increase its sale price, provided the additional bedroom has an area
exceeding around 127 square feet. In any case, the regression model is crude enough
that any interpretation is dubious.
 Chapter 3

Norm and distance

In this chapter we focus on the norm of a vector, a measure of its length, and on
related concepts like distance, angle, standard deviation, and correlation.

3.1 Norm
The Euclidean norm of an n-vector x (named after the ancient Greek mathe-
matician Euclid), denoted kxk, is the squareroot of the sum of the squares of its
elements, q
kxk = x21 + x22 + + x2n .

The Euclidean norm is sometimes written with a subscript 2, as kxk2 . (The sub-
script 2 indicates that the entries of x are raised to the 2nd power.) Other less
widely used terms for the Euclidean norm of a vector are the magnitude, or length,
of a vector. (The term length should be avoided, since it is also often used to refer
to dimension of the vector.) We can express the Euclidean norm in terms of the
inner product of x with itself:
kxk = xT x.
We use the same notation for the norm of vectors of different dimensions.
As simple examples, we have

2
 

1 = 9 = 3,
0
1 = 1.


2

When x is a scalar, i.e., a 1-vector, the Euclidean norm is the same as the
absolute value of x. Indeed, the Euclidean norm can be considered a generalization
or extension of the absolute value or magnitude, that applies to vectors. The double
bar notation is meant to suggest this. Like the absolute value of a number, the
norm of a vector is a (numerical) measure of its size. We say a vector is small if
its norm is a small number, and we say it is large if its norm is a large number.
38 3 Norm and distance

(The numerical values of the norm that qualify for small or large depend on the
particular application and context.)
When we form the norm of a vector x, the entries all have equal status. This
makes sense when the entries of the vector x represent the same type of quantity,
using the same units (say, at different times or locations), for example meters
or dollars. When the entries of a vector represent different types of quantities,
the units should be chosen so the numerical values of the entries are comparable.
(Another approach is to use a weighted norm, described later on page 40.)

Properties of norm. Some important properties of the Euclidean norm are given
below. Here x and y are vectors of the same size, and is a scalar.
Homogeneity. kxk = ||kxk. Multiplying a vector by a scalar multiplies the
norm by the absolute value of the scalar.
Triangle inequality. kx + yk kxk + kyk. The Euclidean norm of a sum
of two vectors is no more than the sum of their norms. (The name of this
property will be explained later.)
Nonnegativity. kxk 0.
Definiteness. kxk = 0 only if x = 0.
The last two properties together, which state that the norm is always nonnegative,
and zero only when the vector is zero, are called positive definiteness. The first,
third, and fourth properties are easy to show directly from the definition of the
norm. As an example, lets verify the definiteness property. If kxk = 0, then
we also have kxk2 = 0, which means that x21 + + x2n = 0. This is a sum of n
nonnegative numbers, which is zero. We can conclude that each of the n numbers is
zero, since if any of them were nonzero the sum would be positive. So we conclude
that x2i = 0 for i = 1, . . . , n, and therefore xi = 0 for i = 1, . . . , n; and thus, x = 0.
Establishing the second property, the triangle inequality, is not as easy; we will
give a derivation a bit later.
Any real-valued function of an n-vector that satisfies the four properties listed
above is called a (general) norm. But in this book we will only use the Euclidean
norm, so from now on, we refer to the Euclidean norm as the norm.

Root-mean-square value. The norm is related to the root-mean-square (RMS)

value of an n-vector x, defined as
r
x21 + + x2n kxk
rms(x) = = .
n n
The argument of the squareroot in the middle expression is called the mean-square
value of x, denoted ms(x), and the RMS value is the squareroot of the mean-square
value. The RMS value of a vector x is useful when comparing norms of vectors
with different dimensions; the RMS value tells us whata typical value of |xi | is.
For example, the norm of 1, the n-vector of all ones, is n, but its RMS value is 1,
independent of n. More generally, if all the entries of a vector are the same, say,
, then the RMS value of the vector is ||.
3.1 Norm 39

Norm of a sum. A useful formula for the norm of the sum of two vectors x and
y is q
kx + yk = kxk2 + 2xT y + kyk2 . (3.1)
To derive this formula, we start with the square of the norm of x+y and use various
properties of the inner product:

kx + yk2 = (x + y)T (x + y)
= xT x + xT y + y T x + y T y
= kxk2 + 2xT y + kyk2 .

Taking the squareroot of both sides yields the formula (3.1) above. In the first line,
we use the definition of the norm. In the second line, we expand the inner product.
In the fourth line we use the definition of the norm, and the fact that xT y = y T x.

Norm of block vectors. The norm-squared of a stacked vector is the sum of the
norm-squared values of its subvectors. For example, with d = (a, b, c) (where a, b,
and c are vectors), we have

kdk2 = dT d = aT a + bT b + cT c = kak2 + kbk2 + kck2 .

This idea is often used in reverse, to express the sum of the norm-squared values
of some vectors as the norm-square value of a block vector formed from them.
We can write the equality above in terms of norms as
p
k(a, b, c)k = kak2 + kbk2 + kck2 = k(kak, kbk, kck)k.

In words: The norm of a stacked vector is the norm of the vector formed from
the norms of the subvectors. (Which is quite a mouthful.) The right-hand side of
the equation above should be carefully read. The outer norm symbols enclose a
3-vector, with (scalar) entries kak, kbk, and kck.

Chebyshev inequality. Suppose that x is an n-vector, and that k of its entries

satisfy |xi | a, where a > 0. Then k of its entries satisfy x2i a2 . It follows that

kxk2 = x21 + + x2n ka2 ,

since k of the numbers in the sum are at least a2 , and the other n k numbers are
nonnegative. We can conclude that k kxk2 /a2 , which is called the Chebyshev
inequality. When kxk2 /a2 n, the inequality tells us nothing, since we always
have k n. In other cases it limits the number of entries in a vector that can
be large. For a > kxk, the inequality is k kxk2 /a2 < 1, so we conclude that
k = 0 (since k is an integer). In other words, no entry of a vector can be larger in
magnitude than the norm of the vector.
The Chebyshev inequality is easier to interpret in terms of the RMS value of a
vector. We can write it as  2
k rms(x)
, (3.2)
n a
40 3 Norm and distance

where k is, as above, the number of entries of x with absolute value at least a. The
left-hand side is the fraction of entries of the vector that are at least a in absolute
value. The right-hand side is the inverse square of the ratio of a to rms(x). It says,
for example, that no more than 1/25 = 4% of the entries of a vector can exceed
its RMS value by more than a factor of 5. The Chebyshev inequality partially
justifies the idea that the RMS value of a vector gives an idea of the size of a
typical entry: It states that not too many of the entries of a vector can be much
bigger (in absolute value) than its RMS value. (A converse statement can also be
made: At least one entry of a vector has absolute value as large as the RMS value
of the vector.)

Weighted norm. The weighted norm of a vector x is defined as

p
kxkw = (x1 /w1 )2 + + (xn /wn )2 ,

where w1 , . . . , wn are given positive weights, used to assign more or less importance
to the different elements of the n-vector x. If all the weights are one, the weighted
norm reduces to the usual (unweighted) norm.
Weighted norms arise naturally when the elements of the vector x have different
physical units, or natural ranges of values. One common rule of thumb is to choose
wi equal to the typical value of |xi | in the application or setting. This choice of
weights bring all the terms in the sum to the same order, one. We can also imagine
that the weights contain the same physical units as the elements xi , which makes
the terms in the sum (and therefore the norm as well) unitless.

3.2 Distance
Euclidean distance. We can use the norm to define the Euclidean distance be-
tween two vectors a and b as the norm of their difference:

dist(a, b) = ka bk.

For one, two, and three dimensions, this distance is exactly the usual distance
between points with coordinates a and b, as illustrated in figure 3.1. But the
Euclidean distance is defined for vectors of any dimension; we can refer to the
distance between two vectors of dimension 100. Since we only use the Euclidean
norm in this book, we will refer to the Euclidean distance between vectors as,
simply, the distance between the vectors.
If a and b are n-vectors, we refer to the
RMS value of the difference, ka bk/ n, as the RMS deviation between the two
vectors.
When the distance between two n-vectors x and y is small, we say they are
close or nearby, and when the distance kx yk is large, we say they are far.
The particular numerical values of kxyk that correspond to close or far depend
on the particular application.
3.2 Distance 41

Figure 3.1 The norm of the displacement b a is the distance between the
points with coordinates a and b.

As an example, consider the 4-vectors

1.8 0.6 2.0
2.0 2.1 1.9
a= 3.7 ,
b= 1.9 ,

4.0 .
c=

4.7 1.4 4.6

The distances between pairs of them are
ka bk = 8.368, ka ck = 0.387, kb ck = 8.533,
so we can say that a is much nearer (or closer) to c than it is to b. We can also say
that c is much nearer to a than it is to b.

Triangle inequality. We can now explain where the triangle inequality gets its
name. Consider a triangle in two or three dimensions, whose vertices have coordi-
nates a, b, and c. The lengths of the sides are the distances between the vertices,
dist(a, b) = ka bk, dist(b, c) = kb ck, dist(a, c) = ka ck.
Geometric intuition tells us that the length of any side of a triangle cannot exceed
the sum of the lengths of the other two sides. For example, we have
ka ck ka bk + kb ck. (3.3)
This follows from the triangle inequality, since
ka ck = k(a b) + (b c)k ka bk + kb ck.
This is illustrated in figure 3.2.

Examples.
Feature distance. If x and y represent vectors of n features of two objects,
the quantity kx yk is called the feature distance, and gives a measure of
how different the objects are (in terms of the feature values). Suppose for
example the feature vectors are associated with patients in a hospital, with
entries such as weight, age, presence of chest pain, difficulty breathing, and
the results of tests. We can use feature vector distance to say that one patient
case is near another one (at least in terms of their feature vectors).
42 3 Norm and distance

ka ck
kb ck

a ka bk b

Figure 3.2 Triangle inequality.

z4

x z6
z5

z3

z2
z1

Figure 3.3 The point z3 is the nearest neighbor of x among the points z1 ,
. . . , z6 .

RMS prediction error. Suppose that the n-vector y represents a time series
of some quantity, for example, hourly temperature at some location, and y is
another n-vector that represents an estimate or prediction of the time series
y, based on other information. The difference y y is called the prediction
error, and its RMS value rms(y y) is called the RMS prediction error. If
this value is small (say, compared to rms(y)) the prediction is good.

Nearest neighbor. Suppose z1 , . . . , zm is a collection of m n-vectors, and that

x is another n-vector. We say that zj is the nearest neighbor of x (among z1 ,
. . . , zm ) if
kx zj k kx zi k, i = 1, . . . , m.
In words: zj is the closest vector to x among the vectors z1 , . . . , zm . This
is illustrated in figure 3.3. The idea of nearest neighbor, and generalizations
such as the k-nearest neighbors, are used in many applications.

Document dissimilarity. Suppose n-vectors x and y represent the histograms

of word occurrences for two documents. Then kx yk represents a measure
of the dissimilarity of the two documents. We might expect the dissimilarity
to be smaller when the two documents have the same genre, topic, or author;
we would expect it to be larger when they are on different topics, or have
 3.3 Standard deviation 43

Veterans Memorial Academy Golden Globe Super Bowl

Day Day Awards Awards
Veterans Day 0 0.095 0.130 0.153 0.170
Memorial Day 0.095 0 0.122 0.147 0.164
Academy A. 0.130 0.122 0 0.108 0.164
Golden Globe A. 0.153 0.147 0.108 0 0.181
Super Bowl 0.170 0.164 0.164 0.181 0

Table 3.1 Pairwise word count histogram distances between five Wikipedia
articles.

different authors. As an example we form the word count histograms for the
5 Wikipedia articles with titles Veterans Day, Memorial Day, Academy
Awards, Golden Globe Awards, and Super Bowl, using a dictionary of
4423 words. (More detail is given in 4.4.) The pairwise distances between
the word count histograms is shown in table 3.1. We can see that pairs of
related articles have smaller word count histogram distances than less related
pairs of articles.

3.3 Standard deviation

For any vector x, the vector x = x avg(x)1 is called the associated de-meaned
vector, obtained by subtracting from each entry of x the mean value of the entries.
(This is not standard notation; i.e., x is not generally used to denote the de-meaned
vector.) The mean value of the entries of x is zero, i.e., avg(x) = 0. This explains
why x is called the de-meaned version of x; roughly speaking, it is x with its mean
removed. The de-meaned vector is useful for understanding how the entries of a
vector deviate from their mean value. It is zero if all the entries in the original
vector x are the same.
The standard deviation of an n-vector x is defined as the RMS value of the
de-meaned vector x avg(x)1, i.e.,
r
(x1 avg(x))2 + + (xn avg(x))2
std(x) = .
n
This is the same as the RMS deviation between a vector x and the vector all of
whose entries are avg(x). It can be written using the inner product and norm as
kx (1T x/n)1k
std(x) = . (3.4)
n
The standard deviation of a vector x tells us the typical amount by which its entries
deviate from their average value. The standard deviation of a vector is zero only
when all its entries are equal. The standard deviation of a vector is small when the
entries of the vector are nearly the same.
44 3 Norm and distance

As a simple example consider the vector x = (1, 2, 3, 2). Its mean or average
value is avg(x) = 1, so the de-meaned vector is x = (0, 3, 2, 1). Its standard
deviation is std(x) = 1.872. We interpret this number as a typical value by which
the entries differ from the mean of the entries. These numbers are 0, 3, 2, and 1,
so 1.872 is reasonable.
We should warn the reader that another slightly different definition
of the stan-
dard deviationof a vector is widely used, in which the denominator n in (3.4) is
replaced with n 1 (for n 2). In this book we will only use the definition (3.4).
The average, RMS value, and standard deviation of a vector are related by the
formula
rms(x)2 = avg(x)2 + std(x)2 . (3.5)
This formula makes sense: rms(x)2 is the mean square value of the entries of x,
which can be expressed as the square of the mean value, plus the mean square
fluctuation of the entries of x around their mean value. We can derive this formula
from our vector notation formula for std(x) given above. We have

std(x)2 = (1/n)kx (1T x/n)1k2

= (1/n)(xT x 2xT (1T x/n)1 + ((1T x/n)1)T ((1T x/n)1))
= (1/n)(xT x (2/n)(1T x)2 + n(1T x/n)2 )
= (1/n)xT x (1T x/n)2
= rms(x)2 avg(x)2 ,

which can be re-arranged to obtain the identity above. This derivation uses many
of the properties for norms and inner products, and should be read carefully to
understand every step. In the second line, we expand the norm-square of the sum
of two vectors. In the third line we use the commutative property of scalar-vector
multiplication, moving scalars such as (1T x/n) to the front of each term, and also
the fact that 1T 1 = n.

Examples.
Mean return and risk. Suppose that an n-vector represents a time series of
return on an investment, expressed as a percentage, in n time periods over
some interval of time. Its average gives the mean return over the whole
interval, often shortened to its return. Its standard deviation is a measure of
how variable the return is, from period to period, over the time interval, i.e.,
how much it typically varies from its mean, and is often called the (per period)
risk of the investment. Multiple investments can be compared by plotting
them on a risk-return plot, which gives the mean and standard deviation of
the returns of each of the investments over some interval. A desirable return
history vector has high mean return and low risk; this means that the returns
in the different periods are consistently high. Figure 3.4 shows an example.
Temperature or rainfall. Suppose that an n-vector is a time series of the
daily average temperature at a particular location, over a one year period.
Its average gives the average temperature at that location (over the year) and
its standard deviation is a measure of how much the temperature varied from
3.3 Standard deviation 45

10 10

5 5
ak

bk
0 0

5 5
5 10 5 10
k k
10 10

5 5
dk
ck

0 0

5 5
5 10 5 10
k k
return
3
c
b
2

d
a
1

0 risk
0 2 4

Figure 3.4 The vectors a, b, c, d represent time series of returns on in-

vestments over 10 periods. The bottom plot shows the investments in a
risk-return plane, with return defined as the average value and risk as the
standard deviation of the corresponding vector.
46 3 Norm and distance

its average value. We would expect the average temperature to be high and
the standard deviation to be low in a tropical location, and the opposite for
a location with high latitude.

Chebyshev inequality for standard deviation. The Chebyshev inequality (3.2)

can be transcribed to an inequality expressed in terms of the mean and standard
deviation: If k is the number of entries of x that satisfy |xi avg(x)| a, then
k/n (std(x)/a)2 . (This inequality is only interesting for a > std(x).) For
example, at most 1/9 = 11.1% of the entries of a vector can deviate from the mean
value avg(x) by 3 standard deviations or more. Another way to state this is: The
fraction of entries of x within standard deviations of avg(x) is at least 1 1/2
(for > 1).
As an example, consider a time series of return on an investment, with a mean
return of 8%, and a risk (standard deviation) 3%. By the Chebyshev inequality,
the fraction of periods with a loss (i.e., xi 0) is no more than (3/8)2 = 14.1%.
(In fact, the fraction of periods when the return is either a loss, xi 0, or very
good, xi 16%, is together no more than 14.1%.)

Properties of standard deviation.

Adding a constant. For any vector x and any number a, we have std(x+a1) =
std(x). Adding a constant to every entry of a vector does not change its
standard deviation.

Multiplying by a scalar. For any vector x and any number a, we have

std(ax) = |a| std(x). Multiplying a vector by a scalar multiplies the standard
deviation by the absolute value of the scalar.

Standardization. For any vector x, we refer to x = x avg(x)1 as the de-meaned

version of x, since it has average or mean value zero. If we then divide by the RMS
value of x (which is the standard deviation of x), we obtain the vector

1
z= (x avg(x)1).
std(x)

This vector is called the standardized version of x. It has mean zero, and standard
deviation one. Its entries are sometimes called the z-scores associated with the
original entries of x. For example, z4 = 1.4 means that x4 is 1.4 standard deviations
above the mean of the entries of x. Figure 3.5 shows an example.
The standardized values for a vector give a simple way to interpret the original
values in the vectors. For example, if an n-vector x gives the values of some
medical test of n patients admitted to a hospital, the standardized values or z-
scores tell us how high or low, compared to the population, that patients value is.
A value z6 = 3.2, for example, means that patient 6 has a very low value of the
measurement; whereas z22 = 0.3 says that patient 22s value is quite close to the
average value.
 3.4 Angle 47

4 4 4
xk

xk

zk
0 0 0

4 4 4
2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
k k k

Figure 3.5 A 10-vector x, the de-meaned vector x = x avg(x)1, and the

standardized vector z = (1/ std(x))x. The horizontal dashed lines indicate
the mean and the standard deviation of each vector. The middle line is the
mean; the distance between the three lines is the standard deviation.

3.4 Angle
Cauchy-Schwarz inequality. An important inequality that relates norms and in-
ner products is the Cauchy-Schwarz inequality:

|aT b| kak kbk

for any n-vectors a and b. Written out in terms of the entries, this is

1/2
1/2
|a1 b1 + + an bn | a21 + + a2n b21 + + b2n ,

which looks more intimidating. This inequality is attributed to the mathematician

Augustin-Louis Cauchy; Hermann Schwarz gave the derivation given below.
The Cauchy-Schwarz inequality can be shown as follows. The inequality clearly
holds if a = 0 or b = 0 (in this case, both sides of the inequality are zero). So we
suppose now that a 6= 0, b 6= 0, and define = kak, = kbk. We observe that
2
0 ka bk
2 2
= kak 2(a)T (b) + kbk
= 2 kak2 2(aT b) + 2 kbk2
= kbk2 kak2 2kbkkak(aT b) + kak2 kbk2
= 2kak2 kbk2 2kak kbk(aT b).

Dividing by 2kak kbk yields aT b kak kbk. Applying this inequality to a and
b we obtain aT b kak kbk. Putting these two inequalities together we get the
Cauchy-Schwarz inequality, |aT b| kak kbk.
This argument also reveals the conditions on a and b under which they satisfy
the Cauchy-Schwarz inequality with equality. This occurs only if ka bk = 0,
i.e., a = b. This means that each vector is a scalar multiple of the other (in
the case when they are nnozero). This statement remains true when either a or
b is zero. So the Cauchy-Schwarz inequality holds with equality when one of the
vectors is a multiple of the other; in all other cases, it holds with strict inequality.
48 3 Norm and distance

Verification of triangle inequality. We can use the Cauchy-Schwarz inequality to

verify the triangle inequality. Let a and b be any vectors. Then

ka + bk2 = kak2 + 2aT b + kbk2

kak2 + 2kakkbk + kbk2
2
= (kak + kbk) ,

where we used the Cauchy-Schwarz inequality in the second line. Taking the square-
root we get the triangle inequality, ka + bk kak + kbk.

Angle between vectors. The angle between two nonzero vectors a, b is defined
as  T 
a b
= arccos
kak kbk
where arccos denotes the inverse cosine, normalized to lie in the interval [0, ]. In
other words, we define as the unique number between 0 and that satisfies

aT b = kak kbk cos .

The angle between a and b is written as 6 (a, b), and is sometimes expressed in
degrees. (The default angle unit is radians; 360 is 2 radians.) For example,
6 (a, b) = 60 means 6 (a, b) = /3, i.e., aT b = (1/2)kakkbk.
The angle coincides with the usual notion of angle between vectors, when they
have dimension two or three. For example, the angle between the vectors a =
(1, 2, 1) and b = (2, 0, 3) is
 
5
arccos = arccos(0.5661) = 0.9690 = 55.52
6 13

(to 4 digits). But the definition of angle is more general; we can refer to the angle
between two vectors with dimension 100.
The angle is a symmetric function of a and b: we have 6 (a, b) = 6 (b, a). The
angle is not affected by scaling each of the vectors by a positive scalar: we have,
for any vectors a and b, and any positive numbers and ,

6 (a, b) = 6 (a, b).

Acute and obtuse angles. Angles are classified according to the sign of aT b.

If the angle is /2 = 90 , i.e., aT b = 0, the vectors are said to be orthogonal.

We write a b if a and b are orthogonal.

If the angle is zero, which means aT b = kakkbk, the vectors are aligned. Each
vector is a positive multiple of the other (assuming the vectors are nonzero).

If the angle is = 180 , which means aT b = kak kbk, the vectors are anti-
aligned. Each vector is a negative multiple of the other (assuming the vectors
are nonzero).
3.4 Angle 49

Figure 3.6 From left to right: examples of orthogonal, aligned, and anti-
aligned vectors, vectors that make an acute and an obtuse angle.

0
b

Figure 3.7 Two points a and b on a sphere with radius R and center at the
origin. The spherical distance between the points is equal to R6 (a, b).

If 6 (a, b) /2 = 90 , the vectors are said to make an acute angle. This is

the same as aT b 0, i.e., the vectors have nonnegative inner product.

If 6 (a, b) /2 = 90 , the vectors are said to make an obtuse angle. This is

the same as aT b 0, i.e., the vectors have nonpositive inner product.

These definitions are illustrated in figure 3.6.

Examples.

Spherical distance. Suppose a and b are 3-vectors that represent two points
that lie on a sphere of radius R (for example, locations on earth). The
spherical distance between them, measured along the sphere, is given by
R6 (a, b). This is illustrated in figure 3.7.

Document similarity via angles. If n-vectors x and y represent the word

counts for two documents, their angle 6 (x, y) can be used as a measure of
document dissimilarity. (When using angle to measure document dissimilar-
ity, either word counts or histograms can be used; they produce the same
result.) As an example, table 3.2 gives the angles in degrees between the
word histograms in the example at the end of 3.2.
50 3 Norm and distance

Veterans Memorial Academy Golden Globe Super Bowl

Day Day Awards Awards
Veterans Day 0 60.6 85.7 87.0 87.7
Memorial Day 60.6 0 85.6 87.5 87.5
Academy A. 85.7 85.6 0 58.7 85.7
Golden Globe A. 87.0 87.5 58.7 0 86.0
Super Bowl 87.7 87.5 86.1 86.0 0

Table 3.2 Pairwise angles (in degrees) between word histograms of five
Wikipedia articles.

Norm of sum via angles. For vectors x and y we have

kx + yk2 = kxk2 + 2xT y + kyk2 = kxk2 + 2kxkkyk cos + kyk2 , (3.6)

where = 6 (x, y). (The first equality comes from (3.1).) From this we can make
several observations.
If x and y are aligned ( = 0), we have kx + yk = kxk + kyk. Thus, their
norms add.
If x and y are orthogonal ( = 90 ), we have kx + yk2 = kxk 2 2
p + kyk . In this
case the norm-squared values add, and we have kx + yk = kxk + kyk2 .
2

Correlation coefficient. Suppose a and b are n-vectors, with associated de-meaned

vectors
a = a avg(a)1, b = b avg(b)1.
Assuming these de-meaned vectors are not zero (which occurs when the original
vectors have all equal entries), we define their correlation coefficient as

aT b
= . (3.7)
kak kbk

Thus, = cos , where = 6 (a, b). We can also express the correlation coefficient in
terms of the vectors u and v obtained by standardizing a and b. With u = a/ std(a)
and v = b/ std(b), we have
= uT v/n. (3.8)

(We use kuk = kvk = n.)
This is a symmetric function of the vectors: the correlation between a and b
is the same as the correlation coefficient between b and a. The Cauchy-Schwarz
inequality tells us that the correlation coefficient ranges between 1 and +1. For
this reason, the correlation coefficient is sometimes expressed as a percentage. For
example, = 30% means = 0.3. When = 0, we say the vectors are uncorrelated.

The correlation coefficient tells us how the entries in the two vectors vary to-
gether. Roughly speaking, high correlation (say, = 0.8) means that entries of a
3.4 Angle 51

ak bk bk

k k ak

ak bk bk

k k ak

ak bk bk

k k ak

Figure 3.8 Three pairs of vectors a, b of length 10, with correlation coeffi-
cients 0.968 (top), 0.988 (middle), and 0.004 (bottom).

and b are typically above their mean for many of the same entries. The extreme
case = 1 occurs only if the vectors a and b are aligned, which means that each is a
positive multiple of the other, and the other extreme case = 1 occurs only when
a and b are negative multiples of each other. This idea is illustrated in figure 3.8,
which shows the entries of two vectors, as well as a scatter plot of them, for cases
with correlation near 1, near 1, and near 0.

The correlation coefficient is often used when the vectors represent time series,
such as the returns on two investments over some time interval, or the rainfall in
two locations over some time interval. If they are highly correlated (say, > 0.8),
the two time series are typically above their mean values at the same times. For
example, we would expect the rainfall time series at two nearby locations to be
highly correlated. As another example, we might expect the returns of two similar
companies, in the same business area, to be highly correlated.
52 3 Norm and distance

Standard deviation of sum. We can derive a formula for the standard deviation
of a sum from (3.6):
p
std(a + b) = std(a)2 + 2 std(a) std(b) + std(b)2 . (3.9)

To derive this from (3.6) we let a and b denote the de-meaned versions of a and b.
Then a + b is the de-meaned version of a + b, and std(a + b)2 = ka bk2 /n. Now
using (3.6) and = cos 6 (a, b), we get

n std(a + b)2 = ka + bk2

= kak2 + 2kakkbk + kbk2
= n std(a)2 + 2n std(a) std(b) + n std(b)2 .

Dividing by n and taking the squareroot yields the formula above.

If = 1, the standard deviation of the sum of vectors is the sum of their
standard deviations, i.e.,

std(a + b) = std(a) + std(b).

As decreases, the standard deviation of the sum decreases. When = 0, i.e., a

and b are uncorrelated, the standard deviation of the sum a + b is
p
std(a + b) = std(a)2 + std(b)2 ,

which is smaller than std(a) + std(b) (unless one of them is zero). When = 1,
the standard deviation of the sum is as small as it can be,

std(a + b) = | std(a) std(b)|.

Hedging investments. Suppose that vectors a and b are time series of returns
for two assets with the same return (average) and risk (standard deviation) ,
and correlation coefficient . (These are the traditional symbols used.) The vector
c = (a + b)/2 is the time series of returns for an investment with 50% in each of the
assets. This blended investment has the same return as the original assets, since

avg(c) = avg((a + b)/2) = (avg(a) + avg(b))/2 = .

The risk (standard deviation) of this blended investment is

p p
std(c) = 2 2 + 2 2 /2 = (1 + )/2,

using (3.9). From this we see that the risk of the blended investment is never
more than the risk of the original assets, and is smaller when the correlation of
the original asset
returns is smaller. When the returns are uncorrelated, the risk
is a factor 1/ 2 = 0.707 smaller than the risk of the original assets. If the asset
returns are strongly negatively correlated (i.e., is near 1), the risk of the blended
investment is much smaller than the risk of the original assets. Investing in two
assets with uncorrelated, or negatively correlated, returns is called hedging (which
is short for hedging your bets). Hedging reduces risk.
 3.5 Complexity 53

3.5 Complexity
Computing the norm of an n-vector requires n multiplications (to square each
entry), n 1 additions (to add the squares), and one squareroot. Even though
computing the squareroot typically takes more time than computing the product
or sum of two numbers, it is counted as just one flop. So computing the norm takes
2n flops. The cost of computing the RMS value of an n-vector is the same, since
we can ignore the two flops involved in division by n. Computing the distance
between two vectors costs 3n flops, and computing the angle between them costs
6n flops. All of these operations have order n.
De-meaning an n-vector requires 2n flops (n for forming the average and an-
other n flops for subtracting the average from each entry). Computing the standard
deviation costs 4n flops, and standardizing an n-vector costs 5n flops. The corre-
lation coefficient between two vectors costs 8n flops to compute. These operations
also have order n.
As a slightly more involved computation, suppose that we wish to determine the
nearest neighbor among a collection of k n-vectors z1 , . . . , zk to another n-vector
x. (This will come up in the next chapter.) The simple approach is to compute the
distances kxzi k for i = 1, . . . , k, and then find the minimum of these. (Sometimes
a comparison of two numbers is also counted as a flop.) The cost of this is 3kn flops
to compute the distances, and k 1 comparisons to find the minimum. The latter
term can be ignored, so the flop count is 3kn. The order of finding the nearest
neighbor in a collection of k n-vectors is kn.
54 3 Norm and distance
 Chapter 4

Clustering

In this chapter we consider the task of clustering a collection of vectors into groups
or clusters of vectors that are close to each other, as measured by the distance
between pairs of them. We describe a famous clustering method, called the k-
means algorithm, and give some typical applications.
The material in this chapter will not be used in the sequel. But the ideas, and
the k-means algorithm in particular, are widely used in practical applications, and
rely only on the ideas developed in the previous three chapters. So this chapter
can be considered an interlude that covers useful material that builds on the ideas
developed so far.

4.1 Clustering
Suppose we have N n-vectors, x1 , . . . , xN . The goal of clustering is to group or
partition the vectors (if possible) into k groups or clusters, with the vectors in each
group close to each other. Clustering is very widely used in many application areas,
typically (but not always) when the vectors represent features of objects.
Normally we have k  N , which means that there are many more vectors than
groups. Typical applications use values of k that range from a handful to a few
hundred or more, with values of N that range from hundreds to billions. Part of
the task of clustering a collection of vectors is to determine whether or not the
vectors can be divided into k groups, with vectors in each group near each other.
Of course this depends on k, the number of clusters, and the particular data, i.e.,
the vectors x1 , . . . , xN .
Figure 4.1 shows a simple example, with N = 300 2-vectors, shown as small
circles. We can easily see that this collection of vectors can be divided into k = 3
clusters, shown on the right with the colors representing the different clusters. We
could partition these data into other numbers of clusters, but we can see that k = 3
is a good value.
This example is not typical in several ways. First, the vectors have dimension
n = 2. Clustering any set of 2-vectors is easy: We simply scatter plot the values
56 4 Clustering

Figure 4.1 300 points in a plane. The points can be clustered in the three
groups shown on the right.

and check visually if the data are clustered, and if so, how many clusters there are.
In almost all applications n is larger than 2 (and typically, much larger than 2),
in which case this simple visual method cannot be used. The second way in which
it is not typical is that the points are very well clustered. In most applications,
the data are not as cleanly clustered as in this simple example; there are several or
even many points that lie in between clusters. Finally, in this example, it is clear
that the best choice of k is k = 3. In real examples, it can be less clear what the
best value of k is. But even when the clustering is not as clean as in this example,
and the best value of k is not clear, clustering can be very useful in practice.
Before we delve more deeply into the details of clustering and clustering algo-
rithms, we list some common applications where clustering is used.

Examples.
Topic discovery. Suppose xi are word histograms associated with N docu-
ments. A clustering algorithm partitions the documents into k groups, which
typically can be interpreted as groups of documents with the same or similar
topics, genre, or author. Since the clustering algorithm runs automatically
and without any understanding of what the words in the dictionary mean,
this is sometimes called automatic topic discovery.
Patient clustering. If xi are feature vectors associated with N patients ad-
mitted to a hospital, a clustering algorithm clusters the patients into k groups
of similar patients (at least in terms of their feature vectors).
Customer market segmentation. Suppose the vector xi gives the quantities (or
dollar values) of n items purchased by customer i over some period of time. A
clustering algorithm will group the customers into k market segments, which
are groups of customers with similar purchasing patterns.
ZIP-code clustering. Suppose that xi is a vector giving n quantities or statis-
tics for the residents of ZIP-code i, such as numbers of residents in various
4.1 Clustering 57

age groups, household size, education statistics, and income statistics. (In
this example N is around 40000.) A clustering algorithm might be used to
cluster the 40000 ZIP-codes into, say, k = 100 groups of ZIP-codes with
similar statistics.

Student clustering. Suppose the vector xi gives the detailed grading record
of student i in a course, i.e., her grades on each question in the quizzes,
homework assignments, and exams. A clustering algorithm might be used to
cluster the students into k = 10 groups of students who performed similarly.

Survey response clustering. A group of N people respond to a survey with n

questions. Each question contains a statement, such as The movie was too
long, followed by some ordered options such as

Strongly Disagree, Disagree, Neutral, Agree, Strongly Agree.

(This is called a Likert scale.) Suppose the n-vector xi encodes the selections
of respondent i on the n questions, using the numerical coding 2, 1, 0, +1,
+2 for the responses above. A clustering algorithm can be used to cluster
the respondents into k groups, each with similar responses to the survey.

Weather zones. For each of N counties we have a 24-vector wi that gives the
average monthly temperature in the first 12 entries and the average monthly
rainfall in the last 12 entries. (We can standardize all the temperatures, and
all the rainfall data, so they have a typical range between 1 and +1.) The
vector wi summarizes the annual weather pattern in county i. A clustering
algorithm can be used to cluster the counties into k groups that have similar
weather patterns, called weather zones. This clustering can be shown on a
map, and used to recommend landcape plantings depending on zone.

Daily energy use patterns. The 24-vectors ui give the average (electric) en-
ergy use for N customers over some period (say, a month) for each hour of
the day. A clustering algorithm partitions customers into groups, each with
similar patterns of daily energy consumption. We might expect a clustering
algorithm to discover which customers have a swimming pool, an electric
water heater, or solar panels.

Financial sectors. For each of N companies we have an n-vector whose com-

ponents are financial and business attributes such as total capitalization,
quarterly returns and risks, trading volume, profit and loss, or dividends
paid. (These quantities would typically be scaled so as to have similar ranges
of values.) A clustering algorithm would group companies into sectors, i.e.,
groups of companies with similar attributes.

In each of these examples, it would be quite informative to know that the vectors
can be well clustered into, say, k = 5 or k = 37 groups. This can be used to develop
insight into the data. By examining the clusters we can often understand them,
and assign labels or descriptions to them.
58 4 Clustering

4.2 A clustering objective

In this section we formalize the idea of clustering, and introduce a natural measure
of the quality of a given clustering.

Specifying the cluster assignments. We specify a clustering of the vectors by

saying which cluster or group each vector belongs to. We label the groups 1, . . . ,
k, and specify a clustering or assignment of the N given vectors to groups using an
N -vector c, where ci is the group (number) that vector i is assigned to. As a simple
example with N = 5 vectors and k = 3 groups, c = (3, 1, 1, 1, 2) means that x1 is
assigned to group 3, x2 , x3 , and x4 are assigned to group 1, and x5 is assigned to
group 2. We will also describe the clustering by the sets of indexes for each group.
We let Gj be the set of indexes corresponding to group j. For our simple example
above, we have

G1 = {2, 3, 4}, G2 = {5}, G3 = {1}.

(Here we are using the notation of sets; see appendix A.) Formally, we can express
these index sets in terms of the group assignment vector c as

Gj = {i | ci = j},

which means that Gj is the set of all indexes i for which ci = j.

Group representatives. With each of the groups we associate a group represen-

tative n-vector, which we denote z1 , . . . , zk . These representatives can be any
n-vectors; they do not need to be one of the given vectors. We want each rep-
resentative to be close to the vectors in its associated group, i.e., we want the
quantities
kxi zci k
to be small. (Note that xi is in group j = ci , so zci is the representative vector
associated with data vector xi .)

A clustering objective. We can now give a single number that we use to judge a
choice of clustering, along with a choice of the group representatives. We define

J clust = kx1 zc1 k2 + + kxN zcN k2 /N,

(4.1)

which is the mean squared distance from the vectors to their associated represen-
tatives. Note that J clust depends on the cluster assignments (i.e., c), as well as the
choice of the group representatives z1 , . . . , zk . The smaller J clust is, the better the
clustering. An extreme case is J clust = 0, which means that the distance between
every original vector and its assigned representative is zero. This happens only
when the original collection of vectors only take k different values, and each vector
is assigned to the representative it is equal to. (This extreme case would probably
not occur in practice.)
Our choice of clustering objective J clust makes sense, since it encourages all
points to be near their associated representative, but there are other reasonable
4.2 A clustering objective 59

choices. For example, it is possible to use an objective that encourages more

balanced groupings. But we will stick with this basic (and very common) choice of
clustering objective.

Optimal and suboptimal clustering. We seek a clustering, i.e., a choice of group

assignments c1 , . . . , cN and a choice of representatives z1 , . . . , zk , that minimize the
objective J clust . We call such a clustering optimal. Unfortunately, for all but the
very smallest problems, it is practically impossible to find an optimal clustering. (It
can be done in principle, but the amount of computation needed grows extremely
rapidly with N .) The good news is that the k-means algorithm described in the
next section requires far less computation (and indeed, can be run for problems
with N measured in billions) and often finds a very good, if not the absolute best,
clustering. (Here, very good means a clustering and choice of representatives
that achieves a value of J clust near its smallest possible value.) We say that the
clustering choices found by the k-means algorithm are suboptimal, which means
that they might not give the lowest possible value of J clust .
Even though it is a hard problem to choose the best clustering and the best
representatives, it turns out that we can find the best clustering, if the representa-
tives are fixed, and we can find the best representatives, if the clustering is fixed.
We address these two topics now.

Partitioning the vectors with the representatives fixed. Suppose that the group
representatives z1 , . . . , zk are fixed, and we seek the group assignments c1 , . . . , cN
that achieve the smallest possible value of J clust . It turns out that this problem
can be solved exactly.
The objective J clust is a sum of N terms. The choice of ci (i.e., the group
to which we assign the vector xi ) only affects the ith term in J clust , which is
(1/N )kxi zci k2 . We can choose ci to minimize just this term, since ci does not
affect the other N 1 terms in J clust . How do we choose ci to minimize this term?
This is easy: We simply choose ci to be the value of j that minimizes kxi zj k
over j. In other words, we should assign each data vector xi to its nearest neighbor
among the representatives. This choice of assignment is very natural, and easily
carried out.
So when the group representatives are fixed, we can readily find the best group
assignment (i.e., the one that minimizes J clust ), by assigning each vector to its
nearest representative. With this choice of group assignment, we have (by the way
the assignment is made)

kxi zci k = min kxi zj k,

j=1,...,k

so the value of J clust is given by

 
min kx1 zj k2 + + min kxN zj k2 /N.
j=1,...,k j=1,...,k

This has a simple interpretation: It is the mean of the squared distance from the
data vectors to their closest representative.
60 4 Clustering

Optimizing the group representatives with the assignment fixed. Now we turn
to the problem of choosing the group representatives, with the clustering (group
assignments) fixed, in order to minimize our objective J clust . It turns out that this
problem also has a simple and natural solution.
We start by re-arranging the sum of N terms into k sums, each associated with
one group. We write
J clust = J1 + + Jk ,
where X
Jj = (1/N ) kxi zj k2
iGj

is the contribution to the objective J clust from the vectors in group j. (The sum
here means that we should add up all terms of the form kxi zj k2 , for any i Gj ,
i.e., for any vector xi in group j; see appendix A.)
The choice of group representative zj only affects the term Jj ; it has no effect
on the other terms in J clust . So we can choose each zj to minimize Jj . Thus we
should choose the vector zj so as to minimize the mean square distance to the
vectors in group j. This problem has a very simple solution: We should choose zj
to be the average (or mean or centroid) of the vectors xi in its group:
X
zj = (1/|Gj |) xi ,
iGj

where |Gj | is standard mathematical notation for the number of elements in the
set Gj , i.e., the size of group j.
So if we fix the group assignments, we minimize J clust by choosing each group
representative to be the average or centroid of the vectors assigned to its group.
(This is sometimes called the group centroid or cluster centroid.)

4.3 The k-means algorithm

It might seem that we can now solve the problem of choosing the group assignments
and the group representatives to minimize J clust , since we know how to do this
when one or the other choice is fixed. But the two choices are circular, i.e., each
depends on the other. Instead we rely on a very old idea in computation: We
simply iterate between the two choices. This means that we repeatedly alternate
between updating the group assignments, and then updating the representatives,
using the methods developed above. In each step the objective J clust gets better
(i.e., goes down) unless the step does not change the choice. This is the celebrated
k-means algorithm for clustering a collection of vectors.

Algorithm 4.1 k-means algorithm

given a list of N vectors x1 , . . . , xN , and an initial list of k group representative
vectors z1 , . . . , zk
repeat until convergence
4.3 The k-means algorithm 61

1. Partition the vectors into k groups. For each vector i = 1, . . . , N , assign xi to

the group associated with the nearest representative.
2. Update representatives. For each group j = 1, . . . , k, set zj to be the mean of
the vectors in group j.

Comments and clarifications.

Ties in step 1 can be broken by assigning xi to the group associated with one
of the closest representatives with the smallest value of j.

It is possible that in step 1, one or more of the groups can be empty, i.e.,
contain no vectors. In this case we simply drop this group (and its repre-
sentative). When this occurs, we end up with a partition of the vectors into
fewer than k groups.

If the group assignments found in step 1 are the same in two successive
iterations, the representatives in step 2 will also be the same. It follows that
the groups assignments and group representatives will never change in future
iterations, so we should stop the algorithm. This is what we mean by until
convergence.

We start the algorithm with a choice of initial group representatives. One

simple method is to pick the representatives randomly from the original vec-
tors; another is to start from a random assignment of the original vectors
to k groups, and use the means of the groups as the initial representatives.
(There are more sophisticated methods for choosing an initial representives,
but this topic is beyond the scope of this book.)

One iteration of the k-means algorithm is illustrated in figure 4.2.

The k-means algorithm was first proposed in 1957 by Stuart Lloyd, and inde-
pendently by Hugo Steinhaus. It is sometimes called the Lloyd algorithm. The
name k-means has been used since the 1960s.

Convergence. The fact that J clust decreases in each step implies that the k-means
algorithm converges in a finite number of steps. However, depending on the initial
choice of representatives, the algorithm can, and does, converge to different final
partitions, with different objective values.
The k-means algorithm is a heuristic, which means it cannot guarantee that the
partition it finds minimizes our objective J clust . For this reason it is common to
run the k-means algorithm several times, with different initial representatives, and
choose the one among them with the smallest final value of J clust . Despite the fact
that the k-means algorithm is a heuristic, it is very useful in practical applications,
and very widely used.
Figure 4.3 shows a few iterations generated by the k-means algorithm, applied
to the example of figure 4.1. We take k = 3 and start with randomly chosen group
representatives. The final clustering is shown in figure 4.4. Figure 4.5 shows how
the clustering objective decreases in each step.
62 4 Clustering

Figure 4.2 One iteration of the k-means algorithm. The 30 2-vectors xi

are shown as filled circles, and the 3 group representatives zj are shown as
rectangles. In the left-hand figure the vectors are each assigned to the closest
representative. On the right-hand figure, the representatives are replaced by
the cluster centroids.

Interpreting the representatives. The representatives z1 , . . . , zk associated with

a clustering are quite interpretable. Suppose, for example, that voters in some
election can be well clustered into 7 groups, on the basis of a data set that includes
demographic data and questionnaire or poll data. If the 4th component of our
vectors is the age of the voter, then (z3 )4 = 37.8 tells that the average of voters
in group 3 is 37.8. Insight gained from this data can be used to tune campaign
messages, or choose media outlets for campaign advertising.

Choosing k. It is common to run the k-means algorithm for different values of k,

and compare the results. How to choose a value of k among these depends on how
the clustering will be used, which we discuss a bit more in 4.5. But some general
statements can be made. For example, if the value of J clust with k = 7 is quite a
bit smaller than for the values k = 2, . . . , 6, and not much larger than the values
for k = 8, 9, . . . , we could reasonably choose k = 7, and conclude that our data
(list of vectors) partitions nicely into 7 groups.

Validation. A general method called out of sample validation or just validation

can be used to determine if k, the number of clusters, is too large. We take our
original list of vectors and divide it into two sets, which need not have the same size.
We then carry out k-means clustering using the data set (called the training data
set), and note the clustering objective J clust obtained, i.e., mean square distance of
the data points to the representatives. Then we evaluate the mean square distance
of the vectors in the second set (called the test data set or validation data set) to
4.3 The k-means algorithm 63

Iteration 1

Iteration 2

Iteration 10

Figure 4.3 Three iterations of the k-means algorithm. The group represen-
tatives are shown as large stars. In each row, the left-hand plot shows the
result of partitioning the vectors in the 3 groups (step 1 of algorithm 4.1).
The right-hand plot shows the updated representatives (step 2 of the algo-
rithm).
64 4 Clustering

Figure 4.4 Final clustering.

1.5
J clust

0.5

1 3 5 7 9 11 13 15
Iteration

Figure 4.5 The clustering objective J clust after step 1 of each iteration.
 4.4 Examples 65

the representatives found using the training data set. Roughly speaking, this test
tells us whether our clustering works on data that it has not seen. If the mean
square distance for the training and test sets are reasonably close to other, we
conclude that k is not too big; if the mean square distance for the test set is much
larger, we conclude that k is too large. We will encounter the idea of validation
again, and discuss it in more detail, in 13.2.

Complexity. In step 1 of the k-means algorithm, we find the nearest neighbor

to each of N n-vectors, over the list of k centroids. This requires approximately
3N kn flops. In step 2 we average the n-vectors over each of cluster groups. For
a cluster with p vectors, this requires n(p 1) flops, which we approximate as np
flops; averaging all clusters requires a total of N n flops. This is negligible compared
to the cost of partitioning in step 1. So k-means requires around 3N kn flops per
iteration. Its order is N kn flops.
Each run of k-means typically takes fewer than a few tens of iterations, and
usually k-means is run some modest number of times, like 10. So a very rough
guess of the number of flops required to run k-means 10 times (in order to choose
the best partition found) is 1000N kn flops. As an example, suppose we use k-
means to partition N = 100000 vectors with size n = 100 into k = 10 groups.
On a 1 Gflop/s computer we guess that this will take around 100 seconds, i.e., on
the order of one minute. Given the approximations made here (for example, the
number of iterations that each run of k-means will take) this is obviously a crude
estimate.

4.4 Examples
4.4.1 Image clustering

The MNIST (Mixed National Institute of Standards) database of handwritten dig-

its, available at [Link]/exdb/mnist, is a data set containing N = 60, 000
grayscale images of size 2828, which we represent as n-vectors with n = 2828 =
784. Figure 4.6 shows a few examples from the data set.
We use the k-means algorithm to partition these images into k = 20 clusters,
starting with a random assignment of the vectors to groups, and repeating the
experiment 20 times. Figure 4.7 shows the clustering objective versus iteration
number for three of the 20 initial assignments, including the two that gave the
lowest and the highest final values of the objective. Figure 4.8 shows the repre-
sentatives with the lowest final value of the clustering objective. Figure 4.9 shows
the set with the highest value. We can see that most of the representatives are
recognizable digits, with some reasonable confusion, for example between 4 and
9 or 3 and 8. This is impressive when you consider that the k-means algorithm
knows nothing about digits, handwriting, or even that the 784-vectors represent
28 28 images; it uses only the distances between 784-vectors. One interpretation
is that the k-means algorithm has discovered the digits in the data set.
66 4 Clustering

Figure 4.6 25 images of handwritten digits from the MNIST data set. Each
image has size 28 28, and can be represented by a 784-vector.

42

40
J clust

38

36

1 5 9 13 17 21 25
Iteration

Figure 4.7 Clustering objective J clust after each iteration of the k-means
algorithm, for three initial partitions, on digits of the MNIST set.
4.4 Examples 67

Figure 4.8 Group representatives found by the k-means algorithm applied

to the MNIST set.

Figure 4.9 Group representatives found by the k-means algorithm applied

to the MNIST set, with a different starting point than in figure 4.8.
68 4 Clustering

103

J clust 7.5

1 5 10 15 20
Iteration

Figure 4.10 Clustering objective J clust after each iteration of the k-means
algorithm, for three initial partitions, on Wikipedia word count histograms.

We can validate the clustering found using a separate test set consisting of
10000 images of handwritten digits. The objective value on the training set (60000
images) is J clust = 35.17; the objective value on the test set (10000 images), with
the cluster representatives found from the training set, is 35.09. These numbers
are very close, so we conclude that k = 20 is not too large.

4.4.2 Document topic discovery

We start with a corpus of N = 500 Wikipedia articles, compiled from weekly lists
of the most popular articles between September 6, 2015, and June 11, 2016. We
remove the section titles and reference sections (bibliography, notes, references,
further reading), and convert each document to a list of words. The conversion
removed numbers and stop words, and applied a stemming algorithm to nouns and
verbs. We then formed a dictionary of all the words that appear in at least 20
documents. This resulted in a dictionary of 4423 words. Each document in the
corpus is represented by a word histogram vector of length 4423.
We apply the k-means algorithm with k = 9, and 20 randomly chosen initial
partitions. The k-means algorithm converges to similar but slightly different clus-
terings of the documents in each case. Figure 4.10 shows the clustering objective
versus iteration of the k-means algorithm for three of these, including the one that
gave the lowest final value of J clust , which we use below.
Table 4.1 summarizes the clustering with the lowest value of J clust . For each of
the nine clusters we show the largest ten coefficients of the word histogram of the
cluster representative. Table 4.2 gives the size of each cluster and the titles of the
4.4 Examples 69

Cluster 1 Cluster 2 Cluster 3

Word Coefficient Word Coefficient Word Coefficient
fight 0.038 holiday 0.012 united 0.004
win 0.022 celebrate 0.009 family 0.003
event 0.019 festival 0.007 party 0.003
champion 0.015 celebration 0.007 president 0.003
fighter 0.015 calendar 0.006 government 0.003
bout 0.015 church 0.006 school 0.003
title 0.013 united 0.005 american 0.003
ali 0.012 date 0.005 university 0.003
championship 0.011 moon 0.005 city 0.003
bonus 0.010 event 0.005 attack 0.003

Cluster 4 Cluster 5 Cluster 6

Word Coefficient Word Coefficient Word Coefficient
album 0.031 game 0.023 series 0.029
release 0.016 season 0.020 season 0.027
song 0.015 team 0.018 episode 0.013
music 0.014 win 0.017 character 0.011
single 0.011 player 0.014 film 0.008
record 0.010 play 0.013 television 0.007
band 0.009 league 0.010 cast 0.006
perform 0.007 final 0.010 announce 0.006
tour 0.007 score 0.008 release 0.005
chart 0.007 record 0.007 appear 0.005

Cluster 7 Cluster 8 Cluster 9

Word Coefficient Word Coefficient Word Coefficient
match 0.065 film 0.036 film 0.061
win 0.018 star 0.014 million 0.019
championship 0.016 role 0.014 release 0.013
team 0.015 play 0.010 star 0.010
event 0.015 series 0.009 character 0.006
style 0.014 appear 0.008 role 0.006
raw 0.013 award 0.008 movie 0.005
title 0.011 actor 0.007 weekend 0.005
episode 0.010 character 0.006 story 0.005
perform 0.010 release 0.006 gross 0.005

Table 4.1 The 9 cluster representatives. For each representative we show

the largest 10 coefficients of the word histogram.
70 4 Clustering

ten articles closest to the cluster representative.

Each of the clusters makes sense, and mostly contains documents on similar
topics, or similar themes. The words with largest coefficients in the group rep-
resentatives also make sense. It is interesting to see that k-means clustering has
assigned movies and TV series (mostly) to different clusters (9 and 6). One can
also note that clusters 8 and 9 share several top key words but are on separate
topics (actors and movies, respectively).
The identification of these separate topics among the documents is impressive,
when you consider that the k-means algorithm does not understand the meaning of
the words in the documents (and indeed, does not even know the order of the words
in each document). It uses only the simple concept of document dissimilarity, as
measured by the distance between word count histogram vectors.

4.5 Applications
Clustering, and the k-means algorithm in particular, has many uses and applica-
tions. It can be used for exploratory data analysis, to get an idea of what a large
collection of vectors looks like. When k is small enough, say less than a few tens,
it is common to examine the group representatives, and some of the vectors in the
associated groups, to interpret or label the groups. Clustering can also be used for
more specific directed tasks, a few of which we describe below.

Classification. We cluster a large collection of vectors into k groups, and label

the groups by hand. We can now assign (or classify) new vectors to one of the
k groups by choosing the nearest group representative. In our example of the
handwritten digits above, this would give us a rudimentary digit classifier, which
would automatically guess what a written digit is from its image. In the topic
discovery example, we can automatically classify new documents into one of the k
topics. (We will see better classification methods in chapter 14.)

Recommendation engine. We describe an example of how clustering can be used

to build a recommendation engine, an algorithm that suggests items that a user
or customer might be interested in. Suppose the vectors give the number of times
a user has listened to or streamed each song from a library of n songs, say over
the last month. These vectors are typically sparse, since each user has listened to
only a very small fraction of the total music library. Clustering the vectors reveals
groups of users with similar musical taste (or at least, similar listening habits). The
group representatives have a very nice interpretation: (zj )i is the average number
of times users in group j have listened to song i.
This interpretation allows us to create a set of recommendations for each user.
We first identify which cluster j her music listening vector xi is in. Then we can
suggest to her songs that she has not listened to, but others in her group (i.e.,
those with similar musical tastes) have listened to most often. To recommend 5
songs to her, we find the indices l with (xi )l = 0, with the 5 largest values of (zj )l .
4.5 Applications 71

Cluster Size Titles

1 21 Floyd Mayweather, Jr; Kimbo Slice; Ronda Rousey; Jose
Aldo; Joe Frazier; Wladimir Klitschko; Saul Alvarez;
Gennady Golovkin; Nate Diaz; Conor McGregor.
2 43 Halloween; Guy Fawkes Night; Diwali; Hanukkah;
Groundhog Day; Rosh Hashanah; Yom Kippur;
Seventh-day Adventist Church; Remembrance Day;
Mothers Day.
3 189 Mahatma Gandhi; Sigmund Freud; Carly Fiorina;
Frederick Douglass; Marco Rubio; Christopher Columbus;
Fidel Castro; Jim Webb; Genie (feral child); Pablo
Escobar.
4 46 David Bowie; Kanye West; Celine Dion; Kesha; Ariana
Grande; Adele; Gwen Stefani; Anti (album); Dolly Parton;
Sia Furler.
5 49 Kobe Bryant; Lamar Odom; Johan Cruyff; Yogi Berra;
Jose Mourinho; Halo 5: Guardians; Tom Brady; Eli
Manning; Stephen Curry; Carolina Panthers.
6 39 The X-Files; Game of Thrones; House of Cards (U.S. TV
series); Daredevil (TV series); Supergirl (U.S. TV series);
American Horror Story; The Flash (2014 TV series); The
Man in the High Castle (TV series); Sherlock (TV series);
Scream Queens (2015 TV series).
7 16 Wrestlemania 32; Payback (2016); Survivor Series (2015);
Royal Rumble (2016); Night of Champions (2015);
Fastlane (2016); Extreme Rules (2016); Hell in a Cell
(2015); TLC: Tables, Ladders & Chairs (2015); Shane
McMahon.
8 58 Ben Affleck; Johnny Depp; Maureen OHara; Kate
Beckinsale; Leonardo DiCaprio; Keanu Reeves; Charlie
Sheen; Kate Winslet; Carrie Fisher; Alan Rickman.
9 39 Star Wars: The Force Awakens; Star Wars Episode I: The
Phantom Menace; The Martian (film); The Revenant
(2015 film); The Hateful Eight; Spectre (2015 film); The
Jungle Book (2016 film); Bajirao Mastani (film); Back to
the Future II; Back to the Future.

Table 4.2 Cluster sizes and titles of 10 documents closest to the cluster
representatives.
72 4 Clustering
 Chapter 5

Linear independence

In this chapter we explore the concept of linear independence, which will play an
important role in the sequel.

5.1 Linear dependence

A collection or list of n-vectors a1 , . . . , ak (with k 1) is called linearly dependent
if
1 a1 + + k ak = 0
holds for some 1 , . . . , k that are not all zero. In other words, we can form the
zero vector as a linear combination of the vectors, with coefficients that are not all
zero. Linear dependence of a list of vectors does not depend on the ordering of the
vectors in the list.
When a collection of vectors is linearly dependent, at least one of the vectors
can be expressed as a linear combination of the other vectors: If i 6= 0 in the
equation above (and by definition, this must be true for at least one i), we can
move the term i ai to the other side of the equation and divide by i to get

ai = (1 /i )a1 + + (i1 /i )ai1 + (i+1 /i )ai+1 + + (k /i )ak .

The converse is also true: If any vector in a collection of vectors is a linear combi-
nation of the other vectors, then the collection of vectors is linearly dependent.
Following standard mathematical language usage, we will say The vectors a1 ,
. . . , ak are linearly dependent to mean The list of vectors a1 , . . . , ak is linearly
dependent. But it must be remembered that linear dependence is an attribute of
a collection of vectors, and not individual vectors.

Linearly independent vectors. A collection of n-vectors a1 , . . . , ak (with k 1)

is called linearly independent if it is not linearly dependent, which means that

1 a1 + + k ak = 0 (5.1)
74 5 Linear independence

only holds for 1 = = k = 0. In other words, the only linear combination of

the vectors that equals the zero vector is the linear combination with all coefficients
zero.
As with linear dependence, we will say The vectors a1 , . . . , ak are linearly
independent to mean The list of vectors a1 , . . . , ak is linearly independent. But,
like linear dependence, linear independence is an attribute of a collection of vectors,
and not individual vectors.
It is generally not easy to determine by casual inspection whether or not a list
of vectors is linearly dependent or linearly independent. But we will soon see an
algorithm that does this.

Examples.
A list consisting of a single vector is linearly dependent only if the vector is
zero. It is linearly independent only if the vector is nonzero.
Any list of vectors containing the zero vector is linearly dependent.
A list of two vectors is linearly dependent if and only if one of the vectors
is a multiple of the other one. More generally, a list of vectors is linearly
dependent if any one of the vectors is a multiple of another one.
The vectors

0.2 0.1 0
a1 = 7 , a2 = 2 , a3 = 1
8.6 1 2.2
are linearly dependent, since a1 + 2a2 3a3 = 0. We can express any of
these vectors as a linear combination of the other two. For example, we have
a2 = (1/2)a1 + (3/2)a3 .
The vectors

1 0 1
a1 = 0 , a2 = 1 , a3 = 1
0 1 1
are linearly independent. To see this, suppose 1 a1 + 2 a2 + 3 a3 = 0. This
means that
1 3 = 0, 2 + 3 = 0, 2 + 3 = 0.
Adding the last two equations we find that 23 = 0, so 3 = 0. Using this,
the first equation is then 1 = 0, and the second equation is 2 = 0.
The standard unit n-vectors e1 , . . . , en are linearly independent. To see this,
suppose that (5.1) holds. We have

1
0 = 1 e1 + + n en = ... ,

n
so we conclude that 1 = = n = 0.
 5.2 Basis 75

Linear combinations of linearly independent vectors. Suppose a vector x is a

linear combination of a1 , . . . , ak ,

x = 1 a1 + + k ak .

When the vectors a1 , . . . , ak are linearly independent, the coefficients that form x
are unique: If we also have

x = 1 a1 + + k ak ,

then i = i for i = 1, . . . , k. This tells us that, in principle at least, we can find

the coefficients that form a vector x as a linear combination of linearly independent
vectors.
To see this, we subtract the two equations above to get

0 = (1 1 )a1 + + (k k )ak .

Since a1 , . . . , ak are linearly independent, we conclude that i i are all zero.

The converse is also true: If each linear combination of a list of vectors can
only be expressed as a linear combination with one set of coefficients, then the
list of vectors is linearly independent. This gives a nice interpretation of linear
independence: A list of vectors is linearly independent if and only if for any linear
combination of them, we can infer or deduce the associated coefficients. (We will
see later how to do this.)

Supersets and subsets. If a collection of vectors is linearly dependent, then any

superset of it is linearly dependent. In other words: If we add vectors to a linearly
dependent collection of vectors, the new collection is also linearly dependent. Any
nonempty subset of a linearly independent collection of vectors is linearly inde-
pendent. In other words: Removing vectors from a collection of vectors preserves
linear independence.

5.2 Basis
Independence-dimension inequality. If the n-vectors a1 , . . . , ak are linearly in-
dependent, then k n. In words:

A linearly independent collection of n-vectors can have at most n elements.

Put another way:

Any collection of n + 1 or more n-vectors is linearly dependent.

We will prove this fundamental fact below; but first, we describe the concept of
basis, which relies on the independence-dimension inequality.
76 5 Linear independence

Basis. A collection of n linearly independent n-vectors (i.e., a collection of linearly

independent vectors of the maximum possible size) is called a basis. If the n-vectors
a1 , . . . , an are a basis, then any n-vector b can be written as a linear combination
of them. To see this, consider the collection of n + 1 n-vectors a1 , . . . , an , b. By the
independence-dimension inequality, these vectors are linearly dependent, so there
are 1 , . . . , n+1 , not all zero, that satisfy
1 a1 + + n an + n+1 b = 0.
If n+1 = 0, then we have
1 a1 + + n an = 0,
which, since a1 , . . . , an are linearly independent, implies that 1 = = n = 0.
But then all the i are zero, a contradiction. So we conclude that n+1 6= 0. It
follows that
b = (1 /n+1 )a1 + + (n /n+1 )an ,
i.e., b is a linear combination of a1 , . . . , an .
Combining this result with the observation above that any linear combination
of linearly independent vectors can be expressed in only one way, we conclude:
Any n-vector b can be written in a unique way as a linear combination
of a basis a1 , . . . , an .

Expansion in a basis. When we express an n-vector b as a linear combination of

a basis a1 , . . . , an , we refer to
b = 1 a1 + + n an ,
as the expansion of b in the a1 , . . . , an basis. The numbers 1 , . . . , n are called
the coefficients of the expansion of b in the basis a1 , . . . , an . (We will see later how
to find the coefficients in the expansion of a vector in a basis.)

Examples.
The n standard unit n vectors e1 , . . . , en are a basis. Any n-vector b can be
written as the linear combination
b = b1 e 1 + + bn e n .
This expansion is unique, which means that there is no other linear combi-
nation of e1 , . . . , en that equals b.
The vectors    
1.2 0.3
a1 = , a2 =
2.6 3.7
are a basis. The vector b = (1, 1) can be expressed in only one way as a linear
combination of them:
b = 0.6513 a1 0.7280 a2 .
(The coefficients are given here to 4 significant places. We will see later how
these coefficients can be computed.)
5.2 Basis 77

Cash flows and single period loans. As a practical example, we consider cash
flows over n periods, with positive entries meaning income or cash in and negative
entries meaning payments or cash out. We define the single-period loan cash flow
vectors as
li = (0i1 , 1, (1 + r), 0ni1 ), i = 1, . . . , n 1,

where r 0 is the per-period interest rate. The cash flow li represents a loan of $1
in period i, which is paid back in period i + 1 with interest r. (The subscripts on
the zero vectors above give their dimensions.) Scaling li changes the loan amount;
scaling li by a negative coefficient converts it into a loan to another entity (which
is paid back in period i + 1 with interest).
The vectors e1 , l1 , . . . , ln1 are a basis. (The first vector e1 represents income of
$1 in period 1.) To see this, we show that they are linearly independent. Suppose
that
1 e1 + 2 l2 + + n ln1 = 0.

We can express this as

(1 + 2 , 3 (1 + r)2 , . . . , n (1 + r)n1 , (1 + r)n ) = 0.

The last entry is (1 + r)n = 0, which implies that n = 0 (since 1 + r > 0).
Using n = 0, the second to last entry is (1 + r)n1 = 0, so we conclude that
n1 = 0. Continuing this way we find that n2 , . . . , 2 = 0 are all zero. The first
entry of the equation above, 1 + 2 = 0, then implies 1 = 0. We conclude that
the vectors e1 , l1 , . . . , ln=1 are linearly independent, and therefore a basis.
This means that any cash flow n-vector c can be expressed as a linear combi-
nation of (i.e., replicated by) an initial payment and one period loans:

c = 1 e1 + 2 l1 + + n ln1 .

