Linear Algebra

Linear Algebra through Matrices
Arbind K Lal Sukant Pati

T
July 10, 2018

AF
DR
2
DR
AF
T
Contents
1 Introduction to Matrices 7
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Definition of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Special Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Matrix Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 Transpose and Conjugate Transpose of Matrices . . . . . . . . . . . . . . 11
1.3.2 Sum and Scalar Multiplication of Matrices . . . . . . . . . . . . . . . . . . 11
1.3.3 Multiplication of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.4 Inverse of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4 Some More Special Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
T
1.5 Submatrix of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

AF
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
DR
2 System of Linear Equations 27

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Row-Reduced Echelon Form (RREF) . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.1 Gauss Elimination Method . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.2 Elementary Row Operations . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Initial Results on LU Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Row-Reduced Echelon Form (RREF) . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.1 The Gauss-Jordan Elimination Method . . . . . . . . . . . . . . . . . . . 40
2.4.2 Results on RREF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4.3 Computing the Inverse using GJE . . . . . . . . . . . . . . . . . . . . . . 42
2.5 Rank of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.6 Solution set of a Linear System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.7 Square Matrices and Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.8 Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.8.1 Inverse of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.8.2 Results on the Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.8.3 Cramer’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.9 Miscellaneous Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3
4 CONTENTS
3 Vector Spaces 65
3.1 Vector Spaces: Definition and Examples . . . . . . . . . . . . . . . . . . . . . . . 65
3.1.1 Vector Subspace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.2 Linear Combination and Linear Span . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2.1 Linear Span . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3 Linear Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3.1 Basic Results on Linear Independence . . . . . . . . . . . . . . . . . . . . 79
3.3.2 Application to Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4 Basis of a Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.4.1 Main Results associated with Bases . . . . . . . . . . . . . . . . . . . . . 86
3.4.2 Constructing a Basis of a Finite Dimensional Vector Space . . . . . . . . 86
3.5 Fundamental Subspaces Associated with a Matrix . . . . . . . . . . . . . . . . . 88
3.6 Fundamental Theorem of Linear Algebra and Applications . . . . . . . . . . . . . 90
3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4 Inner Product Spaces 97

4.1 Definition and Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.2 Cauchy-Schwartz Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.3 Normed Linear Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4 Orthogonality in Inner Product Space . . . . . . . . . . . . . . . . . . . . . . . . 103
T
AF
4.4.1 Properties of Orthonormal Vectors . . . . . . . . . . . . . . . . . . . . . . 105

4.5 Gram-Schmidt Orthonormalization Process . . . . . . . . . . . . . . . . . . . . . 109
DR
4.6 QR Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5 Linear Transformations 117

5.1 Definitions and Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2 Rank-Nullity Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.3 Algebra of Linear Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4 Ordered Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.5 Matrix of a linear transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.6 Similarity of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.7 Orthogonal Projections and Applications . . . . . . . . . . . . . . . . . . . . . . . 141
5.8 Orthogonal Operator and Rigid Motion∗ . . . . . . . . . . . . . . . . . . . . . . . 145
5.9 Dual Space* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6 Eigenvalues, Eigenvectors and Diagonalizability 153

6.1 Introduction and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.2 Spectrum of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.3 Basic Results on Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.4 Schur’s Unitary Triangularization and Diagonalizability . . . . . . . . . . . . . . 167
CONTENTS 5
6.4.1 Diagonalizability of some Special Matrices . . . . . . . . . . . . . . . . . . 170

6.4.2 Cayley Hamilton Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
6.5 Quadratic Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.5.1 Sylvester’s law of inertia . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.5.2 Applications in Eculidean Plane . . . . . . . . . . . . . . . . . . . . . . . 179
6.5.3 Applications in Eculidean Space . . . . . . . . . . . . . . . . . . . . . . . 183
6.6 Singular Value Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7 Jordan canonical form 189

7.1 Jordan canonical form theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.2 Minimal polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
7.3 Applications of Jordan Canonical Form . . . . . . . . . . . . . . . . . . . . . . . 201
7.3.1 Coupled system of linear differential equations . . . . . . . . . . . . . . . 201
7.3.2 Commuting matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8 Advanced Topics on Diagonalizability and Triangularization∗ 205

8.1 More on the Spectrum of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 205
8.2 Methods for Tridiagonalization and Diagonalization . . . . . . . . . . . . . . . . 208
8.2.1 Plane Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
8.2.2 Householder Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
T
8.2.3 Schur’s Upper Triangularization Revisited . . . . . . . . . . . . . . . . . . 215

AF
8.3 Commuting Matrices and Simultaneous Diagonalization . . . . . . . . . . . . . . 217