In this example we can work out what the coefficients are. Using a similar
argument as the one above that establishes linear independence we get
cn cn1 cn
n = , n1 = ,
1+r 1 + r (1 + r)2

and following this pattern,

c2 cn
2 = .
1+r (1 + r)n1

Finally, we have
c2 cn
1 = c1 + + + ,
1+r (1 + r)n1
which is exactly the net present value (NPV) of the cash flow, with interest rate r.
Thus we see that any cash flow can be replicated as an income in period 1 equal
to its net present value, plus a linear combination of one-period loans at interest
rate r.
78 5 Linear independence

Proof of independence-dimension inequality. The proof is by induction on the di-

mension n. First consider a linearly independent collection a1 , . . . , ak of 1-vectors.
We must have a1 6= 0. This means that every element ai of the collection can be
expressed as a multiple ai = (ai /a1 )a1 of the first element a1 . This contradicts
linear independence unless k = 1.
Next suppose n 2 and the independence-dimension inequality holds for di-
mension n 1. Let a1 , . . . , ak be a linearly independent list of n-vectors. We need
to show that k n. We partition the vectors as
 
bi
ai = , i = 1, . . . , k,
i

where bi is an (n 1)-vector and i is a scalar.

Pkthat 1 = = k = 0. Then theP
First suppose vectors b1 , . . . , bk are linearly
k
independent:
i=1 i i b = 0 holds if and only if i=1 i ai = 0, which is only
possible for 1 = = k = 0 because the vectors ai are linearly independent.
The vectors b1 , . . . , bk therefore form a linearly independent collection of (n 1)-
vectors. By the induction hypothesis we have k n 1, so certainly k n.
Next suppose that the scalars i are not all zero. Assume j 6= 0. We define a
collection of k 1 vectors ci of length n 1 as follows:
i i+1
c i = bi bj , i = 1, . . . , j 1, ci = bi+1 bj , i = j, . . . , k 1.
j j
Pk1
These k 1 vectors are linearly independent: If i=1 i ci = 0 then
j1     k  
X bi bj X bi
i + + i1 =0 (5.2)
i j i
i=1 i=j+1

with
j1 k
1 X X
= ( i i + i1 i ).
j i=1 i=j+1

Since the vectors ai = (bi , i ) are linearly independent, the equality (5.2) only
holds when all the coefficients i and are all zero. This in turns implies that
the vectors c1 , . . . , ck1 are linearly independent. By the induction hypothesis
k 1 n 1, so we have established that k n.

5.3 Orthonormal vectors

A collection of vectors a1 , . . . , ak is orthogonal or mutually orthogonal if ai aj for
any i, j with i 6= j, i, j = 1, . . . , k. A collection of vectors a1 , . . . , ak is orthonormal
if it is orthogonal and kai k = 1 for i = 1, . . . , k. (A vector of norm one is called
normalized ; dividing a vector by its norm is called normalizing it.) Thus, each
vector in an orthonormal collection of vectors is normalized, and any pair of two
different vectors from the collection are orthogonal. These two conditions can be
5.3 Orthonormal vectors 79

Figure 5.1 Orthonormal vectors in a plane.

combined into one statement about the inner products of pairs of vectors in the
collection: a1 , . . . , ak is orthonormal means that

1 i=j
aTi aj =
0 i 6= j.

Orthonormality, like linear dependence and independence, is an attribute of a col-

lection of vectors, and not an attribute of vectors individually. By convention,
though, we say The vectors a1 , . . . , ak are orthonormal to mean The collection
of vectors a1 , . . . , ak is orthonormal.

Examples. The standard unit n-vectors e1 , . . . , en are orthonormal. As another

example, the 3-vectors

0 1 1
1 1
0 , 1 , 1 , (5.3)
1 2 0 2 0

are orthonormal. Figure 5.1 shows a set of two orthonormal 2-vectors.

Independence of orthonormal vectors. Orthonormal vectors are linearly inde-

pendent. To see this, suppose a1 , . . . , ak are orthonormal, and

1 a1 + + k ak = 0.

Taking the inner product of this equality with ai yields

0 = aTi (1 a1 + + k ak )
= 1 (aTi a1 ) + + k (aTi ak )
= i ,

since aTi aj = 0 for j 6= i and aTi ai = 1. Thus, the only linear combination of a1 ,
. . . , ak that is zero is the one with all coefficients zero.

Linear combinations of orthonormal vectors. Suppose a vector x is a linear com-

bination of a1 , . . . , ak , where a1 , . . . , ak are orthonormal,

x = 1 a1 + + k ak .

Taking the inner product of the left- and right-hand sides of this equation with ai
yields
aTi x = aTi (1 a1 + + k ak ) = i ,
80 5 Linear independence

using the same argument as above. So if a vector x is a linear combination of

orthonormal vectors, we can easily find the coefficients of the linear combination
by taking the inner products with the vectors.
For any x that is a linear combination of orthonormal vectors a1 , . . . , ak , we
have the identity
x = (aT1 x)a1 + + (aTk x)ak . (5.4)
This identity gives us a simple way to check if an n-vector y is a linear combination
of the orthonormal vectors a1 , . . . , ak . If the identity (5.4) holds for y, i.e.,

y = (aT1 y)a1 + + (aTk y)ak ,

then (evidently) y is a linear combination of a1 , . . . , ak ; conversely, if y is a linear

combination of a1 , . . . , ak , the identity (5.4) holds for y.

Orthonormal basis. If the n-vectors a1 , . . . , an are orthonormal, they are linearly

independent, and therefore also a basis. In this case they are called an orthonormal
basis. The three examples above (on page 79) are orthonormal bases.
If a1 , . . . , an is an orthonormal basis, then we have, for any n-vector x, the
identity
x = (aT1 x)a1 + + (aTn x)an . (5.5)
To see this, we note that since a1 , . . . , an are a basis, x can be expressed as a linear
combination of them; hence the identity (5.4) above holds. The equation above is
sometimes called the orthonormal expansion formula; the right-hand side is called
the expansion of x in the basis a1 , . . . , an . It shows that any n-vector can be
expressed as a linear combination of the basis elements, with the coefficients given
by taking the inner product of x with the elements of the basis.
As an example, we express the 3-vector x = (1, 2, 3) as a linear combination of
the orthonormal basis given in (5.3). The inner products of x with these vectors
are
T T T
0 1 1
1 3 1 1
0 x = 3, 1 x= , 1 x = .
1 2 0 2 2 0 2

It can be verifed that the expansion of x in this basis is

0 1 1
3 1 1 1
x = (3) 0 + 1 + 1 .
1 2 2 0 2 2 0

5.4 Gram-Schmidt algorithm

In this section we describe an algorithm can be used to determine if a list of n-
vectors a1 , . . . , ak is linearly independent. In later chapters we will see that it has
many other uses as well. The algorithm is named after the mathematicians Jrgen
5.4 Gram-Schmidt algorithm 81

Pedersen Gram and Erhard Schmidt, although it was already known before their
work.
If the vectors are linearly independent, The Gram-Schmidt algorithm produces
an orthonormal collection of vectors q1 , . . . , qk with the following properties: For
each i = 1, . . . , k, ai is a linear combination of q1 , . . . , qi , and qi is a linear
combination of a1 , . . . , ai . If the vectors a1 ,. . . , aj1 are linearly independent, but
a1 ,. . . , aj are linearly dependent, the algorithm detects this and terminates. In
other words, the Gram-Schmidt algorithm finds the first vector aj that is a linear
combination of previous vectors a1 , . . . , aj1 .

Algorithm 5.1 Gram-Schmidt algorithm

given n-vectors a1 , . . . , ak
for i = 1, . . . , k,
1. Orthogonalization. qi = ai (q1T ai )q1 (qi1
T
ai )qi1
2. Test for dependence. if qi = 0, quit.
3. Normalization. qi = qi /kqi k

The orthogonalization step, with i = 1, reduces to q1 = a1 . If the algorithm

does not quit (in step 2), i.e., q1 , . . . , qk are all nonzero, we can conclude that
the original collection of vectors is linearly independent; if the algorithm does quit
early, say, with qj = 0, we can conclude that the original collection of vectors is
linearly dependent (and indeed, that aj is a linear combination of a1 , . . . , aj1 ).

Analysis of Gram-Schmidt algorithm. Let us show that the following hold, for
i = 1, . . . , k, assuming a1 , . . . , ak are linearly independent.

1. qi 6= 0, so the dependence test in step 2 is not satisfied, and we do not have

a divide-by-zero error in step 3.

2. q1 , . . . , qi are orthonormal.

3. ai is a linear combination of q1 , . . . , qi .

4. qi is a linear combination of a1 , . . . , ai .

We show this by induction. For i = 1, we have q1 = a1 . Since a1 , . . . , ak are

linearly independent, we must have a1 6= 0, and therefore q1 6= 0, so assertion 1
holds. The single vector q1 (considered as a list with one element) is evidently
orthonormal, since kq1 k = 1, so assertion 2 holds. We have a1 = kq1 kq1 , and
q1 = (1/kq1 k)a1 , so assertions 3 and 4 hold.
Suppose our assertion holds for some i 1, with i < k; we will show it holds
for i. If qi = 0, then ai is a linear combination of q1 , . . . , qi1 (from the first
step in the algorithm); but each of these is (by the induction hypothesis) a linear
combination of a1 , . . . , ai1 , so it follows that ai is a linear combination of a1 , . . . ,
ai1 , which contradicts our assumption that a1 , . . . , ak are linearly independent.
So assertion 1 holds for i.
82 5 Linear independence

Step 3 of the algorithm ensures that q1 , . . . , qi are normalized; to show they are
orthogonal we will show that qi qj for j = 1, . . . , i 1. (Our induction hypothesis
tells us that qr qs for r, s < i.) For any j = 1, . . . , i 1, we have (using step 1 of
the algorithm)

qjT qi = qjT ai (q1T ai )(qjT q1 ) (qi1

T
ai )(qjT qi1 )
= qjT ai qjT ai = 0,

using qjT qk = 0 for j 6= k and qjT qj = 1. (This explains why step 1 is called the
orthogonalization step: We subtract from ai a linear combination of q1 , . . . , qi1
that ensures qi qj for j < i.) Since qi = (1/kqi k)qi , we have qiT qj = 0 for
j = 1, . . . , i 1. So assertion 2 holds for i.
It is immediate that ai is a linear combination of q1 , . . . , qi :

ai = qi + (q1T ai )q1 + + (qi1

T
ai )qi1
= (q1T ai )q1 + + (qi1
T
ai )qi1 + kqi kqi .

From step 1 of the algorithm, we see that qi is a linear combination of a1 , q1 , . . . ,

qi1 . By the induction hypothesis, each of q1 , . . . , qi1 is a linear combination of
a1 , . . . , ai1 , so qi (and therefore also qi ) is a linear combination of a1 , . . . , ai .
Thus assertions 3 and 4 hold.

Gram-Schmidt completion implies independence. From the properties 14 above,

we can argue that the original collection of vectors a1 , . . . , ak is linearly indepen-
dent. To see this, suppose that

1 a1 + + k ak = 0 (5.6)

holds for some 1 , . . . , k . We will show that 1 = = k = 0.

We first note that any linear combination of q1 , . . . , qk1 is orthogonal to any
multiple of qk , since q1T qk = = qk1
T
qk = 0 (by definition). But each of a1 , . . . ,
ak1 is a linear combination of q1 , . . . , qk1 , so we have qkT a1 = = qkT ak1 = 0.
Taking the inner product of qk with the left- and right-hand sides of (5.6) we obtain

0 = qkT (1 a1 + + k ak )
= 1 qkT a1 + + k1 qkT ak1 + k qkT ak
= k kqk k,

where we use qkT ak = kqk k in the last line. We conclude that k = 0.

From (5.6) and k = 0 we have

1 a1 + + k1 ak1 = 0.

We now repeat the argument above to conclude that k1 = 0. Repeating it k

times we conclude that all i are zero.
5.4 Gram-Schmidt algorithm 83

Early termination. Suppose that the Gram-Schmidt algorithm terminates pre-

maturely, in iteration j, because qj = 0. The conclusions 14 above hold for
i = 1, . . . , j 1, since in those steps qi is nonzero. Since qj = 0, we have

aj = (q1T aj )q1 + + (qj1

T
aj )qj1 ,

which shows that aj is a linear combination of q1 , . . . , qj1 . But each of these

vectors is in turn a linear combination of a1 , . . . , aj1 , by conclusion 3 above.
Then aj is a linear combination of a1 , . . . , aj1 , since it is a linear combination of
linear combinations of them (see page 16). This means that a1 , . . . , aj are linearly
dependent, which implies that the larger set a1 , . . . , ak are linearly dependent.
In summary, the Gram-Schmidt algorithm gives us an explicit method for de-
termining if a list of vectors is linearly dependent or independent.

Example. We define three vectors

a1 = (1, 1, 1, 1), a2 = (1, 3, 1, 3), a3 = (1, 3, 5, 7).

Applying the Gram-Schmidt algorithm gives the following results.

i = 1. We have kq1 k = 2, so
1
q1 = q1 = (1/2, 1/2, 1/2, 1/2),
kq1 k
which is simply a1 normalized.
i = 2. We have q1T a2 = 4, so

1 1/2 1
3 1/2 1
q2 = a2 (q1T a2 )q1 =

1 4 1/2 = 1
,

3 1/2 1

which is indeed orthogonal to q1 (and a1 ). It has norm kq2 k = 2; normalizing

it gives
1
q2 = q2 = (1/2, 1/2, 1/2, 1/2).
kq2 k

i = 3. We have q1T a3 = 2 and q2T a3 = 8, so

q3 = a3 (q1T a3 )q1 (q2T a3 )q2

1 1/2 1/2
3 1/2 1/2
5 2 1/2 8
=
1/2
7 1/2 1/2

2
2
=
,
2
2
84 5 Linear independence

which is orthogonal to q1 and q2 (and a1 and a2 ). We have kq3 k = 4, so the

normalized vector is
1
q3 = q3 = (1/2, 1/2, 1/2, 1/2).
kq3 k
Completion of the Gram-Schmidt algorithm without early termination tells us that
the vectors a1 , a2 , a3 are linearly independent.

Determining if a vector is a linear combination of linearly independent vectors.

Suppose the vectors a1 , . . . , ak are linearly independent, and we wish to determine
if another vector b is a linear combination of them. (We have already noted on
page 75 that if it is a linear combination of them, the coefficients are unique.)
The Gram-Schmidt algorithm provides an explicit way to do this. We apply the
Gram-Schmidt algorithm to the list of k + 1 vectors
a1 , . . . , ak , b.
These vectors are linearly dependent if b is a linear combination of a1 , . . . , ak ;
they are linearly independent if b is not a linear combination of a1 , . . . , ak . The
Gram-Schmidt algorithm will determine which of these two cases holds. It cannot
terminate in the first k steps, since we assume that a1 , . . . , ak are linearly indepen-
dent. It will terminate in the (k+1)st step with qk+1 = 0 if b is a linear combination
of a1 , . . . , ak . It will not terminate in the (k + 1)st step (i.e., qk+1 6= 0), otherwise.

Checking if a collection of vectors is a basis. To check if the n-vectors a1 , . . . ,

an are a basis, we run the Gram-Schmidt algorithm on them. If Gram-Schmidt
terminates early, they are not a basis; if it runs to completion, we know they are a
basis.

Complexity of the Gram-Schmidt algorithm. We now derive an operation count

for the Gram-Schmidt algorithm. In the first step of iteration i of the algorithm,
i 1 inner products
q1T ai , . . . , qi1
T
ai
between vectors of length n are computed. This takes (i 1)(2n 1) flops. We
then use these inner products as the coefficients in i 1 scalar multiplications with
the vectors q1 , . . . , qi1 . This requires n(i 1) flops. We then subtract the i 1
resulting vectors from ai , which requires another n(i1) flops. The total flop count
for step 1 is
(i 1)(2n 1) + n(i 1) + n(i 1) = (4n 1)(i 1)
flops. In step 3 we compute the norm of qi , which takes approximately 2n flops.
We then divide qi by its norm, which requires n scalar divisions. So the total flop
count for the ith iteration is (4n 1)(i 1) + 3n flops.
The total flop count for all k iterations of the algorithm is obtained by summing
our counts for i = 1, . . . , k:
k
X k(k 1)
((4n 1)(i 1) + 3n) = (4n 1) + 3nk 2nk 2 ,
i=1
2
5.4 Gram-Schmidt algorithm 85

Pk
where we use the fact that i=1 (i 1) = k(k 1)/2. The complexity of the
Gram-Schmidt algorithm is 2nk 2 ; its order is nk 2 . We can guess that its running
time grows linearly with the lengths of the vectors n, and quadratically with the
number of vectors k.
In the special case of k = n, the complexity of the Gram-Schmidt method is
2n3 . For example, if the Gram-Schmidt algorithm is used to determine whether a
collection of 1000 1000-vectors is linearly independent (and therefore a basis), the
computational cost is around 2 109 flops. On a modern computer, can we can
expect this to take on the order of one second.
When the Gram-Schmidt algorithm is implemented, a variation on it called the
modified Gram-Schmidt algorithm is typically used. This algorithm produces the
same results as the Gram-Schmidt algorithm (5.1), but is less sensitive to the small
round-off errors that occur when arithmetic calculations are done using floating
point numbers. (We do not consider round-off error in this book.)
86 5 Linear independence
 Part II

Matrices
 Chapter 6

Matrices

In this chapter we introduce matrices and some basic operations on them. We give
some applications in which they arise.

6.1 Matrices
A matrix is a rectangular array of numbers written between rectangular brackets,
as in
0 1 2.3 0.1
1.3 4 0.1 0 .
4.1 1 0 1.7
It is also common to use large parentheses instead of rectangular brackets, as in

0 1 2.3 0.1
1.3 4 0.1 0 .
4.1 1 0 1.7
An important attribute of a matrix is its size or dimensions, i.e., the numbers of
rows and columns. The matrix above has 3 rows and 4 columns, so its size is 3 4.
A matrix of size m n is called an m n matrix.
The elements (or entries or coefficients) of a matrix are the values in the array.
The i, j element is the value in the ith row and jth column, denoted by double
subscripts: the i, j element of a matrix A is denoted Aij (or Ai,j , when i or j is
more than one digit or character). The positive integers i and j are called the (row
and column) indices. If A is an m n matrix, then the row index i runs from 1 to
m and the column index j runs from 1 to n. Row indices go from top to bottom,
so row 1 is the top row and row m is the bottom row. Column indices go from left
to right, so column 1 is the left column and column n is the right column.
If the matrix above is B, then we have B13 = 2.3, B32 = 1. The row index
of the bottom left element (which has value 4.1) is 3; its column index is 1.
Two matrices are equal if they have the same size, and the corresponding entries
are all equal. As with vectors, we normally deal with matrices with entries that
90 6 Matrices

are real numbers, which will be our assumption unless we state otherwise. The set
of real m n matrices is denoted Rmn . But matrices with complex entries, for
example, do arise in some applications.

Matrix indexing. As with vectors, standard mathematical notation indexes the

rows and columns of a matrix starting from 1. In computer languages, matrices
are often (but not always) stored as 2-dimensional arrays, which can be indexed in
a variety of ways, depending on the language. Lower level languages typically use
indexes starting from 0; higher level languages and packages that support matrix
operations usually use standard mathematical indexing, starting from 1.

Square, tall, and wide matrices. A square matrix has an equal number of rows
and columns. A square matrix of size n n is said to be of order n. A tall matrix
has more rows than columns (size m n with m > n). A wide matrix has more
columns than rows (size m n with n > m).

Column and row vectors. An n-vector can be interpreted as an n 1 matrix; we

do not distinguish between vectors and matrices with one column. A matrix with
only one row, i.e., with size 1 n, is called a row vector ; to give its size, we can
refer to it as an n-row-vector. As an example,
 
2.1 3 0
is a row vector (or 1 3 matrix). To distinguish them from row vectors, vectors
are sometimes called column vectors. A 1 1 matrix is considered to be the same
as a scalar.

Notational conventions. Many authors (including us) tend to use capital letters
to denote matrices, and lower case letters for (column or row) vectors. But this
convention is not standardized, so you should be prepared to figure out whether
a symbol represents a matrix, column vector, row vector, or a scalar, from con-
text. (The more considerate authors will tell you what the symbols represent, for
example, by referring to the matrix A when introducing it.)

Columns and rows of a matrix. An m n matrix A has n columns, given by (the

m-vectors)
A1j
aj = ... ,

Amj
for j = 1, . . . , n. The same matrix has m rows, given by the (1 n row vectors)
 
bi = Ai1 Ain ,
for i = 1, . . . , m.
As a specific example, the 2 3 matrix
 
1 2 3
4 5 6
6.1 Matrices 91

has first row  

1 2 3
(which is a 3-row-vector or a 1 3 matrix), and second column
 
2
5

(which is a 2-vector or 2 1 matrix), also written compactly as (2, 5).

Block matrices and submatrices. It is useful to consider matrices whose entries

are themselves matrices, as in
 
B C
A= ,
D E

where B, C, D, and E are matrices. Such matrices are called block matrices; the
elements B, C, and D are called blocks or submatrices of A. The submatrices can
be referred to by their block row and column indices; for example, C is the 1,2
block of A.
Block matrices must have the right dimensions to fit together. Matrices in the
same (block) row must have the same number of rows (i.e., the same height);
matrices in the same (block) column must have the same number of columns (i.e.,
the same width). In the example above, B and C must have the same number of
rows, and C and E must have the same number of columns. Matrix blocks placed
next to each other in the same row are said to be concatenated ; matrix blocks
placed above each other are called stacked.
As an example, consider
   
    2 2 1 4
B= 0 2 3 , C = 1 , D= , E= .
1 3 5 4

Then the block matrix A above is given by

0 2 3 1
A= 2 2 1 4 . (6.1)
1 3 5 4

(Note that we have dropped the left and right brackets that delimit the blocks.
This is similar to the way we drop the brackets in a 1 1 matrix to get a scalar.)
We can also divide a larger matrix (or vector) into blocks. In this context the
blocks are called submatrices of the big matrix. As with vectors, we can use colon
notation to denote submatrices. If A is an m n-matrix, and p, q, r, s are integers
with 1 p q m and 1 r s n, then Ap:q,r:s denotes the submatrix

Apr Ap,r+1 Aps
Ap+1,r Ap+1,r+1 Ap+1,s
Ap:q,r:s = .. .. .. .

. . .
Aqr Aq,r+1 Aqs
92 6 Matrices

This submatrix has size (q p + 1) (s r + 1) and is obtained by extracting from

A the elements in rows p through q and columns r through s.
For the specific matrix A in (6.1), we have
 
1 4
A2:3,3:4 = .
5 4

Column and row representation of a matrix. Using block matrix notation we

can write an m n matrix A as a block matrix with one block row and n block
columns,  
A = a1 a2 an ,
where aj , which is an m-vector, is the jth column of A. Thus, an m n matrix
can be viewed as its n columns, concatenated.
Similarly, an m n matrix A can be written as a block matrix with one block
column and m block rows:
b1
b2
A = . ,

..
bm
where bi , which is a row n-vector, is the ith row of A. In this notation, the matrix
A is interpreted as its m rows, stacked.

Examples
Table interpretation. The most direct interpretation of a matrix is as a table of
numbers that depend on two indexes, i and j. (A vector is a list of numbers that
depend on only one index.) In this case the rows and columns of the matrix usually
have some simple interpretation. Some examples are given below.

Images. A black and white image with M N pixels is naturally represented

as an M N matrix. The row index i gives the vertical position of the pixel,
the column index j gives the horizontal position of the pixel, and the i, j
entry gives the pixel value.

Rainfall data. An m n matrix A gives the rainfall at m different locations

on n consecutive days, so A42 (which is a number) is the rainfall at location
4 on day 2. The jth column of A, which is an m-vector, gives the rainfall at
the m locations on day j. The ith row of A, which is an n-row-vector, is the
time series of rainfall at location i.

Asset returns. A T n matrix R gives the returns of a collection of n assets

(called the universe of assets) over T periods, with Rij giving the return of
asset j in period i. So R12,7 = 0.03 means that asset 7 had a 3% loss in
period 12. The 4th column of R is a T -vector that is the return time series
6.1 Matrices 93

Date AAPL GOOG MMM AMZN

March 1 2016 0.00219 0.00006 0.00113 0.00202
March 2 2016 0.00744 0.00894 0.00019 0.00468
March 3 2016 0.01488 0.00215 0.00433 0.00407

Table 6.1 Daily returns of Apple (AAPL), Google (GOOG), 3M (MMM),

and Amazon (AMZN), on March 1, 2, and 3, 2016 (based on closing prices).

for asset 4. The 3rd row of R is an n-row-vector that gives the returns of all
assets in the universe in period 3.
An example of an asset return matrix, with a universe of n = 4 assets over
T = 3 periods, is shown in table 6.1.

Prices from multiple suppliers. An m n matrix P gives the prices of n

different goods from m different suppliers (or locations): Pij is the price that
supplier i charges for good j. The jth column of P is the m-vector of supplier
prices for good j; the ith row gives the prices for all goods from supplier i.

Contingency table. Suppose we have a collection of objects with two at-

tributes, the first attribute with m possible values and the second with n
possible values. An m n matrix A can be used to hold the counts of the
numbers of objects with the different pairs of attributes: Aij is the number
of objects with first attribute i and second attribute j. (This is the analog
of a count n-vector, that records the counts of one attribute in a collection.)
For example, a population of college students can be described by a 4 50
matrix, with the i, j entry the number of students in year i of their studies,
from state j (with the states ordered in, say, alphabetical order). The ith row
of A gives the geographic distribution of students in year i of their studies;
the jth column of A is a 4-vector giving the numbers of student from state j
in their first through fourth years of study.

Customer purchase history. An n N matrix P can be used to store a set of

N customers purchase histories of n products, items, or services, over some
period. The entry Pij represents the dollar value of product i that customer j
purchased over the period (or as an alternative, the number or quantity of the
product). The jth column of P is the purchase history vector for customer
j; the ith row gives the sales report for product i across the N customers.

Matrix representation of a collection of vectors. Matrices are very often used

as a compact way to give a set of indexed vectors of the same size. For example, if
x1 , . . . , xN are n-vectors that give the n feature values for each of N objects, we
can collect them all into one n N matrix
 
X= x1 x2 xN ,
94 6 Matrices

2 3

1 4

Figure 6.1 The relation (6.2) as a directed graph.

often called a data matrix or feature matrix. Its jth column is the feature n-vector
for the jth object (in this context sometimes called the jth example). The ith row
of the data matrix X is an N -row-vector whose entries are the values of the ith
feature across the examples. We can also directly interpret the entries of the data
matrix: Xij (which is a number) is the value of the ith feature for the jth example.
As another example, a 3 M matrix can be used to represent a collection of
M locations or positions in 3-D space, with its jth column giving the jth position.

Matrix representation of a relation or graph. Suppose we have n objects labeled

1, . . . , n. A relation R on the set of objects {1, . . . , n} is a subset of ordered pairs
of objects. As an example, R can represent a preference relation among n possible
products or choices, with (i, j) R meaning that choice i is preferred to choice j.
A relation can also be viewed as a directed graph, with nodes (or vertices) labeled
1, . . . , n, and a directed edge from j to i for each (i, j) R. This is typically drawn
as a graph, with arrows indicating the direction of the edge, as shown in figure 6.1,
for the relation on 4 objects

R = {(1, 2), (1, 3), (2, 1), (2, 4), (3, 4), (4, 1)}. (6.2)

A relation R on {1, . . . , n} is represented by the n n matrix A with


1 (i, j) R
Aij =
0 (i, j) 6 R.

This matrix is called the adjacency matrix associated with the graph. The rela-
tion (6.2), for example, is represented by the matrix

0 1 1 0
1 0 0 1
A=
0
.
0 1 0
1 0 0 0

This is the adjacency matrix of the associated graph, shown in figure 6.1. (We will
encounter another matrix associated with a directed graph in 7.3.)
 6.2 Zero and identity matrices 95

6.2 Zero and identity matrices

Zero matrix. A zero matrix is a matrix with all elements equal to zero. The zero
matrix of size m n is sometimes written as 0mn , but usually a zero matrix is
denoted just 0, the same symbol used to denote the number 0 or zero vectors. In
this case the size of the zero matrix must be determined from the context.

Identity matrix. An identity matrix is another common matrix. It is always

square. Its diagonal elements, i.e., those with equal row and column index, are all
equal to one, and its off-diagonal elements (those with unequal row and column
indices) are zero. Identity matrices are denoted by the letter I. Formally, the
identity matrix of size n is defined by

1 i=j
Iij =
0 i 6= j,

for i, j = 1, . . . , n. For example,

  1 0 0 0
1 0 0 1 0 0
,
0 1 0 0 1 0
0 0 0 1

are the 2 2 and 4 4 identity matrices.

The column vectors of the n n identity matrix are the unit vectors of size n.
Using block matrix notation, we can write
 
I= e1 e2 en ,

where ek is the kth unit vector of size n.

Sometimes a subscript is used to denote the size of an identity matrix, as in
I4 or I22 . But more often the size is omitted and follows from the context. For
example, if
 
1 2 3
A= ,
4 5 6

then
1 0 1 2 3
  0 1 4 5 6
I A
= 0 0 1 0 0 .
0 I
0 0 0 1 0
0 0 0 0 1

The dimensions of the two identity matrices follow from the size of A. The identity
matrix in the 1,1 position must be 2 2, and the identity matrix in the 2,2 position
must be 3 3. This also determines the size of the zero matrix in the 2,1 position.
The importance of the identity matrix will become clear later, in 10.1.
96 6 Matrices

Sparse matrices. A matrix A is said to be sparse if many of its entries are zero,
or (put another way) just a few of its entries are nonzero. Its sparsity pattern is
the set of indices (i, j) for which Aij 6= 0. The number of nonzeros of a sparse
matrix A is the number of entries in its sparsity pattern, and denoted nnz(A). If
A is m n we have nnz(A) mn. Its density is nnz(A)/(mn), which is no more
than one. Densities of sparse matrices that arise in applications are typically small
or very small, as in 102 or 104 . There is no precise definition of how small the
density must be for a matrix to qualify as sparse. A famous definition of sparse
matrix due to Wilkinson is: A matrix is sparse if it has enough zero entries that it
pays to take advantage of them. Sparse matrices can be stored and manipulated
efficiently on a computer.
Many common matrices are sparse. An n n identity matrix is sparse, since
it has only n nonzeros, so its density is 1/n. The zero matrix is the sparsest
possible matrix, since it has no nonzero entries. Several special sparsity patterns
have names; we describe some important ones below.
Like sparse vectors, sparse matrices arise in many applications. A typical cus-
tomer purchase history matrix (see page 93) is sparse, since each customer has
likely only purchased a small fraction of all the products available.

Diagonal matrices. A square n n matrix A is diagonal if Aij = 0 for i 6= j.

(The entries of a matrix with i = j are called the diagonal entries; those with i 6= j
are its off-diagonal entries.) A diagonal matrix is one for which all off-diagonal
entries are zero. Examples of diagonal matrices we have already seen are square
zero matrices and identity matrices. Other examples are

  0.2 0 0
3 0 0 3
, 0 .
0 0
0 0 1.2

(Note that in the first example, one of the diagonal elements is also zero.)
The notation diag(a1 , . . . , an ) is used to compactly describe the n n diagonal
matrix A with diagonal entries A11 = a1 , . . . , Ann = an . This notation is not yet
standard, but is coming into more prevalent use. As examples, the matrices above
would be expressed as

diag(3, 0), diag(0.2, 3, 1.2),

respectively. We also allow diag to take one n-vector argument, as in I = diag(1).

Triangular matrices. A square n n matrix A is upper triangular if Aij = 0 for

i > j, and it is lower triangular if Aij = 0 for i < j. (So a diagonal matrix is
one that is both lower and upper triangular.) If a matrix is either lower or upper
triangular, it is called triangular.
For example, the matrices

1 1 0.7  
0 1.2 1.1 , 0.6 0
,
0.3 3.5
0 0 3.2
 6.3 Transpose and addition 97

are upper and lower triangular, respectively.

A triangular n n matrix A has up to n(n + 1)/2 nonzero entries, i.e., around
half its entries are zero. Triangular matrices are generally not considered sparse
matrices, since their density is around 50%, but their special sparsity pattern will
be important in the sequel.

6.3 Transpose and addition

6.3.1 Matrix transpose

If A is an m n matrix, its transpose, denoted AT (or sometimes A0 or A ), is the

n m matrix given by (AT )ij = Aji . In words, the rows and columns of A are
transposed in AT . For example,
T
0 4  
7 0 7 3
0 = .
4 0 1
3 1

If we transpose a matrix twice, we get back the original matrix: (AT )T = A. (The
superscript T in the transpose is the same one used to denote the inner product of
two n-vectors; we will soon see how they are related.)

Row and column vectors. Transposition converts row vectors into column vectors
and vice versa. It is sometimes convenient to express a row vector as aT , where a
is a column vector. For example, we might refer to the m rows of an m n matrix
A as aTi , . . . , aTm , where a1 , . . . , am are (column) n-vectors. As an example, the
second row of the matrix  
0 7 3
4 0 1
can be written as (the row vector) (4, 0, 1)T .
It is common to extend concepts from (column) vectors to row vectors, by
applying the concept to the transposed row vectors. We say that a set of row
vectors is linearly dependent (or independent) if their transposes (which are column
vectors) are linearly dependent (or independent). For example, the rows of a
matrix A are linearly independent means that the columns of AT are linearly
independent. As another example, the rows of a matrix A are orthonormal means
that their transposes, the columns of AT , are orthonormal. Clustering the rows of
a matrix X means clustering the columns of X T .

Transpose of block matrix. The transpose of a block matrix has the simple form
(shown here for a 2 2 block matrix)
T
AT CT
  
A B
= ,
C D BT DT
98 6 Matrices

where A, B, C, and D are matrices with compatible sizes. Roughly speaking,

the transpose of a block matrix is the transposed block matrix, with each element
transposed.

Document-term matrix. Consider a corpus (collection) of N documents, with

word count vectors for a dictionary with n words. The document-term matrix
associated with the corpus is the N n matrix A, with Aij the number of times word
j appears in document i. The rows of the document-term matrix are aT1 , . . . , aTN ,
where the n-vectors a1 , . . . , aN are the word count vectors for documents 1, . . . , N ,
respectively. The columns of the document-term matrix are also interesting. The
jth column of A, which is an N -vector, gives the number of times word j appears
in the corpus of N documents.

Symmetric matrix. A square matrix A is symmetric if A = AT , i.e., Aij = Aji

for all i, j. Symmetric matrices arise in several applications. For example, suppose
that A is the adjacency matrix of a graph or relation (see page 94). The matrix A
is symmetric when the relation is symmetric, i.e., whenever (i, j) R, we also have
(j, i) R. An example is the friend relation on a set of n people, where (i, j) R
means that person i and person j are friends. (In this case the associated graph is
called the social network graph.)

6.3.2 Matrix addition

Two matrices of the same size can be added together. The result is another matrix
of the same size, obtained by adding the corresponding elements of the two matrices.
For example,
0 4 1 2 1 6
7 0 + 2 3 = 9 3 .
3 1 0 4 3 5
Matrix subtraction is similar. As an example,
   
1 6 0 6
I = .
9 3 9 2

(This gives another example where we have to figure out the size of the identity
matrix. Since we can only add or subtract matrices of the same size, I refers to a
2 2 identity matrix.)

Properties of matrix addition. The following important properties of matrix ad-

dition can be verified directly from the definition. We assume here that A, B, and
C are matrices of the same size.

Commutativity. A + B = B + A.

Associativity. (A + B) + C = A + (B + C). We therefore write both as

A + B + C.
 6.4 Matrix-vector multiplication 99

Addition with zero matrix. A + 0 = 0 + A = A. Adding the zero matrix to a

matrix has no effect.
Transpose of sum. (A + B)T = AT + B T . The transpose of a sum of two
matrices is the sum of their transposes.

6.3.3 Scalar-matrix multiplication

Scalar multiplication of matrices is defined in a similar way as for vectors, and is
done by multiplying every element of the matrix by the scalar. For example

1 6 2 12
(2) 9 3 = 18 6 .
6 0 12 0
As with scalar-vector multiplication, the scalar can also appear on the right. Note
that 0 A = 0 (where the left-hand zero is the scalar zero, and the right-hand 0 is
the zero matrix).
Several useful properties of scalar multiplication follow directly from the defini-
tion. For example, (A)T = (AT ) for a scalar and a matrix A. If A is a matrix
and , are scalars, then
( + )A = A + A, ()A = (A).
It is useful to identify the symbols appearing in these two equations. The + symbol
on the left of the left-hand equation is addition of scalars, while the + symbol on
the right of the left-hand equation denotes matrix addition. On the left side of
the right-hand equation we see scalar-scalar multiplication () and scalar-matrix
multiplication; on the right we see two cases of scalar-matrix multiplication.
Finally, we mention that scalar-matrix multiplication has higher precedence
than matrix addition, which means that we should carry out multiplication before
addition (when there are no parantheses to fix the order). So the right-hand side
of the left equation above is to be interpreted as (A) + (A).

6.4 Matrix-vector multiplication

If A is an mn matrix and x is an n-vector, then the matrix-vector product y = Ax
is the m-vector y with elements
n
X
yi = Aik xk = Ai1 x1 + + Ain xn , i = 1, . . . , m. (6.3)
k=1

As a simple example, we have

  2  
0 2 1 1 = 3
.
2 1 1 4
1
100 6 Matrices

(The reader should check that the two sums in (6.3) give the entries of the vector
on the right-hand side.)

Row and column interpretations. We can express the matrix-vector product in

terms of the rows or columns of the matrix. From (6.3) we see that yi is the inner
product of x with the ith row of A:

yi = bTi x, i = 1, . . . , m,

where bTi is the row i of A. The matrix-vector product can also be interpreted in
terms of the columns of A. If ak is the kth column of A, then y = Ax can be
written
y = x1 a1 + x2 a2 + + xn an .
This shows that y = Ax is a linear combination of the columns of A; the coefficients
in the linear combination are the elements of x.

General examples. In the examples below, A is an m n matrix and x is an

n-vector.
Zero matrix. When A = 0, we have Ax = 0. In other words, 0x = 0. (The
left-hand 0 is an m n matrix, and the right-hand zero is an m-vector.)
Identity. We have Ix = x for any vector x. (The identity matrix here has
dimension n n.) In other words, multplying a vector by the identity matrix
gives the same vector.
Picking out columns and rows. An important identity is Aej = aj , the jth
column of A. Multiplying a unit vector by a matrix picks out one of the
columns of the matrix. AT ei , which is an n-vector, is the ith row of A,
transposed. (In other words, (AT ei )T is the ith row of A.)
Summing or averaging columns or rows. The m-vector A1 is the sum of the
columns of A; its ith entry is the sum of the entries in the ith row of A.
The m-vector A(1/n) is the average of the columns of A; its ith entry is
the average of the entries in the ith row of A. In a similar way, AT 1 is an
n-vector, whose jth entry is the sum of the entries in the jth column of A.
Difference matrix. The (n 1) n matrix

1 0

1 0 0 0
0 1 1 0 0 0


. .. ...

D= (6.4)
.. ..

. .

0 0 0 1 1 0
0 0 0 0 1 1

(where entries not shown are zero, and entries with diagonal dots are 1 or
1, continuing the pattern) is called the difference matrix. The vector Dx is
6.4 Matrix-vector multiplication 101

the (n 1)-vector of differences of consecutive entries of x:

x2 x1
x3 x2
Dx = .. .

.
xn xn1

Running sum matrix. The n n matrix

1 0 0 0 0
1 1
0 0 0

.. ..
. .
S= (6.5)
.. ..

. .

1 1 1 1 0
1 1 1 1 1
is called the running sum matrix. The ith entry of the n-vector Sx is the
sum of the first i entries of x:

x1

x1 + x2

x1 + x2 + x3
Sx = .
..
.
x1 + + xn

Application examples.
Feature matrix and weight vector. Suppose X is a feature matrix, where its N
columns x1 , . . . , xN are the n-feature vectors for N objects or examples. Let
the n-vector w be a weight vector, and let si = xTi w be the score associated
with object i using the weight vector w. Then we can write s = X T w, where
s is the N -vector of scores of the objects.
Portfolio return time series. Suppose that R is a T n asset return matrix,
that gives the returns of n assets over T periods. Let h denote the n-vector
of dollar value investments in the assets over the T periods, so, e.g., h3 = 200
means that we have invested $200 in asset 3. (Short positions are denoted
by negative entries in h.) Then Rh, which is a T -vector, is the time series of
the portfolio profit (in $) over the periods 1, . . . , T . If w is a set of portfolio
weights (with 1T w = 1), then Rw is the time series of portfolio returns, given
as a fraction.
As an example, consider a portfolio of the 4 assets in table 6.1, with weights
w = (0.4, 0.3, 0.2, 0.5). The product Rw = (0.00213, 0.00201, 0.00241)
gives the portofio returns over the three periods in the example.
Polynomial evaluation at multiple points. Suppose the entries of the n-vector
c are the coefficients of a polynomial p of degree n 1 or less:
p(t) = c1 + c2 t + + cn1 tn2 + cn tn1 .
102 6 Matrices

Let t1 , . . . , tm be m numbers, and define the m-vector y as yi = p(ti ). Then

we have y = Ac, where A is the m n matrix
1 t1 tn2 tn1

1 1
1 t2 t n2 n1
2 t2
A= . . (6.6)

. .
. .
. .
.
. . . .
1 tm tn2
m tn1
m

So multiplying a vector c by the matrix A is the same as evaluating a poly-

nomial with coefficients c at m points. The matrix A in (6.6) comes up often,
and is called a Vandermonde matrix (of degree n1, at the points t1 , . . . , tm ).
Total price from multiple suppliers. Suppose the m n matrix P gives the
prices of n goods from m suppliers (or in m different locations). If q is an
n-vector of quantities of the n goods (sometimes called a basket of goods),
then c = P q is an N -vector that gives the total cost of the goods, from each
of the N suppliers.
Document scoring. Suppose A in an N n document-term matrix, which gives
the word counts of a corpus of N documents using a dictionary of n words,
so the rows of A are the word count vectors for the documents. Suppose that
w in an n-vector that gives a set of weights for the words in the dictionary.
Then s = Aw is an N -vector that gives the scores of the documents, using
the weights and the word counts. A search engine, for example, might choose
w (based on the search query) so that the scores are predictions of relevance
of the documents (to the search).
Audio mixing. Suppose the k columns of A are vectors representing audio
signals or tracks of length T , and w is a k-vector. Then b = Aw is a T -vector
representing the mix of the audio signals, with track weights given by the
vector w.

Inner product. When a and b are n-vectors, aT b is exactly the inner product of a
and b, obtained from the rules for transposing matrices and matrix-vector product.
We start with the n-(column) vector a, consider it as an n1 matrix, and transpose
it to obtain the n-row-vector aT . Now we multiply this 1n matrix by the n-vector
b, which we consider an n 1 matrix, to obtain the 1 1 matrix aT b, which we also
consider a scalar. So the notation aT b for the inner product is just a very special
case of matrix-vector multiplication.

Linear dependence of columns. We can express the concepts of linear depen-

dence and independence in a compact form using matrix-vector multiplication.
The columns of a matrix A are linearly dependent if Ax = 0 for some x 6= 0. The
columns of a matrix A are linearly independent if Ax = 0 implies x = 0.

Expansion in a basis. If the columns of A are a basis, which means A is square

with linearly independent columns a1 , . . . , an , then for any n-vector b there is a
unique n-vector x that satisfies Ax = b. In this case the vector x gives the coeffi-
cients in the expansion of b in the basis a1 , . . . , an .
 6.5 Complexity 103

Properties of matrix-vector multiplication. Matrix-vector multiplication satisfies

several properties that are readily verified. First, it distributes across the vector
argument: For any m n matrix A and any n-vectors u and v, we have

A(u + v) = Au + Av.

Matrix-vector multiplication, like ordinary multiplication of numbers, has higher

precedence than addition, which means that when there are no parentheses to force
the order of evaluation, multiplications are to be carried out before additions. This
means that the right-hand side above is to be interpreted as (Au) + (Av). The
equation above looks innocent and natural, but must be read carefully. On the
left-hand side, we first add the vectors u and v, which is the addition of n-vectors.
We then multiply the resulting n-vector by the matrix A. On the right-hand
side, we first multiply each of n-vectors by the matrix A (this is two matrix-vector
multiplies); and then add the two resulting m-vectors together. The left- and right-
hand sides of the equation above involve very different steps and operations, but
the final result of each is the same m-vector.
Another basic property is, for any m n matrix A, any n-vector u, and any
scalar , we have
(A)u = (Au)
(and so we can write this as Au). On the left-hand side, we have scalar-matrix
multiplication, followed by matrix-vector multiplication; on the right-hand side, we
start with matrix-vector multiplication, and then perform scalar-vector multiplica-
tion.

Input-output interpretation. We can interpret the relation y = Ax, with A an

m n matrix, as a mapping from the n-vector x to the m-vector y. In this
context we might think of x as an input, and y as the corresponding output. From
equation (6.3), we can interpret Aij as the factor by which yi depends on xj . Some
examples of conclusions we can draw are given below.
If A23 is positive and large, then y2 depends strongly on x3 , and increases as
x3 increases.
If A32 is much larger than the other entries in the third row of A, then y3
depends much more on x2 than the other inputs.
If A is square and lower triangular, then yi only depends on x1 , . . . , xi .

6.5 Complexity
Computer representation of matrices. An m n matrix is usually represented
on a computer as an m n array of floating point numbers, which requires 8mn
bytes. In some software systems symmetric matrices are represented in a more
efficient way, by only storing the upper triangular elements in the matrix, in some
specific order. This reduces the memory requirement by around a factor of two.
104 6 Matrices

Sparse matrices are represented by various methods that encode for each nonzero
element its row index i (an integer), its column index j (an integer) and its value
Aij (a floating point number). When the row and column indexes are represented
using 4 bytes, this requires a total of around 16 nnz(A) bytes.

Complexity of matrix addition, scalar multiplication, and transposition. The

addition of two m n matrices or a scalar multiplication of an m n matrix take
mn flops. When A is sparse, scalar multiplication requires nnz(A) flops. When at
least one of A and B is sparse, computing A + B requires min{nnz(A), nnz(B)}
flops. (For any entry i, j for which one of Aij or Bij is zero, no arithmetic operations
are needed to find (A + B)ij .) Matrix transposition, i.e., computing AT , requires
zero flops, since we simply copy entries of A to those of AT . (Copying the entries
does take time to carry out, but this is not reflected in the flop count.)

Complexity of matrix-vector multiplication. A matrix-vector multiplication of

an m n matrix A with an n-vector x requires m(2n 1) flops, which we simplify
to 2mn flops. This can be seen as follows. The result y = Ax of the product is an
m-vector, so there are m numbers to compute. The ith element of y is the inner
product of the ith row of A and the vector x, which takes 2n 1 flops.
If A is sparse, computing Ax requires nnz(A) multiplies (of Aij and xj , for
each nonzero entry of A) and a number of additions that is no more than nnz(A).
Thus, the complexity is between nnz(A) and 2 nnz(A) flops. As a special example,
suppose A is nn and diagonal. Then computing Ax can be done with n multiplies
(Aii times xi ) and no additions, a total of n flops.
 Chapter 7

Matrix examples

In this chapter we describe some special matrices that occur often in applications.

7.1 Geometric transformations

Suppose the 2-vector (or 3-vector) x represents a position in 2-D (or 3-D) space.
Several important geometric transformations or mappings from points to points
can be expressed as matrix-vector products y = Ax, with A a 2 2 (or 3 3)
matrix. In the examples below, we consider the mapping from x to y, and focus
on the 2-D case (for which some of the matrices are simpler to describe).

Scaling. Scaling is the mapping y = ax, where a is a scalar. This can be expressed
as y = Ax with A = aI. This mapping stretches a vector by the factor |a| (or
shrinks it when |a| < 1), and it flips the vector (reverses its direction) if a < 0.

Dilation. Dilation is the mapping y = Dx, where D is a diagonal matrix, D =

diag(d1 , d2 ). This mapping stretches the vector x by different factors along the
two different axes. (Or shrinks, if |di | < 1, and flips, if di < 0.)

Rotation. Suppose that y is the vector obtained by rotating x by radians coun-

terclockwise. Then we have
 
cos sin
y= x. (7.1)
sin cos
This matrix is called (for obvious reasons) a rotation matrix.

Reflection. Suppose that y is the vector obtained by reflecting x through the line
that passes through the origin, inclined radians with respect to horizontal. Then
we have  
cos(2) sin(2)
y= x.
sin(2) cos(2)
106 7 Matrix examples

Ax
x Ax
Ax
x
x

Figure 7.1 From left to right: A dilation with A = diag(2, 2/3), a coun-
terclockwise rotation by /6 radians, and a reflection through a line that
makes an angle of /4 radians with the horizontal line.

Projection onto a line. The projection of the point x onto a set is the point in
the set that is closest to x. Suppose y is the projection of x onto the line that
passes through the origin, inclined radians with respect to horizontal. Then we
have  
(1/2)(1 + cos(2)) (1/2) sin(2)
y= x.
(1/2) sin(2) (1/2)(1 cos(2))
Some of these geometric transformations are illustrated in figure 7.1.

Finding the matrix. One simple method to find the matrix associated with a
linear geometric transformation is to find its columns. The ith column is the vector
obtained by applying the transformation to ei . As a simple example consider
(clockwise) rotation by 90 in 2-D. Rotating the vector e1 = (1, 0) by 90 gives
(0, 1); rotating e2 = (0, 1) by 90 gives (1, 0). So rotation by 90 is given by
 
0 1
y= x.
1 0

Change of coordinates. In many applications multiple coordinate systems are

used to describe locations or positions in 2-D or 3-D. For example in aerospace
engineering we can describe a position using earth-fixed coordinates or body-fixed
coordinates, where the body refers to an aircraft. Earth-fixed coordinates are with
respect to a specific origin, with the three axes pointing East, North, and straight
up, respectively. The origin of the body-fixed coordinates is a specific location on
the aircraft (typically the center of gravity), and the three axes point forward (along
the aircraft body), left (with respect to the aircraft body), and up (with respect
to the aircraft body). Suppose the 3-vector xbody describes a location using the
body coordinates, and xearth describes the same location in earth-fixed coordinates.
These are related by
xearth = p + Qxbody ,
where p is the location of the airplane center (in earth-fixed coordinates) and Q is
a 3 3 matrix. The ith column of Q gives the earth-fixed coordinates for the ith
axis of the airplane. For an airplane in level flight, heading due South, we have

0 1 0
Q = 1 0 0 .
0 0 1
 7.2 Selectors 107

7.2 Selectors
An m n selector matrix A is one in which each row is a unit vector (transposed):
T
ek 1
..
A = . ,
eTkm
where k1 , . . . , km are integers in the range 1, . . . , n. When it multiplies a vector,
it simply copies the ki th entry of x into the ith entry of y = Ax:
y = (xk1 , xk2 , . . . , xkm ).
In words, each entry of Ax is a selection of an entry of x.
The identity matrix, and the reverser matrix

T 0 0 0 1
en 0 0 1 0
A = ... = ... ... . . . ... ..


.

eT1 0 1 0 0
1 0 0 0
are special cases of selector matrices. (The reverser matrix reverses the order of
the entries of a vector: Ax = (xn , xn1 , . . . , x2 , x1 ).) Another one is the r : s slicing
matrix, which can be described as the block matrix
 
A = 0m(r1) Imm 0m(ns) ,
where m = s r + 1. (We show the dimensions of the blocks for clarity.) We have
Ax = xr:s , i.e., multiplying by A gives the r : s slice of a vector.

Down-sampling. Another example is the n/2 n matrix (with n even)

1 0 0 0 0 0 0 0 0 0
0 0 1 0
0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0
A= . . . . .. .. .. .. .. .. .

.. .. .. .. . . . . . .

0 0 0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0 1 0
If y = Ax, we have y = (x1 , x3 , x5 , . . . , xn3 , xn1 ). When x is a time series, y is
called the 2 down-sampled version of x. If x is a quantity sampled every hour,
then y is the same quantity, sampled every 2 hours.

Image cropping. As a more interesting example, suppose that x is an image with

M N pixels, with M and N even. (That is, x is an M N -vector, with its entries
giving the pixels values in some specific order.) Let y be the (M/2) (N/2) image
that is the upper left corner of the image x, i.e., a cropped version. Then we have
y = Ax, where A is an (M N )/4 (M N ) selector matrix. The ith row of A is eTki ,
where ki is the index of the pixel in x that corresponds to the ith pixel in y.
108 7 Matrix examples

Permutation matrices. An nn permutation matrix is one in which each column

is a unit vector, and each row is the transpose of a unit vector. (In other words, A
and AT are both selector matrices.) Thus, exactly one entry of each row is one, and
exactly one entry of each column is one. This means that y = Ax can be expressed
as yi = xi , where is a permutation of 1, 2, . . . , n, i.e., each integer from 1 to n
appears exactly once in 1 , . . . , n .
As a simple example consider the permutation = (3, 1, 2). The associated
permutation matrix is
0 0 1
A = 1 0 0 .
0 1 0
Multiplying a 3-vector by A re-orders its entries: Ax = (x3 , x1 , x2 ).

7.3 Incidence matrix

A directed graph consists of a set of vertices (or nodes), labeled 1, . . . , n, and a
set of directed edges (or branches), labeled 1, . . . , m. Each edge is connected from
one of the nodes and into another one, in which case we say the two nodes are
connected or adjacent. Directed graphs are often drawn with the vertices as circles
or dots, and the edges as arrows, as in figure 7.2. A directed graph can be described
by its n m incidence matrix, defined as

1 edge j points to node i
Aij = 1 edge j points from node i
0 otherwise.

The incidence matrix is evidently sparse, since it has only two nonzero entries
in each column (one with value 1 and other with value 1). The jth column is
associated with the jth edge; the indices of its two nonzero entries give the nodes
that the edge connects. The ith row of A corresponds to node i; its nonzero entries
tell us which edges connect to the node, and whether they point into or away from
the node. The incidence matrix for the graph shown in figure 7.2 is

1 1 0 1 0
1 0 1 0 0
A= .
0 0 1 1 1
0 1 0 0 1

A directed graph can also be described by its adjacency matrix, described on

page 94. The adjacency and incidence matrices for a directed graph are closely
related, but not the same. The adjacency matrix does not explicitly label the
edges j = 1, . . . , m. There are also some small difference in the graphs that can
be represented using incidence and adjacency matrices. For example, self edges
(that connect from and to the same vertex) cannot be represented in an incidence
matrix.
7.3 Incidence matrix 109

3
2 3

1 4 5

2
1 4

Figure 7.2 Directed graph with four vertices and five edges.

Edge flows and flow conservation. The matrix-vector product y = Ax can be

given a very simple interpretation in terms of flows of some quantity along the edges
of the graph. We interpret xj as the flow (rate) along the edge j, with a positive
value meaning the flow is in the direction of edge j, and negative meaning the flow
is in the opposite direction of edge j. The n-vector y = Ax can be interpreted as
the vector of net flows, from the edges, into the node: yi is equal to the total of the
flows that come in to node i, minus the total of the flows that go out from node i.
If Ax = 0, we say that flow conservation occurs, since at each node, the total in-
flow matches the total out-flow. In this case the flow vector x is called a circulation.
This could be used as a model of traffic flow (in a closed system), with the nodes
representing intersections and the edges representing road segments (one for each
direction).
For the directed graph example above, the vector

x = (1, 1, 1, 0, 1)

is a circulation, since Ax = 0. This flow corresponds to a unit clockwise flow on

the outer edges (1, 3, 5, and 2) and no flow on the diagonal edge (4). (Visualizing
this explains why such vectors are called circulations.)

Flow conservation with sources. Another common equation is Ax + s = 0, where

s is an n-vector of source flows, which we can think of as an exogeneous flow that
enters node i (or leaves, when si < 0). The equation Ax + s = 0 means that the
flow is conserved at each node: the total of all incoming flow from the in-coming
edges and the source minus the total out-going flow from out-going edges is zero.
This might be used as an approximate model of a power grid (ignoring losses), with
x being the vector of power flows along the transmission lines, si > 0 representing
a generator injecting power into the grid at node i, si < 0 representing a load that
consumes power at node i, and si = 0 representing a substation where power is
exchanged among transmission lines, with no generation or load attached.
For the example above, consider the source vector s = (1, 0, 1, 0), which cor-
responds to an injection of one unit of flow into node 1, and the removal of one
unit of flow at node 3. (We say that node 1 is a source and node 3 is a sink.) Flow
is conserved at nodes 2 and 4. For this source, the flow vector

x = (0.6, 0.3, 0.6, 0.1, 0.3)

110 7 Matrix examples

1 2 3
1 2 3 n

Figure 7.3 Chain graph.

satisfies Ax + s = 0. This flow can be explained in words: The unit external flow
entering node 1 splits three way, with 0.6 flowing up, 0.3 flowing right, and 0.1
flowing diagonally up (on edge 4). The upward flow on edge 1 passes through
node 2, where flow is conserved, and proceeds right on edge 3 towards node 3.
The rightward flow on edge 2 passes through node 4, where flow is conserved, and
proceeds up on edge 5 to node 3. The one unit of excess flow arriving at node 3 is
removed as external flow.

Node potentials. We can also give a simple interpretation to the matrix-vector

product u = AT v. Here v is an n-vector, often interpreted as a potential, with vi
the potential value at node i. The m-vector u = AT v gives the potential differences
across the edges: uj = vk vl , where edge j goes from node k to node l.

Laplacian. The function of v given by

L(v) = kAT vk2

arises in many applications, and is called the Laplacian (associated with the graph).
It can be expressed as X
L(v) = (vk vl )2 ,
edges (k,l)

which is the sum of the squares of the differences of v across all edges in the graph.
The Laplacian is small when the potentials of nodes that are connected by edges
are near each other.
The Laplacian is used as a measure the non-smoothness (roughness) of a po-
tential function on a graph. A set of node potentials with small Laplacian can be
thought of as smoothly varying across the graph. Conversely a set of potentials
with large Laplacian can be thought of as non-smooth or rough. The Laplacian
will arise as a measure of roughness in several applications we will encounter later.
As a simple example, consider the potential vector v = (1, 1, 2, 1) for the
graph shown in figure 7.2. For this set of potentials, the differences across the edges
are relatively large, with AT v = (2, 2, 3, 1, 3), and the associated Laplacian
is kAT vk2 = 27. Now consider the potential vector v = (1, 2, 2, 1). The associated
edge differences are AT v = (1, 0, 0, 1, 1), and the Laplacian has the much smaller
value kAT vk2 = 3.

Chain graph. The incidence matrix and the Laplacian function have a particularly
simple form for the simple chain graph shown in figure 7.3, with n vertices and n1
edges. The n (n 1) incidence matrix is the transpose of the difference matrix
D described on page 100, in (6.4). The Laplacian is then

L(v) = kDvk2 = (v2 v1 )2 + + (vn vn1 )2 ,

 7.4 Convolution 111

3 3

2 2
ak

bk
1 1

0 0

0 20 40 60 80 100 0 20 40 60 80 100
k k

Figure 7.4 Two vectors of length 100, with Laplacians L(a) = 1.14 and
L(b) = 8.99.

the sum of squares of the differences between consecutive entries of the n-vector v.
This is used as a measure of the non-smoothness of the vector v. Figure 7.4 shows
an example.

7.4 Convolution
The convolution of an n-vector a and an m-vector b is the (n + m 1)-vector
denoted c = a b, with entries
X
ck = ai bj , k = 1, . . . , n + m 1, (7.2)
i+j=k+1

where the subscript in the sum means that we should sum over all values of i and
j in their index ranges 1, . . . , n and 1, . . . , m, for which the sum i + j is k + 1. For
example with n = 4, m = 3, we have

c1 = a1 b1
c2 = a1 b2 + a2 b1
c3 = a1 b3 + a2 b2 + a3 b1
c4 = a2 b3 + a3 b2 + a4 b1
c5 = a3 b3 + a4 b2
c6 = a4 b3 .

Convolution reduces to ordinary multiplication of numbers when n = m = 1, and

to scalar-vector multiplication when either n = 1 or m = 1. Convolution arises in
many applications and contexts.
112 7 Matrix examples

As a specific numerical example, we have

(1, 0, 1) (2, 1, 1) = (2, 1, 3, 1, 1).

Polynomial multiplication. If a and b represent the coefficients of two polynomi-

als,

p(x) = a1 + a2 x + + an xn1 , q(x) = b1 + b2 x + + bm xm1 ,

then the coefficients of the product polynomial p(x)q(x) are represented by c = ab:

p(x)q(x) = c1 + c2 x + + cn+m1 xn+m2 .

To see this we will show that ck is the coefficient of xk1 in p(x)q(x). We expand the
product polynomial into mn terms, and collect those terms associated with xk1 .
These terms have the form ai bj xi+j2 , for P i and j that satisfy i + j 2 = k 1,
i.e., i + j = k 1. It follows that ck = i+j=k+1 ai bj , which agrees with the
convolution formula (7.2).

Properties of convolution. Convolution is symmetric: we have a b = b a. It is

also associative: we have (a b) c = a (b c), so we can write both as a b c.
Another property is that a b = 0 implies that either a = 0 or b = 0. These
properties follow from the polynomial coefficient property above, and can also be
directly shown.
Another basic property is that for fixed a, the convolution a b is a linear
function of b; and for fixed b, it is a linear function of a. This means we can express
a b as a matrix-vector product:

a b = T (b)a = T (a)b,

where T (b) is the (n + m 1) n matrix with entries


bij+1 1ij+1m
T (b)ij =
0 otherwise

and similarly for T (a). For example, with n = 4 and m = 3, we have

b1 0 0 0 a1 0 0

b2 b1 0 0


a2 a1 0

b3 b2 b1 0 a3 a2 a1
T (b) = , T (a) = .

0 b3 b2 b1


a4 a3 a2

0 0 b3 b2 0 a4 a3
0 0 0 b3 0 0 a4

The matrices T (b) and T (a) are called Toeplitz matrices, since the entries on any
diagonal (i.e., indices with ij constant) are the same. The columns of the Toeplitz
matrix T (a) are simply shifted versions of the vector a, padded with zero entries.
7.4 Convolution 113

Variations. Several slightly different definitions of convolution are used in different

applications. In one variation, a and b are infinite two-sided sequences (and not
vectors) with indices ranging from to . In another variation, the rows of
T (a) at the top and bottom that do not contain all the coefficients of a are dropped.
(In this version, the rows of T (a) are shifted versions of the vector a, reversed.)
For consistency, we will use the one definition (7.2).

Examples.
Time series smoothing. Suppose the n-vector x is a time series, and a =
(1/3, 1/3, 1/3). Then y = a x can be interpreted as a smoothed version of
the original time series: for 3 i n 3, yi is the average of xi , xi1 , xi2 .
(We can drop the qualifier 3 i n 3, by defining xi to be zero outside
the index range 1, . . . , n.) The time series y is called the (3-period) moving
average of the time series x. Figure 7.5 shows an example.
First order differences. If the n-vector x is a time series and a = (1, 1), the
time series y = a x gives the first order differences in the series x:
y = (x1 , x2 x1 , x3 x2 , . . . , xn xn1 , xn ).

Audio filtering. If the n-vector x is an audio signal, and a is a vector (typically

with length less than around 0.1 second of real time) the vector y = a x
is called the filtered audio signal, with filter coefficients a. Depending on
the coefficients a, y will be perceived as enhancing or suppressing different
frequencies, like the familiar audio tone controls.
Communication channel. In a modern data communication system, a time
series u is transmitted or sent over some channel (e.g., electrical, optical, or
radio) to a receiver, which receives the time series y. A very common model
is that y and u are related via convolution: y = c u, where the vector c is
the channel impulse response.

Input-output convolution system. Many physical systems with an input (time

series) u and output time series y are well modeled as y = h u, where the n-vector
h is called the system impulse response. For example, ut might represent the power
level of a heater at time period t, and yt might represent the resulting temperature
rise (above the surrounding temperature). The lengths of u and y, m and m+n1,
are typically large, and not relevant in these applications. We can express the ith
entry of the output y as
Xn
yi = uij+1 hj ,
j=1
where we interpret uk as zero for k < 0 or k > n. This formula states that y at
time i is a linear combination of ui , ui1 , . . . , uim+1 , i.e., a linear combination
of the current input value ui , and the past m 1 input values ui1 , . . . , uim+1 .
The coefficients are precisely h1 , . . . , hm . Thus, h3 can be interpreted as the factor
by which the current output depends on what the input was, 2 times steps before.
Alternatively, we can say that h3 is the factor by which the input at any time will
affect the output 2 steps in the future.
114 7 Matrix examples

2.5

2
xk

1.5

0.5

0
0 20 40 60 80 100
k

2.5

2
(x b)k

1.5

0.5

0
0 20 40 60 80 100
k

Figure 7.5 Top. A time series represented by a vector x of length 100.

Bottom. The 3-period moving average of the time series as a vector of
length 102. This vector is the convolution of x with a = (1/3, 1/3, 1/3).
 7.4 Convolution 115

Complexity of convolution. The nave method to compute the convolution c =

a b of an n-vector a and an m-vector b, using the basic formula (7.2) to calculate
each ck , requires around 2mn flops. The same number of flops is required to
compute the matrix-vector products T (a)b or T (b)a, taking into account the zeros
at the top right and botttom left in the Toeplitz matrices T (b) and T (a). Forming
these matrices requires us to store mn numbers, even though the original data
contains only m + n numbers.
It turns out that the convolution of two vectors can be computed far faster, using
a so-called fast convolution algorithm. By exploiting the special structure of the
convolution equations, this algorithm can compute the convolution of an n-vector
and an m-vector in around 5(m + n) log2 (m + n) flops, and with no additional
memory requirement beyond the original m + n numbers. The fast convolution
algorithm is based on the fast Fourier transform (FFT), which is beyond the scope
of this book.

7.4.1 2-D convolution

Convolution has a natural extension to multiple dimensions. Suppose that A is an

mn matrix, and B is a pq matrix. Their convolution is the (m+p1)(n+q1)
matrix
X
Crs = Aij Bkl , r = 1, . . . , m + p 1, s = 1, . . . , n + q 1,
i+k=r+1, j+l=s+1

where the indices are restricted to their ranges (or alternatively, we assume that
Aij and Bkl are zero, when the indices are out of range). This is not denoted
C = A B, however, in standard mathematical notation. So we will use the
notation C = A ? B.
The same properties that we observed for 1-D convolution hold for 2-D convo-
lution: We have A ? B = B ? A, (A ? B) ? C = A ? (B ? C), and for fixed B, A ? B
is a linear function.

Image blurring. If the mn matrix X represents an image, Y = X ?B represents

the effect of blurring the image by the point spread function (PSF) given by the
entries of the matrix B. If we represent X and Y as vectors, we have y = T (B)x,
for some (m + p 1)(n + q 1) mn-matrix T (B).
As an example, with
 
1/4 1/4
B= , (7.3)
1/4 1/4

Y = X ? B is an image where each pixel value is the average of a 2 2 block of 4

adjacent pixels in X. The image Y would be perceived as the image X, with some
116 7 Matrix examples

Figure 7.6 An 8 9 image and its convolution with the point spread func-
tion (7.3).

blurring of the fine details. This is illustrated in figure 7.6 for the 8 9 matrix

1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1

1 1 0 0 0 0 0 1 1

1 1 1 0 1 1 0 1 1
X= 1 1
(7.4)
1 0 1 1 0 1 1
1 1 1 0 1 1 0 1 1

1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1

and its convolution with B,

1/4 1/2 1/2 1/2 1/2 1/2 1/2 1/2 1/2 1/4
1/2 1 1 1 1 1 1 1 1 1/2

1/2 1 3/4 1/2 1/2 1/2 1/2 3/4 1 1/2

1/2 1 3/4 1/4 1/4 1/2 1/4 1/2 1 1/2

X ?B = 1/2 1 1 1/2 1/2 1 1/2 1/2 1 1/2 .

1/2 1 1 1/2 1/2 1 1/2 1/2 1 1/2

1/2 1 1 3/4 3/4 1 3/4 3/4 1 1/2

1/2 1 1 1 1 1 1 1 1 1/2
1/4 1/2 1/2 1/2 1/2 1/2 1/2 1/2 1/2 1/4

With the point spread function

Dhor =
 
1 1 ,

the pixel values in the image Y = X ? B are the horizontal first order differences
of those in X:

Yij = Xij Xi,j1 , i = 1, . . . , m, j = 2, . . . , n

7.4 Convolution 117

(and Yi1 = Xi1 , Xi,n+1 = Xin for i = 1, . . . , m). With the point spread function
 
1
Dver = ,
1

the pixel values in the image Y = X ? B are the vertical first order differences of
those in X:
Yij = Xij Xi1,j , i = 2, . . . , m, j = 1, . . . , n
(and Y1j = X1j , Xm+1,j = Xmj for j = 1, . . . , n). As an example, the convolu-
tions of the matrix (7.4) with Dhor and Dver are

1 0 0 0 0 0 0 0 0 1
1 0 0
0 0 0 0 0 0 1

1 0 1 0 0 0 0 1 0 1

1 0 0 1 1 0 1 1 0 1
X ? Dhor =


1 0 0 1 1 0 1 1 0 1

1 0 0 1 1 0 1 1 0 1

1 0 0 0 0 0 0 0 0 1
1 0 0 0 0 0 0 0 0 1

and
1 1 1 1 1 1 1 1 1

0 0 0 0 0 0 0 0 0


0 0 1 1 1 1 1 0 0

0 0 1 0 1 1 0 0 0
X ? Dver

=
0 0 0 0 0 0 0 0 0 .


0 0 0 0 0 0 0 0 0


0 0 0 1 0 0 1 0 0

0 0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1 1
Figure 7.7 shows the effect of convolution on a larger image. The figure shows
an image of size 512512 and its convolution with the 88 matrix B with constant
entries Bij = 1/64.
118 7 Matrix examples

Figure 7.7 512 512 image and the 519 519 image that results from the
convolution of the first image with an 8 8 matrix with constant entries
1/64.
 Chapter 8

Linear equations

In this chapter we consider vector-valued linear and affine functions, and systems
of linear equations.

8.1 Linear and affine functions

Vector valued functions of vectors. The notation f : Rn Rm means that f
is a function that maps real n-vectors to real m-vectors. The value of the function
f , evaluated at an n-vector x, is an m-vector f (x) = (f1 (x), f2 (x), . . . , fm (x)).
Each of the components fi of f is itself a scalar-valued function of x. As with
scalar-valued functions, we sometimes write f (x) = f (x1 , x2 , . . . , xn ) to emphasize
that f is a function of n scalar arguments. We use the same notation for each of
the components of f , writing fi (x) = fi (x1 , x2 , . . . , xn ) to emphasize that fi is a
function mapping the scalar arguments x1 , . . . , xn into a scalar.

The matrix-vector product function. Suppose A is an m n matrix. We can

define a function f : Rn Rm by f (x) = Ax. Linear functions from Rn to R,
discussed in 2.1, are a special case with m = 1.

Superposition and linearity. The function f : Rn Rm , defined by f (x) = Ax,

is linear, i.e., it satisfies the superposition property:

f (x + y) = f (x) + f (y) (8.1)

holds for all n-vectors x and y and all scalars and . It is a good exercise to
parse this simple looking equation, since it involves overloading of notation. On
the left-hand side, the scalar-vector multiplications x and y involve n-vectors,
and the sum x+y is the sum of two n-vectors. The function f maps n-vectors to
m-vectors, so f (x + y) is an m-vector. On the right-hand side, the scalar-vector
multiplications and the sum are those for m-vectors. Finally, the equality sign is
equality between two m-vectors.
120 8 Linear equations

We can verify that superposition holds for f using properties of matrix-vector

and scalar-vector multiplication:

f (x + y) = A(x + y)
= A(x) + A(y)
= (Ax) + (Ay)
= f (x) + f (y)

Thus we can associate with every matrix A a linear function f (x) = Ax.
The converse is also true. Suppose f is a function that maps n-vectors to m-
vectors, and is linear, i.e., (8.1) holds for all n-vectors x and y and all scalars
and . Then there exists an m n matrix A such that f (x) = Ax for all x. This
can be shown in the same way as for scalar-valued functions in 2.1, by showing
that if f is linear, then

f (x) = x1 f (e1 ) + x2 f (e2 ) + + xn f (en ), (8.2)

where ek is the kth unit vector of size n. The right-hand side can also be written
as a matrix-vector product Ax, with
 
A = f (e1 ) f (e2 ) f (en ) .

The expression (8.2) is the same as (2.3), but here f (x) and f (ek ) are vectors. The
implications are exactly the same: a linear vector valued function f is completely
characterized by evaluating f at the n unit vectors e1 , . . . , en .
As in 2.1 it is easily shown that the matrix-vector representation of a linear
function is unique. If f : Rn Rm is a linear function, then there exists exactly
one matrix A such that f (x) = Ax for all x.

Examples of linear functions. In the examples below we define functions f that

map n-vectors x to n-vectors f (x). Each function is described in words, in terms of
its effect on an arbitrary x. In each case we give the associated matrix multiplication
representation.

Negation. f changes the sign of x: f (x) = x.

Negation can be expressed as f (x) = Ax with A = I.

Reversal. f reverses the order of the elements of x: f (x) = (xn , xn1 , . . . , x1 ).

The reversal function can be expressed as f (x) = Ax with

0 0 1
0 1 0
A= . . . .

.. . . . .. ..
1 0 0

(This is the n n identity matrix with the order of its columns reversed. It
is sometimes called the reverser matrix.)
8.1 Linear and affine functions 121

Running sum. f forms the running sum of the elements in x:

f (x) = (x1 , x1 + x2 , x1 + x2 + x3 , . . . , x1 + x2 + + xn ).

The running sum function can be expressed as f (x) = Ax with

1 0 0 0
1 1 0 0
A = ... .. . . . . ,

. .. ..

.
1 1 1 0
1 1 1 1

i.e., Aij = 1 if i j and Aij = 0 otherwise.

De-meaning. f subtracts the mean from each entry of a vector x: f (x) =

x avg(x)1.
The de-meaning function can be expressed as f (x) = Ax with

1 1/n 1/n 1/n
1/n 1 1/n 1/n
A= .. .. .. .

..
. . . .
1/n 1/n 1 1/n

Examples of functions that are not linear. Here we list some examples of func-
tions f that map n-vectors x to n-vectors f (x) that are not linear. In each case
we show a superposition counterexample.

Absolute value. f replaces each element of x with its absolute value: f (x) =
(|x1 |, |x2 |, . . . , |xn |).
The absolute value function is not linear. For example, with n = 1, x = 1,
y = 0, = 1, = 0, we have

f (x + y) = 1 6= f (x) + f (y) = 1,

so superposition does not hold.

Sort. f sorts the elements of x in decreasing order.

The sort function is not linear (except when n = 1, in which case f (x) = x).
For example, if n = 2, x = (1, 0), y = (0, 1), = = 1, then

f (x + y) = (1, 1) 6= f (x) + f (y) = (2, 0).

Affine functions. A vector valued function f : Rn Rm is called affine if it can

be expressed as f (x) = Ax + b, with A an m n matrix and b and m-vector. It
can be shown that a function f : Rn Rm is affine if and only if

f (x + y) = f (x) + f (y)
122 8 Linear equations

holds for all n-vectors x, y, and all scalars , that satisfy + = 1. In other
words, superposition holds for affine combinations of vectors. (For linear functions,
superposition holds for any linear combinations of vectors.)
The matrix A and the vector b in the representation of an affine function as
f (x) = Ax + b are unique. These parameters can be obtained by evaluating f at
the vectors 0, e1 , . . . , en , where ek is the kth unit vector in Rn . We have
 
A= f (e1 ) f (0) f (e2 ) f (0) f (en ) f (0) , b = f (0).

Just like affine scalar-valued functions, affine vector valued functions are often
called linear, even though they are linear only when the vector b is zero.

8.2 Linear function models

Many functions or relations between variables that arise in natural science, engi-
neering, and social sciences can be approximated as linear or affine functions. In
these cases we refer to the linear function relating the two sets of variables as a
model or an approximation, to remind us that the relation is only an approximation,
and not exact. We give a few examples here.

Price elasticity of demand. Consider n goods or services with prices given by

the n-vector p, and demands for the goods given by the n-vector d. A change
in prices will induce a change in demands. We let price be the n-vector
that gives the fractional change in the prices, i.e., iprice = (pnew i pi )/pi ,
where pnew is the n-vector of new (changed) prices. We let dem be the n-
vector that gives the fractional change in the product demands, i.e., idem =
(dnew
i di )/di , where dnew is the n-vector of new demands. A linear demand
elasticity model relates these vectors as dem = E d price , where E d is the nn
d d
demand elasticity matrix. For example, suppose E11 = 0.4 and E21 = 0.2.
This means that a 1% increase in the price of the first good, with other
prices kept the same, will cause demand for the first good to drop by 0.4%,
and demand for the second good to increase by 0.2%. (In this example, the
second good is acting as a partial substitute for the first good.)

Elastic deformation. Consider a steel structure like a bridge or the structural

frame of a building. Let f be an n-vector that gives the forces applied to
the structure at n specific places (and in n specific directions), sometimes
called a loading. The structure will deform slightly due to the loading. Let
d be an m-vector that gives the displacements (in specific directions) of m
points in the structure, due to the load, e.g., the amount of sag at a specific
point on a bridge. For small displacements, the relation between displacement
and loading is well approximated as linear: d = Cf , where C is the m n
compliance matrix. The units of the entries of C are m/N.
 8.2 Linear function models 123

8.2.1 Taylor approximation

Suppose f : Rn Rm is differentiable, i.e., has partial derivatives, and z is an
n-vector. The first-order Taylor approximation of f near z is given by
fi fi
f(x)i = fi (z) + (z)(x1 z1 ) + + (z)(xn zn )
x1 xn
= fi (z) + fi (z)T (x z),

for i = 1, . . . , m. (This is just the first-order Taylor approximation of each of the

scalar-valued functions fi , described in 2.2.) For x near z, f(x) is a very good
approximation of f (x). We can express this approximation in compact notation,
using matrix-vector multiplication, as

f(x) = f (z) + Df (z)(x z), (8.3)

where the m n matrix Df (z) is the derivative or Jacobian matrix of f at z

(see C.1). Its components are the partial derivatives of f ,
fi
Df (z)ij = (z), i = 1, . . . , m, j = 1, . . . , n,
xj

evaluated at the point z. The rows of the Jacobian are fi (z)T , for i = 1, . . . , m.
As in the scalar valued case, Taylor approximation is sometimes written with a
second argument as f(x; z) to show the point z around which the approximation is
made. Evidently the Taylor series approximation f is an affine function of x. (It is
often called a linear approximation of f , even though it is not, in general, a linear
function.)

8.2.2 Regression model

Recall the regression model (2.6)

y = xT + v, (8.4)

where the n-vector x is a feature vector for some object, is an n-vector of weights,
v is a constant (the offset), and y is the (scalar) value of the regression model
prediction.
Now suppose we have a set of N objects (also called samples or examples), with
feature vectors x1 , . . . , xN . The regression model predictions associated with the
examples are given by

yi = xTi + v, i = 1, . . . , N.

These numbers usually correspond to predictions of the value of the outputs or

repsonses. If in addition to the example feature vectors xi we are also given the
actual value of the associated response variables, y1 , . . . , yN , then our prediction
errors are
yi yi , i = 1, . . . , N.
124 8 Linear equations

We can express this using compact matrix-vector notation. We form the n N

feature matrix X with columns x1 , . . . , xN . We let y denote the N -vector whose
entries are the actual values of the response for the N examples. We let y denote
the N -vector of regression model predictions for the N examples. (So now y is an
N -vector, whereas in the equation (8.4) above, it represented a scalar.) We can
then express the regression model predictions for this data set in matrix-vector
form as
y = X T + v1.
The vector of N prediction errors for the examples is given by
y y = X T + v1 y.

We can include the offset v in the regression model by including an additional

feature equal to one as the first entry of each feature vector:
 T T  
1 v
y = = X T ,
X

where X is the new feature matrix, with a new first row of ones, and = (v, ) is
vector of regression model parameters. This is often written without the tildes, as
y = X T , by simply including the feature one as the first feature.
The equation above shows that the N -vector of predictions for the N examples
is a linear function of the model parameters (v, ). The N -vector of prediction
errors is an affine function of the model parameters.

8.3 Systems of linear equations

Consider a set (also called a system) of m linear equations in n variables or un-
knowns x1 , . . . , xn :
A11 x1 + A12 x2 + + A1n xn = b1
A21 x1 + A22 x2 + + A2n xn = b2
..
.
Am1 x1 + Am2 x2 + + Amn xn = bm .
The numbers Aij are called the coefficients in the linear equations, and the numbers
bi are called the right-hand sides (since by tradition, they appear on the right-hand
side of the equation). These equations can be written succinctly in matrix notation
as
Ax = b. (8.5)
In this context, the m n matrix A is called the coefficient matrix, and the m-
vector b is called the right-hand side. An n-vector x is called a solution of the
linear equations if Ax = b holds. A set of linear equations can have no solutions,
one solution, or multiple solutions.
 8.3 Systems of linear equations 125

Examples.

The set of linear equations

x1 + x2 = 1, x1 = 1, x1 x2 = 0

is written as Ax = b with

1 1 1
A = 1 0 , b = 1 .
1 1 0

It has no solutions.

The set of linear equations

x1 + x2 = 1, x2 + x3 = 2

is written as Ax = b with
   
1 1 0 1
A= , b= .
0 1 1 2

It has multiple solutions, including x = (1, 0, 2) and x = (0, 1, 1).

Over-determined and under-determined systems of linear equations. The set

of linear equations is called over-determined if m > n, under-determined if m < n,
and square if m = n; these correspond to the coefficient matrix being tall, wide,
and square, respectively. When the system of linear equations is over-determined,
there are more equations than variables or unknowns. When the system of linear
equations is under-determined, there are more unknowns than equations. When
the system of linear equations is square, the numbers of unknowns and equations
is the same. A set of equations with zero right-hand side, Ax = 0, is called a
homogeneous set of equations. Any homogeneous set of equations has x = 0 as a
solution.
In chapter 11 we will address the question of how to determine if a system of
linear equations has a solution, and how to find one when it does. For now, we
give a few interesting examples.

8.3.1 Examples

Coefficients of linear combinations. Let a1 , . . . , an denote the columns of A.

The system of linear equations Ax = b can be expressed as

x1 a1 + + xn an = b,

i.e., b is a linear combination of a1 , . . . , an with coefficients x1 , . . . , xn . So solving

Ax = b is the same as finding coefficients that express b as a linear combination of
the vectors a1 , . . . , an .
126 8 Linear equations

Polynomial interpolation. We seek a polynomial p of degree at most n 1 that

interpolates a set of m given points (ti , yi ), i = 1, . . . , m. (This means that p(ti ) =
yi .) We can express this as a set of m linear equations in the n unknowns c, where
c is the n-vector of coefficients: Ac = y. Here the matrix A is the Vandermonde
matrix (6.6), and the vector c is the vector of polynomial coefficients, as described
in the example on page 101.

Balancing chemical reactions. A chemical reaction involves p reactants (molecules)

and q products, and can be written as

a1 R1 + + ap Rp b1 P1 + + bq Pq .

Here R1 , . . . , Rp are the reactants, P1 , . . . , Pq are the products, and the numbers
a1 , . . . , ap and b1 , . . . , bq are positive numbers that tell us how many of each of
these molecules is involved in the reaction. They are typically integers, but can be
scaled arbitrarily; we could double all of these numbers, for example, and we still
have the same reaction. As a simple example, we have the electrolysis of water,

2H2 O 2H2 + O2 ,

which has one reactant, water (H2 O) and two products: molecular hydrogen (H2 ),
and molecular oxygen (O2 ). The coefficients tell us that 2 water molecules create
2 hydrogen molecules and 1 oxygen molecule. The coefficients in a reaction can
be multiplied by any nonzero numbers; for example, we could write the reaction
above as 3H2 O 3H2 + (3/2)O2 . By convention reactions are written with all
coefficients integers, with least common divisor one.
In a chemical reaction the numbers of constituent atoms must balance. This
means that for each atom appearing in any of the reactants or products, the total
amount on the left-hand side must equal the total amount on the right-hand side.
(If any of the reactants or products is charged, i.e., an ion, then the total charge
must also balance.) In the simple water electrolysis reaction above, for example,
we have 4 hydrogen atoms on the left (2 water molecules, each with 2 hydrogen
atoms), and 4 on the right (2 hydrogen molecules, each with 2 hydrogen atoms).
The oxygen atoms also balance, so this reaction is balanced.
Balancing a chemical reaction with specified reactants and products, i.e., finding
the numbers a1 , . . . , ap and b1 , . . . , bq , can be expressed as a system of linear
equations. We can express the requirement that the reaction balances as a set of
m equations, where m is the number of different atoms appearing in the chemical
reaction. We define the m p matrix R by

Rij = number of atoms of type i in Rj , i = 1, . . . , m, j = 1, . . . , p.

(The entries of R are nonnegative integers.) The matrix R is interesting; for exam-
ple, its jth column gives the chemical formula for reactant Rj . We let a denote the
p-vector with entries a1 , . . . , ap . Then, the m-vector Ra gives the total number of
atoms of each type appearing in the reactants. We define an m q matrix P in a
similar way, so the m-vector P b gives the total number of atoms of each type that
appears in the products.
8.3 Systems of linear equations 127

We write the balance condition using vectors and matrices as Ra = P b. We

can express this as  
  a
R P = 0,
b
which is a set of m homogeneous linear equations.
A simple solution of these equations is a = 0, b = 0. But we seek a nonzero
solution. We can set one of the coefficients, say a1 , to be one. (This might cause
the other quantities to be fractional valued.) We can add the condition that a1 = 1
to our system of linear equations as
  
R P a
= em+1 .
eT1 0 b

Finally, we have a set of m + 1 equations in p + q variables that expresses the

requirement that the chemical reaction balances. Finding a solution of this set of
equations is called balancing the chemical reaction.
For the example of electrolysis of water described above, we have p = 1 reac-
tant (water) and q = 2 products (molecular hydrogen and oxygen). The reaction
involves m = 2 atoms, hydrogen and oxygen. The reactant and product matrices
are    
2 2 0
R= , P = .
1 0 2
The balancing equations are then

2 2 0 a1 0
1 0 2 b1 = 0 .
1 0 0 b2 1

These equations are easily solved, and have the solution (1, 1, 1/2). (Multiplying
these coefficients by 2 gives the reaction given above.)

Diffusion systems. A diffusion system is a common model that arises in many

areas of physics to describe flows and potentials. We start with a directed graph
with n nodes and m edges. (See 6.1.) Some quantity (like electricity, heat, energy,
or mass) can flow across the edges, from one node to another.
With edge j we associate a flow (rate) fj , which is a scalar; the vector of all
m flows is the flow m-vector f . The flows fj can be positive or negative: Positive
fj means the quantity flows in the direction of edge j, and negative fj means
the quantity flows in the opposite direction of edge j. The flows can represent, for
example, heat flow (in units of Watts) in a thermal model, electrical current (Amps)
in an electrical circuit, or movement (diffusion) of mass (such as, for example, a
pollutant). We also have a source (or exogenous) flow si at each node, with si > 0
meaning that an exogenous flow is injected into node i, and si < 0 means that
an exogenous flow is removed from node i. (In some contexts, a node where flow
is removed is called a sink.) In a thermal system, the sources represent thermal
(heat) sources; in an electrical circuit, they represent electrical current sources; in
a system with diffusion, they represent external injection or removal of the mass.
128 8 Linear equations

s1 2
1

Figure 8.1 A node in a diffusion system with label 1, exogeneous flow s1 and
three incident edges.

8
2 3

Figure 8.2 The flow through edge 8 is equal to f8 = r8 (e2 e3 ).

In a diffusion system, the flows must satisfy (flow) conservation, which means
that at each node, the total flow entering each node from adjacent edges and the
exogenous source, must be zero. This is illustrated in figure 8.1, which shows three
edges adjacent to node 1, two entering node 1 (flows 1 and 2), and one (flow 3)
leaving node 1, and an exogeneous flow. Flow conservation at this node is expressed
as
f1 + f2 f3 + s1 = 0.

Flow conservation at every node can be expressed by the simple matrix-vector

equation
Af + s = 0, (8.6)
where A is the incidence matrix described in 7.3. (This is called Kirchhoff s
current law in an electrical circuit; when the flows represent movement of mass, it
is called conservation of mass.)
With node i we associate a potential ei ; the n-vector e gives the potential at all
nodes. The potential might represent the node temperature in a thermal model,
the electrical potential (voltage) in an electrical circuit, and the concentration in a
system that involves mass diffusion.
In a diffusion system the flow on an edge is proportional to the potential differ-
ence across its adjacent nodes. This is typically written as rj fj = ek el , where
edge j goes from node k to node l, and rj (which is typically positive) is called the
resistance of edge j. In a thermal model, rj is called the thermal resistance of the
edge; in an electrical circuit, it is called the electrical resistance. This is illustrated
in figure 8.2, which shows edge 8, connecting node 2 and node 3, corresponding to
an edge flow equation
r8 f8 = e2 e3 .
8.3 Systems of linear equations 129

We can write the edge flow equations in a compact way as

Rf = AT e, (8.7)

where R = diag(r) is called the resistance matrix.

The diffusion model can be expressed as one set of block linear equations in the
variables f , s, and e:
  f
A I 0 s = 0.
R 0 AT
e
This is a set of n + m homogeneous equations in m + 2n variables. To these under-
determined equations we can add others, for example, by specifying some of the
entries of f , s, and e.

Leontief input-ouput model. We consider an economy with n different industrial

sectors. We let xi be the economic activity level, or total production output, of
sector i, for i = 1, . . . , n, measured in a common unit, such as (billions of) dollars.
The output of each sector flows to other sectors, to support their production, and
also to consumers. We denote the total consumer demand for sector i as di , for
i = 1, . . . , n.
Supporting the output level xj for sector j requires Aij xj output for sector i.
We refer to Aij xj as the sector i input that flows to sector j. (We can have Aii 6= 0;
for example, it requires some energy to support the production of energy.) Thus,
Ai1 x1 + + Ain xn is the total sector i output required by, or flowing into, the n
industrial sectors. The matrix A is called the input-output matrix of the economy,
since it describes the flows of sector outputs to the inputs of itself and other sectors.
The vector Ax gives the sector outputs required to support the production levels
given by x. (This sounds circular, but isnt.)
Finally, we require that for each sector, the total production level matches the
demand plus the total amount required to support production. This leads to the
balance equations,
x = Ax + d.
Suppose the demand vector d is given, and we wish to find the sector output
levels that will support it. We can write this as a set of n equations in n unknowns,

(I A)x = d.

This model of the sector inputs and outputs of an economy was developed by
Wassily Leontief in the late 1940s, and is now known as Leontief input-output
analysis. He was awarded the Nobel Prize in Economics for this work in 1973.
130 8 Linear equations
 Chapter 9

Linear dynamical systems

In this chapter we consider a useful application of matrix-vector multiplication,

which is used to describe many systems or phenomena that change or evolve over
time.

9.1 Linear dynamical systems

Suppose x1 , x2 , . . . is a sequence of n-vectors. The index (subscript) denotes time
or period, and is written as t; xt , the value of the sequence at time (or period)
t, is called the state at time t. We can think of xt as a vector that changes over
time, i.e., one that changes dynamically. In this context, the sequence x1 , x2 , . . .
is sometimes called a trajectory or state trajectory. We sometimes refer to xt as the
current state of the system (implicitly assuming the current time is t), and xt+1 as
the next state, xt1 as the previous state, and so on.
The state xt can represent a portfolio that changes daily, or the positions and
velocities of the parts of a mechanical system, or the quarterly activity of an econ-
omy. If xt represents a portfolio that changes daily, (x5 )3 is the amount of asset 3
held in the portfolio on (trading) day 5.
A linear dynamical system is a simple model for the sequence, in which each
xt+1 is a linear function of xt :

xt+1 = At xt , t = 1, 2, . . . . (9.1)

Here the n n matrices At are called the dynamics matrices. The equation above
is called the dynamics or update equation, since it gives us the next value of x, i.e.,
xt+1 , as a function of the current value xt . Often the dynamics matrix does not
depend on t, in which case the linear dynamical system is called time-invariant.
If we know xt (and At , At+1 , . . . ,) we can determine xt+1 , xt+2 , . . . , simply
by iterating the dynamics equation (9.1). In other words: If we know the current
value of x, we can find all future values. In particular, we do not need to know the
past states. This is why xt is called the state of the system. Roughly speaking, it
contains all the information needed to determine its future evolution.
132 9 Linear dynamical systems

Linear dynamical system with input. There are many variations on and exten-
sions of the basic linear dynamical system model (9.1), some of which we will
encounter in the sequel. As an example, we can add an additional term to the
update equation:

xt+1 = At xt + Bt ut + ct , t = 1, 2, . . . . (9.2)

Here ut is an m-vector called the input (at time t) and Bt are n m matrices
called the input matrices. The vector ct is called the offset. The input and offset
are used to model other factors that affect the time evolution of the state. Another
name for the input ut is exogenous variable, since, roughly speaking, it comes from
outside the system.

Markov model. The linear dynamical system (9.1) is sometimes called a Markov
model (after the famous mathematician Andrey Markov). Markov studied systems
in which the next state value depends on the current one, and not on the previous
state values xt1 , xt2 , . . . . The linear dynamical system (9.1) is the special case
of a Markov system where the next state is a linear function of the current state.
In a variation on the Markov model, called a (linear) K-Markov model, the
next state xt+1 depends on the current state and K 1 previous states. Such a
system has the form

xt+1 = A1 xt + + AK xtK+1 , t = K, K + 1, . . . . (9.3)

Models of this form are used in time series analysis and econometrics, where they
are called auto-regressive models. We will see later that the Markov model (9.3)
can be reduced to a standard linear dynamical system (9.1).

Simulation. If we know the dynamics (and input) matrices, and the state at time
t, we can find the future state trajectory xt+1 , xt+2 , . . . by iterating the equa-
tion (9.1) (or (9.2), provided we also know the input sequence ut , ut+1 , . . . ). This
is called simulating the linear dynamical system. Simulation makes predictions
about the future state of a system. (To the extent that (9.1) is only an approx-
imation or model of some real system, we must be careful when interpreting the
results.) We can carry out what-if simulations, to see what would happen if the
system changes in some way, or if a particular set of inputs occurs.

9.2 Population dynamics

Linear dynamical systems can be used to describe the evolution of the age dis-
tribution in some population over time. Suppose xt is a 100-vector, with (xt )i
denoting the number of people in some population (say, a country) with age i 1
(say, on January 1) in year t, where t is measured starting from some base year, for
i = 1, . . . , 100. While (xt )i is an integer, it is large enough that we simply consider
it a real number. In any case, our model certainly is not accurate at the level of
9.2 Population dynamics 133

4
Population (millions)
3

0
0 10 20 30 40 50 60 70 80 90 100
Age

Figure 9.1 Age distribution in the US in 2010. (United States Census Bu-
reau, [Link]).

individual people. Also, note that the model does not track people 100 and older.
The distribution of ages in the US in 2010 is shown in figure 9.1.
The birth rate is given by a 100-vector b, where bi is the average number of
births per person with age i 1, i = 1, . . . , 100. (This is half the average number
of births per woman with age i 1, assuming equal numbers of men and women
in the population.) Of course bi is approximately zero for i < 13 and i > 50. The
approximate birth rates for the US in 2010 are shown in figure 9.2. The death rate
is given by a 100-vector d, where di is the portion of those aged i 1 who will die
this year. The death rates for the US in 2010 are shown in figure 9.3.
To derive the dynamics equation (9.1), we find xt+1 in terms of xt , taking into
account only births and deaths, and not immigration. The number of 0-year olds
next year is the total number of births this year:

(xt+1 )1 = bT xt .

The number of i-year olds next year is the number of (i 1)-year-olds this year,
minus those who die:

(xt+1 )i+1 = (1 di )(xt )i , i = 1, . . . , 99.

We can assemble these equations into the time-invariant linear dynamical system

xt+1 = Axt , t = 1, 2, . . . , (9.4)

134 9 Linear dynamical systems

Birth rate (%) 2

0 10 20 30 40 50 60 70 80 90 100
Age

Figure 9.2 Approximate birth rate versus age in the US in 2010. The figure is
based on statistics for age groups of five years (hence, the piecewise-constant
shape) and assumes an equal number of men and women in each age group.
(Martin J.A., Hamilton B.E., Ventura S.J. et al., Births: Final data for 2010.
National Vital Statistics Reports; vol. 61, no. 1. National Center for Health
Statistics, 2012.)

30
Death rate (%)

20

10

0 10 20 30 40 50 60 70 80 90 100
Age

Figure 9.3 Death rate versus age, for ages 099, in the US in 2010. (Centers
for Disease Control and Prevention, National Center for Health Statistics,
[Link].)
9.2 Population dynamics 135

Population (millions)
3

0
0 10 20 30 40 50 60 70 80 90 100
Age

Figure 9.4 Predicted age distribution in the US in 2020.

where A is given by

b1 b2 b3 b98 b99 b100

1 d1 0 0 0 0 0

0 1 d2 0 0 0 0
A= .. .. .. .. .. .. .


. . . . . .

0 0 0 1 d98 0 0
0 0 0 0 1 d99 0
We can use this model to predict the total population in 10 years (not including
immigration), or to predict the number of school age children, or retirement age
adults. Figure 9.4 shows the predicted age distribution in 2020, computed by
iterating the model xt+1 = Axt for t = 1, . . . , 10, with initial value x1 given by
the 2010 age distribution of figure 9.1. Note that the distribution is based on an
approximate model, since we neglect the effect of immigration, and assume that the
death and birth rates remain constant and equal to the values shown in figures 9.2
and 9.3.
Population dynamics models are used to carry out projections of the future age
distribution, which in turn is used to predict how many retirees there will be in
some future year. They are also used to carry out various what if analyses, to
predict the effect of changes in birth or death rates on the future age distribution.
It is easy to include the effects of immigration and emigration in the population
dynamics model (9.4), by simply adding a 100-vector ut :
xt+1 = Axt + ut ,
which is a time-invariant linear dynamical system of the form (9.2), with input ut
and B = I. The vector ut gives the net immigration in year t over all ages; (ut )i
is the number of immigrants in year t of age i 1. (Negative entries mean net
emigration.)
136 9 Linear dynamical systems

9.3 Epidemic dynamics

The dynamics of infection and the spread of an epidemic can be modeled using a
linear dynamical sustem. (More sophisticated nonlinear epidemic dynamic models
are also used.) In this section we describe a simple example.
A disease is introduced into a population. In each period (say, days) we count
the fraction of the population that is in four different infection states:
Susceptible. These individuals can aquire the disease the next day.
Infected. These individuals have the disease.
Recovered (and immune). These individuals had the disease and survived,
and now have immunity.
Deceased. These individuals had the disease, and unfortunately died from it.
We denote the fractions of each of these as a 4-vector xt , so, for example, xt =
(0.75, 0.10, 0.10, 0.05) means that in day t, 75% of the population is susceptible,
10% is infected, 10% is recovered and immune, and 5% have died from the disease.
There are many mathematical models that predict how the disease state frac-
tions xt evolve over time. One simple model can be expressed as a linear dynamical
system. The model assumes the following happens over each day.
5% of the susceptible population will acquire the disease. (The other 95%
will remain susceptible.)
1% of the infected population will die from the disease, 10% will recover
and acquire immunity, and 4% will recover and not acquire immunity (and
therefore, become susceptible). The remaining 85% will remain infected.
(Those who have have recovered with immunity and those who have died remain
in those states.)
We first determine (xt+1 )1 , the fraction of susceptible individuals in the next
day. These include the susceptible individuals from today, who did not become
infected, which is 0.95(xt )1 , plus the infected individuals today who recovered
without immunity, which is 0.04(xt )2 . All together we have (xt+1 )1 = 0.95(xt )1 +
0.04(xt )2 . We have (xt+1 )2 = 0.85(xt )2 + 0.05(xt )1 ; the first term counts those
who are infected and remain infected, and the second term counts those who are
susceptible and acquire the disease. A similar argument gives (xt+1 )3 = (xt )3 +
0.10(xt )2 , and (xt+1 )4 = (xt )4 + 0.01(xt )2 . We put these together to get

0.95 0.04 0 0
0.05 0.85 0 0
xt+1 = x ,
0 0.10 1 0 t
0 0.01 0 1
which is a time-invariant linear dynamical system of the form (9.1).
Figure 9.5 shows the evolution of the four groups from the initial condition x0 =
(1, 0, 0, 0). The simulation shows that after around 100 days, the state converges
to one with a little under 10% of the population deceased, and the remaining
population immune.
 9.4 Motion of a mass 137

1
Susceptible
Infected
0.8 Recovered
Deceased

0.6
xt

0.4

0.2

0
0 50 100 150 200
Time t

Figure 9.5 Simulation of epidemic dynamics.

f ( )
m

0 p( )

Figure 9.6 Mass moving along a line.

9.4 Motion of a mass

Linear dynamical systems can be used to (approximately) describe the motion of
many mechanical systems, for example, an airplane (that is not undergoing extreme
maneuvers), or the (hopefully not too large) movement of a building during an
earthquake.
Here we describe the simplest example: A single mass moving in 1-D (i.e., a
straight line), with an external force and a drag force acting on it. This is illustrated
in figure 9.6. The (scalar) position of the mass at time is given by p( ). (Here
is continuous, i.e., a real number.) The position satisfies Isaac Newtons law of
motion, the differential equation

d2 p dp
m ( ) = ( ) + f ( ),
d 2 d
where m > 0 is the mass, f ( ) is the external force acting on the mass at time ,
and > 0 is the drag coefficient. The right-hand side is the total force acting on
the mass; the first term is the drag force, which is proportional to the velocity and
in the opposite direction.
Introducing the velocity of the mass, v( ) = dp( )/d , we can write the equation
138 9 Linear dynamical systems

above as two coupled differential equations,

dp dv
( ) = v( ), m ( ) = v( ) + f ( ).
d d
The first equation relates the position and velocity; the second is from the law of
motion.
To develop an (approximate) linear dynamical system model, we first discretize
time. We let h > 0 be a time interval (called the sampling interval) that is small
enough that the velocity and forces do not change very much over h seconds. We
define
pk = p(kh), vk = v(kh), fk = f (kh),
which are the continuous quantities sampled at multiples of h seconds. We now
use the approximations
dp pk+1 pk dv vk+1 vk
(kh) , (kh) , (9.5)
d h d h
which are justified since h is small. This leads to the (approximate) equations
(replacing with =)
pk+1 pk vk+1 vk
= vk , m = fk vk .
h h
Finally, using state xk = (pk , vk ), we write this as
   
1 h 0
xk+1 = xk + fk , k = 1, 2, . . . ,
0 1 h/m h/m
which is a linear dynamical system of the form (9.2), with input fk and dynamics
and input matrices
   
1 h 0
A= , B= .
0 1 h/m h/m
This linear dynamical system gives an approximation of the true motion, due to
our approximation (9.5) of the derivatives. But for h small enough, it is accurate.
This linear dynamical system can be used to simulate the motion of the mass, if
we know the external force applied to it, i.e., u1 , u2 , . . . .

Example. As a simple example, we consider the case with m = 1 (kilogram),

= 1 (Newtons per meter per second) and sampling period h = 0.01 (seconds).
The external force is

0.0 0.0 < 0.5

1.0 0.5 < 1.0

f ( ) =

1.3 1.0 < 1.4
0.0 1.4 .

We simulate this system for a period of 2.5 seconds, starting from initial state
x1 = (0, 0), which corresponds to the mass starting at rest (zero velocity) at position
0. The simulation involves iterating the dynamics equation from k = 1 to k = 250.
Figure 9.7 shows the force, position, and velocity of the mass, with the axes labeled
using continuous time .
9.4 Motion of a mass 139

0.5

0
f ( )

0.5

1.5
0 1 2 3 4 5 6

0.15

0.1
p( )

0.05

0 1 2 3 4 5 6

0.4

0.2
v( )

0.2
0 1 2 3 4 5 6

Figure 9.7 Simulation of mass moving along a line. Applied force (top),
position (middle), and velocity (bottom).
140 9 Linear dynamical systems

9.5 Supply chain dynamics

The dynamics of a supply chain can often be modeled using a linear dynamical
system. (This simple model does not include some important aspects of a real
supply chain, for example limits on storage at the warehouses, or the fact that
demand fluctuates.) We give a simple example here.
We consider a supply chain for a single divisible commodity (say, oil or gravel, or
discrete quantities so small that their quantities can be considered real numbers).
The commodity is stored at n warehouses or storage locations. Each of these
locations has a target (desired) level or amount of the commodity, and we let the
n-vector xt denote the deviations of the levels of the commodities from their target
levels. For example, (x5 )3 is the actual commodity level at location 3, in period 5,
minus the target level for location 3. If this is positive it means we have more than
the target level at the location; if it is negative, we have less than the target level
at the location.
The commodity is moved or transported in each period over a set of m trans-
portation links between the storage locations, and also enters and exits the nodes
through purchases (from suppliers) and sales (to end-users). The purchases and
sales are given by the n-vectors pt and st , respectively. We expect these to be posi-
tive; but they can be negative if we include returns. The net effect of the purchases
and sales is that we add (pt st )i of the commodity at location i. (This number is
negative if we sell more than we purchase at the location.)
We describe the links by the nm incidence matrix Asc (see 7.3). The direction
of each link does not indicate the direction of commodity flow; it only sets the
reference direction for the flow: Commodity flow in the direction of the link is
considered positive and commodity flow in the opposite direction is considered
negative. We describe the commodity flow in period t by the m-vector ft . For
example, (f6 )2 = 1.4 means that in time period 6, 1.4 units of the commodity
are moved along link 2 in the direction opposite the link direction (since the flow
is negative). The n-vector Asc ft gives the net flow of the commodity into the n
locations, due to the transport across the links.
Taking into account the movement of the commodity across the network, and
the purchase and sale of the commodity, we get the dynamics

xt+1 = xt + Asc ft + pt st , t = 1, 2, . . . .

In applications where we control or run a supply chain, st is beyond our control,

but we can manipulate ft (the flow of goods between storage locations) and pt
(purchases at locations). This suggests treating st as the offset, and ut = (ft , pt )
as the input in a linear dynamical system with input (9.2). We can write the
dynamics equations above in this form, with dynamics and input matrices

B = Asc I .
 
A = I,

(Note that Asc refers to the supply chain graph incidence matrix, while A is the
dynamics matrix in (9.2).) This gives

xt+1 = Axt + B(ft , pt ) st , t = 1, 2, . . . , .

9.5 Supply chain dynamics 141

s2

2
f1
s1 p2

1 f3

p1 s3
f2
3

p3

Figure 9.8 A simple supply chain with n = 3 storage locations and m = 3

transportation links.

A simple example is shown in figure 9.8. The supply chain dynamics equation
is
1 1 0 1 0 0  
ft
xt+1 = xt + 1 0 1 0 1 0 st , t = 1, 2, . . . .
pt
0 1 1 0 0 1
It is a good exercise to check that the matrix-vector product (the middle term of
the right-hand side) gives the amount of commodity added at each location, as a
result of shipment and purchasing.
142 9 Linear dynamical systems
 Chapter 10

Matrix multiplication

In this chapter we introduce matrix multiplication, a generalization of matrix-vector

multiplication, and describe several interpretations and applications.

10.1 Matrix-matrix multiplication

It is possible to multiply two matrices using matrix multiplication. You can multiply
two matrices A and B provided their dimensions are compatible, which means the
number of columns of A equals the number of rows of B. Suppose A and B are
compatible, e.g., A has size m p and B has size p n. Then the product matrix
C = AB is the m n matrix with elements
p
X
Cij = Aik Bkj = Ai1 B1j + +Aip Bpj , i = 1, . . . , m, j = 1, . . . , n. (10.1)
k=1

There are several ways to remember this rule. To find the i, j element of the
product C = AB, you need to know the ith row of A and the jth column of B.
The summation above can be interpreted as moving left to right along the ith row
of A while moving top to bottom down the jth column of B. As you go, you
keep a running sum of the product of elements, one from A and one from B.
As a specific example, we have

  1 1  
1.5 3 2 3.5 4.5
0 2 = .
1 1 0 1 1
1 0

To find the 1, 2 entry of the right-hand matrix, we move along the first row of
the left-hand matrix, and down the second column of the middle matrix, to get
(1.5)(1) + (3)(2) + (2)(0) = 4.5.
Matrix-matrix multiplication includes as special cases several other types of
multiplication (or product) we have encountered so far.
144 10 Matrix multiplication

Scalar-vector product. If x is a scalar and a is a number, we can interpret the

scalar-vector product xa, with the scalar appearing on the right, as a special case
of matrix-matrix multiplication. We consider x as an n 1 matrix, and a as a 1 1
matrix. The matrix product xa then makes sense, and coincides with the scalar-
vector product xa, which we usually write (by convention) as ax. But note that ax
cannot be interpreted as matrix-matrix multiplication (except when n = 1), since
the number of columns of a (which is 1) is not equal to the number of rows of x
(which is n).

Inner product. An important special case of matrix-matrix multiplication is the

multiplication of a row vector with a column vector. If a and b are n-vectors, then
the inner product
aT b = a1 b1 + a2 b2 + + an bn
can be interpreted as the matrix-matrix product of the 1 n matrix aT and the
n 1 matrix b. The result is a 1 1 matrix, which we consider to be a scalar. (This
explains the notation aT b for the inner product of vectors a and b, defined in 1.4.)

Matrix-vector multiplication. The matrix-vector product y = Ax defined in (6.3)

can be interpreted as a matrix-matrix product of A with the n 1 matrix x.

Vector outer product. The outer product of an m-vector a and an n-vector b is

given by abT , which is an m n matrix

a1 b1 a1 b2 a1 bn
a2 b1 a2 b2 a2 bn
abT = . .. .. ,

.. . .
am b1 am b2 am bn
whose entries are all products of the entries of a and the entries of b. Note that
the outer product does not satisfy abT = baT , i.e., it is not symmetric (like the
inner product). Indeed, the equation abT = baT does not even make sense, unless
m = n; even then, it is not true in general.

Multiplication by identity. If A is any m n matrix, then AI = A and IA = A,

i.e., when you multiply a matrix by an identity matrix, it has no effect. (Note the
different sizes of the identity matrices in the formulas AI = A and IA = A.)

Matrix multiplication order matters. Matrix multiplication is (in general) not

commutative: we do not (in general) have AB = BA. In fact, BA may not even
make sense, or, if it makes sense, may be a different size than AB. For example, if
A is 2 3 and B is 3 4, then AB makes sense (the dimensions are compatible)
but BA does not even make sense (the dimensions are incompatible). Even when
AB and BA both make sense and are the same size, i.e., when A and B are square,
we do not (in general) have AB = BA. As a simple example, take the matrices
   
1 6 0 1
A= , B= .
9 3 1 2
10.1 Matrix-matrix multiplication 145

We have    
6 11 9 3
AB = , BA = .
3 3 17 0
Two matrices A and B that satisfy AB = BA are said to commute. (Note that for
AB = BA to make sense, A and B must both be square.)

Properties of matrix multiplication. The following properties hold and are easy
to verify from the definition of matrix multiplication. We assume that A, B, and
C are matrices for which all the operations below are valid, and that is a scalar.
Associativity: (AB)C = A(BC). Therefore we can write the product simply
as ABC.
Associativity with scalar multiplication: (AB) = (A)B, where is a scalar
and A and B are matrices (that can be multiplied). This is also equal to
A(B). (Note that the products A and B are defined as scalar-matrix
products, but in general, unless A and B have one row, not as matrix-matrix
products.)
Distributivity with addition: Matrix multiplication distributes across matrix
addition: A(B+C) = AB+AC and (A+B)C = AC +BC. On the right-hand
sides of these equations we use the higher precedence of matrix multiplication
over addition, so, for example, AC + BC is interpreted as (AC) + (BC).
Transpose of product. The transpose of a product is the product of the
transposes, but in the opposite order: (AB)T = B T AT .
From these properties we can derive others. For example, if A, B, C, and D are
square matrices of the same size, we have the identity
(A + B)(C + D) = AC + AD + BC + BD.
This is the same as the usual formula for expanding a product of sums of scalars;
but with matrices, we must be careful to preserve the order of the products.

Inner product and matrix-vector products. As an exercise on matrix-vector prod-

ucts and inner products, one can verify that if A is m n, x is an n-vector, and y
is an m-vector, then
y T (Ax) = (y T A)x = (AT y)T x,
i.e., the inner product of y and Ax is equal to the inner product of x and AT y. (Note
that when m 6= n, these inner products involve vectors with different dimensions.)

Products of block matrices. Suppose A is a block matrix with mp block entries

Aij , and B is a block matrix with p q block entries Bij , and for each k = 1, . . . , p,
the matrix product Aik Bkj makes sense, i.e., the number of columns of Aik equals
the number of rows of Bkj . Then C = AB can be expressed as the m n block
matrix with entries Cij , given by the formula (10.1). For example, we have
    
A B E F AE + BG AF + BH
= ,
C D G H CE + DG CF + DH
146 10 Matrix multiplication

for any matrices A, B, . . . , H for which the matrix products above make sense. This
formula is the same as the formula for multiplying two 2 2 matrices (i.e., with
scalar entries); but when the entries of the matrix are themselves matrices (as in
the block matrix above), we must be careful to preserve the multiplication order.

Column interpretation of matrix-matrix product. We can derive some additional

insight into matrix multiplication by interpreting the operation in terms of the
columns of the second matrix. Consider the matrix product of an m p matrix A
and an p n matrix B, and denote the columns of B by bk , and the rows of A by
aTk . Using block-matrix notation, we can write the product AB as
   
AB = A b1 b2 bn = Ab1 Ab2 Abn .
Thus, the columns of AB are the matrix-vector products of A and the columns of
B. The product AB can be interpreted as the matrix obtained by applying A to
each of the columns of B.

Multiple linear equations. We can use the column interpretation of matrix mul-
tiplication to express a set of k linear equations with the same m n coefficient
matrix F ,
F xi = gi , i = 1, . . . , k,
in the compact form
F X = G,
where X = [x1 xk ] and G = [g1 gk ].

Row interpretation of matrix-matrix product. We can give an analogous row

interpretation of the product AB, by partitioning A and AB as block matrices
with row vector blocks:
T
aT1 B (B T a1 )T

a1
aT2 aT2 B (B T a2 )T
AB = . B = .. = .. .

.
. . .
aTm aTm B (B T am )T

This shows that the rows of AB are obtained by applying B T to the transposed
row vectors ak of A.

Inner product representation. From the definition of the i, j element of AB

in (10.1), we also see that the elements of AB are the inner products of the rows
of A with the columns of B:
T
a1 b1 aT1 b2 aT1 bn

aT2 b1 aT2 b2 aT2 bn
AB = . .. .. .

.. . .
aTm b1 aTm b2 aTm bn

Thus we can interpret the matrix-matrix product as the mn inner products aTi bj
arranged in an m n matrix.
10.1 Matrix-matrix multiplication 147

Gram matrix. For an m n matrix A, with columns a1 , . . . , an , the matrix

product G = AT A is called the Gram matrix associated with the set of m-vectors
a1 , . . . , an . (It is named after the mathematician Jrgen Pedersen Gram.) From
the inner product interpretation above, the Gram matrix can be expressed as
T
a1 a1 aT1 a2 aT1 an

aT2 a1 aT2 a2 aT2 an
G = AT A = . .. .. .

.. ..
. . .
aTn a1 aTn a2 aTn an

The entries of the Gram matrix G give all inner products of pairs of columns of A.
Note that a Gram matrix is symmetric, since aTi aj = aTj ai . This can also be seen
using the transpose of product rule:

GT = (AT A)T = (AT )(AT )T = AT A = G.

The Gram matrix will play an important role later in this book.
As an example, suppose the m n matrix A gives the membership of m items
in n groups, with entries

1 item i is in group j
Aij =
0 item i is not in group j.

(So the jth column of A gives the membership in the jth group, and the ith row
gives the groups that item i is in.) In this case the Gram matrix G has a nice
interpretation: Gij is the number of items that are in both groups i and j, and Gii
is the number of items in group i.

Outer product representation. If we express the m p matrix A in terms of its

columns a1 , . . . , ap and the p n matrix B in terms of its rows bT1 , . . . , bTp ,
T
b1
  ..
A = a1 ap , B = . ,
bTp

then we can express the product matrix AB as

AB = a1 bT1 + + ap bTp ,

a sum of outer products.

Complexity of matrix multiplication. The total number of flops required for a

matrix-matrix product C = AB with A of size m p and B of size p n can be
found several ways. The product matrix C has size mn, so there are mn elements
to compute. The i, j element of C is the inner product of row i of A with column
j of B. This is an inner product of vectors of length p and requires 2p 1 flops.
Therefore the total is mn(2p 1) flops, which we approximate as 2mnp flops. The
order of computing the matrix-matrix product is mnp, the product of the three
dimensions involved.
148 10 Matrix multiplication

In some special cases the complexity is less than 2mnp flops. As an example,
when we compute the m m Gram matrix G = AT A we only need to compute the
entries in the upper (or lower) half of G, since G is symmetric. This saves around
half the flops, so the complexity is around mn2 flops. But the order is the same.

Complexity of sparse matrix multiplication. Multiplying sparse matrices can be

done efficiently, since we dont need to carry out any multiplications in which one
or the other entry is zero. We start by analyzing the complexity of multiplying
a sparse matrix with a non-sparse matrix. Suppose that A is m p and sparse,
and B is p n, but not necessarily sparse. The inner product of the ith row aTi
of A with the jth column of B requires no more than 2 nnz(aTi ) flops. Summing
over i = 1, . . . , m and j = 1, . . . , n we get 2 nnz(A)n flops. If B is sparse, the
total number of flops is no more that 2 nnz(B)m flops. (Note that these formulas
agree with the one given above, 2mnp, when the sparse matrices have all entries
nonzero.)
There is no simple formula for the complexity of multiplying two sparse matri-
ces, but it is certainly no more than 2 min{nnz(A)n, nnz(B)m} flops.

Complexity of matrix triple product. Consider the product of three matrices,

D = ABC

with A of size m n, B of size n p, and C of size p q. The matrix D can be

computed in two ways, as (AB)C and as A(BC). In the first method we start with
AB (2mnp flops) and then form D = (AB)C (2mpq flops), for a total of 2mp(n+q)
flops. In the second method we compute the product BC (2npq flops) and then
form D = A(BC) (2mnq flops), for a total of 2nq(m + p) flops.
You might guess that the total number of flops required is the same with the
two methods, but it turns out it is not. The first method is less expensive when
2mp(n + q) < 2nq(m + p), i.e., when

1 1 1 1
+ < + .
n q m p

For example, if m = p and n = q, the first method has a complexity proportional

to m2 n, while the second method has complexity mn2 , and one would prefer the
first method when m  n.
As a more specific example, consider the product abT c, where a, b, c are n-
vectors. If we first evaluate the outer product abT , the cost is n2 flops, and we
need to store n2 values. We then multiply the vector c by this n n matrix, which
costs 2n2 flops. The total cost is 3n2 flops. On the other hand if we first evaluate
the inner product bT c, the cost is 2n flops, and we only need to store one number
(the result). Multplying the vector a by this number costs n flops, so the total cost
is 3n flops. For n large, there is a dramatic difference between 3n and 3n2 flops.
(The storage requirements are also dramatically different for the two methods of
evaluating abT c: one number versus n2 numbers.)
 10.2 Composition of linear functions 149

10.2 Composition of linear functions

Matrix-matrix products and composition. Suppose A is an m p matrix and B
is p n. We can associate with these matrices two linear functions f : Rp Rm
and g : Rn Rp , defined as f (x) = Ax and g(x) = Bx. The composition of the
two functions is the function h : Rn Rm with

h(x) = f (g(x)) = A(Bx) = (AB)x.

In words: to find h(x), we first apply the function g, to obtain the partial result
g(x) (which is a p-vector); then we apply the function f to this result, to obtain
h(x) (which is an m-vector). In the formula h(x) = f (g(x)), f appears to the left
of g; but when we evaluate h(x), we apply g first. The composition h is evidently
a linear function, that can be written as h(x) = Cx with C = AB.
Using this interpretation of matrix multiplication as composition of linear func-
tions, it is easy to understand why in general AB 6= BA, even when the dimen-
sions are compatible. Evaluating the function h(x) = ABx means we first evaluate
y = Bx, and then z = Ay. Evaluating the function BAx means we first evaluate
y = Ax, and then z = By. In general, the order matters. As an example, take the
2 2 matrices    
1 0 0 1
A= , B= ,
0 1 1 0
for which    
0 1 0 1
AB = , BA = .
1 0 1 0
The mapping f (x) = Ax = (x1 , x2 ) changes the sign of the first element of the
vector x. The mapping g(x) = Bx = (x2 , x1 ) reverses the order of two elements
of x. If we evaluate f (g(x)) = ABx = (x2 , x1 ), we first reverse the order, and
then change the sign of the first element. This result is obviously different from
g(f (x)) = BAx = (x2 , x1 ), obtained by changing the sign of the first element,
and then reversing the order of the elements.

Second difference matrix. As a more interesting example of composition of linear

functions, consider the (n 1) n difference matrix Dn defined in (6.4). (We use
the subscript n here to denote size of D.) Let Dn1 denote the (n 2) (n 1)
difference matrix. Their product Dn1 Dn is called the second difference matrix,
and sometimes denoted .
We can interpret in terms of composition of linear functions. Multiplying an
n-vector x by Dn yields the (n 1)-vector of consecutive differences of the entries:

Dn x = (x2 x1 , . . . , xn xn1 ).

Multiplying this vector by Dn1 gives the (n 2)-vector of consecutive differences

of consecutive differences (or second differences) of x:

Dn1 Dn x = (x1 2x2 + x3 , x2 2x3 + x4 , . . . , xn2 2xn1 + xn ).

The (n 2) n product matrix = Dn1 Dn is the matrix associated with the

second difference function.
150 10 Matrix multiplication

For the case n = 5, = Dn1 Dn has the form

1 1 0 0 0
1 2 1 0 0 1 1 0 0
0 1 1 0 0
0 1 2 1 0 = 0 1 1 0 .
0 0 1 1 0
0 0 1 2 1 0 0 1 1
0 0 0 1 1

The left-hand matrix is associated with the second difference linear function
frothat maps 5-vectors into 3 vectors. The middle matrix D4 is associated with the
difference function that maps 4-vectors into 3-vectors. The right-hand matrix D5
is associated with the difference function that maps 5-vectors into 4-vectors.

Composition of affine functions. The composition of affine functions is an affine

function. Suppose f : Rp Rm is the affine function given by f (x) = Ax + b, and
g : Rn Rp is the affine function given by g(x) = Cx + d. The composition h is
given by

h(x) = f (g(x)) = A(Cx + d) + b = (AC)x + (Ad + b) = Ax + b,

where A = AC, b = Ad + b.

Chain rule of differentiation. Let f : Rp Rm and g : Rn Rp be dif-

ferentiable functions. The composition of f and g is defined as the function
h : Rn Rm with

h(x) = f (g(x)) = f (g1 (x), . . . , gp (x)).

The function h is differentiable and its partial derivatives follow from those of f
and g via the chain rule:
hi fi g1 fi gp
(z) = (g1 (z)) (z) + + (gp (z)) (z)
xj y1 xi yp xi

for i = 1, . . . , m and j = 1, . . . , n. This relation can be expressed concisely as a

matrix-matrix product: the derivative matrix of h at z is the product

Dh(z) = Df (g(z))Dg(z)

of the derivative matrix of f at g(z) and the derivative matrix of g at z. This

compact matrix formula generalizes the chain rule for scalar-valued functions of a
single variable, i.e., h0 (z) = f 0 (g(z))g 0 (z).
The first order Taylor approximation of h at z can therefore be written as

h(x) = h(z) + Dh(z)(x z)

= f (g(z)) + Df (g(z))Dg(z)(x z).

The same result can be interpreted as a composition of two affine functions, the
first order Taylor approximation of f at g(z),

f(y) = f (g(z)) + Df (g(z))(y g(z))

 10.3 Matrix power 151

and the first order Taylor approximation of g at z,

g(x) = g(z) + Dg(z)(x z).
The composition of these two affine functions is
f(g(x)) = f(g(z) + Dg(z)(x z))
= f (g(z)) + Df (g(z))(g(z) + Dg(z)(x z) g(z))
= f (g(z)) + Df (g(z))Dg(z)(x z)

which is equal to h(x).

When h and f are a scalar valued function (m = 1), the derivative matrices
Dh(z) and Df (g(z)) are the transposes of the gradients, and we write the chain
rule as
h(z) = Dg(z)T f (g(z)).
In particular, if g(x) = Ax + b is affine, then the gradient of h(x) = f (g(x)) =
f (Ax + b) is given by h(z) = AT f (Ax + b).

Linear dynamical system with state feedback. We consider a time-invariant lin-

ear dynamical system with n-vector state xt and m-vector input ut , with dynamics
xt+1 = Axt + But , t = 1, 2, . . . .
Here we think of the input ut as something we can manipulate, e.g., the control
surface deflections for an airplane or the amount of material we order or move in a
supply chain. In state feedback control the state xt is measured, and the input ut
is a linear function of the state, expressed as
ut = Kxt ,
where K is the m n state-feedback gain matrix. The term feedback refers to
the idea that the state is measured, and then (after multiplying by K) fed back
into the system, via the input. This leads to a loop, where the state affects the
input, and the input affects the (next) state. State feedback is very widely used
in many applications. (In 17.2.3 we will see methods for choosing or designing an
appropriate state feedback matrix.)
With state feedback, we have
xt+1 = Axt + But = Axt + B(Kxt ) = (A + BK)xt , t = 1, 2, . . . .
This recursion is called the closed-loop system. The matrix A + BK is called the
closed-loop dynamics matrix. (In this context, the recursion xt = Axt is called the
open-loop system. It gives the dynamics when ut = 0.)

10.3 Matrix power

Matrix powers. It makes sense to multiply a square matrix A by itself to form
AA. We refer to this matrix as A2 . Similarly, if k is a positive integer, then k
152 10 Matrix multiplication

2 3

1 4

Figure 10.1 Directed graph.

copies of A multiplied together is denoted Ak . If k and l are positive integers, and

A is square, then Ak Al = Ak+l and (Ak )l = Akl . By convention we take A0 = I,
which makes the formulas above hold for nonnegative integer values of k and l.
Matrix powers Ak with k a negative integer are discussed in 11.2. Non-integer
powers, such as A1/2 (the matrix squareroot), are pretty tricky they might not
make sense, or be ambiguous, unless certain conditions on A hold. This is an
advanced topic in linear algebra that we will not pursue in this book.

Paths in a directed graph. Suppose A is the n n adjacency matrix of a directed

graph with n vertices:

1 there is a edge from vertex j to vertex i
Aij =
0 otherwise

(see page 94). A path of length ` is a sequence of ` + 1 vertices, with an edge

from the first to the second vertex, an edge from the second to third vertex, and
so on. We say the path goes from the first vertex to the last one. An edge can be
considered a path of length one.
The elements of the matrix powers A` have a simple meaning in terms of paths
in the graph. First examine the expression for the i, j element of the square of A:
n
X
(A2 )ij = Aik Akj .
k=1

Each term in the sum is 0 or 1, and equal to one only if there is an edge from vertex
j to vertex k and an edge from vertex k to vertex i, i.e., a path of length exactly
two from vertex j to vertex i via vertex k. By summing over all k, we obtain the
total number of paths of length two from j to i. The adjacency matrix A for the
graph in figure 10.1, for example, and its square are given by

0 1 0 0 1 1 0 1 1 0
1 0 1 0 0 0 1 1 1 2
2

A = 0 0 1 1 1 ,
A = 1 0 1 2 1 .

1 0 0 0 0 0 1 0 0 1
0 0 0 1 0 1 0 0 0 0
10.3 Matrix power 153

We can verify there is exactly one path of length two from vertex 1 to itself, i.e.,
the path (1, 2, 1)), and one path of length two from vertex 3 to vertex 1, i.e., the
path (3, 2, 1). There are two paths of length two from vertex 4 to vertex 3: the
paths (4, 3, 3) and (4, 5, 3), so A34 = 2.
The property extends to higher powers of A. If ` is a positive integer, then
the i, j element of A` is the number of paths of length ` from vertex j to vertex i.
This can be proved by induction on `. We have already shown the result for ` = 2.
Assume that it is true that the elements of A` give the paths of length ` between
the different vertices. Consider the expression for the i, j element of A`+1 :
n
X
(A`+1 )ij = Aik (A` )kj .
k=1

The kth term in the sum is equal to the number of paths of length ` from j to k if
there is an edge from k to i, and is equal to zero otherwise. Therefore it is equal
to the number of paths of length ` + 1 from j to i that end with the edge (k, i),
i.e., of the form (j, . . . , k, i). By summing over all k we obtain the total number of
paths of length ` + 1 from vertex j to i. This can be verified in the example. The
third power of A is

1 1 1 1 2
2 0 2 3 1
3

A = 2 1 1 2 2 .

1 0 1 1 0
0 1 0 0 1

The (A3 )24 = 3 paths of length three from vertex 4 to vertex 2 are (4, 3, 3, 2),
(4, 5, 3, 2), (4, 5, 1, 2).

Linear dynamical systems. Consider a time-invariant linear dynamical system,

described by xt+1 = Axt . We have xt+2 = Axt+1 = A(Axt ) = A2 xt . Continuing
this argument, we have
xt+` = A` xt ,

for ` = 1, 2, . . .. In a linear dynamical system, we can interpret A` as the matrix

that propagates the state forward ` time steps.
For example, in a population dynamics model, A` is the matrix that maps the
current population distribution into the population distribution ` periods in the
future, taking into account births, deaths, and the births and deaths of children,
and so on. The total population ` periods in the future is given by 1T (A` xt ), which
we can write as (1T A` )xt . The row vector 1T A` has an interesting interpretation:
its ith entry is the contribution to the total population in ` periods due to each
person with current age i 1. It is plotted in figure 10.2 for the US data given
above.
Matrix powers also come up in the analysis of a time-invariant linear dynamical
system with an input. We have

xt+2 = Axt+1 + But+1 = A(Axt + But ) = A2 xt + ABut + But+1 .

154 10 Matrix multiplication

1.5

Factor 0.5

0 10 20 30 40 50 60 70 80 90 100
Age

Figure 10.2 Contribution factor per age in 2010 to the total population in
2020. The value for age i 1 is the ith component of the row vector 1T A10 .

Iterating this over ` periods we obtain

xt+` = A` xt + A`1 But + A`2 But+1 + + But+`1 . (10.2)

(The first term agrees with the formula for xt+` with no input.) The other terms
are readily interpreted. The term Aj But+`j is the contribution to the state xt+`
due to the input at time t + ` j.

10.4 QR factorization
Matrices with orthonormal columns. As an application of Gram matrices, we can
express the condition that the n-vectors a1 , . . . , ak are orthonormal in a simple
way using matrix notation:
AT A = I,
where A is the nk matrix with columns a1 , . . . , ak . There is no standard term for
a matrix whose columns are orthonormal: We refer to a matrix whose columns are
orthonormal as a matrix whose columns are orthonormal. But a square matrix
that satisfies AT A = I is called orthogonal ; its columns are an orthonormal basis.
Orthogonal matrices have many uses, and arise in many applications.
We have already encountered some orthogonal matrices, including identity ma-
trices, 2-D reflections and rotations (page 105), and permutation matrices (page 108).