DR
8.3.1 Diagonalization and Real Orthogonal Matrix . . . . . . . . . . . . . . . . 220

8.3.2 Convergent and nilpotent matrices . . . . . . . . . . . . . . . . . . . . . . 221
9 Appendix 223
9.1 Uniqueness of RREF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
9.2 Permutation/Symmetric Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
9.3 Properties of Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
9.4 Dimension of W1 + W2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
9.5 When does Norm imply Inner Product . . . . . . . . . . . . . . . . . . . . . . . . 234
9.6 Roots of a Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
9.7 Variational characterizations of Hermitian Matrices . . . . . . . . . . . . . . . . . 237
Index 241
6 CONTENTS
T
AF
DR
Chapter 1
Introduction to Matrices
1.1 Motivation
Recall that at some stage, we have solved a linear system of 3 equations in 3 unknowns. But,
for clarity, let us start with a few linear systems of 2 equations in 2 unknowns.
Example 1.1.1. 1. Consider the linear system

)
2x + 5y = 7
(1.1.1)
2x + 4y = 6.
T
AF
The two linear systems represent a pair of non-parallel lines in R2 . Note that x = 1, y = 1
is the unique solution of the given system as (1, 1) is the point of intersection of the two
DR
given lines 2x + 5y = 7 and 2x + 4y = 6. But, we also see that

" # " # " #
2 5 7
·1+ ·1= ,
2 4 6
which corresponds to the solution of

" # " # " # ( )
2 5 7 2x + 5y = 7
·x+ ·y = ⇔ (1.1.2)
2 4 6 2x + 4y = 6.
" # " #
7 2
Equation (1.1.2) also implies that we can write the vector as sum of the vectors
6 2
" #
5
and . So, even though we were looking at the point of intersection of two lines, an
4
interpretation of the solution gives information about vectors in R2 .
2. Consider the linear system 

x + 5y + 4z = 11 

x + 6y − 7z = 1 (1.1.3)

2x + 11y − 3z = 12. 

Here, we have three planes in R3 and an easy observation implies that the third equation
is the sum of the first two equations. Hence, the line of intersection of the first two planes
7
8 CHAPTER 1. INTRODUCTION TO MATRICES
is contained in the third plane. Hence, this system has infinite number of solutions given
by
x = 61 − 59k, y = −10 + 11k, z = k with k arbitrary real number.
For example, verify that for k = 1, we get x = 2, y = 1 and z = 1 as a possible solution.
Also,
             
1 5 4 11 1 5 4
             
1 · 2 +  6  · 1 + −7 · 1 =  1  = 1 · 61 +  6  · (−10) + −7 · 0,
             
2 11 −3 12 2 11 −3
where the second

  part corresponds to k = 0 asa possible
   solution.
  Thus, we again see that
11 1 5 4
       
the vector 
 1  is a sum of the three vectors 1,  6  and −7 (which are associated
      
12 2 11 −3
with the unknowns x, y and z, respectively) after multiplying by certain scalars which
itself appear as solutions of the linear system.
Before going to the next example, also note that the numbers −59, 11 and 1, which appear
as coefficients of k in the solution satisfies
       
1 5 4 0
       
1 · (−59) +  6  · 11 + −7 · 1 = 0.
       
T
2 11 −3 0
AF
3. As a last example, consider the linear system

DR

x + 5y + 4z = 11 

x + 6y − 7z = 1 (1.1.4)

2x + 11y − 3z = 13. 

Here, we see that if we add the first two equations and subtract it with the third equation
then we are left with 0x + 0y + 0z = 1, which has no solution. That is, the above system
has no solution. I leave it to the readers to verify that there does not exist any x, y and
z such that        
1 5 4 11
       
1 · x +  6  · y + −7 · z =  1 .
       
2 11 −3 13
Remark 1.1.2. So, what we see above is “each of the linear systems gives us certain ‘relation-
ships’ between vectors which are ‘associated’ with the unknowns”. These relationships will lead
to the study of certain objects when we study “vector spaces”. They are as follows:
1. The first idea of ‘relationship’ that helps us to write a vector "in#terms of other vectors will
7
lead us to the study of ’linear combination’ of vectors. So, is a ‘linear combination’
6
       
" # " # 11 1 5 4
2 5        
of and  1  is a ‘linear combination’ of 1,  6  and −7.
. Similarly,        
2 4
12 2 11 −3
1.2. DEFINITION OF A MATRIX 9
2. Further, it also leads to the study of ‘linear span’ of a set. A positive answer leads to the
vector being an element of the ‘linear span;
 and
 a negative
  answer
 to ‘NOT
 an element of
1  5   4  11

 
  
the linear span’. For example, for S =  1,  6 , −7, the vector  1  belongs to
      

 2 11 −3  12
 
 
11
 
the ‘linear span’ of S, whereas,  1 

 does NOT belong to the ‘linear span’ of S.
13
3. The idea of a unique solution leads us to the statement
(" # " #) that the corresponding vectors are
2 5
‘linearly independent’. For example, the set , ⊆ R2 is ‘linearly independent’.
2 4
     
1  5   4 

 

Whereas, the set  1 ,  6 , −7 ⊆ R3 is NOT ‘linearly independent’ as
     
 2 11 −3 
 
       
1 5 4 0
       
1 · (−59) +  6  · 11 + −7 · 1 = 0.
       