Norm, inner product, and angle properties. Suppose the columns of the m n
matrix A are orthonormal, and x and y are any n-vectors. We let f : Rn Rm
10.4 QR factorization 155

be the function that maps z to Az. Then we have the following:

kAxk = kxk. That is, f is norm preserving.
(Ax)T (Ay) = xT y. f preserves the inner product between vectors.
6 (Ax, Ay) = 6 (x, y). f also preserves angles between vectors.
Note that in each of the three equations above, the vectors appearing in the left-
and right-hand sides have different dimensions, m on the left and n on the right.
We can verify these properties using simple matrix properties. We start with
the second statement, that multiplication by A preserves the inner product. We
have

(Ax)T (Ay) = (xT AT )(Ay)

= xT (AT A)y
= xT Iy
= xT y.

In the first line, we use the transpose-of-product rule; in the second, we re-associate
a product of 4 matrices (considering the row vector xT and column vector x as
matrices); in the third line we use AT A = I, and in the fourth line we use Iy = y.
From the second property we can derive the first one: By taking y = x we get
(Ax)T (Ax) = xT x; taking the squareroot of each side gives kAxk = kxk. The third
property, angle preservation, follows from the first two, since
(Ax)T (Ay)
   T 
x y
6 (Ax, Ay) = arccos = arccos = 6 (x, y).
kAxkkAyk kxkkyk

QR factorization. We can express the Gram-Schmidt factorization described in

5.4 in a compact form using matrices. Let A be an n k matrix with linearly
independent columns a1 , . . . , ak . By the independence-dimension inequality, A is
tall or square.
Let Q be the n k matrix with columns q1 , . . . , qk , the orthonormal vectors
produced by the Gram-Schmidt algorithm applied to the n-vectors a1 , . . . , ak .
Orthonormality of q1 , . . . , qk is expressed in matrix form as QT Q = I.
We express the equation relating ai and qi ,

ai = (q1T ai )q1 + + (qi1

T
ai )qi1 + kqi kqi ,

where qi is the vector obtained in the first step of the Gram-Schmidt algorithm, as

ai = R1i q1 + + Rii qi ,

where Rij = qiT aj for i < j and Rii = kqi k. Defining Rij = 0 for i > j, we can
express the equations above in compact matrix form as

A = QR.

This is called the QR factorization of A, since it expresses the matrix A as a

product of two matrices, Q and R. The n k matrix Q has orthonormal columns,
156 10 Matrix multiplication

and the k k matrix R is upper triangular, with positive diagonal elements. If A

is square of order n, with linearly independent columns, then Q is orthogonal and
the QR factorization expresses A as a product of two square matrices.
The Gram-Schmidt algorithm is not the only algorithm for QR factorization.
Several other QR factorization algorithms exist, that are more reliable in the pres-
ence of round-off errors. (These QR factorization methods also change the order
in which the columns of A are processed.)

Sparse QR factorization. There are algorithms for QR factorization that effi-

ciently handle the case when the matrix A is sparse. In this case the matrix Q is
stored in a special format that requires much less memory than if it were stored
as a generic n k matrix, i.e., nk numbers. The flop count for these sparse QR
factorizations is also much smaller than nk 2 .
 Chapter 11

Matrix inverses

In this chapter we introduce the concept of matrix inverse. We show how matrix
inverses can be used to solve linear equations, and how they can be computed using
the QR factorization.

11.1 Left and right inverses

Recall that for a number a, its (multiplicative) inverse is the number x for which
xa = 1, which we usually denote as x = 1/a or (less frequently) x = a1 . The
inverse x exists provided a is nonzero. For matrices the concept of inverse is more
complicated than for scalars; in the general case, we need to distinguish between
left and right inverses. We start with the left inverse.

Left inverse. A matrix X that satisfies

XA = I
is called a left inverse of A. The matrix A is said to be left-invertible if a left
inverse exists. Note that if A has size m n, a left inverse X will have size n m,
the same dimensions as AT .

Examples.
If A is a number (i.e., a 1 1 matrix), then a left inverse X is the same as
the inverse of the number. In this case, A is left-invertible whenever A is
nonzero, and it has only one left inverse.
Any nonzero n-vector a, considered as an n 1 matrix, is left invertible. For
any index i with ai 6= 0, the row n-vector x = (1/ai )eTi satisfies xa = 1.
The matrix
3 4
A= 4 6
1 1
158 11 Matrix inverses

has two different left inverses:

   
1 11 10 16 1 0 1 6
B= , C= .
9 7 8 11 2 0 1 4

This can be verified by checking that BA = CA = I. The example illustrates

that a left-invertible matrix can have more than one left inverse.

A matrix A with orthonormal columns satisfies AT A = I, so it is left-

invertible; its transpose AT is a left inverse.

Left-invertibility and column independence. If A has a left inverse C then the

columns of A are linearly independent. To see this, suppose that Ax = 0. Multi-
plying on the left by a left inverse C, we get

0 = C(Ax) = (CA)x = Ix = x,

which shows that the only linear combination of the columns of A that is 0 is the
one with all coefficients zero.
We will see below that the converse is also true; a matrix has a left inverse if
and only if its columns are linearly independent. So the generalization of a number
has an inverse if and only if it is nonzero is a matrix has a left inverse if and only
if its columns are linearly independent.

Dimensions of left inverses. Suppose the m n matrix A is wide, i.e., m < n.

By the independence-dimension inequality, its columns are linearly dependent, and
therefore it is not left invertible. Only square or tall matrices can be left invertible.

Solving linear equations with a left inverse. Suppose that Ax = b, where A is

an m n matrix and x is an n-vector. If C is a left inverse of A, we have

Cb = C(Ax) = (CA)x = Ix = x,

which means that x = Cb is a solution of the set of linear equations. The columns
of A are linearly independent (since it has a left inverse), so there is only one
solution of the linear equations Ax = b; in other words, x = Cb is the solution of
Ax = b.
Now suppose there is no x that satisfies the linear equations Ax = b, and
let C be a left inverse of A. Then x = Cb does not satisfy Ax = b, since no
vector satisfies this equation by assumption. This gives a way to check if the linear
equations Ax = b have a solution, and to find one when there is one, provided we
have a left inverse of A. We simply test whether A(Cb) = b. If this holds, then we
have found a solution of the linear equations; if it does not, then we can conclude
that there is no solution of Ax = b.
In summary, a left inverse can be used to determine whether or not a solution of
an over-determined set of linear equations exists, and when it does, find the unique
solution.
11.1 Left and right inverses 159

Right inverse. Now we turn to the closely related concept of right inverse. A
matrix X that satisfies
AX = I
is called a right inverse of A. The matrix A is right-invertible if a right inverse
exists. Any right inverse has the same dimensions as AT .

Left and right inverse of matrix transpose. If A has a right inverse B, then B T
is a left inverse of AT , since B T AT = (AB)T = I. If A has a left inverse C, then
C T is a right inverse of AT , since AT C T = (CA)T = I. This observation allows us
to map all the results for left invertibility given above to similar results for right
invertibility. Some examples are given below.
A matrix is right invertible if and only if its rows are linearly independent.
A tall matrix cannot have a right inverse. Only square or wide matrices can
be right invertible.

Solving linear equations with a right inverse. Consider the set of m linear equa-
tions in n variables Ax = b. Suppose A is right-invertible, with right inverse B.
This implies that A is square or wide, so the linear equations Ax = b are square or
under-determined.
Then for any m-vector b, the n-vector x = Bb satisfies the equation Ax = b.
To see this, we note that

Ax = A(Bb) = (AB)b = Ib = b.

We can conclude that if A is right-invertible, then the linear equations Ax = b can

be solved for any vector b. Indeed, x = Bb is a solution. (There can be other
solutions of Ax = b; the solution x = Bb is simply one of them.)
In summary, a right inverse can be used to find a solution of a square or under-
determined set of linear equations, for any vector b.

Examples. Consider the matrix appearing in the example above on page 157,

3 4
A= 4 6
1 1
and the two left inverses
   
1 11 10 16 1 0 1 6
B= , C= .
9 7 8 11 2 0 1 4

The over-determined linear equations Ax = (1, 2, 0) have the unique solu-

tion x = (1, 1), which can be obtained from either left inverse:

x = B(1, 2, 0) = C(1, 2, 0).

The over-determined linear equations Ax = (1, 1, 0) do not have a solution,

since x = C(1, 1, 0) = (1/2, 1/2) does not satisfy Ax = (1, 1, 0).
160 11 Matrix inverses

The under-determined linear equations AT y = (1, 2) has (different) solutions

B T (1, 2) = (1/3, 2/3, 38/9), C T (1, 2) = (0, 1/2, 1).

(Recall that B T and C T are both right inverses of AT .) We can find a solution
of AT y = b for any vector b.

Left and right inverse of matrix product. Suppose A and D are compatible for
the matrix product AD (i.e., the number of columns in A is equal to the number
of rows in D.) If A has a right inverse B and D has a right inverse E, then EB is
a right inverse of AD. This follows from

(AD)(EB) = A(DE)B = A(IB) = AB = I.

If A has a left inverse C and D has a left inverse F , then F C is a left inverse
of AD. This follows from

(F C)(AD) = F (CA)D = F D = I.

11.2 Inverse
If a matrix is left- and right-invertible, then the left and right inverses are unique
and equal. To see this, suppose that AX = I and Y A = I, i.e., X is any right
inverse and Y is any left inverse of A. Then we have

X = (Y A)X = Y (AX) = Y,

i.e., any left inverse of A is equal to any right inverse of A. This implies that the
left inverse is unique: If we have AX = I, then the argument above tells us that
X = Y , so we have X = X, i.e., there is only one right inverse of A. A similar
argument shows that Y (which is the same as X) is the only left inverse of A.
When a matrix A has both a left inverse Y and a right inverse X, we call the
matrix X = Y simply the inverse of A, and denote it as A1 . We say that A is
invertible or nonsingular. A square matrix that is not invertible is called singular.

Dimensions of invertible matrices. Invertible matrices must be square, since tall

matrices are not right invertible, while wide matrices are not left invertible. A
matrix A and its inverse (if it exists) satisfy

AA1 = A1 A = I.

If A has inverse A1 , then the inverse of A1 is A; in other words, we have

(A1 )1 = A. For this reason we say that A and A1 are inverses (of each other).
11.2 Inverse 161

Solving linear equations with the inverse. Consider the square system of n linear
equations with n variables, Ax = b. If A is invertible, then for any n-vector b,

x = A1 b (11.1)

is a solution of the equations. (This follows since A1 is a right inverse of A.)

Moreover, it is the only solution of Ax = b. (This follows since A1 is a left inverse
of A.) We summarize this very important result as
The square system of linear equations Ax = b, with A invertible, has
the unique solution x = A1 b, for any n-vector b.
One immediate conclusion we can draw from the formula (11.1) is that the
solution of a square set of linear equations is a linear function of the right-hand
side vector b.

Invertibility conditions. For square matrices, left-invertibility, right-invertibility,

and invertibility are equivalent: If a matrix is square and left-invertible, then it is
also right-invertible (and therefore invertible) and vice-versa.
To see this, suppose A is an n n matrix and left-invertible. This implies that
the n columns of A are linearly independent. Therefore they form a basis and so
any n-vector can be expressed as a linear combination of the columns of A. In
particular, each of the n unit vectors ei can be expressed as ei = Abi for some
n-vector bi . The matrix B = b1 b2 bn satisfies
   
AB = Ab1 Ab2 Abn = e1 e2 en = I.

So B is a right inverse of A.
We have just shown that for a square matrix A,

left invertibility = column independence = right invertibility.

(The symbol = means that the left-hand condition implies the right-hand condi-
tion.) Applying the same result to the transpose of A allows us to also conclude
that

right invertibility = row independence = left invertibility.

So all six of these conditions are equivalent; if any one of them holds, so do the
other five.
In summary, for a square matrix A, the following are equivalent.
A is invertible.
The columns of A are linearly independent.
The rows of A are linearly independent.
A has a left inverse.
A has a right inverse.
162 11 Matrix inverses

Examples.
The identity matrix I is invertible, with inverse I 1 = I, since II = I.
A diagonal matrix A is invertible if and only if its diagonal entries are nonzero.
The inverse of an n n diagonal matrix A with nonzero diagonal entries is

1/A11 0 0
0 1/A22 0
A1 = .. .. . ,

. .. ..
. .
0 0 1/Ann

since
A11 /A11 0 0
0 A22 /A22 0
AA1 = .. .. .. = I.

..
. . . .
0 0 Ann /Ann
In compact notation, we have

diag(A11 , . . . , Ann )1 = diag(A1 1

11 , . . . , Ann ).

Note that the inverse on the left-hand side of this equation is the matrix
inverse, while the inverses appearing on the right-hand side are scalar inverses.
As a non-obvious example, the matrix

1 2 3
A= 0 2 2
3 4 4

is invertible, with inverse

0 20 10
1
A1 = 6 5 2 .
30
6 10 2

This can be verified by checking that AA1 = I (or that A1 A = I, since

either of these implies the other).
2 2 matrices. A 2 2 matrix A is invertible if and only if A11 A22 6= A12 A21 ,
with inverse
 1  
A11 A12 1 A22 A12
A1 = = .
A21 A22 A11 A22 A12 A21 A21 A11
(There are similar formulas for the inverse of a matrix of any size, but they
grow very quickly in complexity and so are not very useful in most applica-
tions.)
Orthogonal matrix. If A is square with orthonormal columns, we have AT A =
I, so A is invertible with inverse A1 = AT .
11.2 Inverse 163

Inverse of matrix transpose. If A is invertible, its transpose AT is also invertible

and its inverse is (A1 )T :
(AT )1 = (A1 )T .
Since the order of the transpose and inverse operations does not matter, this matrix
is sometimes written as AT .

Inverse of matrix product. If A and B are invertible (hence, square) and of the
same size, then AB is invertible, and

(AB)1 = B 1 A1 . (11.2)

The inverse of a product is the product of the inverses, in reverse order.

Dual basis. Suppose that A is invertible with inverse B = A1 . Let a1 , . . . , an

be the columns of A, and bT1 , . . . , bTn denote the rows of B, i.e., the columns of
BT : T
b1
  ..
A = a1 an , B = . .
bTn
We know that a1 , . . . , an form a basis, since the columns of A are independent.
The vectors b1 , . . . , bn also form a basis, since the rows of B are independent. They
are called the dual basis of a1 , . . . , an . (The dual basis of b1 , . . . , bn is a1 , . . . , an ,
so they called dual bases.)
The matrix equation BA = I can be expressed as

T 1 i=j
bi aj =
0 i 6= j,

for i, j = 1, . . . , n. In words, bi is orthogonal to aj , for j 6= i, and has inner product

1 for j = i.
Now suppose that x is any n-vector. It can be expressed as a linear combination
of the basis vectors a1 , . . . , an :

x = 1 a1 + + n an .

Multiplying this equation on the left by bTi we get

bTi x = 1 (bTi a1 ) + + n (bTi an ) = i , i = 1, . . . , n.

So the dual basis gives us a simple way to find the coefficients in the expansion of
a vector in the a1 , . . . , an basis. We can summarize this as the identity

x = (bT1 x)a1 + + (bTn x)an ,

which holds for any n-vector x. This explicit formula shows how to express an arbi-
trary vector x as a linear combination of a basis a1 , . . . , an . To get the coefficients,
we take the inner product with the dual basis vectors.
164 11 Matrix inverses

As a simple numerical example, consider the basis

a1 = (1, 1), a2 = (1, 1).

The dual basis consists of the rows of [ a1 a2 ]1 , which are

bT1 = bT2 =
   
1/2 1/2 , 1/2 1/2 .

To express the vector x = (5, 1) as a linear combination of a1 and a2 , we have

x = (bT1 x)a1 + (bT2 x)a2 = (2)a1 + (3)a2 ,

which can be directly verified.

Negative matrix powers. We can now give a meaning to matrix powers with
negative integer exponents. Suppose A is a square invertible matrix and k is a
positive integer. Then by repeatedly applying property (11.2), we get

(Ak )1 = (A1 )k .

We denote this matrix as Ak . For example, if A is square and invertible, then

A2 = A1 A1 = (AA)1 . With A0 defined as A0 = I, the identity Ak+l = Ak Al
holds for all integers k and l.

Triangular matrix. A triangular matrix with nonzero diagonal elements is invert-

ible. We first discuss this for a lower triangular matrix. Let L be n n and lower
triangular with nonzero diagonal elements. We show that the columns are linearly
independent, i.e., Lx = 0 is only possible if x = 0. Expanding the matrix-vector
product, we can write Lx = 0 as

L11 x1 = 0
L21 x1 + L22 x2 = 0
L31 x1 + L32 x2 + L33 x3 = 0
..
.
Ln1 x1 + Ln2 x2 + + Ln,n1 xn1 + Lnn xn = 0.

Since L11 6= 0, the first equation implies x1 = 0. Using x1 = 0, the second equation
reduces to L22 x2 = 0. Since L22 6= 0, we conclude that x2 = 0. Using x1 = x2 = 0,
the third equation now reduces to L33 x3 = 0, and since L33 is assumed to be
nonzero, we have x3 = 0. Continuing this argument, we find that all entries of x
are zero, and this shows that the columns of L are linearly independent. It follows
that L is invertible.
A similar argument can be followed to show that an upper triangular matrix
with nonzero diagonal elements is invertible. One can also simply note that if R
is upper triangular, then L = RT is lower triangular with the same diagonal, and
use the formula (LT )1 = (L1 )T for the inverse of the transpose.
 11.3 Solving linear equations 165

Inverse via QR factorization. The QR factorization gives a simple expression

for the inverse of an invertible matrix. If A is square and invertible, its columns
are linearly independent, so it has a QR factorization A = QR. The matrix Q is
orthogonal and R is upper triangular with positive diagonal entries. Hence Q and
R are invertible, and the formula for the inverse product gives
A1 = (QR)1 = R1 Q1 = R1 QT . (11.3)
In the following section we give an algorithm for computing R1 , or more
directly, the product R1 QT . This gives us a method to compute the matrix
inverse.

11.3 Solving linear equations

Back substitution. We start with an algorithm for solving a set of linear equa-
tions, Rx = b, where the n n matrix R is upper triangular with nonzero diagonal
entries (hence, invertible). We write out the equations as
R11 x1 + R12 x2 + + R1,n1 xn1 + R1n xn = b1
..
.
Rn2,n2 xn2 + Rn2,n1 xn1 + Rn2,n xn = bn2
Rn1,n1 xn1 + Rn1,n xn = bn1
Rnn xn = bn .
From the last equation, we find that xn = bn /Rnn . Now that we know xn , we
substitute it into the second to last equation, which gives us
xn1 = (bn1 Rn1,n xn ))/Rn1,n1 .
We can continue this way to find xn2 , xn3 , . . . , x1 . This algorithm is known
as back substitution, since the variables are found one at a time, starting from xn ,
and we substitute the ones that are known into the remaining equations.

Algorithm 11.1 Back substitution

given an n n upper triangular matrix R with nonzero diagonal entries, and an
n-vector b.
For i = n, . . . , 1,
xi = (bi Ri,i+1 xi+1 Ri,n xn )/Rii .

The back substitution algorithm computes the solution of Rx = b, i.e., x = R1 b.

It cannot fail since the divisions in each step are by the diagonal entries of R, which
are assumed to be nonzero.
Lower triangular matrices with nonzero diagonal elements are also invertible;
we can solve equations with lower triangular invertible matrices using forward sub-
stitution, the obvious analog of the algorithm given above. In forward substitution,
we find x1 first, then x2 , and so on.
166 11 Matrix inverses

Complexity of back substitution. The first step requires 1 flop (division by Rnn ).
The next step requires one multiply, one subtraction, and one division, for a total
of 3 flops. The kth step requires k 1 multiplies, k 1 subtractions, and one
division, for a total of 2k 1 flops. The total number of flops for back substitution
is then
1 + 3 + 5 + + (2n 1) = n2

flops.
(The formula above was allegedly discovered by the mathematician Carl Friedrich
Gauss when he was a child. Here is his argument, for the case when n is even: Lump
the first entry in the sum together with the last entry, the second entry together
with the second-to-last entry, and so on. Each of these pairs add up to 2n; since
there are n/2 such pairs, the total is (n/2)(2n) = n2 . A similar argument works
when n is odd.)

Solving linear equations using the QR factorization. The formula (11.3) for the
inverse of a matrix in terms of its QR factorization suggests a method for solving
a square system of linear equations Ax = b with A invertible. The solution

x = A1 b = R1 QT b (11.4)

can be found by first computing the matrix vector product y = QT b, and then
solving the triangular equation Rx = y by back substitution.

Algorithm 11.2 Solving linear equations via QR factorization

given an n n invertible matrix A and an n-vector b.

1. QR factorization. Compute the QR factorization A = QR.

2. Compute QT b.
3. Back substitution. Solve the triangular equation Rx = QT b using back substi-
tution.

The first step requires 2n3 flops (see 5.4), the second step requires 2n2 flops,
and the third step requires n2 flops. The total number of flops is then

2n3 + 3n2 2n3 ,

so the order is n3 , cubic in the number of variables, which is the same as the number
of equations.
In the complexity analysis above, we found that the first step, the QR factor-
ization, dominates the other two; that is, the cost of the other two is negligible
in comparison to the cost of the first step. This has some interesting practical
implications, which we discuss below.
11.3 Solving linear equations 167

Factor-solve methods. Algorithm 11.2 is similar to many methods for solving a

set of linear equations and is sometimes referred to as a factor-solve scheme. A
factor-solve scheme consists of two steps. In the first (factor) step the coefficient
matrix is factored as a product of matrices with special properties. In the second
(solve) step one or more linear equations that involve the factors in the factorization
are solved. (In algorithm 11.2, the solve step consists of steps 2 and 3.) The
complexity of the solve step is smaller than the complexity of the factor step, and
in many cases, it is negligible by comparison. This is the case in algorithm 11.2,
where the factor step has order n3 and the solve step has order n2 .

Factor-solve methods with multiple right-hand sides. Now suppose that we must
solve several sets of linear equations,

Ax1 = b1 , ..., Axk = bk ,

all with the same coefficient matrix A, but different right-hand sides. Solving the
k problems independently, by applying algorithm 11.2 k times, costs 2kn3 flops. A
more efficient method exploits the fact that A is the same matrix in each problem,
so we can re-use the matrix factorization in step 1 and only need to repeat steps 2
and 3 to compute xk = R1 QT bk for l = 1, . . . , k. (This is sometimes referred to
as factorization caching, since we save or cache the factorization after carrying it
out, for later use.) The cost of this method is 2n3 + 3kn2 flops, or approximately
2n3 flops if k  n. The (surprising) conclusion is that we can solve multiple sets of
linear equations, with the same coefficient matrix A, at essentially the same cost
as solvng one set of linear equations.

Computing the matrix inverse. We can now describe a method to compute the
inverse B = A1 of an (invertible) n n matrix A. We first compute the QR
factorization of A, so A1 = R1 QT . We can write this as RB = QT , which,
written out by columns is

Rbi = qi , i = 1, . . . , n,

where bi is the ith column of B and qi is the ith column of QT . We can solve these
equations using back substitution, to get the columns of the inverse B.

Algorithm 11.3 Computing the inverse via QR factorization

given an n n invertible matrix A.

1. QR factorization. Compute the QR factorization A = QR.

2. For i = 1, . . . , n,
Solve the triangular equation Rbi = qi using back substitution.

The complexity of this method is 2n3 flops (for the QR factorization) and n3 for
n back substitutions, each of which costs n2 flops. So we can compute the matrix
inverse in around 3n3 flops.
This gives an alternative method for solving the square set of linear equations
Ax = b: We first compute the inverse matrix A1 , and then the matrix-vector
168 11 Matrix inverses

product x = (A1 )b. This method has a higher flop count than directly solving
the equations using algorithm 11.2 (3n3 versus 2n3 ), so algorithm 11.2 is the usual
method of choice. While the matrix inverse appears in many formulas (such as the
solution of a set of linear equations), it is computed far less often.

Sparse linear equations. Systems of linear equations with sparse coefficient ma-
trix arise in many applications. By exploiting the sparsity of the coefficient matrix,
these linear equations can be solved far more efficiently than by using the generic
algorithm 11.2. One method is to use the same basic algorithm 11.2, replacing
the QR factorization with a variant that handles sparse matrices (see page 156).
The memory usage and computational complexity of these methods depends in a
complicated way on the sparsity pattern of the coefficient matrix. In order, the
memory usage is typically a modest multiple of nnz(A) + n, the number of scalars
required to specify the problem data A and b, which is typically much smaller than
n2 + n, the number of scalars required to store A and b if they are not sparse.
The flop count for solving sparse linear equations is also typically closer in order
to nnz(A) than n3 , the order when the matrix A is not sparse.

11.4 Examples
Polynomial interpolation. The 4-vector c gives the coefficients of a cubic polyno-
mial,
p(x) = c1 + c2 x + c3 x2 + c4 x3
(see pages 126 and 101). We seek the coefficients that satisfy

p(1.1) = b1 , p(0.4) = b2 , p(0.2) = b3 , p(0.8) = b4 .

We can express this as the system of 4 equations in 4 variables Ac = b, where

(1.1)2 (1.1)3

1 1.1
1 0.4 (0.4)2 (0.4)3
A= .
(0.2)2 (0.2)3

1 0.2
1 0.8 (0.8)2 (0.8)3

The unique solution is c = A1 b, where

0.5784 1.9841 2.1368 0.7310
1
0.3470 0.1984 1.4957 0.9503
A = 0.1388 1.8651

1.6239 0.1023
0.0370 0.3492 0.7521 0.0643

(to 4 decimal places). This is illustrated in figure 11.1, which shows the two cu-
bic polynomials that interpolate the two sets of points shown as filled circles and
squares, respectively.
The columns of A1 are interesting: They give the coefficients of a polynomial
that evaluates to 0 at three of the points, and 1 at the other point. For example,
11.4 Examples 169

p(x)

x
1.5 1 0.5 0 0.5 1

Figure 11.1 Cubic interpolants through two sets of points, shown as circles
and squares.

the first column of A1 , which is A1 e1 , gives the coefficients of the polynomial

that has value 1 at 1.1, and value 0 at 0.4, 0.2, and 0.8. The four polynomials
with coefficients given by the columns of A1 are called the Lagrange polynomials
associated with the points 1.1, 0.4, 0.2, 0.8. These are plotted in figure 11.2.
The rows of A1 are also interesting: The ith row shows how the values b1 ,
. . . , b4 , the polynomial values at the points 1.1, 0.4, 0.2, 0.8, map into the ith
coefficient of the polynomial, ci . For example, we see that the coefficient c4 is not
very sensitive to the value of b1 (since (A1 )41 is small). We can also see that for
each increase of one in b4 , the coefficient c2 increases by around 0.95.

Balancing chemical reactions. (See page 126 for background.) We consider the
problem of balancing the chemical reaction

a1 Cr2 O2
7 + a2 Fe
2+
+ a3 H+ b1 Cr3+ + b2 Fe3+ + b3 H2 O,

where the superscript gives the charge of each reactant and product. There are 4
atoms (Cr, O, Fe, H) and charge to balance. The reactant and product matrices
are (using the order just listed)

2 0 0 1 0 0

7 0 0


0 0 1

R=
0 1 0 ,
P =
0 1 0 .

0 0 1 0 0 2
2 2 1 3 3 0
170 11 Matrix inverses

p(x) p(x)

1 1

0 0

x x
1 0 1 1 0 1

p(x) p(x)

1 1

0 0

x x
1 0 1 1 0 1

Figure 11.2 Lagrange polynomials associated with the points 1.1, 0.4,
0.2, 0.8.
11.4 Examples 171

Imposing the condition that a1 = 1 we obtain a square set of 6 linear equations,

2 0 0 1 0 0 a1 0
7
0 0 0 0 1 a2 0

0 1 0 0 1 0 a3 0

=
.
0
0 1 0 0 2 b1 0

2 2 1 3 3 0 b2 0
1 0 0 0 0 0 b3 1
Solving these equations we obtain
a1 = 1, a2 = 6, a3 = 14, b1 = 2, b2 = 6, b3 = 7.
(Setting a1 = 1 could have yielded fractional values for the other coefficients, but
in this case, it did not.) The balanced reaction is
Cr2 O2
2 + 6Fe
2+
+ 14H+ 2Cr3+ + 6Fe3+ + 7H2 O.

Heat diffusion. We consider a diffusion system as described on page 127. Some of

the nodes have fixed potential, i.e., ei is given; for the other nodes, the associated
external source si is zero. This would model a thermal system in which some
nodes are in contact with the outside world or a heat source, which maintains
their temperatures (via external heat flows) at constant values; the other nodes are
internal, and have no heat sources. This gives us a set of n additional equations:
ei = efix
i , i P, si = 0, i 6 P,
where P is the set of indices of nodes with fixed potential. We can write these n
equations in matrix-vector form as
Bs + Ce = d,
where B and C are the n n diagonal matrices, and d is the n-vector given by
   fix
0 iP 1 iP ei iP
Bii = Cii = di =
1 i 6 P, 0 i 6 P, 0 i 6 P.
We assemble the flow conservation, edge flow, and the boundary conditions into
one set of m + 2n equations in m + 2n variables (f, s, e):

A I 0 f 0
R 0 AT s = 0 .
0 B C e d
(The matrix A is the incidence matrix of the graph, and R is the resistance matrix;
see page 127.) Assuming the coefficient matrix is invertible, we have
1
f A I 0 0
s = R 0 AT 0 .
e 0 B C d
This is illustrated with an example in figure 11.3. The graph is a 100 100 grid,
with 10000 nodes, and edges connecting each node to its horizontal and vertical
neighbors. The resistance on each edge is the same. The nodes at the top and
bottom are held at zero temperature, and the three sets of nodes with rectilinear
shapes are held at temperature one. All other nodes have zero source value.
172 11 Matrix inverses

0.8

0.6

0.4

0.2

Figure 11.3 Temperature distribution on a 100 100 grid of nodes. Nodes

in the top and bottom rows are held at zero temperature. The three sets of
nodes with rectilinear shapes are held at temperature one.

11.5 Pseudo-inverse
Linearly independent columns and Gram invertibility. We first show that an
m n matrix A has linearly independent columns if and only if its n n Gram
matrix AT A is invertible.
First suppose that the columns of A are linearly independent. Let x be an
n-vector which satisfies (AT A)x = 0. Multiplying on the left by xT we get
0 = xT 0 = xT (AT Ax) = xT AT Ax = kAxk2 ,
which implies that Ax = 0. Since the columns of A are linearly independent, we
conclude that x = 0. Since the only solution of (AT A)x = 0 is x = 0, we conclude
that AT A is invertible.
Now lets show the converse. Suppose the columns of A are linearly dependent,
which means there is a nonzero n-vector x which satisfies Ax = 0. Multiply on the
left by AT to get (AT A)x = 0. This shows that the Gram matrix AT A is singular.

Pseudo-inverse of square or tall matrix. We show here that if A has linearly

independent columns (and therefore, is square or tall) then it has a left inverse.
(We already have observed the converse, that a matrix with a left inverse has
linearly independent columns.) Assuming A has linearly independent columns, we
know that AT A is invertible. We now observe that the matrix (AT A)1 AT is a
left inverse of A:
(AT A)1 AT A = (AT A)1 (AT A) = I.

This particular left-inverse of A will come up in the sequel, and has a name,
the pseudo-inverse of A (also called the Moore-Penrose inverse). It is denoted A
11.5 Pseudo-inverse 173

(or A+ ):
A = (AT A)1 AT . (11.5)

When A is square, the pseudo-inverse A reduces to the ordinary inverse:
A = (AT A)1 AT = A1 AT AT = A1 I = A1 .
Note that this equation does not make sense (and certainly is not correct) when A
is not square.

Pseudo-inverse of a square or wide matrix. Transposing all the equations, we

can show that a (square or wide) matrix A has a right inverse if and only its rows
are linearly independent. Indeed, one right inverse is given by
AT (AAT )1 . (11.6)
(The matrix AAT is invertible if and only if the rows of A are linearly independent.)
The matrix in (11.6) is also referred to as the pseudo-inverse of A, and denoted
A . The only possible confusion in defining the pseudo-inverse using the two dif-
ferent formulas (11.5) and (11.6) occurs when the matrix A is square. In this case,
however, they both reduce to the ordinary inverse:
AT (AAT )1 = AT AT A1 = A1 .

Pseudo-inverse in other cases. The pseudo-inverse A is defined for any nonzero

matrix, including the case when A is tall but its columns are linearly dependent,
the case when A is wide but its rows are linearly dependent, and the case when A
is square but not invertible. In these cases, however, it is not a left inverse, right
inverse, or inverse, respectively. Exactly what A means in these cases is beyond
the scope of this book; we mention it here since the reader may encounter it.

Pseudo-inverse via QR factorization. The QR factorization gives a simple for-

mula for the pseudo-inverse. If A is left-invertible, its columns are linearly inde-
pendent and the QR factorization A = QR exists. We have
AT A = (QR)T (QR) = RT QT QR = RT R,
so
A = (AT A)1 AT = (RT R)1 (QR)T = R1 RT RT QT = R1 QT .
We can compute the pseudo-inverse using the QR factorization, followed by back
substitution on the columns of QT . (This is exactly the same as algorithm 11.3
when A is square and invertible.) The complexity of this method is 2n2 m flops (for
the QR factorization), and mn2 flops for the m back substitutions. So the total is
3mn2 flops.
Similarly, if A is right-invertible, the QR factorization AT = QR of its transpose
exists. We have AAT = (QR)T (QR) = RT QT QR = RT R and
A = AT (AAT )1 = QR(RT R)1 = QRR1 RT = QRT .
We can compute it using the method described above, using the formula
(AT ) = (A )T .
174 11 Matrix inverses

Solving over- and under-determined systems of linear equations. The pseudo-

inverse gives us a method for solving over-determined and under-determined sys-
tems of linear equations, provided the columns of the coefficient matrix are linearly
independent (in the over-determined case), or the rows are linearly independent
(in the under-determined case). If the columns of A are linearly independent, and
the over-determined equations Ax = b have a solution, then x = A b is it. If the
rows of A are linearly independent, the under-determined equations Ax = b have
a solution for any vector b, and x = A b is a solution.

Numerical example. We illustrate these ideas with a simple numerical example,

using the 3 2 matrix A used in earlier examples on pages 157 and 159,

3 4
A= 4 6 .
1 1

This matrix has linearly independent columns, and QR factorization with (to 4
digits)

0.5883 0.4576  
5.0990 7.2563
Q= 0.7845 0.5230 R= .
0 0.5883
0.1961 0.7191

It has pseudo-inverse (to 4 digits)

 
1.2222 1.1111 1.7778
A = QRT = .
0.7778 0.8889 1.2222

We can use the pseudo-inverse to check if the over-determined systems of equations

Ax = b, with b = (1, 2, 0), has a solution, and to find a solution if it does. We
compute x = A (1, 2, 0) = (1, 1) and check whether Ax = b holds. It does, so
we have found the unique solution of Ax = b.
 Part III

Least squares
 Chapter 12

Least squares

In this chapter we look at the powerful idea of finding approximate solutions of

over-determined systems of linear equations by minimizing the sum of the squares
of the errors in the equations. The method, and some extensions we describe in
later chapters, are widely used in many application areas. It was discovered inde-
pendently by the mathematicians Carl Friedrich Gauss and Adrien-Marie Legendre
around the beginning of the 19th century.

12.1 Least squares problem

Suppose that the m n matrix A is tall, so the system of linear equations Ax = b,
where b is an m-vector, is over-determined, i.e., there are more equations (m)
than variables to choose (n). These equations have a solution only if b is a linear
combination of the columns of A.
For most choices of b, however, there is no n-vector x for which Ax = b. As a
compromise, we seek an x for which r = Ax b, which we call the residual (for the
equations Ax = b), is as small as possible. This suggests that we should choose x
so as to minimize the norm of the residual kAx bk. If we find an x for which the
residual vector is small, we have Ax b, i.e., x almost satisfies the linear equations
Ax = b.
Minimizing the norm of the residual and its square are the same, so we can just
as well minimize
kAx bk2 = krk2 = r12 + + rm
2
,
the sum of squares of the residuals. The problem of finding an n-vector x that
minimizes kAx bk2 , over all possible choices of x, is called the least squares
problem. It is denoted with the notation

minimize kAx bk2 , (12.1)

where we should specify that the variable is x (meaning that we should choose x).
The matrix A and the vector b are called the data for the problem (12.1), which
178 12 Least squares

means that they are given to us when we are asked to choose x. The quantity to
be minimized, kAx bk2 , is called the objective function (or just objective) of the
least squares problem (12.1).
The problem (12.1) is sometimes called linear least squares to emphasize that
the residual r (whose norm squared we are to minimize) is an affine function of x,
and to distinguish it from the nonlinear least squares problem, in which we allow
the residual r to be an arbitrary function of x. We will study the nonlinear least
squares problem in chapter 18.
Any vector x that satisfies kAxbk2 kAxbk2 for all x is a solution of the least
squares problem (12.1). Such a vector is called a least squares approximate solution
of Ax = b. It is very important to understand that a least squares approximate
solution x of Ax = b need not satisfy the equations Ax = b; it simply makes the
norm of the residual as small as it can be. Some authors use the confusing phrase
x solves Ax = b in the least squares sense, but we emphasize that a least squares
approximate solution x does not, in general, solve the equation Ax = b.
If kAx bk (which we call the optimal residual norm) is small, then we can say
that x approximately solves Ax = b. On the other hand, if there is an n-vector x
that satisfies Ax = b, then it is a solution of the least squares problem, since its
associated residual norm is zero.
Another name for the least squares problem (12.1), typically used in data fitting
applications (the topic of the next chapter), is regression. We say that x, a solution
of the least squares problem, is the result of regressing the vector b onto the columns
of A.

Column interpretation. If the columns of A are the m-vectors a1 , . . . , an , then

the least squares problem (12.1) is the problem of finding a linear combination of
the columns that is closest to the m-vector b; the vector x gives the coefficients:

kAx bk2 = k(x1 a1 + + xn an ) bk2 .

If x is a solution of the least squares problem, then the vector

Ax = x1 a1 + + xn an

is closest to the vector b, among all linear combinations of the vectors a1 , . . . , an .

Row interpretation. Suppose the rows of A are the n-vectors aT1 , . . . , aTm , so the
residual components are given by

ri = aTi x bi , i = 1, . . . , m.

The least squares objective is then

kAx bk2 = (aT1 x b1 )2 + + (aTm x bm )2 ,

the sum of the squares of the residuals in m scalar linear equations. Minimizing
this sum of squares of the residuals is a reasonable compromise if our goal is to
choose x so that all of them are small.
 12.2 Solution 179

Example. We consider the least squares problem with data

2 0 1
A = 1 1 , b = 0 .
0 2 1

The over-determined set of three equations in two variables Ax = b,

2x1 = 1, x1 + x2 = 0, 2x2 = 1,

has no solution. (From the first equation we have x1 = 1/2, and from the last
equation we have x2 = 1/2; but then the second equation does not hold.) The
corresponding least squares problem is

minimize (2x1 1)2 + (x1 + x2 )2 + (2x2 + 1)2 .

This least squares problem can be solved using the methods described in the next
section (or simple calculus). Its unique solution is x = (1/3, 1/3). The least
squares approximate solution x does not satisfy the equations Ax = b; the corre-
sponding residuals are
r = (1/3, 2/3, 1/3),
with sum of squares value kAx bk2 = 2/3. Let us compare this to another choice
of x, x = (1/2, 1/2), which corresponds to (exactly) solving the first and last of
the three equations in Ax = b. It gives residual

r = Ax b = (0, 1, 0),

with sum of squares value kAx bk2 = 1.

The column interpretation tells us that

2 0 2/3
(1/3) 1 + (1/3) 1 = 2/3
0 2 2/3

is the linear combination of the columns of A that is closest to b.

Figure 12.1 shows the values of the least squares objective kAx bk2 versus
x = (x1 , x2 ), with the least squares solution x shown as the dark point, with
objective value kAx bk2 = 2/3. The curves show the points x that have value
kAx bk2 + 1, kAx bk2 + 2, and so on.

12.2 Solution
In this section we derive several expressions for the solution of the least squares
problem (12.1), under one assumption on the data matrix A:

The columns of A are linearly independent. (12.2)

180 12 Least squares

0.5
f (x) + 2
f (x) + 1

x2
0

0.5

1
1 0.5 0 0.5 1
x1

Figure 12.1 Level curves of the function kAx bk2 = (2x1 1)2 + (x1 +
x2 )2 + (2x2 + 1)2 .

Solution via calculus. In this section we find the solution of the least squares
problem using some basic results from calculus, reviewed in C.2. (We will also
give an independent verification of the result, that does not rely on calculus, below.)
We know that any minimizer x of the function f (x) = kAx bk2 must satisfy
f
(x) = 0, i = 1, . . . , n,
xi
which we can express as the vector equation
f (x) = 0,
where f (x) is the gradient of f evaluated at x. This gradient can be expressed
in matrix form as
f (x) = 2AT (Ax b). (12.3)

This formula can be derived from the chain rule given on page 150, and the
gradient of the sum of squares function, given in C.1. For completeness, we will
derive the formula (12.3) from scratch here. Writing the least squares objective
out as a sum, we get
2
Xm Xn
f (x) = kAx bk2 = Aij xj bi .
i=1 j=1

To find f (x)k we take the partial derivative of f with respect to xk . Differenti-

ating the sum term by term, we get
f
f (x)k = (x)
xk
12.2 Solution 181

m
X n
X
= 2 Aij xj bi (Aik )
i=1 j=1
m
X
= 2(AT )ki (Ax b)i
i=1
2AT (Ax b) k .


=

This is our formula (12.3), written out in terms of its components.

Now we continue the derivation of the solution of the least squares problem.
Any minimizer x of kAx bk2 must satisfy

f (x) = 2AT (Ax b) = 0,

which can be written as

AT Ax = AT b. (12.4)

These equations are called the normal equations. The coefficient matrix AT A is
the Gram matrix associated with the columns of A; its entries are inner products
of columns of A.
Our assumption (12.2) that the columns of A are linearly independent implies
that the Gram matrix AT A is invertible (11.5, page 172). This implies that

x = (AT A)1 AT b (12.5)

is the only solution of the normal equations (12.4). So this must be the unique
solution of the least squares problem (12.1).
We have already encountered the matrix (AT A)1 AT that appears in (12.5): It
is the pseudo-inverse of the matrix A, given in (11.5). So we can write the solution
of the least squares problem in the simple form

x = A b. (12.6)

We observed in 11.5 that A is a left inverse of A, which means that x = A b

solves Ax = b if this set of over-determined equations has a solution. But now
we see that x = A b is the least squares approximate solution, i.e., it minimizes
kAx bk2 . (And if there is a solution of Ax = b, then x = A b is it.)
The equation (12.6) looks very much like the formula for solution of the linear
equations Ax = b, when A is square and invertible, i.e., x = A1 b. It is very
important to understand the difference between the formula (12.6) for the least
squares approximate solution, and the formula for the solution of a square set
of linear equations, x = A1 b. In the case of linear equations and the inverse,
x = A1 b actually satisfies Ax = b. In the case of the least squares approximate
solution, x = A b generally does not satisfy Ax = b.
The formula (12.6) shows us that the solution x of the least squares problem
is a linear function of b. This generalizes the fact that the solution of a square
invertible set of linear equations is a linear function of its right-hand side.
182 12 Least squares

Direct verification of least squares solution. In this section we directly show

that x = (AT A)1 AT b is the solution of the least squares problem (12.1), without
relying on calculus. We will show that for any x 6= x, we have

kAx bk2 < kAx bk2 ,

establishing that x is the unique vector that minimizes kAx bk2 .

We start by writing

kAx bk2 = k(Ax Ax) + (Ax b)k2

= kAx Axk2 + kAx bk2 + 2(Ax Ax)T (Ax b), (12.7)

where we use the identity

ku + vk2 = (u + v)T (u + v) = kuk2 + kvk2 + 2uT v.

The third term in (12.7) is zero:

(Ax Ax)T (Ax b) = (x x)T AT (Ax b)

= (x x)T (AT Ax AT b)
= (x x)T 0
= 0,

where we use (AT A)x = AT b in the third line. With this simplification, (12.7)
reduces to
kAx bk2 = kA(x x)k2 + kAx bk2 .

The first term on the right-hand side is nonnegative and therefore

kAx bk2 kAx bk2 .

This shows that x minimizes kAxbk2 ; we now show that it is the unique minimizer.
Suppose equality holds above, that is, kAx bk2 = kAx bk2 . Then we have
kA(x x)k2 = 0, which implies A(x x) = 0. Since A has linearly independent
columns, we conclude that x x = 0, i.e., x = x. So the only x with kAx bk2 =
kAx bk2 is x = x; for all x 6= x, we have kAx bk2 > kAx bk2 .

Row form. The formula for the least squares approximate solution can be ex-
pressed in a useful form in terms of the rows aTi of the matrix A.

m
!1 m
!
X X
x = (AT A)1 AT b = ai aTi bi ai . (12.8)
i=1 i=1

In this formula we express the n n Gram matrix AT A as a sum of m outer

products, and the n-vector AT b as a sum of m n-vectors.
 12.3 Solving least squares problems 183

12.3 Solving least squares problems

We can use the QR factorization to compute the least squares approximate so-
lution (12.5). Let A = QR be the QR factorization of A (which exists by our
assumption (12.2) that its columns are linearly independent). We have already
seen that the pseudo-inverse A can be expressed as A = R1 QT , so we have

x = R1 QT b. (12.9)

To compute x we first multiply b by QT ; then we compute R1 (QT b) using back

substitution. This is summarized in the following algorithm, which computes the
least squares approximate solution x, given A and b.

Algorithm 12.1 Least squares via QR factorization

given an m n matrix A with linearly independent columns and an m-vector b.

1. QR factorization. Compute the QR factorization A = QR.

2. Compute QT b.

3. Back subsitution. Solve the triangular equation Rx = QT b.

Comparison to solving a square system of linear equations. Recall that the

solution of the square invertible system of linear equations Ax = b is x = A1 b.
We can express x using the QR factorization of A as x = R1 QT b (see (11.4)).
This equation is formally identical to (12.9). The only difference is that in (12.9), A
and Q need not be square, and R1 QT b is the least squares approximate solution,
which is not (in general) a solution of Ax = b.
Indeed, algorithm 12.1 is formally the same as algorithm 11.2, the QR fac-
torization method for solving linear equations. (The only difference is that in
algorithm 12.1, A and Q can be tall.)
When A is square, solving the linear equations Ax = b and the least squares
problem of minimizing kAx bk2 are the same, and algorithm 11.2 and algo-
rithm 12.1 are the same. So we can think of algorithm 12.1 as a generalization of
algorithm 11.2, which solves the equation Ax = b when A is square, and computes
the least squares approximate solution when A is tall.

Backslash notation. The very close relation between solving a square set of linear
equations, and finding the least squares approximate solution of an over-determined
set of equations, has inspired a common notation for both that is used in some
software packages for manipulating matrices: A\b is taken to mean the solution
A1 b when A is square (and invertible), and the least squares approximate solution
A b when A is tall (with linearly independent columns). This backslash notation
is not standard mathematical notation, so we will not use it in this book.
184 12 Least squares

Complexity. The complexity of the first step of algorithm 12.1 is 2mn2 flops. The
second step involves a matrix-vector multiplication, which takes 2mn flops. The
third step requires n2 flops. The total number of flops is

2mn2 + 2mn + n2 2mn2 ,

neglecting the second and third terms, which are smaller than the first by factors
of n and 2m, respectively. The order of the algorithm is mn2 . The complexity is
linear in the row dimension of A and quadratic in the number of variables.
Algorithm 12.1 is another example of a factor-solve algorithm. Suppose we need
to solve several least squares problems

minimize kAxk bk k2 ,

k = 1, . . . , l, which have the same data matrix A but different vectors bk . If we

re-use the QR factorization of A the cost is 2mn2 + l(2mn + n2 ) flops. When l
is small compared to n this is roughly 2mn2 flops, the same cost as a single least
squares problem.

Sparse least squares. Least squares problems with sparse A arise in several appli-
cations and can be solved more efficiently, for example by using a QR factorization
tailored for sparse matrices (see page 156) in the generic algorithm 12.1.
Another simple approach for exploiting sparsity of A is to solve the normal
equations AT Ax = AT b by solving a larger (but sparse) system of equations,

0 AT
    
x 0
= . (12.10)
A I y b

This is a square set of m+n linear equations. Its coefficient matrix is sparse when A
is sparse. If (x, y) satisfies these equations, it is easy to see that x satisfies (12.10);
conversely, if x satisfies the normal equations, (x, y) satisfies (12.10) with y =
Ax b. Any method for solving a sparse system of linear equations can be used to
solve (12.10).

12.4 Examples
Advertising purchases. We have m demographic groups or audiences that we
want to advertise to, with a target number of impressions or views for each group,
which we give as a vector v des . (The entries are positive.) To reach these audiences,
we purchase advertising in n different channels (say, different web publishers, radio,
print, . . . ), in amounts that we give as an n-vector s. (The entries of s are non-
negative, which we ignore.) The m n matrix R gives the number of impressions
in each demographic group per dollar spending in the channels: Rij is the num-
ber of impressions in group i per dollar spent on advertising in channel j. (These
entries are estimated, and are all nonnegative.) The jth column of R gives the
effectiveness or reach (in impressions per dollar) for channel i. The ith row of R
12.4 Examples 185

Channel 1
Channel 2
2 Channel 3
Impressions

1 2 3 4 5 6 7 8 9 10
Group

Figure 12.2 Number of impressions in ten demographic groups, per dollar

spent on advertising in three channels. The units are 1000 views per dollar.

shows which media demographic i is exposed to. The total number of impressions
in each demographic group is the m-vector v, which is given by v = Rs. The goal
is to find s so that v = Rs v des . We can do this using least squares, by choosing
s to minimize kRs v des k2 . (We are not guaranteed that the resulting channel
spend vector will be nonnegative.) This least squares formulation does not take
into account the total cost of the advertising; we will see in chapter 16 how this
can be done.
We consider a simple numerical example, with n = 3 channels and m = 10
demographic groups, and matrix

0.97 1.86 0.41

1.23 2.18 0.53


0.80 1.24 0.62


1.29 0.98 0.51

1.10 1.23 0.69
R= ,

0.67 0.34 0.54


0.87 0.26 0.62


1.10 0.16 0.48

1.92 0.22 0.71
1.29 0.12 0.62

with units of dollars per 1000 views. The entries of this matrix range over an 18:1
range, so the 3 channels are quite different in terms of their audience reach; see
figure 12.2.
We take v des = 103 1, i.e., our goal is to reach one million customers in each of
186 12 Least squares

1,000

Impressions
500

0
1 2 3 4 5 6 7 8 9 10
Group

Figure 12.3 Views vector that best approximates the target of one million
impressions in each group.

the 10 demographic groups. Least squares gives the advertising budget allocation

s = (62, 100, 1443),

which achieves a views vector with RMS error 132, or 13.2% of the target values.
The views vector is shown in figure 12.3.

Illumination. A set of n lamps illuminates an area that we divide into m regions

or pixels. We let li denote the lighting level in region i, so the m-vector l gives the
illumination levels across all regions. We let pi denote the power at which lamp
i operates, so the n-vector p gives the set of lamp powers. (The lamp powers are
nonnegative and also must not exceed a maximum allowed power, but we ignore
these issues here.)
The illumination levels are a linear function of the lamp powers, so we have
l = Ap for some m n matrix A. The jth column of A gives the illumination
pattern for lamp j, i.e., the illumination when lamp j has power 1 and all other
lamps are off. We will assume that A has linearly independent columns (and
therefore is tall). The ith row of A gives the sensitivity of pixel i to the n lamp
powers.
The goal is to find lamp powers that result in a desired illumination pattern
ldes , such as ldes = 1, which is uniform illumination with value across the area.
In other words, we seek p so that Ap ldes . We can use least squares to find p that
minimizes the sum square deviation from the desired illumination, kAp ldes k2 .
This gives the lamp power levels

p = A ldes = (AT A)1 AT ldes .

12.4 Examples 187

25m
1 (4.0m)
2 (3.5m)
1 (4.0m)
2 (3.5m) 1.4
3 (6.0m) 3 (6.0m)
1.2
6 (6.0m) 6 (6.0m)
4 (4.0m) 4 (4.0m)
7 (5.5m) 7 (5.5m) 1
5 (4.0m) 5 (4.0m)

0.8
9 (5.0m) 10 (4.5m) 9 (5.0m) 10 (4.5m)
8 (5.0m) 8 (5.0m)
0.6
0
0 25m 0 25m

Figure 12.4 A square area divided in a 25 25 grid. The circles show the
positions of 10 lamps, the number in parentheses next to each circle is the
height of the lamp. The left-hand plot shows the illumination pattern with
lamps set to power one. The plot on the right shows the illumination pattern
for the lamp powers that minimize the sum square deviation with a desired
uniform illumination of one.

(We are not guaranteed that these powers are nonnegative, or less than the maxi-
mum allowed power level.)
An example is shown in figure 12.4. The area is a 25 25 grid with m = 625
pixels, each (say) 1m square. The lamps are at various heights ranging from 3m
to 6m, and at the positions shown in the figure. The illumination decays with an
inverse square law, so Aij is proportional to d2ij , where dij is the (3-D) distance
between the center of the pixel and the lamp position. The matrix A is scaled so
that when all lamps have power one, the average illumination level is one. The
desired illumination pattern is 1, i.e., uniform with value 1.
With p = 1, the resulting illumination pattern is shown is shown on the left of
figure 12.4. The RMS illumination error is 0.24. We can see that the corners are
quite a bit darker than the center, and there are pronounced bright spots directly
beneath each lamp. Using least squares we find the lamp powers

p = (1.46, 0.79, 2.97, 0.74, 0.08, 0.21, 0.21, 2.05, 0.91, 1.47).

The resulting illumination pattern has an RMS error of 0.14, about half of the RMS
error with all lamp powers set to one. The illumination pattern is shown on the
right of figure 12.4; we can see that the illumination is more uniform than when all
lamps have power 1. Most illumination values are near the target level 1, with the
corners a bit darker and the illumination a bit brighter directly below each lamp,
but less so than when all lamps have power one. This is clear from figure 12.5,
which shows the histogram of patch illumination values for all lamp powers one,
and for lamp powers p.
188 12 Least squares

120

Number of pixels 100

80

60

40

20

0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Intensity
120

100
Number of pixels

80

60

40

20

0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Intensity

Figure 12.5 Histograms of pixel illumination values using p = 1 and p.

 Chapter 13

Least squares data fitting

In this chapter we introduce one of the most important applications of least squares
methods, to the problem of data fitting. The goal is to find a mathematical model,
or an approximate model, of some relation, given some observed data.

13.1 Least squares data fitting

Least squares is widely used as a method to construct a mathematical model from
some data, experiments, or observations. Suppose we have an n-vector x, and a
scalar y, and we believe that they are related, perhaps approximately, by some
function f : Rn R:
y f (x).
The vector x might represent a set of n feature values, and is called the feature
vector or the vector of independent variables, depending on the context. The scalar
y represents some outcome (also called response variable) that we are interested
in. Or x could represent the previous n values of a time series, and y represents
the next value.

Data. We dont know f , but we might have some idea about its general form.
But we do have some data, given by

x1 , . . . , x N , y1 , . . . , yN ,

where the n-vector xi is the feature vector and the scalar yi is the associated value
of the outcome for data sample i. These data are also called observations, examples,
or measurements, depending on the context.

Model. We will form a model of the relationship between x and y, given by

y f(x),
190 13 Least squares data fitting

where f : Rn R. We write y = f(x), where y is the (scalar) prediction (of the

outcome y), given the independent variable (vector) x.
The model has the form

f(x) = 1 f1 (x) + + p fp (x),

where fi : Rn R are basis functions or feature mappings that we choose, and

i are the model parameters that we choose. This form of model is called linear in
the parameters, since for each x, f(x) is a linear function of the model parameter
p-vector . The basis functions are usually chosen based on our idea of what f
looks like. (We will see examples of this below.) Once the basis functions have
been chosen, there is the question of how to choose the model parameters, given
our set of data.

Prediction error. Our goal is to choose the model f so that it is consistent with
the data, i.e., we have yi f(xi ), for i = 1, . . . , N . (There is another goal in
choosing f, that we will discuss in 13.2.) For data sample i, our model predicts
the value yi = f(xi ), so the prediction error or residual for this data point is

ri = yi yi .

Let r, y, and y denote the N -vectors with entries ri , yi , and yi , respectively.

(In the notation above, the symbols y and y refer to generic scalar values; but now
we use them to denote N -vectors of those values, for the observed data.) With this
notation we can express the (vector of) residuals as r = y y. A natural measure
of how well the model predicts the observed data, or how consistent it is with the
observed data, is the RMS prediction error rms(r). The ratio rms(r)/ rms(y)
gives a relative prediction error. For example, if the relative prediction error is 0.1,
we might say that the model predicts the outcomes, or fits the data, within 10%.

Least squares model fitting. A very common method for choosing the model pa-
rameters is to minimize the RMS prediction error, which is the same as minimizing
the sum of squares of the prediction errors, krk2 . We now show that this is a least
squares problem. Expressing f(xi ) in terms of the model parameters, we have

yi = Ai1 1 + + Aip p , i = 1, . . . , N,

where we define the N p matrix A as

Aij = fj (xi ), i = 1, . . . , N, j = 1, . . . , p,

and the p-vector as = (1 , . . . , p ). The jth column of A is the jth basis

function, evaluated at each of the data points x1 , . . . , xN . Its ith row gives the
values of the p basis functions on the ith data point xi . In matrix-vector notation
we have
y = A.
This simple equation shows how our choice of model parameters maps into the
vector of predicted values of the outcomes in the N different experiments.
13.1 Least squares data fitting 191

The sum of squares of the residuals is then

krk2 = kA yk2 .

Choosing to minimize this is evidently a least squares problem, of the same form
as (12.1). Provided the columns of A are linearly independent, we can solve this
least squares problem to find , the model parameter values that minimize the
norm of the prediction error on our data set, as

= (AT A)1 AT y. (13.1)

We say that the model parameter values are obtained by least squares fitting on
the data set.
We can interpret each term in kA yk2 . The term y = A is the N -vector
of measurements or outcomes that is predicted by our model, with the parameter
vector . The term y is the N -vector of actual observed or measured outcomes.
The difference A y is the N -vector of prediction errors. Finally, kA yk2 is
the sum of squares of the prediction errors. This is minimized by the least squares
fit = .
The number kA yk2 is called the minimum sum square error (for the given
model basis and data set). The number kA yk2 /N is called the minimum
mean square error (MMSE) (of our model, on the data set). Its squareroot is
the minimum RMS fitting error. The model performance on the data set can be
visualized by plotting yi versus yi on a scatter plot, with a dashed line showing
y = y for reference.

Least squares fit with a constant. We start with the simplest possible fit: We
take p = 1, with f1 (x) = 1 for all x. In this case the model f is a constant function,
with f(x) = 1 for all x. Least squares fitting in this case is the same as choosing
the best constant value 1 to approximate the data y1 , . . . , yN .
In this simple case, A is the N 1 matrix 1, which always has linearly inde-
pendent columns (since it has one column, which is nonzero). The formula (13.1)
is then
1 = (AT A)1 AT y = N 1 1T y = avg(y),
where we use 1T 1 = N . So the best constant fit to the data is simply its mean,

f(x) = avg(y).

The RMS fit to the data (i.e., the RMS value of the optimal residual) is

rms(avg(y)1 y) = std(y),

the standard deviation of the data. This gives a nice interpretation of the average
value and the standard deviation of the outcomes.
It is common to compare the RMS fitting error for a more sophisticated model
with the standard deviation of the outcomes, which is the optimal RMS fitting
error for a constant model.
192 13 Least squares data fitting

1 1

0.8 0.8

yi = avg(y)
0.6 0.6
f(x)

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x yi

Figure 13.1 The constant fit f(x) = avg(y) to N = 20 data points and a
scatter plot of y versus y.

Independent column assumption. To use least squares fitting we assume that

the columns of the matrix A are linearly independent. We can give an interesting
interpretation of what it means when this assumption fails. If the columns of A
are linearly dependent, it means that one of the columns can be expressed as a
linear combination of the others. Suppose, for example, that the last column can
be expressed as a linear combination of the first p 1 columns. Using Aij = fj (xi ),
this means

fp (xi ) = 1 f1 (xi ) + + p1 fp1 (xi ), i = 1, . . . , N.

This says that the value of pth basis function can be expressed as a linear combina-
tion of the values of the first p 1 basis functions on the given data set. Evidently,
then, the pth basis function is redundant (on the given data set).

13.1.1 Fitting univariate functions

Suppose that n = 1, so both x and y are scalars. The relationship y f (x)

says that y is approximately a (univariate) function f of x. We can plot the data
(xi , yi ) as points in the (x, y) plane, and we can plot the model f as a curve in the
(x, y)-plane. This allows us to visualize the fit of our model to the data.

Straight-line fit. We take basis functions f1 (x) = 1 and f2 (x) = x. Our model
has the form
f(x) = 1 + 2 x,

which is a straight line when plotted. (This is perhaps why f is sometimes called
a linear model, even though it is in general an affine, and not linear, function of
x.) Figure 13.2 shows an example. The matrix A is given by
13.1 Least squares data fitting 193

f(x)

Figure 13.2 Least squares fit of a straight line to 50 points (xi , yi ) in a plane.

1 x1
1 x2  
A= .. .. = 1 x ,

. .
1 xN
where x is the N -vector of values (x1 , . . . , xN ). Provided that there are at least
two different values appearing in x1 , . . . , xN , this matrix has linearly independent
columns. The parameters in the optimal straight-line fit to the data are given by
 
1
= (AT A)1 AT y.
2
This expression is simple enough for us to work it out explicitly, although there is
no computational advantage to doing so. The Gram matrix is
1T x
 
N
AT A = .
1T x xT x
The 2-vector AT y is
1T y
 
T
A y= ,
xT y
so we have (using the formula for the inverse of a 2 2 matrix)
 T
x x 1T x
   T 
1 1 1 y
= .
2 N xT x (1T x)2 1T x N xT y
Multiplying the scalar term by N 2 , and dividing the matrix and vector terms by
N , we can express this as
rms(x)2 avg(x)
    
1 1 avg(y)
= .
2 rms(x)2 avg(x)2 avg(x) 1 xT y/N
194 13 Least squares data fitting

The optimal slope 2 of the straight line fit can be expressed more simply in terms
of the correlation coefficient between the data vectors x and y, and their standard
deviations. We have
N xT y (1T x)(1T y)
2 =
N xT x (1T x)2
(x avg(x)1)T (y avg(y)1)
=
kx avg(x)1k2
std(y)
= .
std(x)
In the last step we used the definitions
(x avg(x)1)T (y avg(y)1) kx avg(x)1k
= , std(x) =
N std(x) std(y) N
from chapter 3. From the first of the two normal equations, N 1 + (1T x)2 = 1T y,
we also obtain a simple expression for 1 :
1 = avg(y) 2 avg(x).
Putting these results together, we can write the least squares fit as
std(y)
f(u) = avg(y) + (u avg(x)).
std(x)
This can be expressed in the more symmetric form
f(u) avg(y) u avg(x)
= ,
std(y) std(x)
which has a nice interpretation. The left-hand side is the difference between the
predicted response value and the mean response value, divided by its standard de-
viation. The right-hand side is the correlation coefficient times the same quantity,
computed for the dependent variable.
The least squares straight-line fit is used in many application areas.

Asset and in finance. In finance, the straight-line fit is used to compare the
returns of an individual asset, yi , to the returns of the whole market, xi . (The
return of the whole market is typically taken to be a sum of the individual asset
returns, weighted by their capitalizations.) The straight-line model for predicting
the asset return y from the market return x is typically written in the form
y = (rrf + ) + (x avg(x)),
where rrf is the risk-free interest rate over the period. (Comparing this to our
straight-line model, we find that 2 = , and 1 = rrf + avg(x).) The
parameter tells us how much of the return of the particular asset is explained,
or predicted, by the market return, and the parameter tells us what the average
return is, over and above the risk-free interest rate. This model is so common that
the terms Alpha and Beta are widely used in finance. (Though not always with
exactly the same meaning, since there a few variations on how the parameters are
defined.)
13.1 Least squares data fitting 195

90

(million barrels per day)

Petroleum consumption
80

70

60

1980 1985 1990 1995 2000 2005 2010

Year

Figure 13.3 World petroleum consumption between 1980 and 2013 (dots)
and least squares straight-line fit (data from [Link]).

Time series trend. Suppose the data represents a series of samples of a quantity
y at time (epoch) xi = i. The straight-line fit to time series data,

yi = 1 + 2 i, i = 1, . . . , N,

is called the trend line. Its slope, which is 2 , is interpreted as the trend in the
quantity over time. Subtracting the trend line from the original time series we
get the de-trended time series, y y. The de-trended time series shows how the
time series compares with its straight-line fit: When it is positive, it means the
time series is above its straight-line fit, and when it is negative, it is below the
straight-line fit.
An example is shown in figures 13.3 and 13.4. Figure 13.3 shows world petroleum
consumption versus year, along with the straight-line fit. Figure 13.4 shows the
de-trended world petroleum consumption.

Estimation of trend and seasonal component. In the previous example, we used

least squares to approximate a time series y = (y1 , . . . , yN ) of length N by a sum
of two components: y y = y const + y lin where

1
2
y const = 1 1, y lin = 2 . .

.
.
N

In many applications, the de-trended time series has a clear periodic component,
i.e., a component that repeats itself periodically. As an example, figure 13.5 shows
an estimate of the road traffic (total number of miles traveled in vehicles) in the
196 13 Least squares data fitting

(million barrels per day)

Petroleum consumption
4

1980 1985 1990 1995 2000 2005 2010

Year

Figure 13.4 De-trended world petroleum consumption.

US, for each month between January 2000 and December 2014. The most striking
aspect of the time series is the pattern that is repeated every year, with a peak in
the summer and a minimum in the winter. In addition there is a slowly increasing
long term trend. The bottom figure shows the least squares fit of a sum of three
components

y y = y const + y lin + y seas ,

where y const and y lin are defined as before, and the third component is periodic
with period P = 12. This periodic, or seasonal, component can be expressed as

3:2+P
3:2+P
y seas = .. ,

.
3:2+P
13.1 Least squares data fitting 197

105

2.6
Miles (millions)

2.4

2.2

2
2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015
Month
5
10

2.6
Miles (millions)

2.4

2.2

2
2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

Month

Figure 13.5 Top. Vehicle miles traveled in the US, per month, in the period
January 2000 December 2014 (U.S. Department of Transportation, Bu-
reau of Transportation Statistics, [Link]). Bottom. Least
squares fit of a sum of three time series: a constant, a linear trend, and a
seasonal component with a 12-month period.
198 13 Least squares data fitting

which consists of the pattern (3 , . . . , 2+p ), repeated N times. The least squares
fit is computed by minimizing kA yk2 where is a (P + 2)-vector and

1 1 1 0 0

1 2 0 1 0
.. .. .. .. . . ..

.


. . . . .


1 P 0 0 1


1 P +1 1 0 0

1 P +2 0 1 0
.. .. .. .. . . ..


A=
. . . . . . .

1 2P 0 0 1
.. .. .. .. ..



. . . . .


1 N P +1 1 0 0


1 N P +2 0 1 0

.. .. .. .. .. ..
. . . . . .
1 N 0 0 1

In this example, N = 15P = 180. The residual or prediction error in this case is
called the de-trended, seasonally adjusted series.

Polynomial fit. A simple extension beyond the straight-line fit is a polynomial fit,
with
fi (x) = xi1 , i = 1, . . . , p,

so f is a polynomial of degree at most p 1,

f(x) = 1 + 2 x + + p xp1 .

In this case the matrix A has the form

x1p1

1 x1
1 x2 x2p1
A= . ,

.. ..
.. . .
1 xN xp1
N

i.e., it is a Vandermonde matrix (see (6.6)). Its columns are linearly independent
provided the numbers x1 , . . . , xN include at least p different values. Figure 13.6
shows an example of the least squares fit of polynomials of degree 2, 6, 10, and 15
to a set of 100 data points. Since any polynomial of degree less than r is also a
polynomial of degree less than s, for r s, it follows that the RMS fit attained
by a polynomial with a larger degree is smaller (or at least, no larger) than that
obtained by a fit with a smaller degree polynomial. This suggests that we should
use the largest degree polynomial that we can, since this results in the smallest
residual and the best RMS fit. But we will see in 13.2 that this is not true, and
explore rational methods for choosing a model from among several candidates.
13.1 Least squares data fitting 199

degree 2 degree 6
f(x) f(x)

x x

degree 10 degree 15
f(x) f(x)

x x

Figure 13.6 Least squares polynomial fits of degree 2, 6, 10, and 15 to 100
points.
200 13 Least squares data fitting

(x + 1)+ (x + 1)+

3 3

2 2

1 1

0 0
x x
3 1 1 3 3 1 1 3

Figure 13.7 The functions (x + 1)+ = max{x + 1, 0} and (x 1)+ = max{x

1, 0}.

f(x)

x
2 1 0 1 2

Figure 13.8 Piecewise-linear fit to 100 points.

Piecewise-linear fit. A piecewise-linear function, with knot points or kink points

a1 < a2 < < ak , is a continuous function that is affine in between the knot
points. (Such functions should be called piecewise-affine.) We can describe any
piecewise-linear function with k knot points using the p = k + 2 basis functions

f1 (x) = 1, f2 (x) = x, fi (x) = (x ai )+ , i = 3, . . . , k,

where (u)+ = max{u, 0}. These basis functions are shown in figure 13.7 for k = 2
knot points at a1 = 1, a2 = 1. An example of a piecewise-linear fit with these
knot points is shown in figure 13.8.
 13.1 Least squares data fitting 201

13.1.2 Regression
Recall that the regression model has the form
y = xT + v,
where is the weight vector and v is the offset. We can put this model in our
general data fitting form using the basis functions f1 (x) = 1, and
fi (x) = xi1 , i = 2, . . . , n + 1,
so p = n + 1. The regression model can then be expressed as
y = xT 2:n+1 + 1 ,
and we see that = 2:n+1 and v = 1 .
The N (n + 1) matrix A in our general data fitting form is given by
A = 1 XT ,
 

where X is the feature matrix with columns x1 , . . . , xN . So the regression model

is a special case of our general linear in the parameters model.

General fitting model as regression. The regression model is a special case of

our general data fitting model. Conversely, we can think of our linear in the
parameters model as regression, with a different set of feature vectors of dimension
p 1. Assuming the first basis element f1 is the constant function with value one,
we consider the new or generated feature vectors x given by

f2 (x)
x = ..
, i = 1, . . . , N.

.
fp (x)
A regression model for the outcome y and the new generated or mapped features
x has the form
y = xT + v,
where has dimension p 1, and v is a number. Comparing this to our linear in
the parameters model
y = 1 f1 (x) + + p fp (x),
we see that they are the same, with v = 1 , = 2:p . So we can think of the general
linear in the parameters model as nothing more than simple regression, but applied
to the transformed, mapped, or generated features f1 (x), . . . , fp (x). (This idea is
discussed more in 13.3.)

House price regression. In 2.3 we described a simple regression model for the
selling price of a house based on two attributes, area and number of bedrooms.
The values of the parameters and the offset v given in (2.8) were computed by
least squares fitting on a set of data consisting of 774 house sales in Sacramento
over a 5 day period. The RMS fitting error for the model is 74.8 (in thousands of
dollars). For comparison, the standard deviation of the prices in the data set is
112.8. So this very basic regression model predicts the prices substantially better
than a constant model (i.e., the mean price of the houses in the data set).
202 13 Least squares data fitting

Auto-regressive time series model. Suppose that z1 , z2 , . . . is a time series. An

auto-regressive model (also called AR model ) for the time series has the form
zt+1 = 1 zt + + M ztM , t = M + 1, M + 2, . . .
where M is the memory of the model. Here zt+1 is the prediction of zt+1 made at
time t (when zt , . . . , ztM are known). This prediction is an affine function of the
previous M values of the time series. With good choice of model parameters, the
AR model can be used to predict the next value in a time zeries, given the current
and previous M values. This has many practical uses.
We can use least squares (or regression) to choose the AR model parameters,
based on the observed data z1 , . . . , zT , by minimizing the sum of squares of the
prediction errors zt+1 zt+1 over t = M, . . . , T , i.e.,
(zM +1 zM +1 )2 + + (zT zT )2 .
(We must start the predictions at t = M + 1, since each prediction involves the
previous M time series values, and we dont know z0 , z1 , . . . .)
The AR model can be put into the general regression model form by taking
yi = zM +i , xi = (zM +i1 , . . . , zi ) i = 1, . . . , T M.
We have N = T M examples, and n = M regressors.
As an example, consider the time series of hourly temperature at Los Angeles
International Airport, May 131 2016, with length 31 24 = 744. The simple
constant prediction zt+1 = 61.76 F (the average temperature) has RMS prediction
error 3.05 F (the standard deviation). The very simple predictor zt+1 = zt , i.e.,
guessing that the temperature next hour is the same as the current temperature, has
RMS error 1.16 F. The predictor zt+1 = zt23 , i.e., guessing that the temperature
next hour is what is was yesterday at the same time, has RMS error 1.73 F.
We fit an AR model with memory M = 8 using least squares, with N =
31 24 8 = 736 samples. The RMS error of this predictor is 0.98 F, smaller than
the RMS errors for the simple predictors described above. Figure 13.9 shows the
temperature and the predictions for the first five days.

13.1.3 Log transform of dependent variable

When the dependent variable y is positive and varies over a large range, it is
common to replace it with its logarithm w = log y, and then use least squares to
develop a model for w, w = g(x). We then form our estimate of y using y = eg(x) .
When we fit a model w = g(x) to the logarithm w = log y, the fitting error for w
can be interpreted in terms of the percentage or relative error between y and y,
defined as
= max{y/y, y/y} 1.
So = 0.1 means either y = 1.1y (i.e., we over-estimate by 10%) or y = (1/1.1)y
(i.e., we under-estimate by 10%). The connection between the relative error be-
tween y and y, and the residual r in predicting w, is
= e|r| 1.
 13.2 Validation 203

70

68

66
Temperature ( F)

64

62

60

58

56

20 40 60 80 100 120
k

Figure 13.9 Hourly temperature at Los Angeles International Airport be-

tween 12:53AM on May 1, 2016, and 11:53PM on May 5, 2016, shown as
circles. The solid line is the prediction of an autoregressive model with eight
coefficients.

For example, a residual with |r| = 0.05 corresponds (approximately) to a relative

error in our prediction y of y of 5%. (Here we use the approximation e|r| 1 |r|
for r smaller than, say, 0.15.) So if our RMS error in predicting w = log y across
our examples is 0.05, our predictions of y are typically within 5%.
As an example suppose we wish to model house sale prices over an area and
period that includes sale prices varying from under $100k to over $10M. Fitting the
prices directly means that we care equally about absolute errors in price predictions,
so, for example, a prediction error of $20k should bother us equally for a house that
sells for $70k and one that sells for $6.5M. But (at least for some purposes) in the
first case we have made a very poor estimate, and the second case, a remarkably
good one. When we fit the logarithm of the house sale price, we are seeking low
percentage prediction errors, not necesarily low absolute errors.
Whether or not to use a logarithmic transform on the dependent variable (when
it is positive) is a judgment call that depends on whether we seek small absolute
prediction errors or small relative or percentage prediction errors.

13.2 Validation
Generalization ability. In this section we address a key point in model fitting:
The goal of model fitting is typically not to just achieve a good fit on the given
204 13 Least squares data fitting

data set, but rather to achieve a good fit on new data that we have not yet seen.
This leads us to a basic question: How well can we expect a model to predict y for
future or other unknown values of x? Without some assumptions about the future
data, there is no good way to answer this question.
One very common assumption is that the data are described by a formal prob-
ability model. With this assumption, techniques from probability and statistics
can be used to predict how well a model will work on new, unseen data. This
approach has been very successful in many applications, and we hope that you will
learn about these methods in another course. In this book, however, we will take
a simple intuitive approach to this issue.
The ability of a model to predict the outcomes for new unseen data values is
called its generalization ability. If it predicts outcomes well, it is said to have good
generalization ability; in the opposite case, it is said to have poor generalization
ability. So our question is: How can we assess the generalization ability of a model?

Validation on a test set. A simple but effective method for assessing the gener-
alization ability of a model is called out-of-sample validation. We divide the data
we have into two sets: a training set and a test set (also called a validation set).
This is often done randomly, with 80% of the data put into the training set and
20% put into the test set. Another common choice for the split ratio is 90%10%.
A common way to describe this is to say that 20% of the data were reserved for
validation.
To fit our model, we use only the data in the training set. The model that we
come up with is based only on the data in the training set; the data in the test set
has never been seen by the model. Then we judge the model by its RMS fit on the
test set. Since the model was developed without any knowledge of the test set data,
the test data is effectively data that are new and unseen, and the performance of
our model on this data gives us at least an idea of how our model will perform on
new, unseen data. If the RMS prediction error on the test set is large, then we can
conclude that our model has poor generalization ability. Assuming that the test
data is typical of future data, the RMS prediction error on the test set is what
we might guess our RMS prediction error will be on new data.
If the RMS prediction error of the model on the training set is similar to the
RMS prediction error on the test set, we have increased confidence that our model
has reasonable generalization ability. (A more sophisticated validation method
called cross-validation, described below, can be used to gain even more confidence.)
For example, if our model achieves an RMS prediction error of 10% (compared
to rms(y)) on the training set and 11% on the test set, we can guess that it will have
a similar RMS prediction error on other unseen data. But there is no guarantee
of this, without further assumptions about the data. The basic assumption we are
making here is that the future data will look like the test data, or that the test
data were typical. Ideas from statistics can make this idea more precise, but we
will leave this idea informal and intuitive.

Over-fitting. When the RMS prediction error on the training set is much smaller
than the RMS prediction error on the test set, we say that the model is over-fit.
13.2 Validation 205

It tells us that, for the purposes of making predictions on new, unseen data, the
model is much less valuable than its performance on the training data suggests.
Good models, which perform well on new data, do not suffer from over-fit.
Roughly speaking, an over-fit model trusts the data it has seen (i.e., the training
set) too much; it is too sensitive to the changes in the data that will likely be seen
in the future data. One method for avoiding over-fit is to keep the model simple;
another technique, called regularization, is discussed in chapter 15.

Choosing among different models. We can use least squares fitting to fit multiple
models to the same data. For example, in univariate fitting, we can fit a constant,
an affine function, a quadratic, or a higher order polynomial. Which is the best
model among these? Assuming that the goal is to make good predictions on new,
unseen data, we should choose the model with the smallest RMS prediction error on
the test set. Since the RMS prediction error on the test set is only a guess about
what we might expect for performance on new, unseen data, we can soften this
advice to we should choose a model that has test set RMS error that is near the
minimum over the candidates.
We observed earlier that when we add basis functions to a model, our fitting
error on the training data can only decrease (or stay the same). But this is not
true for the test error. The test error need not decrease when we add more basis
functions. Indeed, when we have too many basis functions, we can expect over-fit,
i.e., larger error on the test set.
If we have a sequence of basis functions f1 , f2 , . . . , we can consider models
based on using just f1 (which is typically the constant function 1), then f1 and
f2 , and so on. As we increase p, the number of basis functions, our training error
will go down (or stay the same). But the test error typically decreases, and then
starts to increase when p is too large, and the resulting models suffer from over-
fit. The intuition for this typical behavior is that for p too small, our model is
too simple to fit the data well, and so cannot make good predictions; when p is
too large, our model is too complex and suffers from over-fit, and so makes poor
predictions. Somewhere in the middle, where the model achieves near minimum
test set performance, is a good choice (or several good choices) of p.
To illustrate these ideas, we consider the example shown in figure 13.6. Using a
training set of 100 points, we find least squares fits of polynomials of degrees 0, 1,
. . . , 20. (The polynomial fits of degrees 2, 6, 10, and 15 are shown in the figure.)
We now obtain a new set of data for validation, also with 100 points. These test
data are plotted along with the polynomial fits obtained from the training data
in figure 13.10. This is a real check of our models, since these data points were
not used to develop the models. Figure 13.11 shows the RMS training and test
errors for polynomial fits of different degrees. We can see that the RMS training
error decreases with every increase in degree. The RMS test error decreases until
degree 6 and starts to increase for degrees larger than 6. This plot suggests that a
polynomial fit of degree 6 is a reasonable choice.
With a 6th degree polynomial, the relative RMS test error for both training
and test sets is around 0.3. It is a good sign, in terms of generalization ability,
that the training and test errors are similar. While there are no guarantees, we can
guess that the 6th degree polynomial model will have a relative RMS error around
206 13 Least squares data fitting

degree 2 degree 6
f(x) f(x)

x x

degree 10 degree 15
f(x) f(x)

x x

Figure 13.10 The polynomial fits of figure 13.6 evaluated on a test set of 100
points.
13.2 Validation 207

1
Train
Test

Relative RMS error

0.8

0.6

0.4

0.2
0 5 10 15 20
Degree

Figure 13.11 RMS error versus polynomial degree for the fitting example
in figures 13.6 and 13.10. Circles indicate RMS errors on the training set.
Squares show RMS errors on the test set.

0.3 on new, unseen data, provided the new, unseen data is sufficiently similar to
the test set data.

Cross-validation. Cross-validation is an extension of out-of-sample validation that

can be used to get even more confidence in the generalization ability of a model.
We divide the original data set into 10 sets, called folds. We then fit the model
using folds 1, 2, . . . , 9 as training data, and fold 10 as test data. (So far, this is
the same as out-of-sample validation.) We then fit the model using folds 1, 2, . . . ,
8, 10 as training data and fold 9 as the test data. We continue, fitting a model
for each choice of one of the folds as the test set. (This is also called leave-one-out
validation.) We end up with 10 (presumably different) models, and 10 assessments
of these models using the fold that was not used to fit the model. (We have
described 10-fold cross-validation here; 5-fold validation is also commonly used.)
If the test fit performance of these 10 models is similar, we can expect the
same, or at least similar, performance on new unseen data. In cross-validation we
can also check for stability of the model coefficients. This means that the model
coefficients found in the different folds are similar to each other. Stability of the
model coefficients further enhances our confidence in the model.

House price regression model. As an example, we apply cross-validation to assess

the generalization ability of the simple regression model of the house sales data
discussed in 2.3 and on page 201. The simple regression model described there,
based on house area and number of bedrooms, has an RMS fitting error of 74.8
thousand dollars. Cross-validation will help us answer the question of how the
model might do on different, unseen houses.
208 13 Least squares data fitting

Model parameters RMS error

Fold v 1 2 Train Test
1 60.65 143.36 18.00 74.00 78.44
2 54.00 151.11 20.30 75.11 73.89
3 49.06 157.75 21.10 76.22 69.93
4 47.96 142.65 14.35 71.16 88.35
5 60.24 150.13 21.11 77.28 64.20

Table 13.1 Five-fold validation for the simple regression model of the house
sales data set.

Fold v RMS error (train) RMS error (test)

1 230.11 110.93 119.91
2 230.25 113.49 109.96
3 228.04 114.47 105.79
4 225.23 110.35 122.27
5 230.23 114.51 105.59

Table 13.2 Five-fold validation for the constant model of the house sales
data set.

We randomly partition the data set of 774 sales records into five folds, four of
size 155 and one of size 154. Then we fit five regression models, each of the form

y = v + 1 x1 + 2 x2

to the data set after removing one of the folds. Table 13.1 summarizes the results.
The model parameters for the 5 different regression models are not exactly the
same, but quite similar. The train and test RMS errors are reasonably similar,
which suggests that our model does not suffer from over-fit. Scanning the RMS
error on the test sets, we can expect that our prediction error on new houses will be
around 70-80 (thousand dollars) RMS. We can also see that the model parameters
change a bit, but not drastically, in each of the folds. This gives us more confidence
that, for example, 2 being negative is not a fluke of the data.
For comparison, table 13.2 shows the RMS errors for the constant model y = v,
where v is the mean price of the training set. The results suggest that the constant
model can predict house prices with a prediction error around 105-120 (thousand
dollars).
Figure 13.12 shows the scatter plots of actual and regression model predicted
prices for each of the five training and test sets. The results for training and
test sets are reasonably similar in each case, which gives us confidence that the
regression model will have similar performance on new, unseen houses.
13.2 Validation 209

Fold 1 Fold 2
800 800

600 600

400 400

200 200

0 0
0 200 400 600 800 0 200 400 600 800

Fold 3 Fold 4
800 800

600 600

400 400

200 200

0 0
0 200 400 600 800 0 200 400 600 800

Fold 5
800

600

400

200

0
0 200 400 600 800

Figure 13.12 Scatter plots of actual and predicted prices for the five simple
regression models of table 13.1. The horizontal axis is the actual selling
price the vertical axis is the predicted price, both in thousand dollars. Blue
circles are samples in the training set, red circles samples in the test set.
210 13 Least squares data fitting

Validating time series predictions. When the original data are unordered, for
example, patient records or customer purchase histories, the division of the data
into a training and test set is typically done randomly. This same method can
be used to validate a time series prediction model, such as an AR model, but it
does not give the best emulation of how the model will ultimately be used. In
practice, the model will be trained on past data and then used to make predictions
on future data. When the training data in a time series prediction model are
randomly chosen, the model is being built with some knowledge of the future, a
phenomemon called look-ahead or peak-ahead. Look-ahead can make a model look
better than it is really is at making predictions.
To avoid look-ahead, the training set for a time series prediction model is typ-
ically taken to be the data examples up to some point in time, and the test data
are chosen as points that are past that time (and sometimes, at least M samples
past that time, taking into account the memory of the predictor). In this way we
can say that the model is being tested by making predictions on data it has never
seen. As an example, we might train an AR model for some daily quantity using
data from the years 2006 through 2008, and then test the resulting AR model on
the data from year 2009.
As an example, we return to the AR model of hourly temperatures at Los
Angeles International Airport described on page 202 and shown in figure 13.9. We
divide the one month of data into a training set (May 124) and a test set (May
2531). The coefficients in an AR model are computed using the 24 24 8 = 568
samples in the training set. The RMS error on the training set is 1.03 F. The RMS
prediction error on the test set is 0.98 F, which is similar to the RMS prediction
error on the training set, giving us confidence that the AR model is not over-fit.
(The fact that the test RMS error is very slightly smaller than the training RMS
error has no significance.) Figure 13.13 shows the prediction on the first five days
of the test set. The predictions look very similar to those shown in figure 13.9.

Limitations of out-of-sample and cross-validation. Here we mention a few limi-

tations of out-of-sample and cross-validation. First, the basic assumption that the
test data and future data are similar can (and does) fail in some applications. For
example, a model that predicts consumer demand, trained and validated on last
years data, can make much poorer predictions next year, simply because consumer
tastes shift. In finance, patterns of asset returns periodically shift, so models that
predict well on test data from this year need not predict well next year.
Another limitation arises when the data set is small, which makes it harder to
interpret out-of-sample and cross validation results. In this case the out-of-sample
test performance might be small due to good luck, or large due to bad luck, in the
selection of the test set. In cross-validation the test results can vary considerably,
due to luck of which data points fall into the different folds. Here too concepts
from statistics can make this idea more precise, but we leave it as an informal idea:
With small data sets, we can expect to see more variation in test RMS prediction
error than with larger data sets.
Despite these limitations, out-of-sample and cross-validation are powerful and
useful tools for assessing the generalization ability of a model.
 13.3 Feature engineering 211

68

66
Temperature ( F)

64

62

60

58

56

20 40 60 80 100 120
k

Figure 13.13 Hourly temperature at Los Angeles International Airport be-

tween 12:53AM on May 25, 2016, and 11:53PM on May 29, 2016, shown as
circles. The solid line is the prediction of an autoregressive model with eight
coefficients, developed using training data from May 1 to May 24.

13.3 Feature engineering

In this section we discuss some methods used to find appropriate basis functions
or feature mappings f1 , . . . , fp . We observed above that fitting a linear in the
parameters model reduces to regression with new features which are the original
features x mapped through the basis (or feature mapping) functions f1 , . . . , fp .
Choosing the feature mapping functions is sometimes called feature engineering,
since we are generating features to use in regression.
For a given data set we may consider several, or even many, candidate choices
of the basis functions. To choose among these candidate choices of basis functions,
we use out of sample validation or cross-validation.

Adding new features to get a richer model. In many cases the basis functions
include the constant one, i.e., we have f1 (x) = 1. (This is equivalent to having the
offset in the basic regression model.) It is also very common to include the original
features as well, as in fi (x) = xi1 , i = 2, . . . , n + 1. If we do this, we are effectively
starting with the basic regression model; we can then add new features to get a
richer model. In this case we have p > n, so there are more mapped features that
original features. (Whether or not it is a good idea to add the new features can be
determined by out of sample validation or cross-validation.)
212 13 Least squares data fitting

Dimension reduction. In some cases, and especially when the number n of the
original features is very large, the feature mappings are used to construct a smaller
set of p < n features. In this case we can think of the feature mappings or basis
functions as a dimension reduction or data aggregation procedure.

13.3.1 Transforming features

Standardizing features. Instead of using the original features directly, it is com-

mon to apply a scaling and offset to each original feature, say,

fi (x) = (xi bi )/ai , i = 2, . . . , n + 1,

so that across the data set, the average value of fi (x) is near zero, and the standard
deviation is around one. (This is done by choosing bi to be near the mean of
the feature i values over the data set, and choosing ai to be near the standard
deviation of the values.) This is called standardizing or z-scoring the features.
The standardized feature values are easily interpretable since they correspond to
z-values; for example, f3 (x) = +3.3 means that the value of original feature 2 is
quite a bit above the typical value. The standardization of each original feature
is typically the first step in feature engineeering. The constant feature is not
standardized. (In fact, it cannot be standardized since its standard deviation across
the data set is zero.)

Winsorizing features. When the data include some very large values that are
thought to be errors (say, in collecting the data), it is common to clip or winsorize
the data. This means that we set any values that exceed some chosen maximum
absolute value to that value. Assuming, for example, that a feature entry x5
has already been standardized (so it represents z-scores across the examples), we
replace x5 with its winsorized value (with threshold 3),

x5 |x5 | 3
x5 = 3 x5 > 3
3 x5 < 3.

Log transform. When feature values are positive and vary over a wide range,
it is common to replace them with their logarithms. If the feature value also
includes the value 0 (so the logarithm is undefined) a common variation on the log
transformation is to use xk = log(xk + 1). This compresses the range of values that
we encounter. As an example, suppose the original features record the number
of visits to websites over some time period. These can easily vary over a range of
10000:1 (or even more) for a very popular website and a less popular one; taking the
logarithm of the visit counts gives a feature with less variation, which is possibly
more interpretable. (The decision as to whether to use the original feature values
or their logarithms can be decided by validation.)
 13.3 Feature engineering 213

13.3.2 Creating new features

Expanding categoricals. Some features take on only a few values, such as 1 and
1 or 0 and 1, which might represent some value like presence or absence of some
symptom. (Such features are called Boolean.) A Likert scale response (see page 57)
naturally only takes on a small number of values, such as 2, 1, 0, 1, 2. Another
example is an original feature that takes on the values 1, 2, . . . , 7, representing
the day of the week. Such features are called categorical in statistics, since they
specify which category the example is in, and not some real number.
Expanding a categorical feature with l values means replacing it with a set of
l 1 new features, each of which is Boolean, and simply records whether or not
the original feature has the associated value. (When all these features are zero,
it means the original feature had the default value.) As an example, suppose the
original feature x1 takes on only the values 1, 0, and 1. Using the feature value
0 as the default feature value, we replace x1 with the two mapped features
 
1 x1 = 1 1 x1 = 1
f1 (x) = f2 (x) =
0 otherwise, 0 otherwise.

In words, f1 (x) tells us if x1 has the value 1, and f2 (x) tells us if x1 has the value
1. (We do not need a new feature for the default value x1 = 0; this corresponds to
f1 (x) = f2 (x) = 0.) There is no need to expand an original feature that is Boolean
(i.e., takes on two values).
As an example, consider a model that is used to predict house prices based on
various features that include the number of bedrooms, that ranges from 1 to 5 (say).
In the basic regression model, we use the number of bedrooms directly as a feature.
If we expand this categorical feature, using 2 bedrooms as the default, we have 4
Boolean features that correspond to a house having 1, 3, 4, and 5 bedrooms. In the
basic model there is one parameter value that corresponds to value per bedroom;
we multiply this parameter by the number of bedrooms to get the contribution to
our price prediction. When we expand the bedroom categorical feature, we have
4 parameters in our model, which assign the amounts to add to our prediction for
houses with 1, 3, 4, and 5 bedrooms, respectively.

Generalized additive model. We introduce new features that are nonlinear func-
tions of the original features, such as, for each xi , the functions min{xi + a, 0} and
max{xi b, 0}, where a and b are parameters. A common choice, assuming that
xi has already been standardized, is a = b = 1. This leads to the predictor

y = 1 (x1 ) + + n (xn ),

where i is the piecewise-linear function

i (xi ) = n+i min{xi + a, 0} + i xi + 2n+i max{xi b, 0},

which has kink or knot points at the values a and +b. This model has 3n pa-
rameters. This model is a sum of functions of the original features, and is called a
generalized additive model.
214 13 Least squares data fitting

Products and interactions. New features can be developed from pairs of original
feautures, such as their products. From the original features we can add xi xj , for
i, j = 1, . . . , n, i j. Products are used to model interactions among the features.
Product features are easily interpretable when the original features are Boolean,
i.e., take the values 0 or 1. Thus xi = 1 means that feature i is present or has
occured, and xi xj = 1 exactly when both feature i and j have occured.

Stratified models. In a stratified model, we have several different sub-models,

and choose the one to use depending on the values of the regressors. For example,
instead of treating gender as a regressor in a single model of some medical outcome,
we build two different sub-models, one for male patients and one for female patients.
In this case we choose the sub-model to use based on one of the original features,
gender. As a more general example, we can carry out clustering of the original
feature vectors, and fit a separate model within each each cluster. To evaluate y
for a new x, we first determine which cluster x is in, and then use the associated
model. Whether or not a stratified model is a good idea is checked using out of
sample validation.

13.3.3 Advanced feature generation methods

Custom mappings. In many applications custom mappings of the raw data are
used as additional features, in addition to the original features given. For example
in a model meant to predict an assets future price using prior prices, we might also
use the highest and lowest prices over the last week. Another well known example
in financial models is the price-to-earnings ratio, constructed from the price and
(last) earnings features.
In document analysis applications word count features are typically replaced
with term frequency inverse document frequency (TFID) values, which scale the
raw count values by a function of the frequency with which the word appears
across the given set of documents, usually in such a way that uncommon words are
given more weight. (There are many variations on the particular scaling function
to use. Which one to use in a given application can be determined by out of sample
or cross-validation.)

Predictions from other models. In many applications there are existing models
for the data. A common trick is to use the predictions of these models as features
in your model. In this case you can describe your model as one that combines
or blends the raw data available with predictions made from one or more existing
models to create a new prediction.

Distance to cluster representatives. We can build new features from a clustering

of the data into k groups. One simple method uses the cluster representatives z1 ,
2 2
. . . , zk , and gives k new features, given by f (x) = ekxzi k / , where is a
parameter.
 13.3 Feature engineering 215

Random features. The new features are given by a nonlinear function of a random
linear combination of the original features. To add K new features of this type,
we first generate a random K n matrix R. We then generate new features as
(Rx)+ or |Rx|, where ()+ and || are applied elementwise to the vector Rx. (Other
nonlinear functions can be used as well.)
This approach to generating new features is quite counter-intuitive, since you
would imagine that feature engineering should be done using detailed knowledge
of, and intuition about, the particular application. Nevertheless this method can
be very effective in some applications.

Neural network features. A neural network computes transformed features using

compositions of linear transformations interspersed with nonlinear mappings such
as the absolute value. This architecture was originally inspired by biology, as a
crude model of how human and animal brains work. The ideas behind neural
networks are very old, but their use has accelerated over the last few years due to
a combination of new techniques, greatly increased computing power, and access
to large amounts of data. Neural networks can find good feature mappings directly
from the data, provided there is a very large amount of data available.

13.3.4 Summary

The discussion above makes it clear that there is much art in choosing features to
use in a model. But it is important to keep several things in mind when creating
new features:

Try simple models first. Start with a constant, then a simple regression model,
and so on. You can compare more sophisticated models against these.

Compare competing candidate models using validation. Adding new features

will always reduce the RMS error on the training data, but the important
question is whether or not it substantially reduces the RMS error on the test
or validation data sets. (We add the qualifier substantially here because a
small reduction in test set error is not meaningful.)

Adding new features can easily lead to over-fit. (This will show up when
validating the model.) The most straightforward way to avoid over-fit is to
keep the model simple. We mention here that another approach to avoiding
over-fit, called regularization (covered in chapter 15), can be very effective
when combined with feature engineering.

13.3.5 House price prediction

We use feature engineering to develop a more complicated regression model for

the house sales data. As mentioned in 2.3, the data set contains records of 774
house sales in the Sacramento area. For our more complex model we will use four
attributes:
216 13 Least squares data fitting

x4 f6 (x) f7 (x) f8 (x)

95811, 95814, 95816, 95817, 95818, 95819 0 0 0
95608, 95610, 95621, 95626, 95628, 95655, 1 0 0
95660, 95662, 95670, 95673, 95683, 95691,
95742, 95815, 95821, 95825, 95827, 95833,
95834, 95835, 95838, 95841, 95842, 95843,
95864
95624, 95632, 95690, 95693, 95820, 95822, 0 1 0
95823, 95824, 95826, 95828, 95829, 95831,
95832, 95757, 95758

95603, 95614, 95630, 95635, 95648, 95650, 0 0 1

95661, 95663, 95677, 95678, 95682, 95722,
95746, 95747, 95762, 95765

Table 13.3 Definition of f6 , f7 , f8 as functions of x4 (5-digit ZIP code).

x1 is the area of the house (in 1000 square feet),

x2 is the number of bedrooms,
x3 is equal to one if the property is a condominium, and zero otherwise,
x4 is the five-digit ZIP code.
Only the first two attributes were used in the simple regression model

y = 1 x1 + 2 x2 + v

given in 2.3. Here we examine a more complicated model, with 8 basis functions,
8
X
y = i fi (x).
i=1

The first function is the constant f1 (x) = 1. The next two are functions of x1 , the
area of the house,

f2 (x) = x1 , f3 (x) = max {x1 1.5, 0}.

In words, f2 (x) is the area of the house, and f3 (x) is the amount by which the
area exceeds 1.5 (i.e., 1500 square feet). The weighted sum 2 f2 (x) + 3 f3 (x) is
a piecewise-linear function of the house area, with one knot at 1.5. The function
f4 (x) is equal to the number of bedrooms x2 . The function f5 (x) is equal to x3 , i.e.,
one if the property is a condominium, and zero otherwise. The last three functions
are again Boolean, and indicate the location of the house. We partition the 62
ZIP codes present in the data set into four groups, corresponding to different areas
around the center of Sacramento, as shown in table 13.3.
The coefficients in the least squares fit are

1 = 115.62, 2 = 175.41, 3 = 42.75, 4 = 17.88,

13.3 Feature engineering 217

800
Predicted price y (thousand dollars)

600

400

200

0
0 200 400 600 800
Actual price y (thousand dollars)

Figure 13.14 Scatter plot of actual and predicted prices for a model with
eight parameters.

5 = 19.05, 6 = 100.91, 7 = 108.79, 8 = 24.77.

The RMS fitting error is 68.3, a bit better than the simple regression fit, which
achieves RMS fitting error of 74.8. Figure 13.14 shows a scatter plot of predicted
and actual prices.
To validate the model, we use 5-fold cross-validation, using the same folds as in
table 13.1 and figure 13.12. The results are shown in table 13.4 and figure 13.15.
We conclude that our more complex model, that uses feature engineering, gives a
modest improvement in prediction ability over the simple regression model based
on only house area and number of bedrooms. (With more data, more features,
and more feature engineering, a much more accurate model of house price can be
developed.)
218 13 Least squares data fitting

Fold 1 Fold 2
800 800

600 600

400 400

200 200

0 0
0 200 400 600 800 0 200 400 600 800

Fold 3 Fold 4
800 800

600 600

400 400

200 200

0 0
0 200 400 600 800 0 200 400 600 800

Fold 5
800

600

400

200

0
0 200 400 600 800

Figure 13.15 Scatter plots of actual and predicted prices for the five models
of table 13.4. The horizontal axis is the actual selling price the vertical axis
is the predicted price, both in thousand dollars. Blue circles are samples in
the training set, red circles samples in the test set.
13.3 Feature engineering 219

Model parameters RMS error

Fold 1 2 3 4 5 6 7 8 Train Test
1 122.35 166.87 39.27 16.31 23.97 100.42 106.66 25.98 67.29 72.78
2 100.95 186.65 55.80 18.66 14.81 99.10 109.62 17.94 67.83 70.81
3 133.61 167.15 23.62 18.66 14.71 109.32 114.41 28.46 69.70 63.80
4 108.43 171.21 41.25 15.42 17.68 94.17 103.63 29.83 65.58 78.91
5 114.45 185.69 52.71 20.87 23.26 102.84 110.46 23.43 70.69 58.27

Table 13.4 Five-fold validation on the house sales data set.

220 13 Least squares data fitting
 Chapter 14

Least squares classification

In this chapter we consider the problem of fitting a model to data where the outcome
takes on values like true or false (as opposed to being numbers, as in chapter 13).
We will see that least squares can be used for this problem as well.

14.1 Classification
In the data fitting problem of chapter 13, the goal is to reproduce or predict the
outcome y, which is a (scalar) number. In a classification problem, the outcome or
dependent variable y takes on only a finite number of values, and for this reason is
sometimes called a label, or in statistics, a categorical. In the simplest case, y has
only two values, for example true or false, or spam or not spam. This is called
the 2-way classification problem, or the Boolean classification problem. We start
by considering the Boolean classification problem.
We will encode y as a real number, taking y = +1 to mean true and y = 1
to mean false. As in real-valued data fitting, we assume that an approximate
relationship of the form y f (x) holds, where f : Rn {1, +1}. (This notation
means that the function f takes an n-vector argument, and gives a resulting value
that is either +1 or 1.) Our model will have the form y = f(x), where f : Rn
{1, +1}. The model f is also called a classifier, since it classifies n-vectors into
those for which f(x) = +1 and those for which f(x) = 1.

Examples. Boolean classifiers are widely used in many application areas.

Email spam detection. The vector x contains features of an email message.

It can include word counts in the body of the email message, other features
such as the number of exclamation points and all-capital words, and features
related to the origin of the email. The outcome is +1 if the message is spam,
and 1 otherwise. The data used to create the classifier comes from users
who have explicitly marked some messages as junk.
222 14 Least squares classification

Fraud detection. The vector x gives a set of features associated with a credit
card holder, such as her average monthly spending levels, median price of
purchases over the last week, number of purchases in different categories, av-
erage balance, and so on, as well as some features associated with a particular
proposed transaction. The outcome y is +1 for a fraudulent transaction, and
1 otherwise. The data used to create the classifier is taken from historical
data, that includes (some) examples of transactions that were later verified
to be fraudulent and (many) that were verified to be bona fide.
Boolean document classification. The vector x is a word count (or histogram)
vector for a document, and the outcome y is +1 if the document has some
specific topic (say, politics) and 1 otherwise. The data used to construct the
classifier might come from a corpus of documents with their topics labeled.
Disease detection. The examples correspond to patients, with outcome y =
+1 meaning the patient has a particular disease, and y = 1 meaning they do
not. The vector x contains relevant medical features associated with the pa-
tient, including for example age, sex, results of tests, and specific symptoms.
The data used to build the model come from hospital records or a medical
study; the outcome is the associated diagnosis, confirmed by a doctor.

Prediction errors. For a given data point (x, y), with predicted outcome y = f(x),
there are only four possibilities:
True positive. y = +1 and y = +1.
True negative. y = 1 and y = 1.
False positive. y = 1 and y = +1.
False negative. y = +1 and y = 1.
In the first two cases the predicted label is correct, and in the last two cases, the
predicted label is an error. We refer to the third case as a false positive or type I
error, and we refer to the fourth case as a false negative or type II error. In some
applications we care equally about making the two types of errors; in others we
may care more about making one type of error than another.

Error rate and confusion matrix. For a given data set (x1 , y1 ), . . . , (xN , yN ) and
model f, we can count the numbers of each of the four possibilities that occur
across the data set, and display them in a contingency table or confusion matrix,
which is a 2 2 table with the columns corresponding to the value of yi and the
rows corresponding to the value of yi . (This is the convention used in machine
learning; in statistics, the rows and columns are sometimes reversed.) The entries
give the total number of each of the four cases listed above, as shown in table 14.1.
The diagonal entries correspond to correct decisions, with the upper left entry the
number of true positives, and the lower right entry the number of true negatives.
The off-diagonal entries correspond to errors, with the upper right entry the number
of false negatives, and the lower left entry the number of false positives. The total
of the four numbers is N , the number of examples in the data set. Sometimes the
totals of the rows and columns are shown, as in table 14.1.
14.1 Classification 223

Prediction
Outcome y = +1 y = 1 Total
y = +1 Ntp Nfn Np
y = 1 Nfp Ntn Nn
All Ntp + Nfp Nfn + Ntp N

Table 14.1 Confusion matrix.

Prediction
Outcome y = +1 (spam) y = 1 (not spam) Total
y = +1 (spam) 95 32 127
y = 1 (not spam) 19 1120 1139
All 114 1152 1266

Table 14.2 Confusion matrix of a spam detector on a data set of 1266 ex-
amples.

Various performance metrics are expressed in terms of the numbers in the con-
fusion matrix.

The error rate is the total number of errors (of both kinds) divided by the
total number of examples, i.e., (Nfp + Nfn )/N .

The true positive rate (also known as the sensitivity or recall rate) is Ntp /Np .
This gives the fraction of the data points with y = +1 for which we correctly
guessed y = +1.

The false positive rate (also known as the false alarm rate) is Nfp /Nn . The
false positive rate is the fraction of data points with y = 1 for which we
incorrectly guess y = +1.

The specificity or true negative rate is one minus the false positive rate, i.e.,
Ntn /Nn . The true negative rate is the fraction of the data points with y = 1
for which we correctly guess y = 1.

A good classifier will have small (near zero) error rate and false positive rate, and
high (near one) true positive rate and true negative rate. Which of these metrics
is more important depends on the particular application.
An example confusion matrix is given in table 14.2 for the performance of a
spam detector on a data set of N = 1266 examples (emails) of which 127 are spam
(y = +1) and the remaining 1139 are not spam (y = 1). On the data set, this
classifier has 95 true positives and 1120 true negatives, 19 false positives, and 32
false negatives. Its error rate is (19 + 32)/1266 = 4.03%. Its true positive rate
is 95/127 = 74.8% (meaning it is detecting around 75% of the spam in the data
224 14 Least squares classification

set), and its false positive rate is 19/1139 = 1.67% (meaning it incorrectly labeled
around 1.7% of the non-spam messages as spam).

Validation in classification problems. In classification problems we are concerned

with the error, true positive, and false positive rates. So out-of-sample validation
and cross-validation are carried out using the performance metric or metrics that
we care about, i.e., the error rate or some combination of true positive and false
negative rates. We may care more about one of these metrics than the others.

14.2 Least squares classifier

Many sophisticated methods have been developed for constructing a Boolean model
or classifier from a data set. Logistic regression and support vector machine are
two methods that are widely used, but beyond the scope of this book. Here we
discuss a very simple method, based on least squares, that can work quite well,
though not as well as the more sophisticated methods.
We first carry out ordinary real-valued least squares fitting of the outcome,
ignoring for the moment that the outcome y takes on only the values 1 and +1.
We choose basis functions f1 , . . . , fp , and then choose the parameters 1 , . . . , p
so as to minimize the sum squared error
(f(x1 ) y1 )2 + + (f(xN ) yN )2 ,
where f(x) = 1 f1 (x) + + p fp (x). We use the notation f, since this function
is not our final model f. The function f is the least squares fit over our data set,
and f(x), for a general vector x, is a number.
Our final classifier is then taken to be
f(x) = sign(f(x)), (14.1)
where sign(a) = +1 for a 0 and 1 for a < 0. We call this classifier the least
squares classifier.
The intuition behind the least squares classifier is simple. The value f(xi ) is a
number, which (ideally) is near +1 when yi = +1, and near 1 when yi = 1. If
we are forced to guess one of the two possible outcomes +1 or 1, it is natural to
choose sign(f(xi )). Intuition suggests that the number f(xi ) can be related to our
confidence in our guess yi = sign(f(xi )): When f(xi ) is near 1 we have confidence
in our guess yi = +1; when it is small and negative (say, f(xi ) = 0.03), we guess
yi = 1, but our confidence in the guess will be low. We wont pursue this idea
further in this book, except in multi-class classifiers, which we discuss in 14.3.
The least squares classifier is often used with a regression model, i.e., f(x) =
xT + v, in which case the classifier has the form
f(x) = sign(xT + v). (14.2)
We can easily interpret the coefficients in this model. For example, if 7 is negative,
it means that the larger the value of x7 is, the more likely we are to guess y = 1.
 14.2 Least squares classifier 225

Prediction
Outcome y = +1 y = 1 Total
y = +1 46 4 50
y = 1 7 93 100
All 53 97 150

Table 14.3 Confusion matrix for a Boolean classifier of the Iris data set.

If 4 is the coefficient with the largest magnitude, then we can say that x4 is the
feature that contributes the most to our classification decision.

14.2.1 Iris flower classification

We illustrate least squares classification with a famous data set, first used in the
1930s by the statistician Ronald Fisher. The data are measurements of four at-
tributes of three types of iris flowers: Iris Setosa, Iris Versicolour, and Iris Vir-
ginica. The data set contains 50 examples of each class. The four attributes are:

x1 is the sepal length in cm,

x2 is the sepal width in cm,

x3 is the petal length in cm,

x4 is the petal width in cm.

We compute a Boolean classifier of the form (14.2) that distinguishes the class Iris
Virginica from the other two classes. Using the entire set of 150 examples we find
the coefficients

v = 2.39, 1 = 0.0918, 2 = 0.406, 3 = 0.00798, 4 = 1.10.

The confusion matrix associated with this classifier is shown in table 14.3. The
error rate is 7.3%.

Validation. To test our least squares classification method, we apply 5-fold cross-
validation. We randomly divide the data set into 5 folds of 30 examples (10 for
each class). The results are shown in table 14.4. The test data sets contain only
30 examples, so a single prediction error changes the test error rate significantly
(i.e., by 3.3%). This explains what would seem to be large variation seen in the
test set error rates. We might guess that the classifier will perform on new unseen
data with an error rate in the 710% range, but our test sets are not large enough
to predict future performance more accurately than this. (This is an example of
the limitation of cross-validation when the data set is small; see the discussion on
page 210.)
226 14 Least squares classification

Model parameters Error rate (%)

Fold v 1 2 3 4 Train Test
1 2.45 0.0240 0.264 0.00571 0.994 6.7 3.3
2 2.38 0.0657 0.398 0.07593 1.251 5.8 10.0
3 2.63 0.0340 0.326 0.08869 1.189 7.5 3.3
4 1.89 0.3338 0.577 0.09902 1.151 6.7 16.7
5 2.42 0.1464 0.456 0.11200 0.944 8.3 3.3

Table 14.4 Five-fold validation for the Boolean classifier of the Iris data set.

Figure 14.1 Location of the pixels used as features in the handwritten digit
classification example.

14.2.2 Handwritten digit classification

We now consider a much larger example, the MNIST data set described in 4.4.1.
The (training) data set contains 60,000 images of size 28 by 28. (A few samples are
shown in figure 4.6.) The number of examples per digit varies between 5421 (for
digit five) and 6742 (for digit one). The pixel intensities are scaled to lie between 0
and 1. We remove the pixels that are nonzero in fewer than 600 training examples.
The remaining 493 pixels are shown as the white area in figure 14.1. There is also
a separate test set containing 10, 000 images. Here we will consider classifiers to
distinguish the digit zero from the other nine digits.
In this first experiment, we use the 493 pixel intensities, plus an additional
feature with value 1, as the n = 494 features in the least squares classifier (14.1).
The performance on the (training) data set is shown in the confusion matrix in
table 14.5. The error rate is 1.6%, the true positive rate is 87.1%, and the false
positive rate is 0.3%.
Figure 14.2 shows the distribution of the values of f(xi ) for the two classes in
14.2 Least squares classifier 227

Prediction
Outcome y = +1 y = 1 Total
y = +1 5158 765 5923
y = 1 169 53910 54077
All 5325 54675 60000

Table 14.5 Confusion matrix for a classifier for recognizing the digit zero,
on a training set of 60000 examples.

Prediction
Outcome y = +1 y = 1 Total
y = +1 864 116 980
y = 1 42 8978 9020
All 906 9094 10000

Table 14.6 Confusion matrix for the classier for recognizing the digit zero,
on a test set of 10000 examples.

the training set. The interval [2.1, 2.1] is divided in 100 intervals of equal width.
For each interval, the height of the blue bar is the fraction of the total number
of training examples xi from class +1 (digit zero) that have a value f(xi ) in the
interval. The height of the red bar is the fraction of the total number of training
examples from class 1 (digits 19) with f(xi ) in the interval. The vertical dashed
line shows the decision boundary: For f(xi ) to the left (i.e., negative) we guess
that digit i is from class 1, i.e., digit 19; for f(xi ) to the right of the dashed
line, we guess that digit i is from class +1, i.e., digit 0. False positives correspond
to red bars to the right of the dashed line, and false negatives correspond to blue
bars to the left of the line.
Figure 14.3 shows the values of the coefficients k , displayed as an image. We
can interpret this image as a map of the sensitivity of our classifier to the pixel
values. Pixels with i = 0 are not used at all; pixels with larger positive values of
i are locations where the larger the image pixel value, the more likely we are to
guess that the image represents the digit zero.

Validation. The performance of the least squares classifier on the test set is shown
in the confusion matrix in table 14.6. For the test set the error rate is 1.6%, the
true positive rate is 88.2%, and the false positive rate is 0.5%. These performance
metrics are similar to those for the training data, which suggests that our classifier
is not over-fit, and gives us some confidence in our classifier.
228 14 Least squares classification

0.1 Positive
Negative
0.08

Fraction 0.06

0.04

0.02

0
2 1.5 1 0.5 0 0.5 1 1.5 2

f (xi )

Figure 14.2 The distribution of the values of f(xi ) in the Boolean classi-
fier (14.1) for recognizing the digit zero, over all elements xi of the training
set. The red bars correspond to the digits from class 1, i.e., the digits 19;
the blue bars correspond to the digits from class +1, i.e., the digit 0.

0.1

0.05

0.05

0.1

Figure 14.3 The coefficients k in the least squares classifier that distin-
guishes the digit zero from the other nine digits.
 14.2 Least squares classifier 229

Prediction Prediction
Outcome y = +1 y = 1 Total Outcome y = +1 y = 1 Total
y = +1 5813 110 5923 y = +1 963 17 980
y = 1 15 54062 54077 y = 1 7 9013 9020
All 5828 54172 60000 All 970 9030 10000

Table 14.7 Confusion matrices for the Boolean classifier to recognize the
digit zero after addition of 5000 new features. The table on the left is for
the training set. The table on the right is for the test set.

Feature engineeering. We now do some simple feature engineering (as in 13.3)

to improve our classifier. We add 5000 new features, calculated as follows. We first
generate 5000 493-vectors rj and numbers sj with randomly chosen entries 1, and
use them to define nonlinear basis functions max{0, rjT x+sj }. After the addition of
the 5000 new features we get the confusion matrices for the training and test data
sets shown in table 14.7. The error rates are consistent, and equal to 0.21% for the
training set and 0.24% for the test set, a substantial improvement compared to the
1.6% in the first experiment. A comparison of the distributions in figure 14.4 and
figure 14.2 also shows how much better the new classifier distinguishes between the
two classes of the training set. We conclude that this was a successful exercise in
feature engineering.

14.2.3 Receiver operating characteristic

One useful modification of the least squares classifier (14.1) is to skew the decision
boundary, by subtracting a constant from f(x) before taking the sign:

f(x) = sign(f(x) ). (14.3)

The classifier is then

f(x)

+1
f(x) =
1 f(x) < .
We call the decision threshold for the modified classifier. The basic least squares
classifier (14.1) has decision threshold = 0.
By choosing positive, we make the guess f(x) = +1 less frequently, so the
numbers in the first column of the confusion matrix go down, and the numbers in
the second column go up (since the sum of the numbers in each row is always the
same). This means that choosing positive decreases the true positive rate (which
is bad), but it also decreases the false positive rate (which is good). Choosing
negative has the opposite effect, increasing the true positive rate (which is good)
and increasing the false positive rate (which is bad). The parameter is chosen
depending on how much we care about the two competing metrics, in the particular
application.
230 14 Least squares classification

Positive
Negative
0.15

Fraction
0.10

0.05

0
2 1.5 1 0.5 0 0.5 1 1.5 2
f(xi )

Figure 14.4 The distribution of the values of f(xi ) in the Boolean classi-
fier (14.1) for recognizing the digit zero, after addition of 5000 new features.

By sweeping over a range, we obtain a family of classifiers that vary in their

true positive and false positive rates. We can plot the false positive and negative
rates, as well as the error rate, as a function of . A more common way to plot
this data has the strange name receiver operating characteristic (ROC). The ROC
shows the true positive rate on the vertical axis and false positive rate on the
horizontal axis. The name comes from radar systems deployed during World War
II, where y = +1 means that an enemy vehicle (or ship or airplane) is present, and
y = +1 means that an enemy vehicle is detected.

Example. We examine the skewed threshold least squares classifier (14.3) for the
example described above, where we attempt to detect whether or not a handwritten
digit is zero. Figure 14.5 shows how the error, true positive, and false positive rates
depend on the decision threshold , for the training set data. We can see that as
increases, the true positive rate decreases, as does the false positive rate. We
can see that for this particular case the total error rate is minimized by choosing
= 0.1, which gives error rate 1.4%, slightly lower than the basic least squares
classifier. The limiting cases when is negative enough, or positive enough, are
readily understood. When is very negative, the prediction is always y = +1; our
error rate is then the fraction of the data set with y = 1. When is very positive,
the prediction is always y = 1, which gives an error rate equal to the fraction of
the data set with y = +1.
The same information (without the total error rate) is plotted in the traditional
ROC curve shown in figure 14.6. The dots show the basic least squares classifier,
14.2 Least squares classifier 231

0.1 Positive
Negative
0.08
Fraction

0.06

0.04

0.02

0
2 1.5 1 0.5 0 0.5 1 1.5 2
f(xi )

1 True positive
False positive
Total error
0.8

0.6
Rate

0.4

0.2

2 1.5 1 0.5 0 0.5 1 1.5 2

Figure 14.5 True positive, false positive, and total error rate versus decision
threshold . The vertical dashed line is shown for decision threshold =
0.25.
232 14 Least squares classification

0.95 = 0.25

True positive rate

0.9
=0
0.85

0.8

0.75
= 0.25
0.7
0 0.05 0.1 0.15 0.2
False positive rate

Figure 14.6 ROC curve.

with = 0, and the skewed threshold least squars classifiers for = 0.25 and
= 0.25. These curves are for the training data; the same curves for the test
data look similar, giving us some confidence that our classifiers will have similar
performance on new, unseen data.

14.3 Multi-class classifiers

In a multi-class classification problem, we have K > 2 possible labels. This is
sometimes referred to more compactly as K-class classification. (The case K = 2
is Boolean classification, discussed above.) For our generic discussion of multi-class
classifiers, we will encode the labels as y = 1, 2, . . . , K. In some applications there
are more natural encodings; for example, the Likert scale labels Strongly Disagree,
Disagree, Neutral, Agree, and Strongly Agree are typically encoded as 2, 1, 0, 1, 2,
respectively.
A multi-class classifier is a function f : Rn {1, . . . , K}. Given a feature
vector x, f(x) (which is an integer between 1 and K) is our prediction of the
associated outcome. A multi-class classifier classifies n-vectors into those that take
on each of the values 1, . . . , K.

Examples. Multi-class classifiers are used in many application areas.

Handwritten digit classification. We are given an image of a hand-written

digit (and possibly other features generated from the images), and wish to
14.3 Multi-class classifiers 233

guess which of ten digits it represents. This classifier is used to do automatic

(computer-based) reading of handwritten digits.

Marketing demographic classification. Data from purchases made, or web

sites visited, is used to train a multi-class classifier for a set of market seg-
ments, such as college-educated women aged 2530, men without college
degrees aged 4555, and so on. This classifier guesses which demographic
segment a new customer belongs to, based only on their purchase history.
This can be used to select which promotions to offer a customer for whom
we only have purchase data. The classifier is trained using data from known
customers.

Disease diagnosis. The labels are a set of diseases (including one label that
corresponds to disease-free), and the features are medically relevant values,
such as patient attributes and the results of tests. Such a classifier carries
out diagnosis (of the diseases corresponding to the labels). The classifier is
trained on cases in which a definitive diagnosis has been made.

Translation word choice. A machine translation system translates a word in

the source language to one of several possible words in the target language.
The label corresponds to a particular choice of translation for the word in the
source language. The features contain information about the context around
the word, for example, words counts or occurrences in the same paragraph.
As an example, the English word bank might be translated into another
language one way if the word river appears nearby, and another way if
financial or reserves appears nearby. The classifier is trained on data taken
from translations carried out by (human) experts.

Document topic prediction. Each example corresponds to a document or

article, with the feature vector containing word counts or histograms, and
the label corresponding to the topic or category, such as Politics, Sports,
Entertainment, and so on.

Prediction errors and confusion matrix. For a multi-class classifier f and a given
data point (x, y), with predicted outcome y = f(x), there are K 2 possibilities, cor-
responding to the all pairs of values of y, the actual outcome, and y, the predicted
outcome. For a given data set (training or validation set) with N elements, the
numbers of each of the K 2 occurrences are arranged into a K K confusion matrix,
where Nij is the number of data points for which y = i and y = j.
The K diagonal entries N11 , . . . , NKK correspond to the cases when the predic-
tion is correct; the K 2 K off-diagonal entries Nij , i 6= j, correspond to prediction
errors. For each i, Nii is the number of data points with label i for which we cor-
rectly guessed y = i. For i 6= j, Nij is the number of data points for which we have
mistaken label i (its true value) for the label j (our incorrect guess). For K = 2
(Boolean classification) there are only two types of prediction errors, false positive
and false negative. For K > 2 the situation is more complicated, since there are
many more types of errors a predictor can make. From the entries of the confusion
matrix we can derive various measures of the accuracy of the predictions. We let
234 14 Least squares classification

y
y Dislike Neutral Like
Dislike 183 10 5
Neutral 7 61 8
Like 3 13 210

Table 14.8 Example confusion matrix of a multi-class classifier with three

classes.

Ni (with one index) denote the total number of data points for which y = i, i.e.,
Ni = Ni1 + + NiK . We have N = N1 + + NK .
The simplest measure is the overall error rate, which is the total number of
errors (the sum of all off-diagonal entries in the confusion matrix) divided by the
data set size (the sum of all entries in the confusion matrix):
X X
(1/N ) Nij = 1 (1/N ) Nii .
i6=j i

This measure implicitly assumes that all errors are equally bad. In many applica-
tions this is not the case; for example, some medical mis-diagnoses might be worse
for a patient than others.
We can also look at the rate with which we predict each label correctly. The
quantity Nii /Ni is called the true label i rate. It is the fraction of data points with
label y = i for which we correctly predicted y = i. (The true label i rates reduce
to the true positive and true negative rates for Boolean classifiers.)
A simple example, with K = 3 labels (Dislike, Neutral, and Like), and a total
number N = 500 data points, is shown in table 14.8. Out of 500 data points, 454
(the sum of the diagonal entries) were classified correctly. The remaining 46 data
points (the sum of the off-diagonal entries) correspond to the 6 different types of
errors. The overall error rate is 46/500 = 9.2%. The true label Dislike rate is
183/(183 + 10 + 5) = 92.4%, i.e., among the data points with label Dislike, we
correctly predicted the label on 92.4% of the data. The true label Neutral rate is
61/(7 + 16 + 8) = 80.3%, and the true label Like rate is 210/(3 + 13 + 210) = 92.9%.

14.3.1 Least squares multi-class classifier

The idea behind the least squares Boolean classifier can be extended to handle
multi-class classification problems. For each possible label value, we construct a
new data set with the Boolean label +1 if the label has the given value, and 1
otherwise. (This is sometimes called a one-versus-others classifier.) From these K
Boolean classifiers we must create a classifier that chooses one of the K possible
labels. We do this by selecting the label for which the least squares regression fit
has the highest value, which roughly speaking is the one with the highest level of
14.3 Multi-class classifiers 235

confidence. Our classifier is then

f(x) = argmax fk (x),

k=1,...,K

where fk is the least squares regression model for label k against the others. The
notation argmax means the index of the largest value among the numbers fk (x),
for k = 1, . . . , K. Note that fk (x) is the real-valued prediction for the Boolean
classifier for class k versus not class k; it is not the Boolean classifier, which is
sign(fk (x)).
As an example consider a multi-class classification problem with 3 labels. We
construct 3 different least squares classifiers, for 1 versus 2 or 3, for 2 versus 1 or
3, and for 3 versus 1 or 2. Suppose for a given feature vector x, we find that

f1 (x) = 0.7, f2 (x) = +0.2, f3 (x) = +0.8.

The largest of these three numbers is f3 (x), so our prediction is f(x) = 3. We can
interpret these numbers and our final decision. The first classifier is fairly confident
that the label is not 1. Accordinging to the second classifier, the label could be 2,
but it does not have high confidence in this prediction. Finally, the third classifier
predicts the label is 3, and moreover has relatively high confidence in this guess.
So our final guess is label 3. (This interpretation suggests that if we had to make a
second guess, it should be label 2.) Of course here we are anthropomorphizing the
individual label classifiers, since they do not have beliefs or levels of confidence in
their predictions. But the story is helpful in understanding the motivation behind
the classifier above.

Skewed decisions. In a Boolean classifier we can skew the decision threshold (see
14.2.3) to trade off the true positive and false positive rates. In a K-class classifier,
an analogous method can be used to trade off the K true label i rates. We apply
an offset k to fl (x) before finding the largest value. This gives the predictor
 
f(x) = argmax fk (x) k ,
k=1,...,K

where k are constants chosen to trade off the true label k rates. If we decrease k ,
we predict f(x) = k more often, so all entries of the kth column in the confusion
matrix increase. This increases our rate of true positives for label k (since Nkk
increases), which is good. But it can decrease the true positive rates for the other
labels.

Complexity. In least squares multi-class classification, we solve K least squares

problems, each with N rows and p variables. The nave method of computing
1 , . . . , K , the coefficients in our one-versus-others classifiers, costs 2KN p2 flops.
But the K least squares problems we solve all involve the same matrix; only the
right-hand side vector changes. This means that we can carry out the QR fac-
torization just once, and use it for computing all K classifier coefficients. (See
page 184.) When K (the number of classes or labels) is small compared to p (the
236 14 Least squares classification

Prediction
Class Setosa Versicolour Virginica Total
Setosa 40 0 0 40
Versicolour 0 27 13 40
Virginica 0 4 36 40
All 40 31 49 120

Table 14.9 Confusion matrix for a 3-class classifier of the Iris data set, on a
training set of 120 examples.

number of basis functions or coefficients in the classifier), the cost is reduced to

that of solving just one least squares problem.
Another simplification in K-class least squares classification arises due to the
special form of the right-hand sides in the K least squares problems to be solved.
The right-hand sides in these K problems are Boolean vectors with entries +1 for
one of the classes and 1 for all others. It follows that the sum of these K right-
hand sides is the vector with all entries equal to K 2, i.e., (K 2)1. Since the
mapping from the right-hand sides to the least squares approximate solutions k is
linear (see page 181), we have 1 + + k = (K 2)a, where a is the least squares
approximate solution when the right-hand side is 1. Assuming that the first basis
function is f1 (x) = 1, we have a = e1 . So we have

1 + + K = (K 2)e1 ,

where k is the coefficient vector for distinguishing class k from the others. Once
we have computed 1 , . . . , K1 , we can find K by simple vector subtraction.
This explains why for the Boolean classification case we have K = 2, but we
only have to solve one least squares problem. In 14.2 we compute one coefficient
vector ; if the same problem were to be considered a K-class problem with K = 2,
we would have 1 = . (This one distinguishes class 1 versus class 2.) The other
coefficient vector is then 2 = 1 . (This one distinguishes class 2 versus class 1.)

14.3.2 Iris flower classification

We compute a 3-class classifier for the Iris data set described on page 225. The
examples are randomly partitioned into a training set of size 120, containing 40
examples of each class, and a training set of size 30, with 10 examples of each class.
The 3 3 confusion matrix for the training set is given in table 14.9. The error
rate is 14.2%. The results for the test set are in table 14.10. The error rate is
13.3%, similar enough to the training error rate to give us some confidence in our
classifier. The true Setosa rate is 100% for both train and test sets, suggesting
that our classifier can detect this type well. The true Versicolour rate is 67.5% for
the train data, and 60% for the test set. The true Virginica rate is 90% for the
 14.3 Multi-class classifiers 237

Prediction
Class Setosa Versicolour Virginica Total
Setosa 10 0 0 10
Versicolour 0 6 4 10
Virginica 0 0 10 10
All 10 6 14 30

Table 14.10 Confusion matrix for a 3-class classifier of the Iris data set, on
a test set of 30 examples.

train data, and 100% for the test set. This suggests that our classifier can detect
Virginica well, but perhaps not as well as Setosa. (The 100% true Virginica rate
on the test set is a matter of luck, due to the very small number of test examples
of each type; see the discussion on page 210.)

14.3.3 Handwritten digit classification

We illustrate the least squares multi-class classification method by applying it to

the MNIST data set. For each of the ten digits 0, . . . , 9 (which we encode as
k = 1, . . . , 10) we compute a least squares Boolean classifier

fk (x) = sign(xT k + vk ),

to distinguish digit k from the other digits. The ten Boolean classifiers are combined
into a multi-class classifier

f(x) = argmax (xT k + vk ).

k=1,...,10

The 10 10 confusion matrix for the data set and the test set are given in ta-
bles 14.11 and 14.12.
The error rate on the training set is 14.5%; on the test set it is 13.9%. The true
label rates on the test set range from 73.5% for digit 5 to 97.5% for digit 1.
Many of the entries of the confusion matrix make sense. From the first row of
the matrix, we see a handwritten 0 was rarely mistakenly classified as a 1, 2, or
9; presumably these digits look different enough that they are easily distinguished.
The most common error (80) corresponds to y = 9, y = 4, i.e., mistakenly identi-
fying a handwritten 9 as a 4. This makes sense since these two digits can look very
similar.

Feature engineering. After adding the 5000 randomly generated new features the
training set error is reduced to about 1.5%, and the test set error to 2.6%. The
confusion matrices are given in tables 14.13 and 14.14. Since we have (substantially)
reduced the error in the test set, we conclude that adding these 5000 new features
was a successful exercise in feature engineering.
238 14 Least squares classification

Prediction
Digit 0 1 2 3 4 5 6 7 8 9 Total
0 5669 8 21 19 25 46 65 4 60 6 5923
1 2 6543 36 17 20 30 14 14 60 6 6742
2 99 278 4757 153 116 17 234 92 190 22 5958
3 38 172 174 5150 31 122 59 122 135 128 6131
4 13 104 41 5 5189 52 45 24 60 309 5842
5 164 94 30 448 103 3974 185 44 237 142 5421
6 104 78 77 2 64 106 5448 0 36 3 5918
7 55 191 36 48 165 9 4 5443 13 301 6265
8 69 492 64 225 102 220 64 21 4417 177 5851
9 67 66 26 115 365 12 4 513 39 4742 5949
All 6280 8026 5262 6182 6180 4588 6122 6277 5247 5836 60000

Table 14.11 Confusion matrix for least squares multi-class classification of

handwritten digits (training set).

Prediction
Digit 0 1 2 3 4 5 6 7 8 9 Total
0 944 0 1 2 2 8 13 2 7 1 980
1 0 1107 2 2 3 1 5 1 14 0 1135
2 18 54 815 26 16 0 38 22 39 4 1032
3 4 18 22 884 5 16 10 22 20 9 1010
4 0 22 6 0 883 3 9 1 12 46 982
5 24 19 3 74 24 656 24 13 38 17 892
6 17 9 10 0 22 17 876 0 7 0 958
7 5 43 14 6 25 1 1 883 1 49 1028
8 14 48 11 31 26 40 17 13 756 18 974
9 16 10 3 17 80 0 1 75 4 803 1009
All 1042 1330 887 1042 1086 742 994 1032 898 947 10000

Table 14.12 Confusion matrix for least squares multi-class classification of

handwritten digits (test set).
14.3 Multi-class classifiers 239

Prediction
Digit 0 1 2 3 4 5 6 7 8 9 Total
0 5888 1 2 1 3 2 10 0 14 2 5923
1 1 6679 27 6 11 0 0 10 6 2 6742
2 11 7 5866 6 12 0 3 22 26 5 5958
3 1 4 31 5988 0 27 0 24 34 22 6131
4 1 15 3 0 5748 1 13 4 5 52 5842
5 6 2 4 26 7 5335 23 2 9 7 5421
6 8 5 0 0 3 15 5875 0 11 1 5918
7 3 25 23 4 8 0 1 6159 5 37 6265
8 5 16 11 12 9 17 11 7 5749 14 5851
9 10 5 1 29 41 16 2 35 25 5785 5949
All 5934 6759 5968 6072 5842 5413 5938 6263 5884 5927 60000

Table 14.13 Confusion matrix for least squares multi-class classification of

handwritten digits, after addition of 5000 features (training set).

Prediction
Digit 0 1 2 3 4 5 6 7 8 9 Total
0 972 0 0 2 0 1 1 1 3 0 980
1 0 1126 3 1 1 0 3 0 1 0 1135
2 6 0 998 3 2 0 4 7 11 1 1032
3 0 0 3 977 0 13 0 5 8 4 1010
4 2 1 3 0 953 0 6 3 1 13 982
5 2 0 1 5 0 875 5 0 3 1 892
6 8 3 0 0 4 6 933 0 4 0 958
7 0 8 12 0 2 0 1 992 3 10 1028
8 3 1 3 6 4 3 2 2 946 4 974
9 4 3 1 12 11 7 1 3 3 964 1009
All 997 1142 1024 1006 977 905 956 1013 983 997 10000

Table 14.14 Confusion matrix for least squares multi-class classification of

handwritten digits, after addition of 5000 features (test set).
240 14 Least squares classification

You could reasonably wonder how much performance improvement is possible

for this example, using feature engineering. For the handwritten digit data set,
humans have an error rate around 2% (with the true digits verified by checking
actual addresses, zip codes, and so on). Further feature engineering (i.e., introduc-
ing even more additional random features, or using neural network features) brings
the error rate down well below 2%, i.e., well below human ability. This should give
you some idea of how powerful the ideas in this book are.
 Chapter 15

Multi-objective least squares

In this chapter we consider the problem of choosing a vector that achieves a com-
promise in making two or more norm squared objectives small. The idea is widely
used in data fitting, image reconstruction, control, and other applications.

15.1 Multi-objective least squares

In the basic least squares problem (12.1), we seek the vector x that minimizes
the single objective function kAx bk2 . In some applications we have multiple
objectives, all of which we would like to be small:

J1 = kA1 x b1 k2 , ..., Jk = kAk x bk k2 .

Here Ai is an mi n matrix, and bi is an mi -vector. We can use least squares to

find the x that makes any one of these objectives as small as possible (provided the
associated matrix has linearly independent columns). This will give us (in general)
k different least squares approximate solutions. But we seek a single x that gives
a compromise, and makes them all small, to the extent possible. We call this the
multi-objective (or multi-criterion) least squares problem, and refer to J1 , . . . , Jk
as the k objectives.

Multi-objective least squares via weighted sum. A standard method for finding
a value of x that gives a compromise in making all the objectives small is to choose
x to minimize a weighted sum objective:

J = 1 J1 + + k Jk = 1 kA1 x b1 k2 + + k kAk x bk k2 , (15.1)

where 1 , . . . , k are positive weights, that express our relative desire for the terms
to be small. If we choose all i to be one, the weighted sum objective is the sum
of the objective terms; we give each of them equal weight. If 2 is twice as large
as 1 , it means that we attach twice as much weight to the objective J2 as to J1 .
Roughly speaking, we care twice as strongly that J2 should be small, compared
242 15 Multi-objective least squares

to our desire that J1 should be small. We will discuss later how to choose these
weights.
Scaling all the weights in the weighted sum objective (15.1) by any positive
number is the same as scaling the weighted sum objective J by the number, which
does not change its minimizers. Since we can scale the weights by any positive
number, it is common to choose 1 = 1. This makes the first objective term J1
our primary objective; we can interpret the other weights as being relative to the
primary objective.

Weighted sum least squares via stacking. We can minimize the weighted sum
objective function (15.1) by expressing it as a standard least squares problem. We
start by expressing J as the norm squared of a single vector:
2
1 (A1 x b1 )
..

J = ,

.

k (Ak x bk )

where we use k(a1 , . . . , ak )k2 = ka1 k2 + + kak k2 for any vectors a1 , . . . , ak . So

we have 2
1 A1 1 b1
.. ..

J = x = kAx bk2 ,

. .


k Ak k bk
where A and b are the matrix and vector

1 A1 1 b1
A = .. ..
, b = . (15.2)

. .
k Ak k bk

The matrix A is m n, and the vector b has length m, where m = m1 + + mk .

We have now reduced the problem of minimizing the weighted sum least squares
objective to a standard least squares problem. Provided the columns of A are
linearly independent, the minimizer is unique, and given by

x = (AT A)1 AT b
= (1 AT1 A1 + + k ATk Ak )1 (1 AT1 b1 + + k ATk bk ). (15.3)

This reduces to our standard formula for the solution of a least squares problem
when k = 1 and 1 = 1. (In fact, when k = 1, 1 does not matter.) We can
compute x via the QR factorization of A.

Independent columns of stacked matrix. Our assumption (12.2) that the columns
of A in (15.2) are linearly independent is not the same as assuming that each of
A1 , . . . , Ak has linearly independent columns. We can state the condition that A
has linearly independent columns as: There is no nonzero vector x that satisfies
Ai x = 0 for i = 1, . . . , k. This implies that if just one of the matrices A1 , . . . , Ak
has linearly independent columns, then A does.
15.1 Multi-objective least squares 243

The stacked matrix A can have linearly independent columns even when none
of the matrices A1 , . . . , Ak do. This can happen when mi < n for all i, i.e., all
Ai are wide. However, we must have m1 + + mk n, since A must be tall (or
square) for the linearly independent columns assumption to hold.

Optimal trade-off curve. We start with the special case of two objectives (also
called the bi-criterion problem), and write the weighted sum objective as
J = J1 + J2 = kA1 x b1 k2 + kA2 x b2 k2 ,
where > 0 is the relative weight put on the second objective, compared to the
first. For small , we care much more about J1 being small than J2 being small;
for large, we care much less about J1 being small than J2 being small.
Let x() denote the weighted sum least squares solution x as a function of ,
assuming the stacked matrices have linearly independent columns. These points
are called Pareto optimal, which means there is no point z that satisfies
kA1 z b1 k2 kA1 x() b1 k2 , kA2 z b2 k2 kA2 x() b2 k2 ,
with one of the inequalities holding strictly. Roughly speaking, there is no point z
that is as good as x() in one of the objectives, and beats it on the other one. To
see why this is the case, we note that any such z would have a value of J that is
less than that achieved by x(), which minimizes J, a contradiction.
We can plot the two objectives kA1 x() b1 k2 and kA2 x() b2 k2 against each
other, as varies over (0, ), to understand the trade-off of the two objectives.
This curve is called the optimal trade-off curve of the two objectives. There is no
point z that achieves values of J1 and J2 that lies below and to the left of the
optimal trade-off curve.

Simple example. We consider a simple example with two objectives, with A1 and
A2 both 10 5 matrices. The entries of the weighted least squares solution x()
are plotted against in figure 15.1. On the left, where is small, x() is very close
to the least squares approximate solution for A1 , b1 . On the right, where is large,
x() is very close to the least squares approximate solution for A2 , b2 . In between
the behavior of x() is very interesting; for instance, we can see that x()3 first
increases with increasing before eventually decreasing.
Figure 15.2 shows the values of the two objectives J1 and J2 versus . As
expected, J1 increases as increases, and J2 decreases as increases. (It can
be shown that this always holds.) Roughly speaking, as increases we put more
emphasis on making J2 small, which comes at the expense of making J1 bigger.
The optimal trade-off curve for this bi-criterion problem is plotted in figure 15.3.
The left end-point corresponds to minimizing kA1 x b1 k2 , and the right end-point
corresponds to minimizing kA2 x b2 k2 . We can conclude, for example, that there
is no vector z that achieves kA1 z b1 k2 2.60 and kA2 z b2 k2 1.96.
The steep slope of the optimal trade-off curve near the left end-point means
that we can achieve a substantial reduction in J2 with only a small increase in J1 .
The small slope of the optimal trade-off curve near the right end-point means that
we can achieve a substantial reduction in J1 with only a small increase in J2 . This
is quite typical, and indeed, is why multi-criterion least squares is useful.
244 15 Multi-objective least squares

x1 ()
0.6
x2 ()
x3 ()
0.4 x4 ()
x5 ()
0.2

0.2

104 102 100 102 104

Figure 15.1 Weighted-sum least squares solution x() as a function of for

a bi-criterion least squares problem with five variables.

4
J1 ()
J2 ()
3.5

2.5

104 102 100 102 104

Figure 15.2 Objective functions J1 = kA1 x() b1 k2 (blue line) and J2 =

kA2 x() b2 k2 (red line) as functions of for the bi-criterion problem in
figure 15.1.
15.1 Multi-objective least squares 245

3.5
= 0.1

3
J2 ()

2.5
=1

2 = 10

2.5 3 3.5 4
J1 ()

Figure 15.3 Optimal trade-off curve for the bi-criterion least squares problem
of figures 15.1 and 15.2.

Optimal trade-off surface. Above we described the case with k = 2 objectives.

When we have more than 2 objectives, the interpretation is similar, although it is
harder to plot the objectives, or the values of x, versus the weights. For example
with k = 3 objectives, we have two weights, 2 and 3 , which give the relative
weight of J2 and J3 compared to J1 . Any solution x() of the weighted least squares
problem is Pareto optimal, which means that there is no point that achieves values
of J1 , J2 , J3 less than or equal to those obtained by x(), with strict inequality
holding for at least one of them. As the parameters 2 and 3 vary over (0, ),
the values of J1 , J2 , J3 sweep out the optimal trade-off surface.

Using multi-objective least squares. In the rest of this chapter we will see several
specific applications of multi-objective least squares. Here we give some general
remarks on how it is used in applications.
First we identify a primary objective J1 that we would like to be small. The
objective J1 is typically the one that would be used in an ordinary single-objective
least squares approach, such as the mean square error of a model on some training
data, or the mean-square deviation from some target or goal.
We also identify one or more secondary objectives J2 , J3 , . . . , Jk , that we would
also like to be small. These secondary objectives are typically generic ones, like
the desire that some parameters be small or smooth, or close to some previous
or prior value. In estimation applications these secondary objectives typically cor-
respond to some kind of prior knowledge or assumption about the vector x that
we seek. We wish to minimize our primary objective, but are willing to accept an
increase in it, if this gives a sufficient decrease in the secondary objectives.
The weights are treated like knobs in our method, that we change (turn or
tune or tweak) to achieve a value of x that we like (or can live with). For given
246 15 Multi-objective least squares

candidate values of we evaluate the objectives; if we decide that J2 is larger

than we would like, but we can tolerate a somewhat larger J3 , then we increase
2 and decrease 3 , and find x and the associated values of J1 , J2 , J3 using the
new weights. This is repeated until a reasonable trade-off among them has been
obtained. In some cases we can be principled in how we adjust the weights; for
example, in data fitting, we can use validation to help guide us in the choice of
the weights. In many other applications, it comes down to (application-specific)
judgment or even taste.
The additional terms 2 J2 , . . . , k Jk that we add to the primary objective
J1 , are sometimes called regularization (terms). The secondary objectives are the
sometimes described by name, as in least squares fitting with smoothness regular-
ization.

15.2 Control
In control applications, the goal is to decide on a set of actions or inputs, specified
by an n-vector x, that achieve some goals. The actions result in some outputs or
effects, given by an m-vector y. We consider here the case when the inputs and
outputs are related by an affine model

y = Ax + b.

The m n matrix A and m-vector b characterize the input-output mapping of

the system. The model parameters A and b are found from analytical models,
experiments, computer simulations, or fit to past (observed) data. Typically the
input or action x = 0 has some special meaning. The m-vector b gives the output
when the input is zero. In many cases the vectors x and y represent deviations of
the inputs and outputs from some standard values.
We typically have a desired or target output, denoted by the m-vector y des .
The primary objective is
J1 = kAx + b y des k2 ,

the norm squared deviation of the output from the desired output. The main
objective is to choose an action x so that the output is as close as possible to the
desired value.
There are many possible secondary objectives. The simplest one is the norm
squared value of the input, J2 = kxk2 , so the problem is to optimally trade off
missing the target output (measured by ky y des k2 ), and keeping the input small
(measured by kxk2 ).
Another common secondary objective has the form J2 = kx xnom k2 , where
nom
x is a nominal or standard value for the input. In this case the secondary ob-
jective it to keep the input close to the nominal value. This objective is sometimes
used when x represents a new choice for the input, and xnom is the current value.
In this case the goal is to get the output near its target, while not changing the
input much from its current value.
15.2 Control 247

Control of heating and cooling. As an example, x could give the vector of n

heating (or cooling) power levels in a commercial building with n air handling
units (with xi > 0 meaning heating and xi < 0 meaning cooling) and y could
represent the resulting temperature at m locations in the building. The matrix
A captures the effect of each of n heating/cooling units on the temperatures in
the building at each of m locations; the vector b gives the temperatures at the m
locations when no heating or cooling is applied. The desired or target output might
be y des = T des 1, assuming the target temperature is the same at all locations. The
primary objective ky y des k2 is the sum of squares of the deviations of the location
temperatures from the target temperature. The secondary objective J2 = kxk2 ,
the norm squared of the vector of heating/cooling powers, would be reasonable,
since it is at least roughly related to the energy cost of the heating and cooling.
We find tentative choices of the input by minimizing J1 +2 J2 for various values
of 2 . If for the current value of 2 the heating/cooling powers are larger than wed
like, we increase 2 and re-compute x.

Product demand shaping. In demand shaping, we adjust or change the prices of

a set of n products in order to move the demand for the products towards some
given target demand vector, perhaps to better match the available supply of the
products. The standard price elasticity of demand model is dem = E d price , where
dem is the vector of fractional demand changes, price is the vector of fractional
price changes, and E d is the price elasticity of demand matrix. (These are all
described on page 122.) In this example the price change vector price represents
the action that we take; the result is the change in demand, dem . The primary
objective could be

J1 = k dem tar k2 = kE d price tar k2 ,

where tar is the target change in demand.

At the same time, we want the price changes to be small. This suggests the
secondary objective J2 = k price k2 . We then minimize J1 + J2 for various values
of , which trades off how close we come to the target change in demand with how
much we change the prices.

Dynamics. The system can also be dynamic, meaning that we take into account
time variation of the input and output. In the simplest case x is the time series of
a scalar input, so xi is the action taken in period i, and yi is the (scalar) output in
period i. In this setting, y des is a desired trajectory for the output. A very common
model for modeling dynamic systems, with x and y representing scalar input and
output time series, is a convolution: y = h x. In this case, A is Toeplitz, and b
represents a time series, which is what the output would be with x = 0.
As a typical example in this category, the input xi can represent the torque
applied to the drive wheels of a locomotive (say, over 1 second intervals), and yi is
the locomotive speed.
In addition to the usual secondary objective J2 = kxk2 , it is common to have
an objective that the input should be smooth, i.e., not vary too rapidly over time.
This is achieved with the objective kDxk2 , where D is the (n1)n first difference
248 15 Multi-objective least squares

matrix
1 1 0 0 0 0
0 1 1 0 0 0
.. .. .. .. .. ..

D= . (15.4)

. . . . . .
0 0 0 1 1 0
0 0 0 0 1 1
In 17.2 we will see another way to carry out control with a dynamic system.

15.3 Estimation and inversion

In the broad application area of estimation (also called inversion), the goal is to
estimate a set of n values (also called parameters), the entries of the n-vector x. We
are given a set of m measurements, the entries of an m-vector y. The parameters
and measurements are related by
y = Ax + v,
where A is a known m n matrix, and v is an unknown m-vector. The matrix A
describes how the measured values (i.e., yi ) depend on the unknown parameters
(i.e., xj ). The m-vector v is the measurement error or measurement noise, and is
unknown but presumed to be small. The estimation problem is to make a sensible
guess as to what x is, given y (and A), and prior knowledge about x.
If the measurement noise were zero, and A has linearly independent columns,
we could recover x exactly, using x = A y. (This is called exact inversion.) Our
job here is to guess x, even when these strong assumptions do not hold. Of course
we cannot expect to find x exactly, when the measurement noise is nonzero, or
when A does not have linearly independent columns. This is called approximate
inversion, or in some contexts, just inversion.
The matrix A can be wide, square, or tall; the same methods are used to esti-
mate x in all three cases. When A is wide, we would not have enough measurements
to determine x from y, even without the noise (i.e., with v = 0). In this case we
have to also rely on our prior information about x to make a reasonable guess.
When A is square or tall, we would have enough measurements to determine x, if
there were no noise present. Even in this case, judicious use of multiple-objective
least squares can incorporate our prior knowledge in the estimation, and yield far
better results.

15.3.1 Regularized inversion

If we guess that x has the value x, then we are implicitly making the guess that v
has the value y Ax. If we assume that smaller values of v (measured by kvk) are
more plausible than larger values, then a sensible choice for x is the least squares
approximate solution, which minimizes kAxyk2 . We will take this as our primary
objective.
15.3 Estimation and inversion 249

Our prior information about x enters in one or more secondary objectives.

Simple examples are listed below.
kxk2 : x should be small. This corresponds to the (prior) assumption that x
is more likely to be small than large.
kx xprior k2 : x should be near xprior . This corresponds to the assumption
that x is near some known vector xprior .
kDxk2 , where D is the first difference matrix (15.4). This corresponds to
the assumption that x should smooth, i.e., xi+1 should be near xi . This
regularization is often used when x represents a time series.
The Laplacian L(x) = kAT xk2 , where A is the incidence matrix of a graph
(see page 110). This corresponds to the assumption that x varies smoothly
across the graph, i.e., xi is near xj when i and j are connected by an edge
of the graph.
Finally, we will choose our estimate x by minimizing
kAx yk2 + 2 J2 (x) + + p Jp (x),
where i > 0 are weights, and J2 , . . . , Jp are the regularization terms. This
is called regularized inversion or regularized estimation. We may repeat this for
several choices of the weights, and choose the estimate that is best for the particular
application.

Tikhonov regularized inversion. Choosing x to minimize

kAx yk2 + kxk2
for some choice of > 0 is called Tikhonov regularized inversion. Here we seek a
guess x that is consistent with the measurements (i.e., kAx yk2 is small), but not
too big.
The stacked matrix in this case,
 
A
A = ,
I
always has linearly independent columns, without any assumption about A, which
can have any dimensions, andneed not have linearlyindependent columns. To see
this we note that Ax = (Ax, x) = 0 implies that x = 0, which implies x = 0.
The Gram matrix associated with A,
AT A = AT A + I,
is therefore always invertible (provided > 0).

Equalization. The vector x represents a transmitted signal or message, consisting

of n real values. The matrix A represents the mapping from the transmitted signal
to what is received (called the channel ); y = Ax + v includes noise as well as the
action of the channel. Guessing what x is, given y, can be thought of as un-doing
the effects of the channel. In this context, estimation is called equalization.
250 15 Multi-objective least squares

15.3.2 Estimating a periodic time series

Suppose that the T -vector y is a (measured) time series, that we believe is a noisy
version of a periodic time series, i.e., one that repeats itself every P periods. We
might also know or assume that the periodic time series is smooth, i.e., its adjacent
values are not too far apart.
Periodicity arises in many time series. For example, we would expect a time
series of hourly temperature at some location to approximately repeat itself every
24 hours, or the monthly snowfall in some region to approximately repeat itself
every 12 months. (Periodicity with a 24 hour period is called diurnal ; periodicity
with a yearly period is called seasonal or annual.) As another example, we might
expect daily total sales at a restaurant to approximately repeat itself weekly. The
goal is to get an estimate of Tuesdays total sales, given some historical daily sales
data.
The periodic time series will be represented by a P -vector x, which gives its
values over one period. It corresponds to the full time series

y = (x, x, . . . , x)

which just repeats x, where we assume here for simplicity that T is a multiple of
P . (If this is not the case, the last x is replaced with a slice of the form x1:k .) We
can express y as y = Ax, where A is the T P selector matrix

I
A = ... .

Our total square estimation error is kAx yk2 .

We can minimize this objective analytically. The solution x is found by aver-
aging the values of y associated with the different entries in x. For example, we
estimate Tuesday sales by averaging all the entries in y that correspond to Tues-
days. This simple averaging works well if we have many periods worth of data, i.e.,
if T /P is large.
A more sophisticated estimate can be found by adding regularization for x to
be smooth, based on the assumption that

x1 x2 , ..., xP 1 xP , xP x1 .

(Note that we include the wrap-around pair xP and x1 here.) We measure non-
smoothness as kDcirc xk2 , where Dcirc is the P P circular difference matrix

1 1 0 0 0 0
0 1 1 0 0 0
.. .. .. .. .. ..


Dcirc =
. . . . . .
.

0 0 0 1 1 0

0 0 0 0 1 1
1 0 0 0 0 1
 15.3 Estimation and inversion 251

We estimate the periodic time series by minimizing

kAx yk2 + kDcirc xk2 .

For = 0 we recover the simple averaging mentioned above; as gets bigger, the
estimated signal becomes smoother, ultimately converging to a constant (which is
the mean of the original time series data).
The time series Ax is called the extracted seasonal component of the given time
series data y (assuming we are considering yearly variation). Subtracting this from
the original data yields the time series yAx, which is called the seasonally adjusted
time series.
The parameter can be chosen using validation. This can be done by selecting
a time interval over which to build the estimate, and another one to validate it.
For example, with 4 years of data, we might train our model on the first 3 years of
data, and test it on the last year of data.

Example. In figure 15.4 we apply this method to a series of hourly ozone mea-
surements. The top figure shows hourly measurements over a period of 14 days
(July 114, 2014). We represent these values by a 336-vector c, with c24(j1)+i ,
i = 1, . . . , 24, defined as the hourly values on day j, for j = 1, . . . , 14. As indicated
by the gaps in the graph, a number of measurements are missing from the record
(only 275 of the 336 = 24 14 measurements are available). We use the notation
Mj {1, 2, . . . , 24} to denote the set containing the indices of the available mea-
surements on day j. For example, M8 = {1, 2, 3, 4, 6, 7, 8, 23, 24}, because on July
8, the measurements at 4AM, and from 8AM to 9PM are missing. The middle and
bottom figures show two periodic time series. The time series are parametrized
by a 24-vector x, repeated 14 times to get the full series (x, x, . . . , x). The two
estimates of x in the figure were computed by minimizing
14 X 23
!
X X
2 2 2
(xi log(c24(j1)+i )) + (xi+1 xi ) + (x1 x24 )
j=1 iMj i=1

for = 1 and = 100.

15.3.3 Image de-blurring

The vector x is an image, and the matrix A gives blurring, so y = Ax+v is a blurred,
noisy image. Our prior information about x is that it is smooth; neighboring pixels
values are not very different from each other. Estimation is the problem of guessing
what x is, and is called de-blurring.
In least squares image deblurring we form an estimate x by minimizing a cost
function of the form

kAx yk2 + (kDh xk2 + kDv xk2 ). (15.5)

Here Dv and Dh represent vertical and horizontal differencing operations, and

the role of the second term in the weighted sum is to penalize non-smoothness in
252 15 Multi-objective least squares

101

Ozone level (ppm)

102

July 1
July 2
July 3
July 4
July 5
July 6
July 7
July 8
July 9
July 10
July 11
July 12
July 13
July 14
July 15
101
Ozone level (ppm)

102
July 1
July 2
July 3
July 4
July 5
July 6
July 7
July 8
July 9
July 10
July 11
July 12
July 13
July 14

101 July 15
Ozone level (ppm)

102
July 1
July 2
July 3
July 4
July 5
July 6
July 7
July 8
July 9
July 10
July 11
July 12
July 13
July 14
July 15

Time

Figure 15.4 Top. Hourly ozone level at Azusa, California, during the first
14 days of July 2014 (California Environmental Protection Agency, Air Re-
sources Board, [Link]). Measurements start at 12AM on July 1st,
and end at 11PM on July 14. Note the large number of missing measure-
ments. In particular, all 4AM measurements are missing. Middle. Smooth
periodic least squares fit to logarithmically transformed measurements, using
= 1. Bottom. Smooth periodic least squares fit using = 100.
 15.3 Estimation and inversion 253

the reconstructed image. Specifically, suppose the vector x has length M N and
contains the pixel intensities of an M N image X stored column-wise. Let Dh
be the (N 1)M matrix

I I 0 0 0 0
0 I I 0 0 0
.. . . . . .. ,

Dh = . .. .. .. .. .

0 0 0 I I 0
0 0 0 0 I I

where all blocks have size M M , and let Dv be the (M 1)N matrix

D 0 0
0 D 0
Dv = . .. . . .. ,

.. . . .
0 0 D

where each of the N diagonal blocks D is an (M 1) M difference matrix

1 1 0 0 0 0
0 1 1 0 0 0
D = ... .. .. .. .. .. .


. . . . .

0 0 0 1 1 0
0 0 0 0 1 1

With these definitions the penalty term in (15.5) is the sum of squared differences
of intensities at adjacent pixels in a row or column:
M N
X X 1 M
X 1 X
N
kDh xk2 + kDv xk2 = (Xij Xi,j+1 )2 + (Xij Xi+1,j )2 .
i=1 j=1 i=1 j=1

This quantity is the Laplacian (see page 110), for the graph that connects each
pixel to its left and right, and up and down, neighbors.

Example. In figures 15.5 and 15.6 we illustrate this this method for an image of
size 512 512. The blurred, noisy image is shown in the left part of figure 15.5.
Figure 15.6 shows the estimates x, obtained by minimizing (15.5), for four different
values of the parameter . The best result (in this case, judged by eye) is obtained
for around 0.007 and is shown on the right in figure 15.5.

15.3.4 Tomography

In tomography, the vector x represents the values of some quantity (such as density)
in a region of interest in n voxels (or pixels) over a 3-D (or 2-D) region. The entries
of the vector y are measurements obtained by passing a beam of radiation through
254 15 Multi-objective least squares

Figure 15.5 Left: Blurred, noisy image. Right: Result of regularized least
squares deblurring with = 0.007.

Figure 15.6 Deblurred images for = 106 , 104 , 102 , 1.

15.3 Estimation and inversion 255

the region of interest, and measuring the intensity of the beam after it exits the
region.
A familiar application is the computer-aided tomography (CAT) scan used in
medicine. In this application, beams of X-rays are sent through a patient, and
an array of detectors measure the intensity of the beams after passing through
the patient. These intensity measurements are related to the integral of the X-
ray absorption along the beam. Tomography is also used in applications such as
manufacturing, to assess internal damage or certify quality of a welded joint.

Line integral measurements. For simplicity we will assume that each beam is a
single line, and that the received value yi is the integral of the quantity over the
region, plus some measurement noise. (The same method can be used when more
complex beam shapes are used.) We consider the 2-D case.
Let d(x, y) denote the density (say) at the position (x, y) in the region. (Here
x and y are the scalar 2-D coordinates, not the vectors x and y in the estimation
problem.) We assume that d(x, y) = 0 outside the region of interest. A line through
the region is defined by the set of points

p(t) = (x0 , y0 ) + t(cos , sin ),

where (x0 , y0 ) denotes a (base) point on the line, and is the angle of the line with
respect to the x-axis. The parameter t gives the distance along the line from the
point (x0 , y0 ). The line integral of d is given by
Z
d(p(t)) dt.

We assume that m lines are specified (i.e., by their base points and angles), and
the measurement yi is the line integral of d, plus some noise, which is presumed
small.
We divide the region of interest into n pixels (or voxels in the 3-D case), and
assume that the density has the constant value xi over pixel i. Figure 15.7 illus-
trates this for a simple example with n = 25 pixels. (In applications the number
of pixels or voxels is in the thousands or millions.) The line integral is then given
by the sum of xi (the density in pixel i) times the length of the intersection of the
line with pixel i. In figure 15.7, with the pixels numbered row-wise starting at the
top left corner, with width and height one, the line integral for the line shown is

1.06x16 + 0.80x17 + 0.27x12 + 1.06x13 + 1.06x14 + 0.53x15 + 0.54x10 .

The coefficient of xi is the length of the intersection of the line with pixel i.

Measurement model. We can express the vector of m line integral measurements,

without the noise, as Ax, where the m n matrix A has entries

Aij = length of line i in pixel j, i = 1, . . . , m, j = 1, . . . , n,

with Aij = 0 if line i does not intersect voxel j.

256 15 Multi-objective least squares

x1 x2

x6

(x0 , y0 )

Figure 15.7 A square region of interest divided into 25 pixels, and a line
passing through it.

Tomographic reconstruction. In tomography, estimation or inversion is often

called tomographic reconstruction or tomographic inversion.
The objective term kAx yk2 is the sum of squares of the residual between
the predicted (noise-free) line integrals Ax and the actual measured line integrals
y. Regularization terms capture prior information or assumptions about the voxel
values, for example, that they vary smoothly over the region. A simple regularizer
commonly used is the Laplacian (see page 110) associated the graph that connects
each voxel to its 6 neighbors (in the 3-D case) or its 4 neighbors (in the 2-D case).

Example. A simple 2-D example is shown in figures 15.815.10. Figure 15.8 shows
the geometry of the m = 4000 lines and the square region, shown as the square.
The square is divided into 100 100 pixels, so n = 10000.
The density of the object we are imaging is shown in figure figure 15.9. In
this object the density of each pixel is either 0 or 1 (shown as white or black,
respectively). We reconstruct or estimate the object density from the 4000 line
(noisy) line integral measurements by solving the regularized least squares problem

minimize kAx yk2 + kDxk2 ,

where kDxk2 is the sum of squares of the differences of the pixel values from their
neighbors. Figure 15.10 shows the results for six different value of . We can see
that for small the reconstruction is relatively sharp, but suffers from noise. For
large the noise in the reconstruction is smaller, but it is too smooth.
15.3 Estimation and inversion 257

Figure 15.8 The square region at the center of the picture is surrounded
by 100 points shown as circles. 40 lines (beams) emanate from each point.
(The lines are shown for two points only.) This gives a total of 4000 lines
that intersect the region.
.

Figure 15.9 Density of object used in the tomography example.

258 15 Multi-objective least squares

= 102 = 101

=1 =5

= 10 = 100

Figure 15.10 Regularized least squares reconstruction for six values of the
regularization parameter.
 15.4 Regularized data fitting 259

15.4 Regularized data fitting

We consider least squares data fitting, as described in chapter 13. In 13.2 we con-
sidered the issue of over-fitting, where the model performs poorly on new, unseen
data, which occurs when the model is too complicated for the given data set. The
remedy is to keep the model simple, e.g., by fitting with a polynomial of not too
high a degree.
Regularization is another way to avoid over-fitting, different from simply choos-
ing a model that is simple (i.e., does not have too many basis functions). Regular-
ization is also called de-tuning, shrinkage, or ridge regression, for reasons we will
explain below.

Motivation. To motivate regularization, consider the model

f(x) = 1 f1 (x) + + p fp (x). (15.6)

We can interpret i as the amount by which our prediction depends on fi (x), so

if i is large, our prediction will be very sensitive to changes or variations in the
value of fi (x), such as those we might expect in new, unseen data. This suggests
that we should prefer that i be small, so our model is not too sensitive. There
is one exception here: if fi (x) is constant (for example, the number one), then we
should not worry about the size of i , since fi (x) never varies. But we would like
all the others to be small, if possible.
This suggests the bi-criterion least squares problem with primary objective
kA yk2 , the sum of squares of the prediction errors, and secondary objective
k2:p k2 , assuming that f1 is the constant function one. Thus we should minimize

kA yk2 + k2:p k2 ,

where > 0 is called the regularization parameter.

For the regression model, this weighted objective can be expressed as

kX T + v1 yk2 + kk2 .

Here we penalize being large (because this leads to sensitivity of the model), but
not the offset v. Choosing to minimize this weighted objective is called ridge
regression.

Effect of regularization. The effect of the regularization is to accept a worse value

of sum square fit (kAyk2 ) in return for a smaller value of k2:p k2 , which measures
the size of the parameters (except 1 , which is associated with the constant basis
function). This explains the name shrinkage: The parameters are smaller than they
would be without regularization, i.e., they are shrunk. The term de-tuned suggests
that with regularization, the model is not excessively tuned to the training data
(which would lead to over-fit).

Regularization path. We get a different model for every choice of . The way
the parameters change with is called the regularization path. When p is small
260 15 Multi-objective least squares

enough (say, less than 15 or so) the parameter values can be plotted, with on the
horizontal axis. Usually only 30 or 50 values of are considered, typically spaced
logarithmically over a large range.
An appropriate value of can be chosen via out-of-sample or cross validation.
As increases, the RMS fit on the training data worsens (increases). But (as with
model order) the test set RMS prediction error typically decreases as increases,
and then, when gets too big, it increases. A good choice of regularization parame-
ter is one which approximately minimizes the test set RMS prediction error. When
a range of values of approximately minimize the RMS error, common practice
is to take the largest value of . The idea here is to use the model of minimum
sensitivity, as measured by kk2 , among those that make good predictions on the
test set.

Example. We illustrate these ideas with a small example with synthetic (simu-
lated) data. We start with a signal, shown in figure 15.11 consisting of a constant
plus four sinusoids:
X 4
s(t) = c + k cos(k t + k ),
k=1

with coefficients

c = 1.54, 1 = 0.66, 2 = 0.90, 3 = 0.66, 4 = 0.89. (15.7)

(The other parameters are 1 = 13.69, 2 = 3.55, 3 = 23.25, 4 = 6.03, and

1 = 0.21, 2 = 0.02, 3 = 1.87, 4 = 1.72.) We will fit a model of the
form (15.6) with p = 5 and

f1 (x) = 1, fk+1 (x) = cos(k x + k ), k = 1, . . . , 4.

(Note that the model is exact when the parameters are chosen as 1 = c, k = k1 ,
k = 2, . . . , 5. This rarely occurs in practice.) We fit our model using regularized
least squares on 10 noisy samples of the function, shown as the blue circles in
figure 15.11. We will test the model obtained on the 20 noisy data points shown
as the red circles in figure 15.11.
Figure 15.12 shows the regularization path and the RMS training and test errors
as functions of the regularization parameter , as it varies over a large range. The
regularization path shows that as increases, the parameters 2 , . . . , 5 get smaller
(i.e., shrink), converging towards zero as gets very large. We can see that the
training prediction error increases with increasing (since we are trading off model
sensitivity for sum square fitting error). The test error follows a typical pattern:
It first decreases to a minimum value, then increases again. The minimum test
error occurs around = 0.079; any choice between around = 0.065 and 0.100
(say) would be reasonable. The horizontal dashed lines show the true values of
the coefficients (i.e., the ones we used to synthesize the data) given in (15.7). We
can see that for near 0.079, our estimated parameters are very close to the true
values.
15.4 Regularized data fitting 261

s(t) 2

Train
1 Test

0 0.2 0.4 0.6 0.8 1

t

Figure 15.11 A signal s(t) and 30 noisy samples. Ten of the samples are
used as training set, the other 20 as test set.

Linear independence of columns. One side benefit of adding regularization to

the basic least squares data fitting method is that the columns of the associated
stacked matrix are always linearly independent, even if the columns of the matrix
A are not. To see this, suppose that
 
A x = 0,
B

where B is the (p 1) p selector matrix

B = (eT2 , . . . , eTp ),

so B = 2:p . From the last p 1 entries in the equation above, we get xi = 0
for i = 2, . . . , p, which implies that x2 = = xp = 0. Using these values of x2 ,
. . . , xp , and the fact that the first column of A is 1, the top m equations become
1x1 = 0, and we conclude that x1 = 0 as well. So the columns of the stacked
matrix are always linearly independent.

Feature engineering and regularized least squares. The simplest method of

avoiding over-fit is to keep the model simple, which usually means that we should
not use too many features. A typical and rough rule of thumb is that the number
of features should be small compared to the number of data points (say, no more
than 10% or 20%). The presence of over-fit can be detected using out-of-sample or
cross-validation, which is always done when you fit a model to data.
Regularization is a powerful alternative method to avoid over-fitting a model.
With regularization, you can fit a model with more features than would be appro-
priate without regularization. You can even fit a model using more features than
262 15 Multi-objective least squares

1.2
Train
1 Test

0.8
RMS error

0.6

0.4

0.2

0
105 103 101 101 103 105

2 1 2
3 4
5
1
Coefficients

105 103 101 101 103 105

Figure 15.12 Top. RMS training and test errors as a function of the reg-
ularization parameter . Bottom. The regularization path. The dashed
horizontal lines show the values of the coefficients used to generate the data.
 15.5 Complexity 263

you have data points, in which case the matrix A is wide. Regularization is often
the key to success in feature engineering, which can greatly increase the number of
features.

15.5 Complexity
In the general case we can minimize the weighted sum objective (15.1) by creating
the stacked matrix and vector A and b in (15.2), and then using the QR factorization
to solve the resulting least squares problem. The computational cost of this method
is order mn2 flops, where m = m1 + + mk is the sum of heights of the matrices
A1 , . . . , Ak .
When using multi-objective least squares, it is common to minimize the weighted
sum objective for some, or even many, different choices of weights. Assuming that
the weighted sum objective is minimized for L different values of the weights, the
total computational cost is order Lmn2 flops.

15.5.1 Gram caching

We start from the formula (15.3) for the minimizer of the weighted sum objective,

x = (1 AT1 A1 + + k ATk Ak )1 (1 AT1 b1 + + k ATk bk ).

The matrix appearing in the inverse is a weighted sum of the Gram matrices Gi =
ATi Ai associated with the matrices Ai . We can compute x by forming these Gram
matrices Gi , along with the vectors hi = ATi bi , then forming the weighted sums

G = 1 G1 + + k Gk , h = 1 h1 + + k hk ,

and finally, solving the n n set of equations Gx = h. Forming Gi and hi for

i = 1, . . . , m costs order mn2 flops. Forming G and h costs order kn2 flops. Solving
Gx = h costs order n3 flops.
Gram caching is the simple trick of computing Gi (and hi ) just once, and re-
using these matrices and vectors for the L different choices of weights. This leads
to a computational cost of order

mn2 + L(k + n)n2

flops. When m is much larger than k + n, which is a common occurrence, this cost
is smaller than Lmn2 , the cost for the simple method.
As a simple example consider Tychonov regularization. We will compute

x(i) = (AT A + (i) I)1 AT b

for i = 1, . . . , L, where A is m n. The cost of the simple method is 2Lmn2 flops;

using Gram caching the cost is order 2(m + Ln)n2 flops. With m = 100n and
L = 100, Gram caching reduces the computational cost by around a factor of 50.
264 15 Multi-objective least squares

15.5.2 The kernel trick

In this section we focus on another special case, which arises in many applications:

J = kAx bk2 + kx xdes k2 ,

where the m n matrix A is wide, i.e., m < n, and > 0. (Here we drop the
subscript on A, b, and m since we have only one matrix in this problem.) The
associated (m + n) n stacked matrix (see (15.2))
 
A
A =
I
always has linearly independent columns. Using the QR factorization to solve the
stacked least squares problem requires order (m + n)n2 flops, which grows like n3 .
We will show now how this special problem can be solved far more efficiently when
m is much smaller than n, using something called the kernel trick. Recall that the
minimizer of J is given by (see (15.3))

x = (AT A + I)1 (AT b + xdes )

= (AT A + I)1 (AT b + (I + AT A)xdes (AT A)xdes )
= (AT A + I)1 AT (b Axdes ) + xdes .

The matrix inverse here has size n n.

We will use the identity

(AT A + I)1 AT = AT (AAT + I)1 (15.8)

which holds for any matrix A and any > 0. Note that the left-hand side of
the identity involves the inverse of an n n matrix, whereas the right-hand side
involves the inverse of a (smaller) m m matrix.
To show the identity (15.8), we first observe that the matrices AT A + I and
T
AA + I are invertible. We start with the equation

AT (AAT + I) = (AT A + I)AT ,

and multiply each side by (AT A+I)1 on the left and (AAT +I)1 on the right,
which yields the identity above.
Using (15.8) we can express the minimizer of J as

x = AT (AAT + I)1 (b Axdes ) + xdes .

We can compute the term (AAT + I)1 (b Axdes ) by computing the QR factor-
ization of the (m + n) m matrix

AT
 
A = ,
I

which has a cost of (m + n)m2 flops. The other operations involve matrix-vector
products and have order (at most) mn flops, so we can use this method to compute
x in order (m + n)m2 flops. This complexity grows only linearly in n.
 Chapter 16

Constrained least squares

In this chapter we discuss a useful extension of the least squares problem that
includes linear equality constraints. Like least squares, the constrained least squares
problem can be reduced to a set of linear equations, which can be solved using the
QR factorization.

16.1 Constrained least squares problem

In the basic least squares problem, we seek x that minimizes the objective function
kAx bk2 . We now add constraints to this problem, by insisting that x satisfy
the linear equations Cx = d, where the matrix C and the vector d are given.
The linearly constrained least squares problem (or just constrained least squares
problem) is written as
minimize kAx bk2
(16.1)
subject to Cx = d.
Here x, the variable to be found, is an n-vector. The problem data (which are
given) are the m n matrix A, the m-vector b, the p n matrix C, and the
p-vector d.
We refer to the function kAx bk2 as the objective of the problem, and the set
of p linear equality constraints Cx = d as the constraints of the problem. They
can be written out as p scalar constraints (equations)
cTi x = di , i = 1, . . . , p,
where cTi is the ith row of C.
An n-vector x is called feasible (for the problem (16.1)) if it satisfies the con-
straints, i.e., Cx = d. An n-vector x is called an optimal point or solution of the
optimization problem (16.1) if it is feasible, and if kAx bk2 kAx bk2 holds for
any feasible x. In other words, x solves the problem (16.1) if it is feasible and has
the smallest possible value of the objective function among all feasible vectors.
The constrained least squares problem combines the problems of solving a set of
linear equations (find x that satisfies Cx = d) with the least squares problem (find
266 16 Constrained least squares

f(x)

q(x)

p(x)

x
a

Figure 16.1 Least squares fit of two cubic polynomials to 140 points, with
continuity constraints p(a) = q(a) and p0 (a) = q 0 (a).

x that minimizes kAx bk2 ). Indeed each of these problems can be considered a
special case of the constrained least squares problem (16.1).
The constrained least squares problem can also be thought of as a limit of a bi-
objective least squares problem, with primary objective kAx bk2 and secondary
objective kCx dk2 . Roughly speaking, we put infinite weight on the second
objective, so that any nonzero value is unacceptable (which forces x to satisfy
Cx = d). So we would expect (and it can be verified) that minimizing the weighted
objective
kAx bk2 + kCx dk2 ,
for a very large value of yields a vector close to a solution of the constrained least
squares problem (16.1). We will encounter this idea again in chapter 19, when we
consider the nonlinear constrained least squares problem.

Example. In figure 16.1 we fit a piecewise-polynomial function f(x) to a set of

N = 140 points (xi , yi ) points in the plane. The function f(x) is defined as

p(x) x a
f(x) =
q(x) x > a,

with a given, and p(x) and q(x) polynomials of degree three or less,

p(x) = 1 + 2 x + 3 x2 + 4 x3 , q(x) = 5 + 6 x + 7 x2 + 8 x3 .

We also impose the condition that p(a) = q(a) and p0 (a) = q 0 (a), so that f(x) is
continuous and has a continuous first derivative at x = a. Suppose the N data
16.1 Constrained least squares problem 267

points (xi , yi ) are numbered so that x1 , . . . , xM a and xM +1 , . . . , xN > a. The

sum of squares of the prediction errors is
M
X N
X
(1 + 2 xi + 3 x2i + 4 x3i yi )2 + (5 + 6 xi + 7 x2i + 8 x3i yi )2 .
i=1 i=M +1

The conditions p(a) q(a) = 0 and p0 (a) q 0 (a) = 0 are two linear equations
1 + 2 a + 3 a2 + 4 a3 5 6 a 7 a2 8 a3 = 0
2 + 23 a + 34 a2 6 27 a 38 a2 = 0.
We can determine the coefficients = (1 , . . . , 10 ) that minimize the sum of squares
of the prediction errors, subject to the continuity constraints, by solving a con-
strained least squares problem
minimize kA bk2
subject to C = d.
The matrices and vectors A, b, C, d are defined as
x21 x31

1 x1 0 0 0 0 y1
1 x2 x22 x32 0 0 0 0 y2
.. .. .. .. .. .. .. .. ..


. . . . . . . . .
x2M x3M

1 xM 0 0 0 0 yM
A= , b= ,
0 0 0 0 1 xM +1 x2M +1 x3M +1 yM +1
x2M +2 x3M +2

0 0 0 0 1 xM +2 yM +2
. .. .. .. .. .. .. .. ..

..

. . . . . . . .
0 0 0 0 1 xN x2N x3N yN
and
a2 a3 a2 a3
   
1 a 1 a 0
C= , d= .
0 1 2a 3a2 0 1 2a 3a2 0
This method is easily extended to piecewise-polynomial functions with more than
two intervals. Functions of this kind are called splines.

Advertising budget allocation. We continue the example described on page 184,

where the goal is to purchase advertising in n different channels so as to achieve
(or approximately achieve) a target set of customer views or impressions in m
different demographic groups. We denote the n-vector of channel spending as s;
this spending results in a set of views (across the demographic groups) given by the
m-vector Rs. We will minimize the sum of squares of the deviation from the target
set of views, given by v des . In addition, we fix our total advertising spending, with
the constraint 1T s = B, where B is a given total advertising budget. (This can
also be described as allocating a total budget B across the n different channels.)
This leads to the constrained least squares problem
minimize kRs v des k2
subject to 1T s = B.
268 16 Constrained least squares

Optimal
Scaled

1,000

Impressions
500

0
1 2 3 4 5 6 7 8 9 10
Group

Figure 16.2 Advertising with budget constraint. The optimal views vector
is the solution of the constrained least squares problem with budget con-
straint. The scaled views vector is obtained by scaling the unconstrained
least squares solution so that it satisfies the budget constraint. Hence this
is a scalar multiple of the views vector of figure 12.3.

(The solution s of this problem is not guaranteed to have nonnegative entries, as it

must to make sense in this application. But we ignore this aspect of the problem
here.)
We consider the same problem instance as on page 184, with m = 10 demo-
graphic groups and n = 3 channels, and reach matrix R given there. The least
squares method yields an RMS error of 133 (around 13.3%), with a total budget
of 1T sls = 1605. We seek a spending plan with a budget that is 20% smaller,
B = 1284. Solving the associated constrained least squares problem yields the
spending vector scls = (315, 110, 859), which has RMS error of 161 in the target
views. We can compare this spending vector to the one obtained by simply scaling
the least squares spending vector by 0.80. The RMS error for this allocation is 239.
The resulting impressions for both spending plans are shown in figure 16.2.

16.1.1 Least norm problem

An important special case of the constrained least squares problem (16.1) is when
A = I and b = 0:
minimize kxk2
(16.2)
subject to Cx = d.
In this problem we seek the vector of smallest or least norm that satisfies the linear
equations Cx = d. For this reason the problem (16.2) is called the least norm
problem or minimum-norm problem.
16.1 Constrained least squares problem 269

1 1

Position
Force

0 0.5

1 0
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

Figure 16.3 Left: A force sequence f bb = (1, 1, 0, . . . , 0) that transfers the

mass over a unit distance in 10 seconds. Right: The resulting position of
the mass p(t).

Example. The 10-vector f represents a series of forces applied, each for one sec-
ond, to a unit mass on a surface with no friction. The mass starts with zero velocity
and position. By Newtons laws, its final velocity and position are given by

v fin = f1 + f2 + + f10
fin
p = (19/2)f1 + (17/2)f2 + + (1/2)f10 .

Now suppose we want to choose a force sequence that results in v fin = 0, pfin = 1,
i.e., a force sequence that moves the mass to a resting position one meter to the
right. There are many such force sequences; for example f bb = (1, 1, 0, . . . , 0).
This force sequence accelerates the mass to velocity 0.5 after one second, then
decelerates it over the next second, so it arrives after two seconds with velocity 0,
at the destination position 1. After that it applies zero force, so the mass stays
where it is, at rest at position 1. The superscript bb refers to bang-bang, which
means that the force applies a large force to get the mass moving (the first bang)
and another large force (the second bang) to slow it to zero velocity. The force and
position versus time for this choice of f is shown in figure 16.3.
Now we ask, what is the smallest force sequence that can achieve v fin = 0,
fin
p = 1, where smallest is measured by the sum of squares of the applied forces,
kf k2 = f12 + + f10
2
? This problem can be posed as a least norm problem,

2
minimize kf
 k   
1 1 1 1 0
subject to f= ,
19/2 17/2 3/2 1/2 1

with variable f . The solution f ln , and the resulting position, are shown in fig-
ure 16.4. The norm square of the least norm solution f ln is 0.0121; in contrast,
the norm square of the bang-bang force sequence is 2, a factor of 165 times larger.
(Note the very different vertical axis scales in figures 16.4 and 16.3.)
270 16 Constrained least squares

0.05 1

Position
Force 0 0.5

0.05 0
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

Figure 16.4 Left: The smallest force sequence f ln that transfers the mass
over a unit distance in 10 steps. Right: The resulting position of the mass
p(t).

16.2 Solution
Optimality conditions via Lagrange multipliers. We will use the method of La-
grange multipliers (see C.3) to solve the constrained least squares problem (16.1).
Later we give an independent verification, that does not rely on calculus or La-
grange multipliers, that the solution we derive is correct.
We first write the contrained least squares problem with the constraints given
as a list of p scalar equality constraints:

minimize kAx bk2

subject to cTi x = di , i = 1, . . . , p,

where cTi are the rows of C. We form the Lagrangian function

L(x, z) = kAx bk2 + z1 (cT1 x d1 ) + + zp (cTp x dp ),

where z is the p-vector of Lagrange multipliers. The method of Lagrange multipliers

tells us that if x is a solution of the constrained least squares problem, then there
is a set of Lagrange multipliers z that satisfy
L L
(x, z) = 0, i = 1, . . . , n, (x, z) = 0, i = 1, . . . , p. (16.3)
xi zi
These are the optimality conditions for the constrained least squares problem. Any
solution of the constrained least squares problem must satisfy them. We will now
see that the optimality conditions can be expressed as a set of linear equations.
The second set of equations in the optimality conditions can be written as
L
(x, z) = cTi x di = 0, i = 1, . . . , p,
zi
which states that x satisfies the equality constraints C x = d (which we already
knew). The first set of equations, however, is more informative. Expanding the
16.2 Solution 271

objective kAx bk2 as a sum of terms involving the entries of x (as was done on
page 181) and taking the partial derivative of L with respect to xi we obtain
n p
L X X
(x, z) = 2 (AT A)ij xj 2(AT b)i + zj ci = 0.
xi j=1 j=1

We can write these equations in compact matrix-vector form as

2(AT A)x 2AT b + C T z = 0.
Combining this set of linear equations with the feasibility conditions C x = d,
we can write the optimality conditions (16.3) as one set of n + p linear equations
in the variables (x, z):
2AT A C T 2AT b
    
x
= . (16.4)
C 0 z d
These equations are called the KKT equations for the constrained least squares
problem. (KKT stands for Karush, Kuhn, and Tucker, the names of three re-
searchers who derived the optimality conditions for a much more general form of
constrained optimization problem.) The KKT equations (16.4) are an extension
of the normal equations (12.4) for a least squares problem with no constraints. So
we have reduced the constrained least squares problem to the problem of solving a
(square) set of n + p linear equations in n + p variables (x, z).

Invertibility of KKT matrix. The (n + p) (n + p) coefficient matrix in (16.4) is

called the KKT matrix. It is invertible if and only if
 
A
C has linearly independent rows, and has linearly independent columns.
C
(16.5)
The first condition requires that C is wide (or square), i.e., that there are fewer
constraints than variables. The second condition depends on both A and C, and it
can be satisfied even when the columns of A are dependent. The condition (16.5)
is the generalization of our assumption (12.2) for unconstrained least squares (i.e.,
that A has linearly independent columns).
Before proceeding, let us verify that the KKT matrix is invertible if and only
if (16.5) holds. First suppose that the KKT matrix is not invertible. This means
that there is a nonzero vector (x, z) with
2AT A C T
  
x
= 0.
C 0 z

Multiply the top block equation 2AT Ax + C T z = 0 on the left by xT to get

2kAxk2 + xT C T z = 0.
The second block equation, C x = 0, implies (by taking the transpose) xT C T = 0,
so the equation above becomes 2kAxk2 = 0, i.e., Ax = 0. We also have C x = 0, so
 
A
x = 0.
C
272 16 Constrained least squares

Since the matrix on the left has linearly independent columns (by assumption),
we conclude that x = 0. The first block equation above then becomes C T z = 0.
But by our assumption that the columns of C T are linearly independent, we have
z = 0. So (x, z) = 0, which is a contradiction.
The converse is also true. First suppose that the rows of C are dependent.
Then there is a nonzero vector z with C T z = 0. Then
2AT A C T
  
0
= 0,
C 0 z
which shows the KKT matrix is not invertible. Now suppose that the stacked
matrix in (16.5) has dependent columns, which means there is a nonzero vector x
for which  
A
x = 0.
C
Direct calculation shows that
2AT A CT
  
x
= 0,
C 0 0
which shows that the KKT matrix is not invertible.
When the conditions (16.5) hold, the constrained least squares problem (16.1)
has the (unique) solution x, given by
1 
2AT A C T 2AT b
   
x
= . (16.6)
z C 0 d

(This formula also gives us z, the set of Lagrange multipliers.) From (16.6), we
observe that the solution x is a linear function of (b, d).

Direct verification of constrained least squares solution. We will now show

directly, without using calculus, that the solution x given in (16.6) is the unique
vector that minimizes kAx bk2 over all x that satisfy the constraints Cx = d,
when the conditions (16.5) hold. Let x and z denote the vectors given in (16.6), so
they satisfy
2AT Ax + C T z = 2AT b, C x = d.
Suppose that x 6= x is any vector that satisfies Cx = d. We will show that
kAx bk2 > kAx bk2 .
We proceed in the same way as for the least squares problem:

kAx bk2 = k(Ax Ax) + (Ax b)k2

= kAx Axk2 + kAx bk2 + 2(Ax Ax)T (Ax b).

Now we expand the last term:

2(Ax Ax)T (Ax b) = 2(x x)T AT (Ax b)

= (x x)T C T z
= (C(x x))T z
= 0,
 16.3 Solving constrained least squares problems 273

where we use 2AT (Ax b) = C T z in the second line and Cx = C x = d in the

last line. So we have, exactly as in the case of unconstrained least squares,

kAx bk2 = kA(x x)k2 + kAx bk2 ,

from which we conclude that kAx bk2 kAx bk2 . So x minimizes kAx bk2
subject to Cx = d.
It remains to show that for x 6= x, we have the strict inequality kAx bk2 >
kAx bk2 , which by the equation above is equivalent to kA(x x)k2 > 0. If this
is not the case, then A(x x) = 0. We also have C(x x) = 0, and so
 
A
(x x) = 0.
C

By our assumption that the matrix on the left has linearly independent columns,
we conclude that x = x.

16.3 Solving constrained least squares problems

We can compute the solution (16.6) of the constrained least squares problem by
forming and solving the KKT equations (16.4).

Algorithm 16.1 Constrained least squares via KKT equations

given an m n matrix A and a p n matrix C that satisfy (16.5), an m-vector b,
and a p-vector d.

1. Form Gram matrix. Compute AT A.

2. Solve KKT equations. Solve KKT equations (16.4) by QR factorization and
back substitution.

The second step cannot fail, provided the assumption (16.5) holds. Let us
analyze the complexity of this algorithm. The first step is multiplying an n m
matrix by an m n matrix, which requires 2mn2 flops. (In fact we can get away
with half this number, since the Gram matrix is symmetric, and we only have to
compute the entries on and above the diagonal.) The second step requires the
solution of a square system of n + p equations, which costs 2(n + p)3 flops, so the
total is
2mn2 + 2(n + p)3
flops. This grows linearly in m and cubicly in n and p. The assumption (16.5)
implies p n, so in terms of order, (n + p)3 can be replaced with n3 .

Solving constrained least squares problems via QR factorization. We now give

a method for solving the constrained least squares problem that generalizes the QR
factorization method for least squares problems (algorithm 12.1). We assume that
A and C satisfy the conditions (16.5).
274 16 Constrained least squares

We start by rewriting the KKT equations (16.4) as

2(AT A + C T C)x + C T w = 2AT b, C x = d (16.7)

with a new variable w = z 2d. To obtain (16.7) we multiplied the equation

C x = d on the left by 2C T , then added the result to the first equation of (16.4),
and replaced the variable z with w + 2d.
Next we use the QR factorization
   
A Q1
= QR = R (16.8)
C Q2

to simplify (16.7). This factorization exists because the stacked matrix has linearly
independent columns, by our assumption (16.5). In (16.8) we also partition Q in
two blocks Q1 and Q2 , of size m n and p n, respectively. If we make the
substitutions A = Q1 R, C = Q2 R, and AT A + C T C = RT R in (16.7) we obtain

2RT Rx + RT QT2 w = 2RT QT1 b, Q2 Rx = d.

We multiply the first equation on the left by RT (which we know exists) to get

Rx = QT1 b (1/2)QT2 w. (16.9)

Substituting this expression into Q2 Rx = d gives an equation in w:

Q2 QT2 w = 2Q2 QT1 b 2d. (16.10)

We now use the second part of the assumption (16.5) to show that the matrix
QT2 = RT C T has linearly independent columns. Suppose QT2 z = RT C T z = 0.
Multiplying with RT gives C T z = 0. Since C has linearly independent rows, this
implies z = 0, and we conclude that the columns of QT2 are linearly independent.
The matrix QT2 therefore has a QR factorization QT2 = QR. Substituting this
into (16.10) gives
RT Rw = 2RT QT QT1 b 2d,
which we can write as
Rw = 2QT QT1 b 2RT d.
We can use this to compute w, first by computing RT d (by forward substitution),
then forming the right-hand side, and then solving for w using back substitution.
Once we know w, we can find x from (16.9). The method is summarized in the
following algorithm.

Algorithm 16.2 Constrained least squares via QR factorization

given an m n matrix A and a p n matrix C that satisfy (16.5), an m-vector b,
and a p-vector d.

1. QR factorizations. Compute the QR factorizations

   
A Q1
= R, QT2 = QR.
C Q2
16.3 Solving constrained least squares problems 275

2. Compute RT d by forward substitution.

3. Form right-hand side and solve

Rw = 2QT QT1 b 2RT d

via back substitution.

4. Compute x. Form right-hand side and solve

Rx = QT1 b (1/2)QT2 w

by back substitution.

In the unconstrained case (when p = 0), step 1 reduces to computing the QR

factorization of A, steps 2 and 3 are not needed, and step 4 reduces to solving
Rx = QT1 b. This is the same as algorithm 12.1 for solving (unconstrained) least
squares problems.
We now give a complexity analysis. Step 1 involves the QR factorizations of
an (m + p) n and an n p matrix, which costs 2(m + p)n2 + 2np2 flops. Step 2
requires p2 flops. In step 3, we first evaluate QT1 b (2mn flops), multiply the result
by QT (2pn flops), and then solve for w using forward substitution (p2 flops). Step 4
requires 2mn + 2pn flops to form the right-hand side, and n2 flops to compute x via
back substitution. The costs of steps 2, 3, and 4 are quadratic in the dimensions,
and so are negligible compared to the cost of step 1, so our final complexity is

2(m + p)n2 + 2np2

flops. The assumption (16.5) implies the inequalities

p n m + p,

and therefore (m + p)n2 np2 . So the flop count above is no more than 3(m + p)n2
flops. In particular, its order is (m + p)n2 .

Sparse constrained least squares. Constrained least squares problems with sparse
matrices A an C arise in many applications; we will see several examples in the
next chapter. Just as for solving linear equations, or (unconstrained) least squares
problems, there are methods that exploit the sparsity in A and C to solve con-
strained least squares problems more efficiently than the generic algorithms 16.1
or 16.2. The simplest such methods follow these basic algorithms, replacing the
QR factorizations with sparse QR factorizations (see page 156).
One potential problem with forming the KKT matrix as in algorithm 16.1 is
that the Gram matrix AT A can be far less sparse than the matrix A. This problem
can be avoided using a trick analogous to the one used on page 184 to solve sparse
(unconstrained) least squares problems. We form the square set of m + n + p linear
equations
AT CT

0 x 0
A (1/2)I 0 y = b .
C 0 0 z d
276 16 Constrained least squares

If (x, y, z) satisfies these equations, it is easy to see that (x, z) satisfies the KKT
equations (16.4); conversely, if (x, z) satisfies the KKT equations (16.4), (x, y, z)
satisfies the equations above, with y = 2(Ax b). Provided A and C are sparse,
the coefficient matrix above is sparse, and any method for solving a sparse system
of linear equations can be used to solve it.

Solution of least norm problem. Here we specialize the solution of the general
constrained least squares problem (16.1) given above to the special case of the least
norm problem (16.2).
We start with the conditions (16.5). The stacked matrix is in this case
 
I
,
C
which always has linearly independent columns. So the conditions (16.5) reduce
to: C has linearly independent rows. We make this assumption now.
For the least norm problem, the KKT equations (16.4) reduce to
2I C T
    
x 0
= .
C 0 z d
We can solve this using the methods for general constrained least squares, or derive
the solution directly, which we do now. The first block row of this equation is
2x + C T z = 0, so
x = (1/2)C T z.
We substitute this into the second block equation, C x = d, to obtain
(1/2)CC T z = d.
Since the rows of C are linearly independent, CC T is invertible, so we have
z = 2(CC T )1 d.
Substituting this expression for z into the formula for x above gives
x = C T (CC T )1 d. (16.11)
We have seen the matrix in this formula before: It is the pseudo-inverse of a wide
matrix with linearly independent rows. So we can express the solution of the least
norm problem (16.2) in the very compact form
x = C d.
In 11.5, we saw that C is a right inverse of C; here we see that not only does
x = C d satisfy Cx = d, but it gives the vector of least norm that satisfies Cx = d.
In 11.5, we also saw that the pseudo-inverse of C can be expressed as C =
QRT , where C T = QR is the QR factorization of C T . The solution of the least
norm problem can therefore be expressed as
x = QRT d
and this leads to an algorithm for solving the least norm problem via the QR
factorization.
16.3 Solving constrained least squares problems 277

Algorithm 16.3 Least norm via QR factorization

given a p n matrix C with linearly independent rows and a p-vector d.

1. QR factorization. Compute the QR factorization C T = QR.

2. Compute x. Solve RT y = d by forward substitution.
3. Compute x = Qy.

The complexity of this algorithm is dominated by the cost of the QR factoriza-

tion in step 1, i.e., 2np2 flops.
278 16 Constrained least squares
 Chapter 17

Constrained least squares

applications

In this chapter we discuss several applications of equality constrained least squares.

17.1 Portfolio optimization

In portfolio optimization (also known as portfolio selection), we invest in different
assets, typically stocks, over some investment periods. The goal is to make invest-
ments so that the combined return on all our investments is consistently high. (We
must accept the idea that for our average return to be high, we must tolerate some
variation in the return, i.e., some risk.) The idea of optimizing a portfolio of assets
was proposed in 1953 by Harry Markowitz, who won the Nobel prize in Economics
for this work in 1990. In this section we will show that a version of this problem
can be formulated and solved as a linearly-constrained least squares problem.

17.1.1 Portfolio risk and return

Portfolio allocation weights. We allocate a total amount of money to be invested

in n different assets. The allocation across the n assets is described by an allocation
n-vector w, which satisfies 1T w = 1, i.e., its entries sum to one. If a total (dollar)
amount V is to be invested in some period, then V wj is the amount invested in
asset j. (This can be negative, meaning a short position of |V wj | dollars on asset
j.) The entries of w are called by various names including fractional allocations,
asset weights, asset allocations, or just weights.
For example, the asset allocation w = ej means that we invest everything in
asset j. (In this way, we can think of the individual assets as simple portfolios.)
The asset allocation w = (0.2, 0.0, 1.2) means that we take a short position in
asset 1 of one fifth of the total amount invested, and put the cash derived from the
280 17 Constrained least squares applications

short position plus our initial amount to be invested into asset 3. We do not invest
in asset 2 at all.
The leverage L of the portfolio is given by

L = |w1 | + + |wn |,

the sum of the absolute values of the weights. If all entries of w are nonnegative
(which is a called a long-only portfolio), we have L = 1; if some entries are negative,
then L > 1. If a portfolio has a leverage of 5, it means that for every $1 of portfolio
value, we have $3 of total long holdings, and $2 of total short holdings. (Other
definitions of leverage are used, for example, (L 1)/2.)

Multi-period investing with allocation weights. The investments are held for T
periods of, say, one day each. (The periods could just as well be hours, weeks, or
months). We describe the investment returns by the T n matrix R, where Rtj
is the fractional return of asset j in period t. Thus R61 = 0.02 means that asset 1
gained 2% in period 6, and R82 = 0.03 means that asset 2 lost 3%, over period
8. The jth column of R is the return time series for asset j; the tth row of R gives
the returns of all assets in period t. It is often assumed that one of the assets is
cash, which has a constant (positive) return rf , where the superscript stands for
risk-free. If the risk-free asset is asset n, then the last column of R is rf 1.
Suppose we invest a total (positive) amount Vt at the beginning of period t,
so we invest Vt wj in asset j. At the end of period t, the dollar value of asset j is
Vt wj (1 + Rtj ), and the dollar value of the whole portfolio is
n
X
Vt+1 = Vt wj (1 + Rtj ) = Vt (1 + rtT w),
j=1

where rtT is the tth row of R. We assume Vt+1 is positive; if the total portfolio
value becomes negative we say that the portfolio has gone bust and stop trading.
The total (fractional) return of the portfolio over period t, i.e., its fractional
increase in value, is

Vt+1 Vt Vt (1 + rtT w) Vt
= = rtT w.
Vt Vt
Note that we invest the total portfolio value in each period according to the weights
w. This entails buying and selling assets so that the dollar value fractions are once
again given by w. This is called re-balancing the portfolio.
The portfolio return in each of the T periods can be expressed compactly using
matrix-vector notation as
r = Rw,
where r is the T -vector of portfolio returns in the T periods, i.e., the time series
of portfolio returns. (Note that r is a T -vector, which represents the time series
of total portfolio return, whereas rt is an n-vector, which gives the returns of the
n assets in period t.) If asset n is risk-free, and we choose the allocation w = en ,
then r = Ren = rf 1, i.e., we obtain a constant return in each period of rf .
17.1 Portfolio optimization 281

We can express the total portfolio value in period t as

Vt = V1 (1 + r1 )(1 + r2 ) (1 + rt1 ), (17.1)

where V1 is the total amount initially invested in period t = 1. This total value time
series is often plotted using V1 = $10000 as the initial investment by convention.
The product in (17.1) arises from re-investing our total portfolio value (including
any past gains or losses) in each period. In the simple case when the last asset is
risk-free and we choose w = en , the total value grows as Vt = V1 (1 + rf )t1 . This
is called compounded interest at rate rf .
When the returns rt are small (say, a few percent), and T is not too big (say, a
few hundred), we can approximate the product above using the sum or average of
the returns. To do this we expand the product in (17.1) into a sum of terms, each
of which involves a product of some of the returns. One term involves none of the
returns, and is V1 . There are t 1 terms that involve just one return, which have
the form V1 rs , for s = 1, . . . , t 1. All other terms in the expanded product involve
the product of at least two returns, and so can be neglected since we assume that
the returns are small. This leads to the approximation

Vt V1 + V1 (r1 + + rt1 ),

which for t = T + 1 can be written as

VT +1 V1 + T avg(r)V1 .

This approximation suggests that to maximize our total final portfolio value, we
should seek high return, i.e., a large value for avg(r).

Portfolio return and risk. The choice of weight vector w is judged by the result-
ing portfolio return time series r = Rw. The portfolio mean return (over the T
periods), often shortened to just the return, is given by avg(r). The portfolio risk
(over the T periods) is the standard deviation of portfolio return, std(r).
The quantities avg(r) and std(r) give the per-period return and risk. They
are often converted to their equivalent values for one year, which are called the
annualized return and risk, and reported as percentages. If there are P periods in
one year, these are given by

P avg(r), P std(r),

respectively. For example, suppose each period is one (trading) day. There are
about 250 trading days in one year, so the annualized return and risk are given
by 250 avg(r) and 15.81 std(r). Thus a daily return sequence r with per-period
(daily) return 0.05% (0.0005) and risk 0.5% (0.005) has an annualized return and
risk of 12.5% and 7.9%, respectively. (The squareroot of P in the risk annualization
comes from the assumption that the fluctuations in the returns vary randomly and
independently from period to period.)
282 17 Constrained least squares applications

17.1.2 Portfolio optimization

We want to choose w so that we achieve high return and low risk. This means
that we seek portfolio returns rt that are consistently high. This is an optimization
problem with two objectives, return and risk. Since there are two objectives, there
is a family of solutions, that trade off return and risk. For example, when the
last asset is risk-free, the portfolio weight w = en achieves zero risk (which is the
smallest possible value), and return rf . We will see that other choices of w will lead
to higher return, but with higher risk as well. Portfolio weights that minimize risk
for a given level of return (or maximize return for a given level of risk) are called
Pareto optimal. The risk and return of this family of weights are typically plotted
on a risk-return plot, with risk on the horizontal axis and return on the vertical
axis. Individual assets can be considered (very simple) portfolios, corresponding
to w = ej . In this case the corresponding portfolio return and risk are simply the
return and risk of asset j (over the same T periods).
One approach is to fix the return of the portfolio to be some given value , and
minimize the risk over all portfolios that achieve the required return. Doing this
for many values of produces (different) portfolio allocation vectors that trade off
risk and return. Requiring that the portfolio return be can be expressed as

avg(r) = (1/T )1T (Rw) = T w = ,

where = RT 1/T is the n-vector of the average asset returns. This is a single
linear equation in w. Assuming that it holds, we can express the square of the risk
as
std(r)2 = (1/T )kr avg(r)1k2 = (1/T )kr 1k2 .
Thus to minimize risk (squared), with return value , we must solve the linearly
constrained least squares problem
2
minimize kRw
 T 1k  
1 1 (17.2)
subject to w= .
T

(We dropped the factor 1/T from the objective, which does not affect the solution.)
This is a constrained least squares problem with two linear equality constraints.
The first constraint sets the sum of the allocation weights to one, and the second
requires that the mean portfolio return is .
The portfolio optimization problem has solution
1
2RT R

w 1 2T
z1 = 1T 0 0 1 , (17.3)
z2 T 0 0

where z1 and z2 are Lagrange multipliers for the equality constraints (which we
dont care about).
As a historical note, the portfolio optimization problem (17.2) is not exactly the
same as the one proposed by Markowitz. His formulation used a statistical model
of returns, where instead we are using a set of actual (or realized ) returns.
17.1 Portfolio optimization 283

Future returns and the big assumption. The portfolio optimization problem (17.2)
suffers from what would appear to be a serious conceptual flaw: It requires us to
know the asset returns over the periods t = 1, . . . , T , in order to compute the op-
timal allocation to use over those periods. This is silly: If we knew any future
returns, we would be able to achieve as large a portfolio return as we like, by sim-
ply putting large positive weights on the assets with positive returns and negative
weights on those with negative returns. The whole challenge in investing is that
we do not know future returns.
Assume the current time is period T , so we know the (so-called realized ) return
matrix R. The portfolio weight w found by solving (17.2), based on the observed
returns in periods t = 1, . . . , T , can still be useful, when we make one (big) as-
sumption:

Future asset returns are similar to past returns. (17.4)

In other words, if the asset returns for future periods T + 1, T + 2, . . . are similar
in nature to the past periods t = 1, . . . , T , then the portfolio allocation w found by
solving (17.2) could be a wise choice to use in future periods.
Every time you invest, you are warned that the assumption (17.4) need not
hold; you are required to acknowledge that past performance is no guarantee of
future performance. The assumption (17.4) often holds well enough to be useful,
but in times of market shift it need not.
This situation is similar to that encountered when fitting models to observed
data, as in chapters 13 and 14. The model is trained on past data that you have
observed; but it will be used to make predictions on future data, that you have not
yet seen. A model is useful only to the extent that future data looks like past data.
And this is an assumption which often (but not always) holds reasonably well.
Just as in model fitting, investment allocation vectors can (and should) be
validated before being used. For example, we determine the weight vector by
solving (17.2) using past returns data over some past training period, and check
the performance on some other past testing period. If the portfolio performance
over the training and testing periods are reasonably consistent, we gain confidence
(but no guarantee) that the weight vector will work in future periods. For example,
we might determine the weights using the realized returns from two years ago, and
then test these weights by the performance of the portfolio over last year. If the test
works out, we use the weights for next year. In portfolio optimization, validation
is sometimes called back-testing, since you are testing the investment method on
previous realized returns, to get an idea of how the method will work on (unknown)
future returns.
The basic assumption (17.4) oftens hold less well than the analogous assumption
in data fitting, i.e., that future data looks like past data. For this reason we expect
less coherence between the training and test performance of a portfolio, compared
to a generic data fitting application. This is especially so when the test period has
a small number of periods in it, like 100; see the discussion on page 210.
284 17 Constrained least squares applications

0.4

0.3
Return
0.2

0.1

1/n
0
risk-free

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Risk

Figure 17.1 The open circles show annualized risk and return for 20 assets
(19 stocks and one risk-free asset with a return of 1%). The solid line shows
risk and return for the Pareto optimal portfolios. The dots show risk and
return for three Pareto optimal portfolios with 10%, 20%, and 40% return,
and the portfolio with weights wi = 1/n.

17.1.3 Example

We use daily return data for 19 stocks over a period of 2000 days. After adding a
risk-free asset with a 1% annual return, we obtain a 2000 20 return matrix R.
The circles in figure 17.1 show the annualized risk and return for the 20 assets, i.e.,
the pairs

( 250 std(Rei ), 250 avg(Rei )), i = 1, . . . , 20.

It also shows the Pareto-optimal risk-return curve, and the risk and return for the
uniform portfolio with equal weights wi = 1/n. The annualized risk, return, and
the leverage for five portfolios (the four Pareto-optimal portfolios indicated in the
figure, and the 1/n portfolio) are given in table 17.1. Figure 17.2 shows the total
portfolio value (17.1) for the five portfolios. Figure 17.3 shows the portfolio values
for a different test period of 500 days.

17.1.4 Variations

There are many variations on the basic portfolio optimization problem (17.2). We
describe a few of them here.
17.1 Portfolio optimization 285

Return Risk
Portfolio Train Test Train Test Leverage
Risk-free 0.01 0.01 0.00 0.00 1.00
10% 0.10 0.08 0.09 0.07 1.96
20% 0.20 0.15 0.18 0.15 3.03
40% 0.40 0.30 0.38 0.31 5.48
1/n 0.10 0.21 0.23 0.13 1.00

Table 17.1 Annualized risk, return, and leverage for five portfolios.

150 Risk-free
1/n
10%
Value (thousand dollars)

20%
100 40%

50

10

0 400 800 1200 1600 2000

Day

Figure 17.2 Total value over time for five portfolios: the risk-free portfolio
with 1% annual return, the Pareto optimal portfolios with 10%, 20%, and
40% return, and the uniform portfolio. The total value is computed using
the 2000 20 daily return matrix R.
286 17 Constrained least squares applications

18
Risk-free
1/n
16 10%

Value (thousand dollars)

20%
40%

14

12

10

0 100 200 300 400 500

Day

Figure 17.3 Value over time for the five portfolios in figure 17.2 over a test
period of 500 days.

Time-varying weights. Markets do shift, so it is not uncommon to periodically

update or change the allocation weights that are used. In one extreme version of
this, a new allocation vector is used in every period. The allocation weight for any
period is obtained by solving the portfolio optimization problem over the preceding
M periods. (This scheme can be modified to include testing periods as well.) The
parameter M in this method would be chosen by validation on previous realized
returns, i.e., back-testing.
When the allocation weights are changed over time, we can add a (regulariza-
tion) term of the form kwcurr wk2 to the objective, where is a positive constant.
Here wcurr is the currently used allocation, and w is the proposed new allocation
vector. The additional regularization term encourages the new allocation vector
to be near the current one. (When this is not the case, the portfolio will require
much buying and selling of assets. This is called turnover, which leads to trading
costs not included in our simple model.) The parameter would be chosen by
back-testing.

Regularization. Just as in data fitting, our formulation of portfolio optimization

can suffer from over-fit, which means that the chosen weights perform very well on
past (realized) returns, but poorly on new (future) returns. This can be avoided
by adding regularization, which here means to penalize investments in assets other
than cash. (This is analogous to penalizing the size of the model coefficients, except
for the parameter associated with the constant feature.) A very natural way to do
 17.1 Portfolio optimization 287

this is to add a positive multiple of the weighted sum of squares term

12 w12 + + n1
2 2
wn1
to the objective in (17.2). Note that we do not penalize wn , which is the weight
associated with the risk-free asset. The constants i are the standard deviations
of the (realized) returns, i.e., i = std(Rei ). This regularization penalizes weights
associated with risky assets more than than those associated with less risky assets.
A good choice of can be found by back-testing.

17.1.5 Two-fund theorem

We can express the solution (17.3) of the portfolio optimization problem in the
form
1 1
2RT R 1 2RT R 1

w 0 2T
z1 = 1T 0 0 1 + 1T 0 0 0 .
T T
z2 0 0 0 0 0 1
Taking the first n components of this, we obtain
w = w0 + v, (17.5)
where w0 and v are the first n components of the (n + 2)-vectors
1 1
2RT R 1 2RT R 1

0 2T
1T 0 0 1 , 1T 0 0 0 ,
T T
0 0 0 0 0 1
respectively. The equation (17.5) shows that the Pareto optimal portfolios form a
line in weight space, parametrized by the required return . The portfolio w0 is a
point on the line, and the vector v, which satisfies 1T v = 0, gives the direction of
the line. This equation tells us that we do not need to solve the equation (17.3) for
each value of . We first compute w0 and v (by factoring the matrix once and using
two solve steps), and then form the optimal portfolio with return as w0 + v.
Any point on a line can be expressed as an affine combination of two different
points on the line. So if we find two different Pareto optimal portfolios, then we
can express a general Pareto optimal portfolio as an affine combination of them.
In other words, all Pareto optimal portfolios are affine combinations of just two
portfolios (indeed, any two different Pareto optimal portfolios). This is the two-
fund theorem. (Fund is another term for portfolio.)
Now suppose that the last asset is risk-free. The portfolio w = en is Pareto
optimal, since it achieves return rf with zero risk. We then find one other Pareto
optimal portfolio, for example, the one w2 that achieves return 2rf , twice the
risk-free return. (We could choose here any return other than rf .) Then we can
express the general Pareto optimal portfolio as
w = (1 )en + w2 ,
where = /rf 1.
288 17 Constrained least squares applications

17.2 Linear quadratic control

We consider a time-varying linear dynamical system with state n-vector xt and
input m-vector ut , with dynamics equations

xt+1 = At xt + Bt ut , t = 1, 2, . . . . (17.6)

The system has an output, the p-vector yt , given by

yt = Ct xt , t = 1, 2, . . . . (17.7)

We usually have m n and p n, i.e., there are fewer inputs and outputs than
states.
In control applications, the input ut represents quantities that we can choose
or manipulate, like control surface deflections or engine thrust on an airplane. The
state xt , input ut , and output yt typically represent deviations from some standard
or desired operating condition, for example, the deviation of aircraft speed and
altitude from the desired values. For this reason it is desirable to have xt , yt , and
ut small.
Linear quadratic control refers to the problem of choosing the input and state
sequences, over a time period t = 1, . . . , T , so as to minimize a sum of squares
objective, subject to the dynamics equations (17.6), the output equations (17.7),
and additional linear equality constraints. (In linear quadratic, linear refers to
the linear dynamics, and quadratic refers to the objective function, which is a
sum of squares.)
Most control problems include an initial state constraint, which has the form
x1 = xinit , where xinit is a given initial state. Some control problems also include a
final state constraint xT = xdes , where xdes is a given (desired) final (also called
terminal or target) state.
The objective function has the form J = Joutput + Jinput , where

Joutput = ky1 k2 + + kyT k2 = kC1 x1 k2 + + kCT xT k2 ,

Jinput = ku1 k2 + + kuT 1 k2 .

The positive parameter weights the input objective Jinput relative to the output
objective Joutput .
The linear quadratic control problem (with initial and final state constraints)
is
minimize Joutput + Jinput
subject to xt+1 = At xt + Bt ut , t = 1, . . . , T 1, (17.8)
x1 = xinit , xT = xdes ,
where the variables to be chosen are x1 , . . . , xT and u1 , . . . , uT 1 .

Formulation as constrained least squares problem. We can solve the linear

quadratic control problem (17.8) by setting it up as a big linearly constrained
least squares problem. We define the vector z of all these variables, stacked:

z = (x1 , . . . , xT , u1 , . . . , uT 1 ).
17.2 Linear quadratic control 289

The dimension of z is T n + (T 1)m. The control objective can be expressed as

kAz bk2 , where b = 0 and A is the block matrix

C1

C2

..

.

A = CT .


I

..
.

I

In this matrix, (block) entries not shown are zero, and the identity matrices in
the lower right corner have dimension m. (The lines in the matrix delineate the
portions related to the states and the inputs.) The dynamics constraints, and the
with
initial and final state constraints, can be expressed as Cz = d,

A1 I B1 0
A2 I B2 0
..

.. .. ..
C =
. . . ,

d =
. ,


AT 1 I BT 1

0



I xinit
I xdes

where (block) entries not shown are zero. (The vertical line separates the portions
of the matrix associated with the states and the inputs, and the horizontal lines
separate the dynamics equations and the initial and final state constraints.)
The solution z of the constrained least squares problem

minimize kAz bk2

(17.9)
subject to Cz = d

gives us the optimal input trajectory and the associated optimal state (and output)
since here b = 0, it is a
trajectory. The solution z is a linear function of b and d;
init des
linear function of x and x .

Complexity. The large constrained least squares problem (17.9) has dimensions

n = T n + (T 1)m, m = T p + (T 1)m, p = (T 1)n + 2n,

so using one of the standard methods described above would require order

(p + m)n2 T 3 (m + p + n)(m + n)2

flops. But the matrices A and C are very sparse, and by exploiting this sparsity
(see page 275), the large constrained least squares problem can be solved in order
T (m + p + n)(m + n)2 flops, which grows only linearly in T .
290 17 Constrained least squares applications

0.4

0.3

Output
0.2

0.1

0 20 40 60 80 100
t

Figure 17.4 Open-loop response CAt1 xinit .

17.2.1 Example

We consider the time-invariant linear dynamical system with

0.855 1.161 0.667 0.076
A = 0.015 1.073 0.053 , B = 0.139 ,
0.084 0.059 1.022 0.342

 
C= 0.218 3.597 1.683 ,

with initial condition xinit = (0.496, 0.745, 1.394), and target or desired final state
xdes = 0, and T = 100. In this example, both the input ut and the output yt have
dimension one, i.e., are scalar.
Figure 17.4 shows the output when the input is zero,

yt = CAt1 xinit , t = 1, . . . , T.

which is called the open-loop output. Figure 17.5 shows the optimal trade-off curve
of the objectives Jinput and Joutput , found by varying the parameter , solving
the problem (17.9), and evaluating the objectives Jinput and Joutput . The points
corresponding to the values = 0.05, = 0.2, and = 1 are shown as circles. As
always, increasing has the effect of decreasing Jinput , at the cost of increasing
Joutput .
The optimal input and output trajectories for these three values of are shown
in figure 17.6. Here too we see that for larger , the input is smaller but the output
is larger.
 17.2 Linear quadratic control 291

4.4

4.2

=1
Joutput

= 0.2
3.8
= 0.05

3.6

0 1 2 3 4
Jinput

Figure 17.5 Optimal trade-off curve of the objectives Jinput and Joutput .

17.2.2 Variations

There are many variations on the basic linear quadratic control problem described
above. We describe some of them here.

Tracking. We replace yt in Joutput with yt ytdes , where ytdes is a given desired

output trajectory. In this case the objective function Joutput is called the tracking
error. Decreasing the parameter leads to better output tracking, at the cost of
larger input trajectory.
This variation on the linear quadratic control problem can be expressed as a
linearly constrained least squares problem with the same big matrices A and C,
the same vector d, and a nonzero vector b. The desired trajectory y des appears in
t
the vector b.

Time-weighted objective. We replace Joutput with

Joutput = w1 ky1 k2 + + wT kyT k2 ,

where w1 , . . . , wT are given positive constants. This allows us to weight earlier or

later output values differently. A common choice, called exponential weighting, is
wt = t , where > 0. For > 1 we weight later values of yt more than earlier
values; the opposite is true for < 1 (in which case is sometimes called the
discount or forgetting factor ).

Way-point constraints. A way-point constraint specifies that y = y wp , where

y wp is a given p-vector, and is a given way-point time. This constraint is typically
used when yt represents a position of a vehicle; it requires that the vehicle pass
292 17 Constrained least squares applications

0.5 0.4

0 0.2
ut

yt
0.5 0

0 20 40 60 80 100 0 20 40 60 80 100
t t

0.5 0.4

0 0.2
ut

yt

0.5 0

0 20 40 60 80 100 0 20 40 60 80 100
t t

0.5 0.4

0 0.2
ut

yt

0.5 0

0 20 40 60 80 100 0 20 40 60 80 100
t t

Figure 17.6 Optimal inputs (left) and outputs (right) for = 0.05 (top),
= 0.2 (center), and = 1 (bottom).
 17.2 Linear quadratic control 293

through the position y wp at time t = . Way-point constraints can be expressed as

linear equality constraints on the big vector z.

17.2.3 Linear state feedback control

In the linear quadratic control problem we work out a sequence of inputs u1 ,

. . . , uT 1 to apply to the system, by solving the constrained least squares prob-
lem (17.8). It is typically used in cases where t = T has some significance, like the
time of landing or docking for a vehicle.
We have already mentioned (on page 151) another simpler approach to the
control of a linear dynamical system. In linear state feedback control we measure
the state in each period and use the input

ut = Kxt

for t = 1, 2, . . .. The matrix K is called the state feeback gain matrix. State feedback
control is very widely used in practical applications, especially ones where there is
no fixed future time T when the state must take on some desired value; instead, it
is desired that both xt and ut should be small and converge to zero. One practical
advantage of linear state feedback control is that we can find the state feedback
matrix K ahead of time; when the system is operating, we determine the input
values using one simple matrix-vector multiply. Here we show how an appropriate
state feedback gain matrix K can be found using linear quadratic control.
Let z denote the solution of the linear quadratic control problem, i.e., the
solution of the linearly constrained least squares problem (17.8), with xdes = 0.
The solution z is a linear function of xinit and xdes ; since here xdes = 0, z is a linear
function of xinit = x1 . Since u1 , the optimal input at t = 1, is a slice or subvector of
z, we conclude that u1 is a linear function of x1 , and so can be written as u1 = Kx1
for some m n matrix K. The columns of K can be found by solving (17.8) with
initial conditions xinit = e1 , . . . , en . This can be done efficiently by factorizing the
coefficient once, and then carrying out n solves.
This matrix generally provides a good choice of state feedback gain matrix.
With this choice, the input u1 with state feeback control and under linear quadratic
control are the same; for t > 1, the two inputs differ. An interesting phenomenon,
beyond the scope of this book, is that the state feedback gain matrix K found this
way does not depend very much on T , provided it is chosen large enough.

Example. For the example described in 17.2.1 the state feedback gain matrix for
= 1 is
 
K = 0.308 2.659 1.446 .
In figure 17.7, we plot the input and output trajectories with linear quadratic
control (in blue) and using the simpler linear state feedback control ut = Kxt .
We can see that the input sequence found using linear quadratic control achieves
yT = 0 exactly; the input sequence found by linear state feedback control makes
yT small, but not zero.
294 17 Constrained least squares applications

0.1 0.4

0
ut

yt
0.2

0.1
0

0 50 100 150 0 50 100 150

t t

Figure 17.7 The blue curves are the solutions of (17.8) for = 1. The
red curves are the inputs and outputs that result from the constant state
feedback ut = Kxt .

17.3 Linear quadratic state estimation

The setting is a linear dynamical system of the form

xt+1 = At xt + Bt wt , yt = Ct xt + vt , t = 1, 2, . . . . (17.10)

Here the n-vector xt is the state of the system, the p-vector yt is the measurement,
the m-vector wt is the input or process noise, and the p-vector vt is the measurement
noise or residual. The matrices At , Bt , and Ct are the dynamics, input, and output
matrices, respectively.
In state estimation, we know the matrices At , Bt , and Ct over the time period
t = 1, . . . , T , as well as the measurements y1 , . . . , yT , but we do not know the
process or measurement noises. The goal is to guess or estimate the state sequence
x1 , . . . , xT . State estimation is widely used in many application areas, including
all guidance and navigation systems, such as GPS (global positioning system).
Since we do not know the process or measurement noises, we cannot exactly
deduce the state sequence. Instead we will guess or estimate the state sequence
x1 , . . . , xT and process noise sequence w1 , . . . , wT 1 , subject to the requirement
that they satisfy the dynamic system model (17.10). When we guess the state
sequence, we implicitly guess that the measurement noise is vt = yt Ct xt . We
make one fundamental assumption: The process and measurement noises are both
small, or at least, not too large.
Our primary objective is the sum of squares of the norms of the measurement
residuals,

Jmeas = kv1 k2 + + kvT k2 = kC1 x1 y1 k2 + + kCT xT yT k2 .

If this quantity is small, it means that the proposed state sequence guess is consis-
tent with our measurements. Note that the quantities in the squared norms above
are the same as vt .
17.3 Linear quadratic state estimation 295

The secondary objective is the sum of squares of the norms of the process noise,
Jproc = kw1 k2 + + kwT 1 k2 .
Our prior assumption that the process noise is small corresponds to this objective
being small.

Least squares state estimation. We will make our guesses of x1 , . . . , xT and

w1 , . . . , wT 1 so as to minimize a weighted sum of our objectives, subject to the
dynamics constraints:
minimize Jmeas + Jproc
(17.11)
subject to xt+1 = At xt + Bt wt , t = 1, . . . , T 1,
where is a positive parameter that allows us to put more emphasis on making
our measurement discrepancies small (by choosing small), or the process noises
small (by choosing large). Roughly speaking, small means that we trust the
measurements more, while large means that we trust the measurements less, and
put more weight on choosing a trajectory consistent with the dynamics, with small
process noise. We will see later how can be chosen using validation.

Estimation versus control. The least squares state estimation problem is very
similar to the linear quadratic control problem, but the interpretation is quite dif-
ferent. In the control problem, we can choose the inputs; they are under our control.
Once we choose the inputs, we know the state sequence. In the control problem,
the inputs are typically actions that we take to affect the state trajectory. In the
estimation problem, the inputs (called process noise in the estimation problem) are
unknown, and the problem is to guess them. Our job is to guess the state sequence,
which we do not know. This is a passive task. We are not choosing inputs to affect
the state; rather, we are observing the outputs and hoping to deduce the state
sequence. The mathematical formulations of the two problems, however, are very
closely related. The close connection between the two problems is sometimes called
control/estimation duality.

Formulation as constrained least squares problem. The least squares state esti-
mation problem (17.11) can be formulated as a linearly constrained least squares
problem, using stacking. We define the stacked vector
z = (x1 , . . . , xT , w1 , . . . , wT 1 ).
The objective in (17.11) can be expressed as kAz bk2 , with
C1

y1
C2 y2
.. ..

.

.

A =
CT ,
yT .
b =

I
0

.. .
..

.


I 0
296 17 Constrained least squares applications

with d = 0 and
The constraints in (17.11) can be expressed as Cz = d,

A1 I B1
A2 I B2
C = .

. .. . .. . ..
AT 1 I BT 1
The constrained least squares problem has dimensions
n = T n + (T 1)m, m = T p + (T 1)m, p = (T 1)n.
so using one of the standard methods described above would require order
(p + m)n2 T 3 (m + p + n)(m + n)2
flops. As in the case of linear quadratic control, the matrices A and C are very
sparse, and by exploiting this sparsity (see page 275), the large constrained least
squares problem can be solved in order T (m + p + n)(m + n)2 flops, which grows
only linearly in T .
The least squares state estimation problem was formulated in around 1960 by
Rudolf Kalman and others (in a statistical framework). He and others developed a
particular recursive algorithm for solving the problem, and the whole method has
come to be known as Kalman filtering. For this work Kalman was awarded the
Kyoto Prize in 1985.

17.3.1 Example
We consider a system with n = 4, p = 2, and m = 2, and time-invariant matrices

1 0 1 0 0 0  
0 1 0 1 0 0 1 0 0 0
A= 0 0 1
, B= , C= .
0 1 0 0 1 0 0
0 0 0 1 0 1
This is a very simple model of motion of a mass moving in 2-D. The first two
components of xt represent the position coordinates; components 3 and 4 represent
the velocity coordinates. The input wt acts like a force on the mass, since it adds to
the velocity. We think of the 2-vector Cxt as the exact or true position of the mass
at period t. The measurement yt = Cxt + vt is a noisy measurement of the mass
position. We will estimate the state trajectory over t = 1, . . . , T , with T = 100.
In figure 17.8 the 100 measured positions yt are shown as circles in 2-D. The
solid black line shows Cxt , i.e., the actual position of the mass. We solve the least
squares state estimation problem (17.11) for a range of values of . The estimated
trajectories C xt for three values of are shown as red lines. We can see that = 1
is too small for this example: The estimated state places too much trust in the
measurements, and is following measurement noise. We can also see that = 105
is too large: The estimated state is very smooth (since the estimated process noise
is small), but the imputed noise measurements are too high. In this example the
choice of is simple, since we have the true position trajectory. We will see later
how can be chosen using validation in the general case.
 17.3 Linear quadratic state estimation 297

=1

0
t=1

500
(xt )2

1000

1500

0 100 200 300

(xt )1

= 103 = 105

Figure 17.8 The circles show 100 noisy measurements in 2-D. Top left. The
black line is the exact position Cxt . Three other plots. The red lines are
estimated trajectories C xt for three values of .
298 17 Constrained least squares applications

17.3.2 Variations
Known initial state. There are several interesting variations on the state estima-
tion problem. For example, we might know the initial state x1 . In this case we
simply add an equality constraint x1 = xknown
1 .

Missing measurements. Another useful variation on the least squares state esti-
mation problem allows for missing measurements, i.e., we only know yt for t T ,
where T is the set of times for which we have a measurement.P We can han-
T
dle
P this variation two (equivalent) ways: We can either replace t=1 kvt k2 with
2
tT kvt k , or we can consider yt for t 6 T to be optimization variables as well.
(Both lead to the same state sequence estimate.) When there are missing mea-
surements, we can estimate what the missing measurements might have been, by
taking
yt = Ct xt , t 6 T .
(Here we assume that vt = 0.)

17.3.3 Validation

The technique of estimating what a missing measurement might have been directly
gives us a method to validate a quadratic state estimation method, and in partic-
ular, to choose . To do this, we remove some of the measurements (say, 20%),
and carry out least squares state estimation pretending that those measurements
are missing. Our state estimate produces predicted values for the missing (really,
held back) measurements, which we can compare to the actual measurements. We
choose a value of which approximately minimizes this (test) prediction error.

Example. Continuing the previous example, we randomly remove 20 of the 100

measurement points. We solve the same problem (17.11) for a range of values of
, but with Jmeas defined as
X
Jmeas = kCxt yt k2 ,
tT

i.e., we only sum the measurement errors over the measurements we have. For each
value of we compute the RMS train and test errors
!1/2 1/2
1 X
2 1 X 2
Etrain = kC xt yt k , Etest = kC xt yt k .
80p tT 20p
t6T

The training error (squared and scaled) appears directly in our minimization prob-
lem. The test error, however, is a good test of our estimation method, since it
compares predictions of positions (in this example) with measurements of position
that were not used to form the estimates. The errors are shown in figure 17.9, as
a function of the parameter . We can clearly see that for < 100 or so, we are
over-fit, since the test RMS error substantially exceeds the train RMS error. We
can also see that around 103 is a good choice.
17.3 Linear quadratic state estimation 299

80
Train
Test
60
RMS error

40

20

103 101 101 103 105

Figure 17.9 Training and test error for the state estimation example.
300 17 Constrained least squares applications
 Chapter 18

Nonlinear least squares

In previous chapters we studied the problems of solving a set of linear equations

or finding a least squares approximate solution to them. In this chapter we study
extensions of these problems in which linear is replaced with nonlinear. These
nonlinear problems are in general hard to solve exactly, but we describe a heuristic
algorithm that often works well in practice.

18.1 Nonlinear equations and least squares

18.1.1 Nonlinear equations

Consider a set of m possibly nonlinear equations in n unknowns (or variables)

x = (x1 , . . . , xn ), written as

fi (x) = 0, i = 1, . . . , m,

where fi : Rn R is a scalar valued function. We refer to fi (x) = 0 as the ith

equation. For any x we call fi (x) the ith residual, since it is a quantity we want to
be zero. Many interesting practical problems can be expressed as the problem of
solving, possibly approximately, a set of nonlinear equations.
We take the right-hand side of the equations to be zero to simplify the problem
notation. If we need to solve fi (x) = bi , i = 1, . . . , m, where bi are some given
nonzero numbers, we define fi (x) = fi (x) bi , and solve fi (x) = 0, i = 1, . . . , m,
which gives us a solution of the original equations. Assuming the right-hand sides
of the equations are zero will simplify formulas and equations.
We often write the set of equations in the compact vector form

f (x) = 0, (18.1)

where f (x) = (f1 (x), . . . , fm (x)) is an m-vector, and the zero vector on the right-
hand side has dimension m. We can think of f as a function that maps n-vectors to
302 18 Nonlinear least squares

m-vectors, i.e., f : Rn Rm . We refer to the m-vector f (x) as the residual (vec-

tor) associated with the choice of n-vector x; our goal is to find x with associated
residual zero.
When f is an affine function, the set of equations (18.1) is a set of m linear
equations in n unknowns, which can be solved (or approximately solved in a least
squares sense when m > n), using the techniques covered in previous chapters. We
are interested here in the case when f is not affine.
We extend the ideas of under-determined, square, and over-determined equa-
tions to the nonlinear case. When m < n, there are fewer equations than unknowns,
and the system of equations (18.1) is called under-determined. When m = n, so
there are as many equations as unknowns, and the system of equations is called
square. When m > n, there are more equations than unknowns, and the system of
equations is called over-determined.

18.1.2 Nonlinear least squares

When we cannot find a solution of the equations (18.1), we can seek an approximate
solution, by finding x that minimizes the sum of squares of the residuals,

f1 (x)2 + + fm (x)2 = kf (x)k2 .

This means finding x for which kf (x)k2 kf (x)k2 holds for all x. We refer to
such a point as a least squares approximate solution of (18.1), or more directly, as
a solution of the nonlinear least squares problem

minimize kf (x)k2 , (18.2)

where the n-vector x is the variable to be found. When the function f is affine, the
nonlinear least squares problem (18.2) reduces to the (linear) least squares problem
from chapter 12.
The nonlinear least squares problem (18.2) includes the problem of solving
nonlinear equations (18.1) as a special case, since any x that satisfies f (x) = 0 is
also a solution of the nonlinear least squares problem. But as in the case of linear
equations, the least squares approximate solution of a set of nonlinear equations is
often very useful even when it does not solve the equations. So we will focus on
the nonlinear least squares problem (18.2).

18.1.3 Optimality condition

Calculus gives us a necessary condition for x to be a solution of (18.2), i.e., to

minimize kf (x)k2 . (This means that the condition must hold for a solution, but
it may also hold for other points that are not solutions.) The partial derivative of
kf (x)k2 with respect to each of x1 , . . . , xn must vanish at x:

kf (x)k2 = 0, i = 1, . . . , n,
xi
 18.1 Nonlinear equations and least squares 303

or, in vector form, kf (x)k2 = 0 (see C.2). This gradient can be expressed as
m m
!
X X
2 2
kf (x)k = fi (x) =2 fi (x)fi (x) = 2Df (x)T f (x),
i=1 i=1

where the mn matrix Df (x) is the derivative or Jacobian matrix of the function f
at the point f , i.e., the matrix of its partial derivatives (see 8.3). So if x minimizes
kf (x)k2 , it must satisfy
2Df (x)T f (x) = 0. (18.3)
This optimality condition must hold for any solution of the nonlinear least squares
problem (18.2). But the optimality condition can also hold for other points that
are not solutions of the nonlinear least squares problem. For this reason the opti-
mality condition (18.3) is called a necessary condition for optimality, because it is
necessarily satisfied for any solution x. It is a not a sufficient condition for opti-
mality, since the optimality condition (18.3) is not enough (i.e., is not sufficient)
to guarantee that the point is a solution of the nonlinear least squares problem.
When the function f is affine, the optimality conditions (18.3) reduce to the
normal equations (12.4), the optimality conditions for the (linear) least squares
problem.

18.1.4 Difficulty of solving nonlinear equations

Solving a set of nonlinear equations (18.1), or solving the nonlinear least squares
problem (18.2), is in general much more difficult than solving a set of linear equa-
tions or a linear least squares problem. For nonlinear equations, there can be no
solution, or any number of solutions, or an infinite number of solutions. Unlike
linear equations, it is a very difficult computational problem to determine which
one of these cases holds for a particular set of equations; there is no analog of the
QR factorization that we can use for linear equations and least squares problems.
Even the simple sounding problem of determining whether or not there are any
solutions to a set of nonlinear equations is very difficult computationally. There
are advanced non-heuristic algorithms for exactly solving nonlinear equations, or
exactly solving nonlinear least squares problems, but they are complicated and
very computationally demanding, and rarely used in applications.
Given the difficulty of solving a set of nonlinear equations, or solving a nonlinear
least squares problem, we must lower our expectations. We can only hope for an
algorithm that often finds a solution (when one exists), or produces a value of x
with small residual norm, if not the smallest that is possible. Algorithms like this,
that often work, or tend to produce a good if not always the best possible point, are
called heuristics. The k-means algorithm of chapter 4 is an example of a heuristic
algorithm. Solving linear equations or linear least squares problems using the QR
factorization are not heuristics; these algorithms always work.
Many heuristic algorithms for the nonlinear least squares problem, including
those we describe later in this chapter, compute a point x that satisfies the op-
timality condition (18.3). Unless f (x) = 0, however, such a point need not be a
solution of the nonlinear least squares problem (18.2).
304 18 Nonlinear least squares

S(p)

D(p)
p

Figure 18.1 Supply and demand as functions of the price, shown on the hori-
zontal axis. They intersect at the point shown as a circle. The corresponding
price is the equilibrium price.

18.1.5 Examples

In this section we list a few applications that reduce to solving a set of nonlinear
equations, or a nonlinear least squares problem.

Computing equilibrium points. The idea of an equilibrium, where some type

of consumption and generation balance each other, arises in many applications.
Consumption and generation depend, often nonlinearly, on the values of some pa-
rameters, and the goal is to find values of the parameters that lead to equilibrium.
These examples typically have m = n, i.e., the system of nonlinear equations is
square.

Equilibrium prices. We consider n commodities or goods, with associated

prices given by the n-vector p. The demand for the n goods (an n-vector)
is a nonlinear function of the prices, given by D(p). (In an example on
page 122 we described an approximate model for demand that is accurate
when the prices change from nominal values by few percent; here we consider
the demand over a large range of prices.) The supply of the goods (an n-
vector) also depends on the prices, and is given by S(p). (When the price for
a good is high, for example, more producers are willing to produce it, so the
supply increases.) A set of commodity prices p is an equilibrium price vector
if it results in supply balancing demand, i.e., S(p) = D(p). Finding a set of
equilibrium prices is the same as solving the square set of nonlinear equations

f (p) = S(p) D(p) = 0.

(The vector f (p) is called the excess supply, at the set of prices p.) This is
shown in figure 18.1 for a simple case with n = 1.
18.1 Nonlinear equations and least squares 305

Chemical equilibrium. We consider n chemical species in a solution. The

n-vector c denotes the concentrations of the n species. Reactions among
the species consume some of them (the reactants) and generate others (the
products). The rate of each reaction is a function of the concentrations of
its reactants (and other parameters we assume are fixed, like temperature or
presence of catalysts). We let C(c) denote the vector of total consumption
of the n reactants, over all the reactions, and we let G(c) denote the vector
of generation of the n reactants, over all reactions. A concentration vector
c is in chemical equilibrium if C(c) = G(c), i.e., the rate of consumption of
all species balances the rate of generation. Computing a set of equilibrium
concentrations is the same as solving the square set of nonlinear equations

f (c) = C(c) G(c) = 0.

Mechanical equilibrium. A mechanical system in 3-D with N nodes is char-

acterized by the positions of the nodes, given by a 3N -vector q of the stacked
node positions, called the generalized position. The net force on each node is
a 3-vector, which depends on q, i.e., the node positions. We describe this as a
3N -vector of forces, F (q). The system is in mechanical equilibrium if the net
force on each node is zero, i.e., F (q) = 0, a set of 3N nonlinear equations in
3N unknowns. (In this example we take into account only the forces at each
node. A more realistic model include torques applied at each node, which
leads to a set of 6N equations in 6N unknowns.)

Nash equilibrium. We consider a simple setup for a mathematical game. Each

of n competing agents or participants chooses a number xi . Each agent is
given a (numerical) reward (say, money) that depends not only on her own
choice, but on the choice of all the agents. The reward for agent i is given
by the function Ri (x), called the payoff function. Each agent wishes to make
a choice that maximizes her reward. This is complicated since the reward
depends not only on her choice, but the choices of the other agents.
A Nash equilibrium is a set of choices given by the n-vector x where no agent
can improve (increase) her reward by changing her choice. Such a choice is
argued to be stable since no agent is incented to change her choice. At a
Nash equilibrium xi maximizes Ri (x), so we must have

Ri
(x) = 0, i = 1, . . . , n.
xi

This necessary condition for a Nash equilibrium is a square set of nonlinear

equations.
The idea of a Nash equilibrium is widely used in economics, social science,
and engineering. Nash was awarded the Nobel Prize in Economics for this
work in 1994.

Nonlinear least squares examples. Nonlinear least squares problems arise in

many of the same settings and applications as linear least squares problems.
306 18 Nonlinear least squares

Location from range measurements. The 3-vector (or 2-vector) x represents

the location of some object or target in 3-D (or 2-D), which we wish to
determine or guess. We are given m range measurements, i.e., the distance
from x to some known locations a1 , . . . , am ,
i = kx ai k + vi , i = 1, . . . , m,
where vi is an unknown measurement error, assumed to be small. Our esti-
mate x of the location is found by minimizing the sum of the squares of the
range residuals,
m
2
X
(kx ai k i ) .
i=1
A similar method is used in Global Positioning System (GPS) devices, where
ai are the known locations of GPS satellites that are in view.
Nonlinear model fitting. We consider a data model y f(x; ) where f is not
an affine function of the model parameter p-vector (as it is in chapter 13).
We choose the model parameter by minimizing the sum of the squares of the
residuals over a data set with N examples,
N
(f(xi ; ) yi )2 .
X
(18.4)
i=1

(As in linear least squares model fitting, we can add a regularization term to
this objective function.)

18.2 Gauss-Newton algorithm

In this section we describe a powerful heuristic algorithm for the nonlinear least
squares problem (18.2) that bears the names of the two famous mathematicians
Carl Friedrich Gauss and Isaac Newton. We also describe a variation of the Gauss-
Newton algorithm known as the Levenberg-Marquardt algorithm, which addresses
some shortcomings of the basic Gauss-Newton algorithm.
The Gauss-Newton and Levenberg-Marquardt algorithms are iterative algo-
rithms that generate a sequence of points x(1) , x(2) , . . . . The vector x(1) is called
the starting point of the algorithm, and x(k) is called the kth iterate. Moving from
x(k) to x(k+1) is called an iteration of the algorithm. We judge the iterates by
the norm of the associated residuals, kf (x(k) )k, or its square. The algorithm is
terminated when kf (x(k) )k is small enough, or x(k+1) is very near x(k) , or when a
maximum number of iterations is reached.

18.2.1 Basic Gauss-Newton algorithm

The idea behind the Gauss-Newton algorithm is simple: We alternate between
finding an affine approximation of the function f at the current iterate, and then
18.2 Gauss-Newton algorithm 307

solving the associated linear least squares problem to find the next iterate. This
combines two of the most powerful ideas in applied mathematics: Calculus is used
to form an affine approximation of a function near a given point, and least squares
is used to compute an approximate solution of the resulting affine equations.
We now describe the algorithm in more detail. At each iteration k, we form
the affine approximation f of f at the current iterate x(k) , given by the Taylor
approximation
f(x; x(k) ) = f (x(k) ) + Df (x(k) )(x x(k) ), (18.5)
where the m n matrix Df (x(k) ) is the Jacobian or derivative matrix of f (see 8.3
and C.1). The affine function f(x; x(k) ) is a very good approximation of f (x)
provided x is near x(k) , i.e., kx x(k) k is small.
The next iterate x(k+1) is then taken to be the minimizer of kf(x; x(k) )k2 , the
norm squared of the affine approximation of f at x(k) . Assuming that the derivative
matrix Df (x(k) ) has linearly independent columns (which requires m n), we have
 1
x(k+1) = x(k) Df (x(k) )T Df (x(k) ) Df (x(k) )T f (x(k) ).

This iteration gives the basic Gauss-Newton algorithm.

Algorithm 18.1 Basic Gauss-Newton algorithm for nonlinear least squares

given a differentiable function f : Rn Rm , an initial point x(1) .
For k = 1, 2, . . . , kmax

1. Form affine approximation at current iterate using calculus. Evaluate the Ja-
cobian Df (x(k) ) and define

f(x; x(k) ) = f (x(k) ) + Df (x(k) )(x x(k) ).

2. Update iterate using linear least squares. Set x(k+1) as the minimizer of kf(x; x(k) )k2 ,

1
x(k+1) = x(k) Df (x(k) )T Df (x(k) ) Df (x(k) )T f (x(k) ).

The Gauss-Newton algorithm is terminated early if f (x) is very small, or x(k+1)

x(k) . It terminates with an error if the columns of Df (x(k) ) are linearly dependent.
The condition x(k+1) = x(k) (the exact form of our stopping condition) holds
when   1
Df (x(k) )T Df (x(k) ) Df (x(k) )T f (x(k) ) = 0,

which occurs if and only if Df (x(k) )T f (x(k) ) = 0 (since we assume that Df (x(k) )
has linearly independent columns). Roughly speaking, the Gauss-Newton algo-
rithm stops only when the optimality condition (18.3) holds.
We can also observe that

kf(x(k+1) ; x(k) )k2 kf(x(k) ; x(k) )k2 = kf (x(k) )k2 (18.6)

308 18 Nonlinear least squares

holds, since x(k+1) minimizes kf(x; x(k) )k2 , and f(x(k) ; x(k) ) = f (x(k) ). Roughly
speaking, the norm of the residual of the approximation goes down in each iteration.
This is not the same as

kf (x(k+1) )k2 kf (x(k) )k2 , (18.7)

i.e., the norm of the residual goes down in each iteration, which is what we would
like.

Shortcomings of the basic Gauss-Newton algorithm. We will see in examples

that the Gauss-Newton algorithm can work well, in the sense that the iterates
x(k) converge very quickly to a point with small residual. But the Gauss-Newton
algorithm has two related serious shortcomings.
The first is that it can fail, by producing a sequence of points with the norm of
the residual kf (x(k) )k increasing to large values, as opposed to decreasing to a small
value, which is what we want. (In this case the algorithm is said to diverge.) The
mechanism behind this failure is related to the difference between (18.6) and (18.7).
The approximation
kf (x)k2 kf(x; x(k) )k2
is guaranteed to hold only when x is near x(k) . So when x(k+1) is not near x(k) ,
kf (x(k+1) )k2 and kf(x(k+1) ; x(k) )k2 can be very different. In particular, the (true)
residual at x(k+1) can be larger than the residual at x(k) .
The second serious shortcoming of the basic Gauss-Newton algorithm is the
assumption that the columns of the derivative matrix Df (x(k) ) are linearly inde-
pendent. In some applications, this assumption never holds; in others, it can fail to
hold at some iterate x(k) , in which case the Gauss-Newton algorithm stops, since
x(k+1) is not defined.
We will see that a simple modification of the Gauss-Newton algorithm, described
below in 18.3, addresses both of these shortcomings.

18.2.2 Newton algorithm

For the special case m = n, the Gauss-Newton algorithm reduces to another famous
algorithm for solving a set of n nonlinear equations in n variables, called the Newton
algorithm. (The algorithm is sometimes called the Newton-Raphson algorithm,
since Newton developed the method only for the special case n = 1, and Joseph
Raphson later extended it to the case n > 1.)

Algorithm 18.2 Newton algorithm for solving nonlinear equations

given a differentiable function f : Rn Rn , an initial point x(1) .
For k = 1, 2, . . . , kmax

1. Form affine approximation at current iterate. Evaluate the Jacobian Df (x(k) )

and define
f(x; x(k) ) = f (x(k) ) + Df (x(k) )(x x(k) ).
18.2 Gauss-Newton algorithm 309

f(x; x(k) )

f (x)

x(k+1)
x(k)

Figure 18.2 One iteration of the Newton algorithm for solving an equation
f (x) = 0 in one variable.

2. Update iterate by solving linear equations. Set x(k+1) as the solution of f(x; x(k) ) =
0,

1
x(k+1) = x(k) Df (x(k) ) f (x(k) ).

The basic Newton algorithm shares the same shortcomings as the basic Gauss-
Newton algorithm, i.e., it can diverge, and the iterations terminate if the derivative
matrix is not invertible.

Newton algorithm for n = 1. The Newton algorithm is easily understood for

n = 1. The iteration is

x(k+1) = x(k) f (x(k) )/f 0 (x(k) ) (18.8)

and is illustrated in figure 18.2. To update x(k) we form the Taylor approximation

f(x; x(k) ) = f (x(k) ) + f 0 (x(k) )(x x(k) )

and set it to zero to find the next iterate x(k+1) . If f 0 (x(k) ) 6= 0, the solution of
f(x; x(k) ) = 0 is given by the right-hand side of (18.8). If f 0 (x(k) ) = 0, the Newton
algorithm terminates with an error.

Example. The function

ex ex
f (x) = (18.9)
ex + ex
has a unique zero at the origin, i.e., the only solution of f (x) = 0 is x = 0. (This
function is called the sigmoid function, and will make another appearance later.)
The Newton iteration started at x(1) = 0.95 converges quickly to the solution x = 0.
With x(1) = 1.15, however, the iterates diverge. This is shown in figures 18.3
and 18.4.
310 18 Nonlinear least squares

f (x) f (x)

1 1

x x
3 2 1 1 2 3 3 2 1 1 2 3

1 1

Figure 18.3 The first iterations in the Newton algorithm for solving f (x) = 0,
for two starting points: x(1) = 0.95 and x(1) = 1.15.

1 1
f (x(k) )

f (x(k) )

0 0

1 1
2 4 6 2 4 6
k k

Figure 18.4 Value of f (x(k) ) versus iteration number k for Newtons method
in the example of figure 18.3, started at x(1) = 0.95 and x(1) = 1.15.
 18.3 Levenberg-Marquardt algorithm 311

18.3 Levenberg-Marquardt algorithm

In this section we describe a variation on the basic Gauss-Newton algorithm (as
well as the Newton algorithm) that addresses the shortcomings described above.
The variation comes directly from ideas we have encountered earlier in this book. It
was first proposed by Kenneth Levenberg and Donald Marquardt, and is called the
Levenberg-Marquardt algorithm. It is also sometimes called the Gauss-Newton
algorithm, since it is a natural extension of the basic Gauss-Newton algorithm
described above.
The main problem with the Gauss-Newton algorithm is that the minimizer
of the approximation kf(x; x(k) )k2 may be far from the current iterate x(k) , in
which case the approximation f(x; x(k) ) f (x) need not hold, which implies that
kf(x; x(k) )k2 kf (x)k2 need not hold. In choosing x(k+1) , then, we have two
objectives: We would like kf(x; x(k) )k2 small, and we would also like kx x(k) k2
small. The first objective is an approximation of what we really want to minimize;
the second objective expresses the idea that we should not move so far that we
cannot trust the affine approximation. This suggests that we should choose x(k+1)
as the minimizer of
kf(x; x(k) )k2 + (k) kx x(k) k2 , (18.10)
where (k) is a positive parameter. (We add an iteration superscript to the param-
eter since it can take different values in different iterations.) For (k) small, we
primarily minimize the first term, the squared norm of the approximation; for (k)
large, we choose x(k+1) near x(k) . (For (k) = 0, this coincides with the next iterate
in the basic Gauss-Newton algorithm.) The second term in (18.10) is sometimes
called a trust penalty term, since it penalizes choices of x that are far from x(k) ,
where we cannot trust the affine approximation.
Computing the minimizer of (18.10) is a regularized least squares problem, and
equivalent to minimizing
 2
Df (x(k) ) Df (x(k) (k) (k)
  
)x f (x ) .

(k)
x (k) (k)
I x

Since (k) is positive, the stacked matrix in this least squares problem has linearly
independent columns, even when Df (x(k) ) does not. It follows that the solution of
the least squares problem exists and is unique.
From the normal equations of the least squares problem we can derive a useful
expression for x(k+1) :
 
Df (x(k) )T Df (x(k) ) + (k) I x(k+1)
 
= Df (x(k) )T Df (x(k) )x(k) f (x(k) ) + (k) x(k)
 
= Df (x(k) )T Df (x(k) ) + (k) I x(k) Df (x(k) )T f (x(k) ),

and therefore
 1
x(k+1) = x(k) Df (x(k) )T Df (x(k) ) + (k) I Df (x(k) )T f (x(k) ). (18.11)
312 18 Nonlinear least squares

The matrix inverse here always exists.

From (18.11), we see that x(k+1) = x(k) only if 2Df (x(k) )T f (x(k) ) = 0, i.e.,
only when the optimality condition (18.3) holds for x(k) . So like the Gauss-Newton
algorithm, the Levenberg-Marquardt algorithm stops only when the optimality
condition (18.3) holds.
The final issue is how to choose the parameter (k) . When (k) is too small,
(k+1)
x can be far enough away from x(k) that kf (x(k+1) )k2 > kf (x(k) k2 can hold,
i.e., our true objective function increases, which is not what we want. When (k)
is too large, the affine approximation is good, so the objective decreases (which is
good), but we have x(k+1) very near x(k) , so the decrease in objective is small, and
it will take many iterations to make progress. We want (k) in between these two
cases, big enough that the approximation holds well enough to get a decrease in
objective, but not much bigger.
Several algorithms can be used to adjust . One simple method forms x(k+1)
using the current value of and checks if the objective has decreased. If it has we
accept the new point and decrease a bit for the next iteration. If the objective
has not decreased, which means is too small, we do not update the point x(k+1) ,
and increase substantially.
These ideas can be formalized as the Levenberg-Marquardt algorithm, given
below.

Algorithm 18.3 Levenberg-Marquardt algorithm for nonlinear least squares

given a differentiable function f : Rn Rm , an initial point x(1) , (1) > 0.
For k = 1, 2, . . . , kmax

1. Form affine approximation at current iterate. Evaluate the Jacobian Df (x(k) )

and define
f(x; x(k) ) = f (x(k) ) + Df (x(k) )(x x(k) ).
2. Compute tentative iterate. Set x(k+1) as minimizer of

kf(x; x(k) )k2 + (k) kx x(k) k2 .

3. Check tentative iterate.

If kf (x(k+1) )k2 < kf (x(k) )k2 , accept iterate and reduce : (k+1) = 0.8(k) .
Otherwise, increase and do not update x: (k+1) = 2(k) and x(k+1) = x(k) .

Stopping criteria. The algorithm is stopped before the maximum number of it-
erations if any of the following conditions hold.

kf (x(k+1) )k2 is small enough. This means we have (almost) solved f (x) = 0
(and also, almost minimized kf (x)k2 ).

kx(k+1) x(k) k is small. This means that the algorithm has (almost) con-
verged, and that the optimality condition (18.3) almost holds.

The test in step 3 fails too many consecutive times, or (k) becomes larger
than some given maximum value.
 18.3 Levenberg-Marquardt algorithm 313

Even when the algorithm converges normally (the second case), we can say very
little for sure about the point computed. The point found may be a minimizer of
kf (x)k2 , or perhaps not. A with many other heuristic algorithms, the point found
is often very useful in applications, even if we cannot be sure that it solves the
nonlinear least squares problem.

Warm start. In many applications a sequence of similar or related nonlinear least

squares problems are solved. In these cases it is common to start the Levenberg-
Marquardt algorithm at the solution of the previously solved problem. If the prob-
lem to be solved is not much different from the previous problem, this can greatly
reduce the number of iterations required to converge. This technique is called warm
starting. It is commonly used in nonlinear model fitting, when multiple models are
fit as we vary a regularization parameter.

Multiple runs. It is common to run the Levenberg-Marquardt algorithm from

several different starting points x(1) . If the final points found by running the algo-
rithm from these different starting points are the same, or very close, it increases
our confidence that we have found a solution of the nonlinear least squares prob-
lem, but we cannot be sure. If the different runs of the algorithm produce different
points, we use the best one found, i.e., the one with the smallest value of kf (x)k2 .
Like the k-means algorithm, which is also a heuristic, the Levenberg-Marquardt
algorithm is widely used in many applications, even if we cannot be sure that it
has found a point that gives the smallest possible residual norm.

Complexity. Each execution of step 1 requires evaluating the derivative matrix

of f . The complexity of this step depends on the particular function f . Each
execution of step 2 requires the solution of a regularized least squares problem.
Using the QR factorization of the stacked matrix this requires 2(m + n)n2 flops
(see 15.5). When m is on the order of n, or larger, this is the same order as mn2 .
When m is much smaller than n, x(k+1) can be computed using the kernel trick
described in 15.5, which requires 2nm2 flops.

18.3.1 Examples
Nonlinear equation. Our first example is the sigmoid function (18.9) from page 309.
We saw in figures 18.3 and 18.4 that the Gauss-Newton method, which reduces to
Newtons method in this case, diverges when the initial value x(1) is 1.15. The
Levenberg-Marquardt algorithm, however, solves this problem. Figure 18.5 shows
the value of the residual f (x(k) ), and the value of (k) , for the Levenberg-Marquardt
algorithm started from x(1) = 1.15 and (1) = 1. It converges to the solution x = 0
in around 10 iterations.

Equilibrium prices. We illustrate algorithm 18.1 with a small instance of the equi-
librium price problem, with supply and demand functions

D(p) = exp E d (log p log pnom ) + dnom ,

314 18 Nonlinear least squares

0.8 1

0.6
f (x(k) )

(k)
0.4 0.5
0.2

0
0
5 10 5 10
k k

Figure 18.5 Values of f (x(k) ) and (k) versus the iteration number k
for the Levenberg-Marquardt algorithm applied to f (x) = (exp(x)
exp(x))/(exp(x)+exp(x)). The starting point is x(1) = 1.15 and (1) = 1.

S(p) = exp (E s (log p log pnom ) + snom ) ,

where E d and E s are the demand and supply elasticity matrices, dnom and snom
are the nominal demand and supply vectors, and the log and exp appearing in
the equations apply to vectors elementwise. Figure 18.6 shows the contour lines of
kf (p)k2 , where f (p) = S(p) D(p) is the excess supply, for

pnom = (2.8, 10), dnom = (3.1, 2.2), snom = (2.2, 0.3)

and    
0.5 0.2 0.5 0.3
Ed = , Es = .
0 0.5 0.15 0.8
Figure 18.7 shows the iterates of the algorithm 18.1, started at p = (3, 9) and
(1) = 1. The values of kf (p(k) )k2 and the regularization parameter (k) versus
iteration k are shown figures 18.8 and figures 18.8.

Location from range measurements. We illustrate algorithm 18.1 with a small

instance of the location from range measurements problem, with five points ai in a
plane, shown in figure 18.9. The range measurements i are the distances of these
points to the true point (1, 1), plus some measurement errors. Figure 18.9 also
shows the level curves of kf (x)k2 , and the point (1.18, 0.82) (marked with a star)
that minimizes kf (x)k2 . (This point is close to, but not equal to, the true value
(1, 1), due to the noise added to the range measurements.) Figure 18.10 shows the
graph of kf (x)k.
We run algorithm 18.1 from three different starting points,

x(1) = (1.8, 3.5), x(1) = (2.2, 3.5), x(1) = (3.0, 1.5),

with (1) = 0.1. Figure 18.11 shows the iterates x(k) for the three starting points.
When started at (1.8, 3.5) (blue circles) or (3.0, 1.5) (brown diamonds) the al-
gorithm converges to (1.18, 0.82), the point that minimizes kf (x)k2 . When the
18.3 Levenberg-Marquardt algorithm 315

10

6
p2

2
2 4 6 8 10
p1

Figure 18.6 Contour lines of the square norm of the excess supply f (p) =
S(p) D(p) for a small example with two commodities. The point marked
with a star are the equilibrium prices, for which f (p) = 0.

10

6
p2

2
2 4 6 8 10
p1

Figure 18.7 Iterates of the Levenberg-Marquardt algorithm started at p =

(3, 9).
316 18 Nonlinear least squares

150

100
kf (p(k) )k2

50

0 5 10 15

0.8

0.6
(k)

0.4

0.2

0
0 5 10 15
k

Figure 18.8 Cost function kf (p(k) k2 and regularization parameter (k) versus
iteration number k in the example of figure 18.7.
18.3 Levenberg-Marquardt algorithm 317

3
x2

0
0 1 2 3 4
x1

Figure 18.9 Contour lines of kf (x)k2 where fi (x) = kx ai k i . The

dots show the points ai , and the point marked with a star is the point that
minimizes kf (x)k2 .

kf (x)k

2
4

2
00 x2
1 2 3 40
x1

Figure 18.10 Graph of kf (x)k in the location from range measurements

example.
318 18 Nonlinear least squares

x2
2

0
0 1 2 3 4
x1

Figure 18.11 Iterates of the Gauss-Newton algorithm started at three dif-

ferent starting points.

algorithm is started at (2.2, 3.5) the algorithm converges to a non-optimal point

(2.98, 2.12) (which gives a poor estimate of the true location (1, 1)). The values
of kf (x(k) )k2 versus iteration k are shown in figure 18.12.
The value of the regularization parameter (k) during the iteration is shown
in figure 18.12. As can be seen from this figure, in the first run of the algorithm
(blue circles), (k) is increased in the third iteration. Correspondingly, x(3) =
x(4) in figure 18.12. For the second starting point (red squares) (k) decreases
monotonically. For the third starting point (brown diamonds) (k) increases in
iterations 2 and 4.

18.4 Nonlinear model fitting

The Levenberg-Marquardt algorithm is widely used for nonlinear model fitting.

Example. Figure (18.13) shows a nonlinear model fitting example. The model is
an exponentially decaying sinusoid

f(x; ) = 1 e2 x cos(3 x + 4 ),

with four parameters 1 , . . . , 4 . We fit this model to 60 points (xi , yi ) by minimiz-

ing the sum of the squared residuals (18.4) over the four parameters.

Orthogonal distance regression. Suppose

f(x; ) = 1 f1 (x) + + p fp (x)

18.4 Nonlinear model fitting 319

3
kf (x(k) )k2

0
1 2 3 4 5 6 7 8 9 10

0.3

0.2
(k)

0.1

0
1 2 3 4 5 6 7 8 9 10
k

Figure 18.12 Cost function kf (x(k) )k2 and regularization parameter (k)
versus iteration number k for the three starting points.
320 18 Nonlinear least squares

f(x; )

Figure 18.13 Least squares fit of a function f(x; ) = 1 e2 x cos(3 x + 4 ) to

60 points (xi , yi ).

is a model from Rn R, with differentiable basis functions fi . The model that

minimizes the sum of the squared distances of N points (xi , yi ) to the graph of f
can be computed by solving a nonlinear least squares problem
N N
(f(ui ; ) yi )2 +
X X
minimize kui xi k2
i=1 i=1

with variables 1 , . . . , p , and u1 , . . . , uN . Figure 18.14 shows a cubic polynomial

fitted to 25 points using this method. The open circles are the points (xi , yi ). The
small circles on the graph of the polynomial are the points (ui , f(ui ; )).

18.5 Nonlinear least squares classification

In this section we describe a nonlinear extension of the least squares classification
method discussed in chapters 14 and 15, that typically out-performs the basic least
squares classifier in practice.
The Boolean classifier of chapter 14 fits a linearly parametrized function

f(x) = 1 f1 (x) + + p fp (x)

to the data points (xi , yi ), i = 1, . . . , N , where yi {1, 1}, using linear least
squares. The parameters 1 , . . . , p are chosen to minimize the sum squares ob-
18.5 Nonlinear least squares classification 321

f(x; )

Figure 18.14 The solid line minimizes the sum of the squares of the orthog-
onal distances of points to the graph of the polynomial.

jective
N
(f(xi ) yi )2 ,
X
(18.12)
i=1

plus, optionally, a regularization term. This (hopefully) results in f(xi ) yi , which

is roughly what we want. We can think of f(x) as the continuous prediction of the
Boolean outcome y. The classifier itself is given by f(x) = sign(f(x)); this is the
Boolean prediction of the outcome.
Instead of the sum square prediction error for the continuous prediction, con-
sider the sum square prediction error for the Boolean prediction,
N N
(f(xi ) yi )2 = (sign(f(xi )) yi )2 .
X X
(18.13)
i=1 i=1

This is 4 times the number of classification errors we make on the training set. To
see this, we note that when f(xi ) = yi , which means that a correct prediction was
made on the ith data point, we have (f(xi ) yi )2 = 0. When f(xi ) 6= yi , which
means that an incorrect prediction was made on the ith data point, one of the
values is +1 and the other is 1, so we have (f(xi ) yi )2 = 4.
The objective (18.13) is what we really want; the least squares objective (18.12)
is a surrogate for what we want. But we cannot use the Levenberg-Marquardt algo-
rithm to minimize the objective (18.13), since the sign function is not differentiable.
To get around this, we replace the sign function with a differentiable approximation,
for example the sigmoid function

eu eu
(u) = , (18.14)
eu + eu
322 18 Nonlinear least squares

(u)

u
4 2 2 4

Figure 18.15 The sigmoid function .

shown in figure 18.15. We choose by solving the nonlinear least squares problem
of minimizing
N
((f(xi )) yi )2 ,
X
(18.15)
i=1

using the Levenberg-Marquardt algorithm. (We can also add regularization to

this objective.) Minimizing the nonlinear least squares objective (18.15) is a good
approximation for choosing the parameter vector so as to minimize the number
of classification errors made on the training set.

Loss function interpretation. We can interpret the objective functions (18.12),

(18.13), and (18.15) in terms of loss functions that depend on the continuous
prediction f(xi ) and the outcome yi . Each of the three objectives has the form
N
`(f(xi ), yi ),
X

i=1

where `(u, y) is a loss function. For the linear least squares objective (18.12),
the loss function is `(u, y) = (u y)2 . For the nonlinear least squares objective
with the sign function (18.13), the loss function is `(u, y) = (sign(u) y)2 . For the
differentiable nonlinear least squares objective (18.13), the loss function is `(u, y) =
((u) y)2 . Roughly speaking, the loss function `(u, y) tells us how bad it is to
have f(xi ) = u when y = yi .
Since the outcome y takes on only two values, 1 and +1, we can plot the loss
functions as functions of u for these two values of y. Figure 18.16 shows these three
functions, with the value for y = 1 in the left column and the value for y = +1
in the right column. We can see that all three loss functions discourage prediction
errors, since their values are higher for sign(u) 6= y than when sign(u) = y.
The loss function for nonlinear least squares classification with the sign function
(shown in the middle row) assesses a cost of 0 for a correct prediction and 4 for
an incorrect prediction. The loss function for nonlinear least squares classification
 18.5 Nonlinear least squares classification 323

Prediction Prediction
Outcome y = +1 y = 1 Total Outcome y = +1 y = 1 Total
y = +1 5627 296 5923 y = +1 945 35 980
y = 1 148 53929 54077 y = 1 40 8980 9020
All 5775 54225 60000 All 985 9015 10000

Table 18.1 Confusion matrices for a Boolean classifier to recognize the digit
zero. The table on the left is for the training set. Thet table on the right is
for the test set.

with the sigmoid function (shown in the bottom row) is a smooth approximation
of this.

18.5.1 Handwritten digit classification

We apply nonlinear least squares classification on the MNIST set of handwritten

digits used in chapter 14. We first consider the Boolean problem of recognizing the
digit 0. We use linear features, i.e.,
f(x) = xT + v,
where x is the 493-vector of pixel intensities. To determine the parameters v and
we solve the nonlinear least squares problem
N
X
minimize ((xTi + v) yi )2 + kk2 , (18.16)
i=1

where is the sigmoid function (18.14) and is a positive regularization parameter.

(This is the regularization parameter in the classification problem; it has no
relation to (k) in the iterates of the Levenberg-Marquardt algorithm.)
Figure 18.17 shows the classification error on the data and test sets as a function
of the regularization parameter . For = 100, the classification errors on data
and training set are about 0.7%. This is less than half the 1.6% error of the Boolean
least squares classifier that used the same features, discussed in chapter 14. This
improvement in performance, by more than a factor of two, comes from minimizing
an objective that is closer to what want (i.e., the number of prediction errors on
the training set) than the surrogate linear least squares objective. The confusion
matrices for the training set and test set are given in table 18.1.
Figure 18.18 shows the distribution of the values of f(xi ) for the two classes of
the data set. The bottom part of the figure shows the true positive, false positive,
and total error rate versus the threshold , when a skewed decision function f(x) =
sign(f(x) ) is used.

Convergence of Levenberg-Marquardt algorithm. The Levenberg-Marquardt al-

gorithm is used to compute the parameters in the nonlinear least squares classifier.
324 18 Nonlinear least squares

(u + 1)2 (u 1)2
4 4

3 3

2 2

1 1

u u
3 1 1 3 3 1 1 3

(sign(u) + 1)2 (sign(u) + 1)2

4 4

3 3

2 2

1 1

u u
3 1 1 3 3 1 1 3

((u) + 1)2 ((u) 1)2

4 4

3 3

2 2

1 1

u u
3 1 1 3 3 1 1 3

Figure 18.16 The loss functions `(u, y) for linear least squares classification
(top), nonlinear least squares classification with the sign function (middle),
and nonlinear least squares classification with the sigmoid function (bottom).
The left column shows `(u, 1) and the right columns shows `(u, +1).
18.5 Nonlinear least squares classification 325

Train
Classification error (%) Test

0
109 106 103 100 103 106

Figure 18.17 Boolean classification error in percent versus .

Positive
0.08 Negative

0.06
Fraction

0.04

0.02

0
10 8 6 4 2 0 2 4 6 8
f(xi )

Figure 18.18 The distribution of the values of f(xi ) used in the Boolean
classifier (14.1) for recognizing the digit zero. The function f was computed
by solving the nonlinear least squares problem (18.15).
326 18 Nonlinear least squares

102 Train
Test

Classification error (%)

101

100

0 2 4 6 8 10
Iteration

Figure 18.19 Training and test error versus Levenberg-Marquardt iteration

for = 100.

Prediction
Outcome y = +1 y = 1 Total
y = +1 967 13 980
y = 1 11 9009 9020
All 978 9022 10000

Table 18.2 Confusion matrix on the test set for the Boolean classifier to
recognize the digit zero after addition of 5000 new features.

In this example the algorithm takes several tens of iterations to converge, i.e., until
the stopping criterion for the nonlinear least squares problem is satisfied. But in
this application we are more interested in the performance of the classifier, and
not minimizing the objective of the nonlinear least squares problem. Figure 18.19
shows the classification error of the classifier (on the train and test data sets) with
parameter (k) , the kth iterate of the Levenberg-Marquardt algorithm. We can
see that the classification errors reach their final values of 0.7% after just a few
iterations. This phenomenon is very typical in nonlinear data fitting problems.

Feature engineering. After adding the 5000 random features used in chapter 14,
the training and test errors we obtain the classification errors shown in figure 18.20.
The error on the training set is zero for small . For = 1000, the error on the
test set is 0.24%, with the confusion matrix in table 18.2. The distribution of f(xi )
on the training set in figure 18.21 shows why the training error is zero.
Figure 18.22 shows the classification errors versus Levenberg-Marquardt iter-
18.5 Nonlinear least squares classification 327

Train
Classification error (%) 0.6 Test

0.4

0.2

102 100 102 104 106

Figure 18.20 Boolean classification error in percent versus , after adding

5000 random features.

0.06 Positive
Negative

0.04
Fraction

0.02

0
10 8 6 4 2 0 2 4 6 8
f(xi )

Figure 18.21 The distribution of the values of f(xi ) used in the Boolean
classifier (14.1) for recognizing the digit zero, after addition of 5000 new
features. The function f was computed by solving the nonlinear least squares
problem (18.15).
328 18 Nonlinear least squares

102 Train
Test

Classification error (%)

101

100

101

102

103
0 2 4 6 8 10
Iteration

Figure 18.22 Training and test error versus Levenberg-Marquardt iteration

for = 1000.

ation, if we start the Levenberg-Marquardt algorithm with = 0, v = 0 (This

implies that the values computed in the first iteration are the coefficients of the
linear least squares classifier.) The error on the training set is exactly zero at itera-
tion 5. The error on the test set is almost equal to its final value after one iteration.

Multi-class classifier. Next we apply the nonlinear least squares method to the
multi-class classification of recognizing the ten digits in the MNIST data set. For
each digit k, we compute a Boolean classifier fk (x) = xT k + vk by solving a
regularized nonlinear least squares problem (18.16). The same value of is used
in the ten nonlinear least squares problems. The Boolean classifiers are combined
into a multi-class classifier

f(x) = argmax (xT k + vk ).

k=1,...,10

Figure 18.23 shows the classification errors versus . The test set confusion matrix
(for = 1) is given in table 18.3. The classification error on the test set is 7.6%,
down from the 13.9% error we obtained for the same set of features with the least
squares method of chapter 14.

Feature engineering. Figure 18.24 shows the error rates when we add the 5000
randomly generated features. The train and test error rates are now 0.02% and 2%.
The test set confusion matrix for = 1000 is given in table 18.4. This classifier
has matched human performance in classifying digits correctly. Further, or more
sophisticated, feature engineering can bring the test performance well below what
humans can achieve.
18.5 Nonlinear least squares classification 329

10
Classification error (%)

7
Train
Test
6
105 103 101 101 103

Figure 18.23 Multiclass classification error in percent versus .

Prediction
Digit 0 1 2 3 4 5 6 7 8 9 Total
0 964 0 0 2 0 2 5 3 3 1 980
1 0 1112 4 3 0 1 4 1 10 0 1135
2 5 5 934 13 7 3 13 10 38 4 1032
3 3 0 19 926 1 21 2 8 21 9 1010
4 1 2 4 2 917 0 7 1 10 38 982
5 10 2 2 31 10 782 17 7 23 8 892
6 8 3 3 1 5 20 910 1 7 0 958
7 2 6 25 5 11 5 0 947 4 23 1028
8 13 10 4 18 16 27 8 9 865 4 974
9 8 6 0 12 43 11 1 19 23 886 1009
All 1014 1146 995 1013 1010 872 967 1006 1004 973 10000

Table 18.3 Confusion matrix for test set. The error rate is 7.6%.
.
330 18 Nonlinear least squares

Train
Test
3

Classification error (%) 2

102 100 102 104

Figure 18.24 Multiclass classification error in percent versus after adding

5000 random features.

Prediction
Digit 0 1 2 3 4 5 6 7 8 9 Total
0 972 1 1 1 0 1 1 1 2 0 980
1 0 1124 2 2 0 0 3 1 3 0 1135
2 5 0 1006 1 3 0 2 6 9 0 1032
3 0 0 3 986 0 5 0 3 7 6 1010
4 0 0 4 1 966 0 4 1 0 6 982
5 2 0 2 5 2 875 5 0 1 0 892
6 7 2 0 1 3 2 941 0 2 0 958
7 1 7 6 1 2 0 0 1003 3 5 1028
8 3 0 0 4 4 5 3 4 949 2 974
9 2 5 0 5 6 4 1 6 2 978 1009
All 992 1139 1024 1007 986 892 960 1025 978 997 10000

Table 18.4 Confusion matrix for test set after adding 5000 features. The
error rate is 2.0%.
 Chapter 19

Constrained nonlinear least

squares

In this chapter we consider an extension of the nonlinear least squares problem

that includes nonlinear constraints. Like the problem of solving a set of nonlinear
equations, or finding a least squares approximate solution to a set of nonlinear
equations, the constrained nonlinear least squares problem is in general hard to
solve exactly. We we describe a heuristic algorithm that often works well in practice.

19.1 Constrained nonlinear least squares

In this section we consider an extension of the nonlinear least squares problem (18.2)
that includes equality constraints:

minimize kf (x)k2
(19.1)
subject to g(x) = 0,

where the n-vector x is the variable to be found. Here f (x) is an m-vector, and
g(x) is a p-vector. We sometimes write out the components of f (x) and g(x), to
express the problem as

minimize f1 (x)2 + + fm (x)2

subject to gi (x) = 0, i = 1, . . . , p.

We refer to fi (x) as the ith (scalar) residual, and gi (x) = 0 as the ith (scalar)
equality constraint. When the functions f and g are affine, the equality constrained
nonlinear least squares problem (19.1) reduces to the (linear) least squares problem
with equality constraints from chapter 16.
We say that a point x is feasible for the problem (19.1) if it satisfies g(x) = 0. A
point x is a solution of the problem (19.1) if it is feasible and has smallest objective
among all feasible points, i.e., if whenever g(x) = 0, we have kf (x)k2 kf (x)k2 .
332 19 Constrained nonlinear least squares

Like the nonlinear least squares problem, or solving a set of nonlinear equations,
the constrained nonlinear least squares problem is in general hard to solve exactly.
But the Levenberg-Marquardt algorithm for solving the (unconstrained) nonlinear
least squares problem (18.2) can be leveraged to handle the problem with equality
constraints. We will describe a basic algorithm below, the penalty algorithm, and
a variation on it that works much better in practice, the augmented Lagrangian al-
gorithm. These algorithms are heuristics for (approximately) solving the nonlinear
least squares problem (19.1).

Linear equality constraints. One special case of the constrained nonlinear least
squares problem (19.1) is when the constraint function g is affine, in which case
the constraints g(x) = 0 can be written Cx = d for some p n matrix C and
a p-vector d. In this case the problem (19.1) is called a nonlinear least squares
problem with linear equality constraints. It can be (approximately) solved by
the Levenberg-Marquardt algorithm described above, by simply adding the linear
equality constraints to the linear least squares problem that is solved in step 2.
The more challenging problem is the case when g is not affine.

19.1.1 Optimality condition

Using Lagrange multipliers (see C.3) we can derive a condition that any solu-
tion of the constrained nonlinear least squares problem (19.1) must satisfy. The
Lagrangian for the problem (19.1) is

L(x, z) = kf (x)k2 + z1 g1 (x) + + zp gp (x) = kf (x)k2 + g(x)T z, (19.2)

where the p-vector z is the vector of Lagrange multipliers. The method of Lagrange
multipliers tells us that for any solution x of (19.1), there is a set of Lagrange
multipliers z that satisfy
L L
(x, z) = 0, i = 1, . . . , n, (x, z) = 0, i = 1, . . . , p
xi zi

(provided the rows of Dg(x) are linearly independent). The p-vector z is called an
optimal Lagrange multiplier.
The second set of equations can be written as gi (x) = 0, i = 1, . . . , p, in vector
form
g(x) = 0, (19.3)
i.e., x is feasible, which we already knew. The first set of equations can be written
in vector form as
2Df (x)T f (x) + Dg(x)T z = 0. (19.4)
This equation is the extension of the condition (18.3) for the unconstrained non-
linear least squares problem (18.2). The equation (19.4), together with (19.3), i.e.,
x is feasible, form the optimality conditions for the problem (19.1).
If x is a solution of the constrained nonlinear least squares problem (19.1), then
it satisfies the optimality condition (19.4) for some Lagrange multiplier vector z
 19.2 Penalty algorithm 333

(provided the rows of Dg(x) are linearly independent). So x and z satisfy the
optimality conditions.
These optimality conditions are not sufficient, however; there can be choices of
x and z that satisfy them, but x is not a solution of the constrained nonlinear least
squares problem.

19.2 Penalty algorithm

We start with the observation (already made on page 266) that the equality con-
strained problem can be thought of as a limit of a bi-objective problem with ob-
jectives kf (x)k2 and kg(x)k2 , as the weight on the second objective increases to
infinity. Let be a positive number, and consider the composite objective

kf (x)k2 + kg(x)k2 . (19.5)

This can be (approximately) minimized using the Levenberg-Marquardt algorithm

applied to
  2
f (x)

. (19.6)
g(x)
By minimizing the composite objective (19.5), we do not insist that g(x) is zero,
but we assess a cost or penalty kg(x)k2 on the residual. If we solve this for large
enough , we should obtain a choice of x for which g(x) is very small, and kf (x)k2
is small, i.e., an approximate solution of (19.1). The second term kg(x)k2 is a
penalty imposed on choices of x with nonzero g(x).
Minimizing the composite objective (19.5) for an increasing sequence of values
of is known as the penalty algorithm.

Algorithm 19.1 Penalty agorithm for constrained nonlinear least squares

given differentiable functions f : Rn Rm and g : Rn Rp , and an initial
point x(1) . Set (1) = 1.
For k = 1, 2, . . . , kmax

1. Solve unconstrained nonlinear least squares problem. Set x(k+1) to be the (ap-
proximate) minimizer of

kf (x)k2 + (k) kg(x)k2

using the Levenberg-Marquardt algorithm, starting from initial point x(k) .

2. Update (k) : (k+1) = 2(k) .

The penalty algorithm is stopped early if kg(x(k) )k is small enough, i.e., the equality
constraint is almost satisfied.
The penalty algorithm is simple and easy to implement, but has an important
drawback: The parameter k rapidly increases with iterations (as it must, to drive
334 19 Constrained nonlinear least squares

g(x) to zero). When the Levenberg-Marquardt algorithm is used to minimize (19.5)

for very high values of , it can take a large number of iterations or simply fail.
The augmented Lagrangian algorithm described below gets around this drawback,
and gives a much more reliable algorithm.
We can connect the penalty algorithm iterates to the optimality condition (19.4).
The iterate x(k) (almost) satisfies the optimality condition for minimizing (19.5),

2Df (x(k+1) )T f (x(k+1) ) + 2(k) Dg(x(k+1) )T g(x(k+1) ) = 0.

Defining
z (k+1) = 2(k) g(x(k+1) )
as our estimate of a suitable Lagrange multiplier in iteration k + 1, we see that the
optimality condition (19.4) (almost) holds for x(k+1) and z (k+1) . (The feasibility
condition g(x(k) ) = 0 only holds in the limit as k .)

19.3 Augmented Lagrangian algorithm

The augmented Lagrangian algorithm is a modification of the penalty algorithm
that addresses the difficulty associated with the penalty parameter (k) becoming
very large. It was proposed by Magnus Hestenes and Michael Powell in the 1960s.

Augmented Lagrangian. The augmented Lagrangian for the problem (19.1), with
parameter > 0, is defined as

L (x, z) = L(x, ) + kg(x)k2 = kf (x)k2 + g(x)T z + kg(x)k2 . (19.7)

This is the Lagrangian, augmented with the new term kg(x)k2 ; alternatively, it
can be interpreted as the composite objective function (19.5) used in the penalty
algorithm, with the Lagrange multiplier term g(x)T z added.
The augmented Lagrangian (19.7) is also the ordinary Lagrangian associated
with the problem
minimize kf (x)k2 + kg(x)k2
subject to g(x) = 0.
This problem is equivalent to the original constrained nonlinear least squares prob-
lem (19.1): A point x is a solution of one if and only if it a solution of the other.
(This follows since the term kg(x)k2 is zero for any feasible x.)

Minimizing the augmented Lagrangian. In the augmented Lagrangian algorithm

we minimize the augmented Lagrangian over the variable x for a sequence of values
of and z. We show here how this can be done using the Levenberg-Marquardt
algorithm. We first establish the identity

L (x, z) = kf (x)k2 + kg(x) + z/(2)k2 kz/(2)k2 . (19.8)

19.3 Augmented Lagrangian algorithm 335

We expand the second term on the right-hand side to get

kg(x) + z/(2)k2
= kg(x)k2 + 2g(x)T (z/(2)) + kz/(2)k2
= g(x)T z + kg(x)k2 + kz/(2)k2 .
Substituting this into the right-hand side of (19.8) verifies the identity.
When we minimize L (x, z) over the variable x, the term kz/(2)k2 in (19.8)
is a constant (i.e., does not depend on x), and does not affect the choice of x. It
follows that we can minimize L (x, z) over x by minimizing the function
kf (x)k2 + kg(x) + z/(2)k2 , (19.9)
which in turn can be expressed as
  2

f (x)
. (19.10)
g(x) + z/(2 )

This can be be (approximately) minimized using the Levenberg-Marquardt algo-

rithm.
Any minimizer x of L (x, z) (or equivalently, (19.9)) satisfies the optimality
condition
0 = 2Df (x)T f (x) + 2Dg(x)T (g(x) + z/(2))
= 2Df (x)T f (x) + Dg(x)T (2g(x) + z).
From this equation we can observe that if x minimizes the augmented Lagrangian
and is also feasible (i.e., g(x) = 0), then it satisfies the optimality condition (19.4)
with the vector z as the Lagrange multiplier. The bottom equation also suggests
a good choice for updating the Lagrange multiplier vector z if x is not feasible. In
this case the choice
z = z + 2g(x) (19.11)
satisfies the optimality condition (19.4) with x and z.
The augmented Lagrangian algorithm alternates between minimizing the aug-
mented Lagrangian (approximately, using the Levenberg-Marquardt algorithm),
and updating the parameter z (our estimate of a suitable Lagrange multiplier) us-
ing the suggestion (19.11) above. The penalty parameter is increased only when
needed, when kg(x)k does not sufficiently decrease.

Algorithm 19.2 Augmented Lagrangian algorithm

given differentiable functions f : Rn Rm and g : Rn Rp , and an initial
point x(1) . Set z (1) = 0, (1) = 1.
For k = 1, 2, . . . , kmax

1. Solve unconstrained nonlinear least squares problem. Set x(k+1) to be the (ap-
proximate) minimizer of

kf (x)k2 + (k) kg(x) + z (k) /(2(k) )k2

using Levenberg-Marquardt algorithm, starting from initial point x(k) .
336 19 Constrained nonlinear least squares

2. Update z (k) .
z (k+1) = z (k) + 2(k) g(x(k+1) ).
3. Update (k) .

(k) kg(x(k+1) )k < 0.25kg(x(k) )k
(k+1) =
2(k) kg(x(k+1) )k 0.25kg(x(k) )k.

The augmented Lagrangian algorithm is stopped early if g(x(k) ) is very small.

Note that due to our particular choice in how z (k) is updated, the iterate x(k+1)
(almost) satisfies the optimality condition (19.4) with z (k+1) .
The augmented Lagrangian algorithm is not much more complicated than the
penalty algorithm, but it works much better in practice. In part this is because
the penalty parameter (k) does not need to increase as much as the algorithm
proceeds.

Example. We consider an example with two variables and

 
x1 + exp(x2 )
f (x1 , x2 ) = , g(x1 , x2 ) = x1 + x31 + x2 + x22 .
x21 + 2x2 + 1

Figure 19.1 shows the contour lines of the cost function kf (x)k2 (solid line) and
the constraint function g(x) (dashed line). The point x = (0, 0) is optimal with
corresponding Lagrange multiplier z = 2. One can verify that g(x) = 0 and
    
1 0 1 1
2Df (x)T f (x) + Dg(x)T z = 2 2 = 0.
1 2 1 1

The circle at x = (0.666, 0.407) indicates the position of the unconstrained

minimizer of kf (x)k2 .
The augmented Lagrangian algorithm is started from the point x(1) = (0.5, 0.5).
Figure 19.2 illustrates the first six iterations. The solid lines are the contour lines
for L (x, z (k) ), the augmented Lagrangian with the current value of the Lagrange
multiplier. For comparison, we also show in figure 19.3 the first six iterations of
the penalty algorithm, started from the same point. The solid lines are the contour
lines of kf (x)k2 + (k) kg(x)k2 .
In figure 19.4 we show how the algorithms converge. The horizontal axis is
the cumulative number of Levenberg-Marquardt iterations. Each of these requires
the solution of one linear least squares problem (minimizing (19.10) and (19.6),
respectively). The two lines show the absolute value of the feasibility residual
|g(x(k) )|, and the norm of the optimality condition residual,

k2Df (x(k) )T f (x(k) ) + Dg(x(k) )T z (k) k.

The vertical jumps in the optimality condition norm occur in steps 2 and 3 of the
augmented Lagrangian algorithm, and in step 2 of the penalty algorithm, when the
parameters and z are updated.
Figure 19.5 shows the value of the penalty parameter versus the cumulative
number of Levenberg-Marquardt iterations in the two algorithms.
 19.4 Nonlinear control 337

g(x) = 1

g(x) = 0
x2

0 g(x) = 1

1
1 0 1
x1

Figure 19.1 Contour lines of the cost function kf (x)k2 (solid line) and the
constraint function g(x) (dashed line) for a nonlinear least squares problem
in two variables with one equality constraint.

19.4 Nonlinear control

A nonlinear dynamical system has the form of an iteration

xk+1 = f (xk , uk ), k = 1, 2, . . . , N,

where the n-vector xk is the state, and the m-vector uk is the input or control, at
time period k. The function f : Rn+m Rn specifies what the next state is, as a
function of the current state and the current input. When f is an affine function,
this reduces to a linear dyanimcal system.
In nonlinear control, the goal is to choose the inputs u1 , . . . , uN 1 to achieve
some goal for the state and input trajectories. In many problems the initial state
x1 is given, and the final state xN is specified. Subject to these constraints, we may
wish the control inputs to be small and smooth, which suggests that we minimize

N
X N
X 1
kuk k2 + kuk+1 uk k2 ,
k=1 k=1

where > 0 is a parameter used to trade off input size and smoothness. (In many
nonlinear control problems the objective also involves the state trajectory.)
We can formulate the nonlinear control problem, with a norm squared objective
that involves the state and input, as a large constrained least problem, and then
solve it using the augmented Lagrangian algorithm. We illustrate this with a
specific example.
338 19 Constrained nonlinear least squares

(1) = 1, z (1) = 0 (2) = 2, z (2) = 0.893

0.5 0.5

0 0
x(3)
(2)
x

0.5 0.5
0.5 0 0.5 0.5 0 0.5

(3) = 4, z (3) = 1.569 (4) = 4, z (4) = 1.898

0.5 0.5

0 0
x(4) x(5)

0.5 0.5
0.5 0 0.5 0.5 0 0.5

(5) = 4, z (5) = 1.976 (6) = 4, z (6) = 1.994

0.5 0.5

0 0
x(6) x(7)

0.5 0.5
0.5 0 0.5 0.5 0 0.5

Figure 19.2 First six iterations of the augmented Lagrangian algorithm.

19.4 Nonlinear control 339

(1) = 1 (2) = 2
0.5 0.5

0 0
x(3)
x(2)

0.5 0.5
0.5 0 0.5 0.5 0 0.5

(3) = 4 (4) = 8
0.5 0.5

0 0
(4)
x x(5)

0.5 0.5
0.5 0 0.5 0.5 0 0.5

(5) = 16 (6) = 32
0.5 0.5

0 0
x(6) x(7)

0.5 0.5
0.5 0 0.5 0.5 0 0.5

Figure 19.3 First six iterations of the penalty algorithm.

340 19 Constrained nonlinear least squares

101 Feasibility
Opt. cond.
100

101
Residual

102

103

104

105

106
0 20 40 60 80 100 120 140

101 Feasibility
Opt. cond.
100

101
Residual

102

103

104

105

106
0 20 40 60 80 100 120 140
Cumulative Levenberg-Marquardt iterations

Figure 19.4 Feasibility and optimality condition error versus the cumulative
number of Levenberg-Marquardt iterations in the augmented Lagrangian
algorithm (top) and the penalty algorithm (bottom).
19.4 Nonlinear control 341

Aug. Lag.
104 Penalty

Penalty parameter
103

102

101

100
0 20 40 60 80 100 120 140
Cumulative Levenberg-Marquardt iterations

Figure 19.5 Penalty parameter versus cumulative number of Levenberg-

Marquardt iterations in the agumented Lagrangian algorithm and the
penalty algorithm.

(p1 , p2 )

Figure 19.6 Simple model of a car.

Control of a car. Consider a car with position p = (p1 , p2 ) and orientation (angle)
. The car has wheelbase (length) L, steering angle , and speed s (which can be
negative, meaning the car moves in reverse). This is illustrated in figure 19.6.
The wheelbase L is a known constant; all of the other quantities p, , , and s
are functions of time. The dynamics of the car motion are given by the differential
equations
dp1
(t) = s(t) cos (t),
dt
dp2
(t) = s(t) sin (t),
dt
d
(t) = (s(t)/L) tan (t).
dt
Here we assume that the steering angle is always less than 90 , so the tangent
term in the last equation makes sense. The first two equations state that the car is
moving in the direction (t) (its orientation) at speed s(t). The last equation gives
342 19 Constrained nonlinear least squares

the change in orientation as a function of the car speed and the steering angle. For
a fixed steering angle and speed, the car moves in a circle.
We can control the speed s and the steering angle ; the goal is to move the car
over some time period from a given initial position and orientation to a specified
final position and orientiation.
We now discretize the equations in time. We take a small time interval h, and
obtain the approximations

p1 (t + h) p1 (t) + hs(t) cos (t),

p2 (t + h) p2 (t) + hs(t) sin (t),
(t + h) (t) + h(s(t)/L) tan (t).

We will use these approximations to derive nonlinear state equations for the car
motion, with state xk = (p1 (kh), p2 (kh), (kh)) and input uk = (s(kh), (kh)). We
have
xk+1 = f (xk , uk ),
with
cos(xk )3
f (xk , uk ) = xk + h(uk )1 sin(xk )3 .
(tan(uk )2 )/L
We now consider the nonlinear optimal control problem
N 1
NP
kuk k2 + kuk+1 uk k2
P
minimize
k=1 k=1
subject to x2 = f (0, u1 ) (19.12)
xk+1 = f (xk , uk ), k = 2, . . . , N 1
xfinal = f (xN , uN ),

with variables u1 , . . . , uN , and x2 , . . . , xN .

Figure 19.7 shows solutions for

L = 0.1, N = 50, h = 0.1, = 10,

and different values of xfinal . They are computed using the augmented Lagrangian
algorithm. The algorithm is started at the same starting point for each example.
The starting point for the input variables uk is randomly chosen, the starting point
for the states xk is zero.
 19.4 Nonlinear control 343

xfinal = (0, 1, 0) xfinal = (0, 1, /2)

1
1

0.5
0.5

0 0

0.5 0 0.5 0.5 0 0.5

xfinal = (0, 0.5, 0) xfinal = (0.5, 0.5, /2)

0.6
0.4

0.4
0.2

0.2 0

0 0.2

0.6 0.4 0.2 0 0.2 0.2 0 0.2 0.4 0.6 0.8

Figure 19.7 Solution trajectories of 19.12 for different end states xfinal . The
outline of the car shows the position (p1 (kh), p2 (kh)), orientation (kh), and
the steering angle (kh) at time kh.
344 19 Constrained nonlinear least squares

Speed
0.5 0.5
Angle
uk

uk
0 0

0.5 0.5

0 10 20 30 40 50 0 10 20 30 40 50
k k

0.5 0.5
uk

uk

0 0

0.5 0.5

0 10 20 30 40 50 0 10 20 30 40 50
k k

Figure 19.8 The two inputs (speed and steering angle) for the trajectories
in figure 19.7.
 19.4 Nonlinear control 345

102 102
Feasibility
Opt. cond.
100 100
Residual

Residual
102 102

104 104

106 106

0 100 200 300 400 0 50 100 150 200 250 300 350
Cumulative LM iterations Cumulative LM iterations

102 102

100 100
Residual

Residual

102 102

104 104

106 106

0 200 400 600 800 0 500 1,000 1,500

Cumulative LM iterations Cumulative LM iterations

Figure 19.9 Feasibility and optimality condition residuals in the augmented

Lagrangian algorithm for computing the trajectories in figure 19.7.
346 19 Constrained nonlinear least squares
Appendices
Appendix A

Notation

Vectors

x1
..
. n-vector with entries x1 , . . . , xn .
xn
(x1 , . . . , xn ) n-vector with entries x1 , . . . , xn .
xi The ith element of a vector x.
xr:s Subvector with entries from r to s.
0 Vector with all entries zero.
1 Vector with all entries one.
ei The ith standard basis vector.
xT y Inner product of vectors x and y.
kxk Norm of vector x.
rms(x) RMS value of a vector x.
avg(x) Average of entries of a vector x.
std(x) Standard deviation of a vector x.
dist(x, y) Distance between vectors x and y.
6 (x, y) Angle between vectors x and y.
xy Vectors x and y are orthogonal.

Matrices

X11 X1n
.. .. m n matrix with entries X , . . . , X .
. . 11 mn
Xm1 Xmn

Xij The i, jth element of a matrix X.

Xr:s,p:q Submatrix with row range r to s and column range p to q.
0 Matrix with all entries 0.
I Identity matrix.
XT Transpose of matrix X.
350 A Notation

Xk (Square) matrix X to the kth power.

X 1 Inverse of (square) matrix X.
X T Inverse of transpose of matrix X.
X Pseudo-inverse of matrix X.
diag(x) Diagonal matrix with diagonal entries x1 , . . . , xn .

Functions and derivatives

f :AB f is a function on the set A into the set B.

f (z) Gradient of function f : Rn R at z.
Df (z) Derivative (Jacobian) matrix of function f : Rn Rm at z.

Ellipsis notation

In this book we use standard mathematical ellipsis notation in lists and sums. We
write k, . . . , l to mean the list of all integers from k to l. For example, 3, . . . , 7
means 3, 4, 5, 6, [Link] notation is used to describe a list of numbers or vectors,
or in sums, as in i=1,...,n ai , which we also write as a1 + + an . Both of these
mean the sum of the n terms a1 , a2 , . . . , an .

Sets

In a few places in this book we encounter the mathematical concept of sets. The
notation {a1 , . . . , an } refers to a set with elements a1 , . . . , an . This is not the
same as the vector with elements a1 ,. . . , an , which is denoted (a1 , . . . , an ). For
sets the order does not matter, so, for example, we have {1, 2, 6} = {6, 1, 2}. We
can also specify a set by giving conditions that its entries must satisfy, using the
notation {x | condition(x)}, which means the set of x that satisfy the condition,
which depends on x. We say that a set contains its elements, or that the elements
are in the set, using the symbol , as in 2 {1, 2, 6}. The symbol 6 means not in,
or not an element of, as in 3 6 {1, 2, 6}.
P We can use sets to describe a sum over some elements in a list. The notation
PiS xi means the sum over all xi for which i is in the set S. As an example,
i{1,2,6} ai means a1 + a2 + a6 .
A few sets have specific names: R is the set of real numbers (or scalars), and
Rn is the set of all n-vectors. So R means that is a number, and x Rn
means that x is an n-vector.
Appendix B

Complexity

Here we summarize approximate complexities or flop counts of various operations

and algorithms encountered in the book. We drop terms of lower order. When
the operands or arguments are sparse, and the operation or algorithm has been
modified to take advantage of sparsity, the flop counts can be dramatically lower
than those given here.

Vector operations

In the table below, x and y are n-vectors and a is a scalar.

ax n
x+y n
xT y 2n
kxk 2n
kx yk 3n
rms(x) 2n
std(x) 4n
6 (x, y) 6n

The convolution a b of an n-vector a and m-vector b can be computed by a special

algorithm that requires 5(m + n) log2 (m + n) flops.

Matrix operations

In the table below, A and B are m n matrices, C is an m p matrix, x is an

n-vector, and a is a scalar.

aA mn
A+B mn
Ax 2mn
AC 2mnp
AT A mn2
352 B Complexity

Factorization and inverses

In the table below, A is a tall or square m n matrix, R is an n n triangular
matrix, and b is an n-vector. We assume the factorization or inverses exist; in
particular in any expression involving A1 , A must be square.
QR factorization of A 2mn2
R1 b n2
A1 b 2n3
A1 3n3
A 2mn2
The pseudo-inverse A of a wide m n matrix (with independent rows) can be
computed in 2m2 n flops.

Solving least squares problems

In the table below, A is an m n matrix, C is a wide p n matrix, and b is an
m-vector. We assume the associated independence conditions hold.
minimize kAx bk2 2mn2
minimize kAx bk2 subject to Cx = d 2(m + p)n2
minimize kxk2 subject to Cx = d 2np2

Big-times-small-squared mnemonic
Many of the complexities listed above that involve two dimensions can be remem-
bered using a simple mnemonic: The cost is
2 (big) (small)2 flops,
where big and small refer to big and small problem dimensions. We list some
specific examples of this rule below.
In the QR factorization of an m n matrix, we have m n, so m is the big
dimension and n is the small dimension. The complexity is 2mn2 flops.
Computing the pseudo-inverse A of an m n matrix A when A is tall
(and has independent columns) costs 2mn2 flops. When A is wide (and has
independent rows), it is 2nm2 flops.
For least squares, we have m n, so m is the big dimension and n is the small
dimension. The cost of computing the least squares approximate solution is
2mn2 flops.
For the least norm problem, we have p n, so n is the big dimension and p
is the small dimension. The cost is 2np2 flops.
The constrained least squares problem involves two matrices A and C, and
three dimensions that satisfy m + p n. The numbers
  m + p and n are the
A
big and small dimensions of the stacked matrix . The cost of solving
C
the constrained least squares problem is 2(m + p)n2 flops.
 Appendix C

Derivatives and optimization

Calculus does not play a big role in this book, except in chapters 18 and 19 (on
nonlinear least squares and constrained nonlinear least squares), where we use
derivatives, Taylor approximations, and the method of Lagrange multipliers. In
this appendix we collect some basic material about derivatives and optimization,
focussing on the few results and formulas we use.

C.1 Derivatives
C.1.1 Scalar-valued function of a scalar

Definition. Suppose f : R R is a real-valued function of a real (scalar) variable.

For any number x, the number f (x) is the value of the function, and x is called
the argument of the function. The number

f (z + t) f (z)
lim ,
t0 t
(if the limit exists) is called the derivative of the function f at the point z. It gives
the slope of the graph of f at the point (z, f (z)). We denote the derivative of f at
z as f 0 (z). We can think of f 0 as a scalar-valued function of a scalar variable; this
function is called the derivative (function) of f .

Taylor approximation. Let us fix the number z. The (first order) Taylor approx-
imation of the function f at the point z is defined as

f(x) = f (z) + f 0 (z)(x z)

for any x. Here f (z) is the value of f at z, x z is the deviation of x from z,

and f 0 (z)(x z) is the approximation of the change in value of the function due
to the deviation of x from z. Sometimes the Taylor approximation is shown with
354 C Derivatives and optimization

a second argument, separated by a semicolon, to denote the point z where the ap-
proximation is made. Using this notation, the left-hand side of the equation above
is written f(x; z). The Taylor approximation is sometimes called the linearized
approximation of f at z. (Here linear uses informal mathematical language, where
affine is sometimes called linear.) The Taylor approximation function f is affine,
i.e., a linear function plus a constant.
The Taylor approximation f satisfies f(z; z) = f (z), i.e., at the point z it agrees
with the function f . For x near z, f(x; z) is a very good approximation of f (x).
For x not close enough to z, however, the approximation can be poor.

Finding derivatives. In a basic calculus course, the derivatives of many common

functions are worked out. For example, with f (x) = x2 , we have f 0 (z) = 2z, and
for f (x) = ex , we have f 0 (z) = ez . Derivatives of more complex functions can be
found using these known derivatives of common functions, along with a few rules
for finding the derivative of various combinations of functions. For example, the
chain rule gives the derivative of a composition of two functions. If f (x) = g(h(x)),
where g and h are scalar-valued functions of a scalar variable, we have
f 0 (z) = g 0 (h(z))h0 (z).
Another useful rule is the derivative of product rule, for f (x) = g(x)h(x), which is
f 0 (z) = g(z)0 h(z) + g(z)h0 (z).
The derivative operation is linear, which means that if f (x) = ag(x) + bh(x), where
a and b are constants, we have
f 0 (z) = ag 0 (z) + b0 h(z).
Knowledge of the derivative of just a few common functions, and a few combination
rules like the ones above, is enough to determine formulas for the derivatives of
many functions.

C.1.2 Scalar-valued function of a vector

Suppose f : Rn R is a scalar-valued function of an n-vector argument. The
number f (x) is the value of the function f at the n-vector (argument) x. We
sometimes write out the argument of f to show that it can be considered a function
of n scalar arguments, x1 , . . . , xn :
f (x) = f (x1 , . . . , xn ).

Partial derivative. The partial derivative of f at the point z, with respect to its
ith argument, is defined as
f f (z1 , . . . , zi1 , zi + t, zi+1 , . . . , zn ) f (z)
(z) = lim
xi t0 t
f (z + tei ) f (z)
= lim ,
t0 t
C.1 Derivatives 355

(if the limit exists). Roughly speaking the partial derivative is the derivative with
respect to the ith argument, with all other arguments fixed.

Gradient. The partial derivatives of f with respect to its n arguments can be

collected into an n vector called the gradient of f (at z):
f
x1 (z)

f (z) = ..
.

.
f
xn (z)

Taylor approximation. The (first-order) Taylor approximation of f at the point

z is the function f : Rn R defined as

f f
f(x) = f (z) + (z)(x1 z1 ) + + (z)(xn zn )
x1 xn

for any x. We interpret xi zi as the deviation of xi from zi , and the term

f
xi (z)(xi zi ) as an approximation of the change in f due to the deviation of xi
from zi . Sometimes f is written with a second vector argument, as f(x; z), to show
the point z at which the approximation is developed. The Taylor approximation
can be written in compact form as

f(x; z) = f (z) + f (z)T (x z).

The Taylor approximation f is an affine function of x.

The Taylor approximation f agrees with the function f at the point z, i.e.,
f (z; z) = f (z). When all xi are near the associated zi , f(x; z) is a very good

approximation of f (z). The Taylor approximation is sometimes called the linear

approximation or linearized approximation of f (at z), even though in general it is
affine, and not linear.

Finding gradients. The gradient of a function can be found by evaluating the

partial derivatives using the common functions and rules for derivatives of scalar-
valued functions, and assembling the result into a vector. In many cases the result
can be expressed in a more compact matrix-vector form. As an example let us find
the gradient of the function

f (x) = kxk2 = x21 + + x2n ,

which is the sum of squares of the arguments. The partial derivatives are

f
(z) = 2zi , i = 1, . . . , n.
xi

This leads to the very simple vector formula

f (z) = 2z.
356 C Derivatives and optimization

(Note the resemblance to the formula for the derivative of the square of a scalar
variable.)
There are rules for the gradient of a combination of functions similar to those
for functions of a scalar. For example if f (x) = ag(x) + bh(x), we have
f (z) = ag(z) + bh(z).

C.1.3 Vector-valued function of a vector

Suppose f : Rn Rm is a vector-valued function of a vector. The n-vector x is

the argument; the m-vector f (x) is the value of the function f at x. We can write
out the m components of f as

f1 (x)
f (x) = ..
,

.
fm (x)
where fi is a scalar-valued function of x = (x1 , . . . , xn ).

Jacobian. The partial derivatives of the components of f (x) with respect to the
components of x, evaluated at z, are arranged into an mn matrix denoted Df (z),
called the derivative or Jacobian matrix of f at z. (In the notation Df (z), the
D and f go together; Df does not represent, say, a matrix-vector product.) The
derivative matrix is defined by
fi
Df (z)ij = (z), i = 1, . . . , m, j = 1, . . . , n.
xj
The rows of the Jacobian are fi (z)T , for i = 1, . . . , m. For m = 1, i.e., when
f is a scalar-valued function, the derivative matrix is a row vector of size n, the
transpose of the gradient of the function. The derivative matrix of a vector-valued
function of a vector is a generalization of the derivative of a scalar-valued function
of a scalar.

Taylor approximation. The (first-order) Taylor approximation of f near z is given

by
fi fi
f(x)i = fi (z) + (z)(x1 z1 ) + + (z)(xn zn )
x1 xn
= fi (z) + fi (z)T (x z),
for i = 1, . . . , m. We can express this approximation in compact notation as
f(x) = f (z) + Df (z)(x z).
For x near z, f(x) is a very good approximation of f (x). As in the scalar case,
Taylor approximation is sometimes written with a second argument as f(x; z) to
show the point z around which the approximation is made. The Taylor approxi-
mation f is an affine function of x, sometimes called a linear approximation of f ,
even though it is not, in general, a linear function.
 C.2 Optimization 357

Finding Jacobians. We can always find the derivative matrix by calculating par-
tial derivatives of the entries of f with respect to the components of the argument
vector. In many cases the result simplifies using matrix vector notation. As an
example, let us find the derivative of the (scalar-valued) function

h(x) = kf (x)k2 = f1 (x)2 + + fm (x)2 ,

where f : Rn Rm . The partial derivative with respect to xj , at z, is

h f1 fm
(z) = 2f1 (z) (z) + + 2fm (z) (z).
xj xj xj

Arranging these to form the row vector Dh(z), we see we can write this using
matrix multiplication as
Dh(z) = 2f (z)T Df (z).

The gradient of h is the transpose of this expression,

h(z) = 2Df (z)T f (z).

(Note the analogy to the formula for the scalar-valued function of a scalar variable
h(x) = f (x)2 , which is h0 (z) = 2f 0 (z)f (z).)
Many of the formulas for derivatives in the scalar case also hold for the vector
case, with scalar multiplication replaced with matrix multiplication (provided the
order of the terms is correct). As an example, consider the composition function
f (x) = g(h(x)), where h : Rn Rk and g : Rk Rm . The Jacobian or derivative
matrix of f at z is given by

Df (z) = Dg(h(z))Dh(z).

(This is matrix multiplication; compare it to composition formula for scalar-valued

functions of scalars given above.) This chain rule is described on page 150.

C.2 Optimization
Derivative condition for minimization. Suppose h is a scalar-valued function of
a scalar argument. If x minimizes h(x), we must have h0 (x) = 0. This fact is easily
understood: If h0 (x) 6= 0, then by taking a point x slightly less than x (if h0 (x) > 0)
or slightly more than x (if h0 (x) < 0), we would obtain h(x) < h(x), which shows
that x does not minimize h(x). This leads to the classic calculus-based method
for finding a minimizer of a function f : Find the derivative, and set it equal to
zero. One subtlety here is that there can be (and generally are) points that satisfy
h0 (z) = 0, but are not minimizers of h. So we generally need to check which of the
solutions of h0 (z) = 0 are in fact minimizers of h.
358 C Derivatives and optimization

Gradient condition for minimization. This basic calculus-based method for find-
ing a minimizer of a scalar-valued function can be generalized to functions with
vector arguments. If the n-vector x minimizes h : Rn R, then we must have
h
(x) = 0, i = 1, . . . , n.
xi
In vector notation, we must have

h(x) = 0.

Like the case of scalar argument, this is easily seen to hold if x minimizes h. Also
as in the case of scalar argument, there can be points that satisfy h(z) = 0 but
are not minimizers of h. So we need to check if points found this way are in fact
minimizers of h.

Nonlinear least squares. As an example, consider the nonlinear least squares

problem, with objective h(x) = kf (x)k2 , where f : Rn Rm . The optimality
condition h(x) = 0 is
2Df (x)T f (x) = 0
(using the expression for the gradient derived above). This equation will hold for a
minimizer, but there can be points that satisfy the equation, but are not solutions
of the nonlinear least squares problem.

C.3 Lagrange multipliers

Constrained optimization. We now consider the problem of minimizing a scalar-
valued function h : Rn R, subject to the requirements, or constraints, that

g1 (x) = 0, ..., gp (x) = 0

must hold, where gi : Rn R are given functions. We can write the constraints
in compact vector form g(x) = 0, where g(x) = (g1 (x), . . . , gp (x)), and express the
problem as
minimize f (x)
subect to g(x) = 0.
We seek a solution of this optimization problem, i.e., a point x that satisfies g(x) =
0 (i.e., is feasible) and, for any other x that satisfies g(x) = 0, we have h(x) h(x).
The method of Lagrange multipliers is an extension of the derivative or gradient
conditions for (unconstrained) minimization, that handles constrained optimization
problems.

Lagrange multipliers. The Lagrangian function associated with the constrained

problem is defined as

L(x, z) = h(x) + z1 g1 (x) + + zp gp (x) = h(x) + g(x)T z,

C.3 Lagrange multipliers 359

with arguments x (the original variable to be determined in the optimization prob-

lem), and a p-vector z, called the (vector of) Lagrange multipliers. The Lagrangian
function is the original objective, with one term added for each constraint func-
tion. Each term is the constraint function value multiplied by zi , hence the name
multiplier.

KKT conditions. The KKT conditions (named for Karush, Kuhn, and Tucker)
state that if x is a solution of the constrained optimization problem, then there is
a vector z that satisfies
L L
(x, z) = 0, i = 1, . . . , n, (x, z) = 0, i = 1, . . . , p.
xi zi
(This is provided the rows of Dg(x) are linearly independent, a technical condition
we ignore.) As in the unconstrained case, there are pairs x, z that satisfy the KKT
conditions but x is not a solution of the constrained optimization problem.
The KKT conditions give us a method for solving the constrained optimization
optimization that is similar to the approach for the unconstrained optimization
problem. We attempt to solve the KKT equations for x and z; then we check to
see if any of the points found are really solutions.
We can simplify the KKT conditions, and express them compactly using matrix
notation. The last p equations can be expressed as gi (x) = 0, which we already
knew. The first n can be expressed as

x L(x, z) = 0,

where x denotes the gradient with respect to the xi arguments. This can be
written as

h(x) + z1 g1 (x) + + zp gp (x) = h(x) + Dg(x)T z = 0.

So the KKT conditions for the constrained optimization problem are

h(x) + Dg(x)T z = 0, g(x) = 0.

This is the extension of the gradient condition for unconstrained optimization to

the constrained case.

Constrained nonlinear least squares. As an example, consider the constrained

least squares problem
minimize kf (x)k2
subect to g(x) = 0,
where f : Rn Rm and g : Rn Rp . Define h(x) = kf (x)k2 . Its gradient at x
is 2Df (x)T f (x) (see above) so the KKT conditions are

2Df (x)T f (x) + Dg(x)T z = 0, g(x) = 0.

These conditions will hold for a solution of the problem (assuming the rows of Dg(x)
are independent). But there are points that satisfy them but are not solutions.
360 C Derivatives and optimization
Appendix D

Whats next

In this appendix we list some further topics of study that are closely related to the
material in this book, give a different perspective on the same material, complement
it, or provide useful extensions.

Mathematics

Probability and statistics. In this book we do not use probability and statistics,
even though we cover multple topics that are traditionally addressed using ideas
from probability and statistics, including data fitting and classification, control,
state estimation, and portfolio optimization. Further study of many of the topics
in this book requires a background in basic probability and statistics, and we
strongly urge you learn this material.

Abstract linear algebra. This book covers some of the most important basic ideas
from linear algebra, such as linear independence. In a more abstract course you
will learn about vector spaces, subspaces, nullspace and range. Eigenvalues and
singular values are very useful topics that we do not cover in this book.

Computational linear algebra. In a course on computational or numerical linear

algebra you will learn more about floating point numbers and how the small round-
off errors made in numerical calculations affect the computed solutions. You will
also learn about methods for sparse matrices, and iterative methods that can solve
linear equations, or compute least squares solutions, for extremely large problems
such as those arising in image processing or in the solution of partial differential
equations.

Mathematical optimization. This book focusses on just a few optimization prob-

lems: Least squares, linearly constrained least squares, and their nonlinear exten-
sions. In an optimization course you will learn about more general optimization
problems, for example ones that include inequality constraints. Convex optimiza-
tion is a particularly useful generalization of the linearly constrained least squares
362 D Whats next

problem. Convex optimization problems can be solved efficiently and exactly, and
include a very wide range of practically useful problems that arise in many ap-
plication areas, including all of the ones we have seen in this book. We would
strongly urge you to learn convex optmization, which is very widely used in many
applications.

Applications

Machine learning and artificial intelligence. This book covers some of the basic
ideas of machine learning and artificial intelligence, including a first exposure to
clustering, data fitting, classification, validation, and feature engineering. In a fur-
ther course on this material, you will learn about unsupervised learning methods
(like k-means) such as principal components analysis and nonnegative matrix fac-
torization, and more sophisticated clustering methods. You will also learn about
more sophisticated regression and classification methods, such as logistic regression
and the support vector machine, as well as methods for computing model parame-
ters that scale to extremely large scale problems. Additional topics might include
feature engineering and deep neural networks.

Linear dynamical systems, control, and estimation. We cover the very basics
of each of these topics; entire courses cover them in much more detail, including
applications in aerospace, navigation, and GPS.

Finance and portfolio optimization. Our coverage of portfolio optimization is

basic. In a further course you would learn about statistical models of returns,
factor models, transaction costs, more sophisticated models of risk, and the use of
convex optimization to handle constraints, for example a limit on leverage.

Signal and image processing. Traditional signal processing, which is still very
widely used, focusses on convolution; more modern approaches use convex opti-
mization, especially in non-real-time applications, like image enhancement or med-
ical image reconstruction. You will find whole courses on signal processing for a
specific application area, like communications, speech, audio, and radar; for im-
age processing, there are whole courses on microscopy, computational photography,
tomography, and medical imaging.

Time series analysis. Time series analysis, and especially prediction, play an
important role in many applications areas, including finance and supply chain op-
timization. It is typically taught in a statistics or operations research course, or as
a specialty course in a specific area such as econometrics.
Index

diag, 96 return, 8
k-means return matrix, 92
algorithm, 60 attribute vector, 10
complexity, 65 audio
validation, 62 addition, 13
z-score, 212 mixing, 17, 102
auto-regressive model, 202
acute angle, 48
addition back substitution, 165
audio, 13 balancing chemical reactions, 126, 169
matrix, 98 basis, 75
vector, 11 dual, 163
adjacency matrix, 94, 108 functions, 190
advertising, 184, 267 orthonormal, 80
affine Ben Afleck, 71
approximation, 30 bill of materials, 13
combination, 16 birth rate, 133
function, 28, 121 block
algorithm matrix, 91, 145
k-means, 60 vector, 4
back substitution, 165 byte, 22, 103
constrained least squares, 273
Gauss-Newton, 306 Carl Friedrich Gauss, 306
Gram-Schmidt, 80, 155 cash flow
least norm, 276 discounted, 21
least squares, 183 replication, 17
Levenberg-Marquardt, 306 vector, 8, 77
Newton, 308 categorical feature, 213
solving linear equations, 166 Cauchy-Schwarz inequality, 47
aligned vectors, 48 centroid, 60
angle, 47 chain graph, 110
document, 49 chain rule, 150, 354, 357
orthogonal, 48 Chebyshev inequality, 39, 46
annualized chemical
return and risk, 281 equilibrium, 305
anti-aligned vectors, 48 reaction balance, 126
approximation circular difference matrix, 251
affine, 30 classification, 221
least squares, 178 handwritten digit, 226, 323
Taylor, 30 iris flower, 225
AR model, 202 closed-loop, 151
argmax, 235 cluster centroid, 60
argument of function, 25 clustering, 55
asset digits, 65
allocation, 279 objective, 58
alpha and beta, 194 optimal, 59
364 Index

validation, 62 directed graph, 94, 108, 152

co-occurrence, 20 discount factor, 291
coefficients discounted cash flow, 21
linear equations, 124 disease dynamics, 136
matrix, 89 distance, 40
vector, 3 spherical, 49
color vector, 6 document
complexity angle, 49
k-means algorithm, 65 dissimilarity, 42
Gram-Schmidt algorithm, 84 scoring, 102
matrix-matrix multiply, 147 topic discovery, 68
matrix-vector multiplication, 104 word count, 9
compliance matrix, 122 document-term matrix, 98
computer representation dot product, 18
matrix, 103 dual basis, 163
conservation of mass, 128 dynamics
constrained least squares epidemic, 136
sparse, 275 matrix, 131
constrained optimization, 358 supply chain, 140
KKT conditions, 359
contingency table, 93 edge, 94
control elastic deformation, 122
linear quadratic, 288 elasticity, 122, 247
state feedback, 151 energy use patterns, 57
convolution, 111 epidemic dynamics, 136
correlation coefficient, 50, 194 equality
cross-validation, 207 matrices, 89
customer purchase vectors, 3
matrix, 93 equations
vector, 10 homogeneous, 125
KKT, 271
data fitting, 189 nonlinear, 301
data matrix, 94 normal, 181
de-meaned vector, 43 equilibrium
de-meaning, 121 Nash, 305
de-trended, 195 equilibrium prices, 304
death rate, 133 Euclidean
decision threshold, 229 distance, 40
deformation, 122 norm, 37
demand, 122 examples, 189
elasticity matrix, 122 expansion in a basis, 76
shaping, 247 exponential weighting, 291
dependent variable, 33
dependent vectors, 73 factor solve method, 167
derivative, 31, 353 feature
chain rule, 150, 354, 357 categorical, 213
partial, 354 distance, 41
diagonal matrix, 96 engineering, 211, 227
difference matrix, 100, 249 Likert, 213
difference of vectors, 11 matrix, 124
diffusion, 127 neural network, 215
digits standardized, 212
handwritten, 65 TFID, 214
dimension vector, 10
matrix, 89 winsorized, 212
vector, 3 feature vector, 189
diplacement, 12 floating point
Index 365

number, 21 independence-dimension inequality, 75

numbers, 85 independent vectors, 73
operation, 22 index
flop, 22 column, 89
flow conservation, 127 range, 4
forgetting factor, 291 row, 89
friend relation, 98 vector, 3
function inequality
affine, 28, 121 Cauchy-Schwarz, 47
argument, 25 Chebyshev, 39, 46
basis, 190 independence-dimension, 75
composition, 149 triangle, 38, 41, 48
inner product, 26 inner product, 18, 144
linear, 26, 119 function, 26
notation, 25 input, 132
objective, 178 intercept, 33
reversal, 120 inverse
running sum, 121 left, 157
sigmoid, 309 matrix, 160
pseudo, 172
Game of Thrones, 71 right, 159
Gauss, 177, 306 invertible matrix, 160
Gauss-Newton algorithm, 306 investment
generalization, 203 hedging, 52
generalized additive model, 213 iris flower classification, 225, 236
gigabyte, 22 Isaac Newton, 306
gigaflop, 22
global positioning system, 306 Jacobian, 123, 356
gone bust, 280
GPS, 294, 306 KKT
gradient, 180, 355 matrix, 271
Gram matrix, 147, 181 KKT conditions, 359
Gram-Schmidt algorithm, 80, 155 Kyoto prize, 296
complexity, 84
graph, 94, 108, 152 label, 33
chain, 110 Lagrange
social network, 98 multipliers, 358
grayscale, 8 polynomial, 169
group representative, 58 Laplacian, 110, 249
least squares, 177
handwritten digit, 226 data fitting, 189
handwritten digits, 65 nonlinear, 301
heat flow, 127 solution method, 183
hedging, 52 sparse, 184
Hestenes, 334 left inverse, 157
histogram, 42 Legendre, 177
vector, 9 Leontief input-output, 129
homogeneous equations, 125 Levenberg-Marquardt algorithm, 306
house price, 207, 215 Likert scale, 57, 213
house price regression, 201 line, 18, 287
segment, 18
identity matrix, 95 linear
illumination, 186 combination, 16
image dynamical system, 131
matrix, 92 equations, 119, 124
vector, 8 function, 26, 119
incidence matrix, 108, 140 least squares problem, 178
366 Index

quadratic control, 288 power, 151

linear dynamical system pseudo-inverse, 172, 181
closed-loop, 151 relation, 94
state feedback, 151 resistance, 129
linear equations return, 92
sparse, 168 running sum, 101
linearity, 119 second difference, 149
linearly independent singular, 160
row vectors, 97 sparse, 96
loan, 8, 77 sqare, 90
location vector, 6 squareroot, 151
logistic regression, 224 stacked, 91
long-only portfolio, 280 state feedback gain, 151
look-ahead, 210 subtraction, 98
lower triangular matrix, 96 tall, 90
Toeplitz, 112
market transpose, 97
clearing, 13 triangular, 96, 164
segmentation, 56 triple product, 148
Markov upper triangular, 96
model, 132 Vandermonde, 198
mass, 137 vector multiplication, 99
matrix, 89 wide, 90
addition, 98 zero, 95
adjacency, 94, 108 matrix-vector product, 119
asset return, 92 mean return, 44
block, 91, 145 minimum mean square error, 191
circular difference, 251 mixing audio, 17
coefficients, 89 mixture of vectors, 16
compliance, 122 MMSE, 191
computer representation, 103 MNIST, 65, 323
data, 94 MNIST data set, 226
demand elasticity, 122 model
diagonal, 96 over-fit, 204
difference, 100, 249 parameter, 190
dimensions, 89 validation, 203
document-term, 98 model fitting
dynamics, 131 nonlinear, 306
elements, 89 monochrome image, 8
equality, 89 motion, 137
feature, 124 multiplication
Gram, 147, 181 matrix-matrix, 143
graph, 94 matrix-vector, 99
identity, 95 scalar-matrix, 99
image, 92 scalar-vector, 15
incidence, 108, 140 sparse matrix, 148
inverse, 157, 160, 167
invertible, 160 Nash equilibrium, 305
Jacobian, 123, 356 navigation, 294
KKT, 271 net present value, 21, 77
left inverse, 157 neural network, 215
lower triangular, 96 Newton, 306
multiplication, 143 algorithm, 308
negative power, 164 law of motion, 137
nonsingular, 160 Nobel prize
orthogonal, 154, 162 Leontief, 129
permutation, 107 Markowitz, 279
Index 367

Nash, 305 polynomial

node, 94 evaluation, 21, 101
nonlinear fit, 198
equations, 301 interpolation, 126, 168
least squares, 301 Lagrange, 169
model fitting, 306 population dynamics, 132, 153
nonsingular matrix, 160 portfolio
norm, 37 gone bust, 280
Euclidean, 37 long-only, 280
weighted, 40 optimization, 279
normal equations, 181 return, 21, 101
notation trading, 14
function, 25 value, 21
overloading, 5 vector, 7
NPV, 21, 77 weights, 279
number potential, 128
floating point, 21 Powell, 334
of nonzeros, 96 power
matrix, 151
objective prediction error, 42, 124
clustering, 58 price
objective function, 178 elasticity, 122
observations, 189 equilibrium, 304
obtuse angle, 48 vector, 20
occurrence vector, 10 probability, 20
offset, 33 product
ones vector, 5 block matrix, 145
open-loop, 290 dot, 18
optimal inner, 18, 144
clustering, 59 matrix-matrix, 143
optimality condition matrix-vector, 119
nonlinear least squares, 302 outer, 144
optimization, 357 projection, 105
constrained, 358 proportions, 7
order, 22 pseudo-inverse, 172, 181
orthogonal
matrix, 154, 162 QR factorization, 154, 155
vectors, 48 sparse, 156
orthonormal
basis, 80 receiver operating characteristic, 229
expansion, 80 recommendation engine, 70
row vectors, 97 regression, 123, 201
vectors, 78 logistic, 224
out-of-sample validation, 204 model, 33
outcome, 189 regressors, 33
outer product, 144 relation, 94
over-determined, 125 friend, 98
over-fit, 204 residual, 177
overloading, 5 resistance, 128
matrix, 129
parameter return, 8, 44
model, 190 annualized, 281
partial derivative, 31, 354 vector, 21
path, 152 reversal function, 120
permutation matrix, 107 RGB, 6
piecewise-linear fit, 200 right inverse, 159
pixel, 8 right-hand side, 124
368 Index

risk, 44 subtraction
annualized, 281 matrix, 98
RMS, 38 subvector, 4
deviation, 40 sum
prediction error, 42 vector, 11
ROC, 229 superposition, 26, 119
Ronald Fisher, 225 supply chain dynamics, 140
root-mean-square, 38 support vector machine, 224
round-off error, 85 survey response, 57
row vector, 90
linearly independent, 97 tall matrix, 90
Rudolph Kalman, 296 Taylor approximation, 30, 123, 353
running sum, 101, 121 term frequency inverse document frequency,
214
scalar, 3 test data set, 204
scalar-matrix multiplication, 99 test set, 204
scalar-vector TFID, 214
multiplication, 15 thermal resistance, 128
score, 20 time series
seasonal component, 198 auto-regressive model, 202
second difference matrix, 149 de-trended, 195
segment, 18 vector, 7
shaping demand, 247 time-invariant, 131
short position, 7, 21 time-series
sigmoid function, 309 prediction validation, 210
sign function, 224 Toeplitz matrix, 112
signal, 7 topic discovery, 56, 68
singular matrix, 160 tracking, 291
skewed classifier, 229 training data set, 204
slice, 4 training set, 204
social network graph, 98 transpose, 97
sparse triangle inequality, 38, 41, 48
constrained least squares, 275 triangular matrix, 96, 164
least squares, 184
linear equations, 168, 276 uncorrelated, 50
matrix, 96 under-determined, 125
QR factorization, 156 unit vector, 5
vector, 23 upper triangular matrix, 96
sparse matrix
multiplication, 148 validation, 62, 203, 204
sparsity, 5 k-means, 62
spherical distance, 49 clustering, 62
square matrix, 90 limitations, 210
stacked set, 204
matrix, 91 Vandermonde matrix, 198
vector, 4 variable, 177
standard deviation, 43, 191 vector, 3
standardization, 46 addition, 11
standardized features, 212 affine combination, 16
state, 131 aligned, 48
state feedback control, 151 angle, 47
stemming, 10, 68 anti-aligned, 48
stop words, 10 basis, 75
straight-line fit, 192 block, 4
submatrix, 91 cash flow, 8, 77
subset clustering, 55
vector, 10 coefficients, 3
Index 369

color, 6 winsorized feature, 212

components, 3 word count
computer representation, 21 TFID, 214
correlation coefficient, 50 vector, 9, 42
customer purchase, 10
de-meaned, 43 Yann Lecun, 65
dependent, 73
dimension, 3 zero
distance, 40 matrix, 95
entries, 3 vector, 5
equality, 3 ZIP code, 56
feature, 10, 20, 189
histogram, 9
image, 8
independence, 73
inner product, 18
linear combination, 16
linear dependence, 73
linear independence, 73
location, 6
matrix multiplication, 99
mixture, 16
occurrence, 10
ones, 5
orthogonal, 48
orthonormal, 78
outer product, 144
portfolio, 7
price, 20
probability, 20
proportions, 7
quantities, 7
return, 21
RMS deviation, 40
RMS value, 38
row, 90
slice, 4
sparse, 5, 23
stacked, 4
standardization, 46
subset, 10
sum, 11
time series, 7
unit, 5
weight, 20, 33
word count, 9
zero, 5
vertex, 94
video, 9

warm start, 313

way-point constraint, 291
weather zones, 57
weight vector, 33
weighted norm, 40
wide matrix, 90
Wikipedia, 68