2 11 −3 0
1.2 Definition of a Matrix

T
AF
Definition 1.2.1. A rectangular array of numbers is called a matrix.

DR
The horizontal arrays of a matrix are called its rows and the vertical arrays are called its
columns. A matrix A having m rows and n columns is said to be a matrix of size/ order
m × n and can be represented in either of the following forms:
   
a11 a12 · · · a1n a11 a12 · · · a1n
   
 a21 a22 · · · a2n   a21 a22 · · · a2n 
A= . or A =  . ,
   
 .. .. .. .. . .. .. .. 
 . . .
 
 . . . . 

am1 am2 · · · amn am1 am2 · · · amn
where aij is the entry at the intersection of the ith row and j th column. One writes A ∈ Mm,n (F)
to mean that A is an m×n matrix with entries from the set F, or in short A = [aij ] or A = (aij ).
We write A[i, :] to denote the i-th row of A, A[:, j] to denote the j-th column of A and aij or
(A)ij or A[i, j], for the (i,"j)-th entry of A.# " #
1 3+i 7 7
For example, if A = then A[1, :] = [1 3 + i 7], A[:, 3] = and
4 5 6 − 5i 6 − 5i
a22 = 5. Sometimes commas are inserted to differentiate between entries of a row vector. Thus,
A[1, :] may also be written as [1, 3 + i, 7]. A matrix having only one column is called a column
vector and a matrix with only one row is called a row vector. All our vectors will be column
vectors and will be represented by bold letters. A matrix of size 1 × 1 is also called a scalar
and is treated as such and hence we may or may not put it under brackets.
Definition 1.2.2. Two matrices A = [aij ], B = [bij ] ∈ Mm,n (C) are said to be equal if aij = bij ,
for each i = 1, 2, . . . , m and j = 1, 2, . . . , n.
In other words, two matrices are said to be equal if they have the same order and their
corresponding entries are equal.
Example 1.2.3. 1. Consider a system of linear
" equations
# 2x + 5y ="7#and 3x + 2y =" 6. #
2 5 7 2 5
Then, we identify it with the matrix A = . Here, A[:, 1] = and A[:, 2] =
3 2 6 3 2
are associated with the variables/ unknowns x and y, respectively.
 
" # " # 0 0
0 0 0 1  
2. A = ,B = Then, A 6= B as a12 6= b12 . Similarly, if C =  0 0  then
0 0 0 0  
0 0
A 6= C as they are of different sizes.
1.2.1 Special Matrices
Definition 1.2.4. Let A = [aij ] be an m × n matrix with aij ∈ F.
1. Then A is called a zero-matrix, denoted 0"(order # is mostly clear

" from the
# context), if
0 0 0 0 0
aij = 0 for all i and j. For example, 02×2 = and 02×3 = .
0 0 0 0 0
T
AF
2. Then A is called a square matrix if m = n and is denoted by A ∈ Mn (F).

DR
3. Let A ∈ Mn (F).
(a) Then, the entries a11 , a22 , . . . , ann are called the diagonal entries of A. They consti-
tute the principal diagonal of A.
(b) Then, A is said to be a diagonal matrix, , denoted
" #diag(a11 , . . . , ann ), if aij = 0
4 0
for i 6= j. For example, the zero matrix 0n and are diagonal matrices.
0 1
(c) Then, A = diag(1, . . . , 1) is called the  matrix, denoted In , or in short I.
 identity
" # 1 0 0
1 0  
For example, I2 = and I3 = 0 1 0

.
0 1
0 0 1
(d) If A = αI, for some α ∈ F, then A is called a scalar matrix.
(e) Then, A is said to be an upper triangular matrix if aij = 0 for i > j.
(f) Then, A is said to be a lower triangular matrix if aij = 0 for i < j.
(g) Then, A is said to be triangular if it is an upper or a lower triangular matrix.
   
0 1 4 0 0 0
   
For example, 0 3 −1 is upper triangular, 1 0 0 is lower triangular and the
  
0 0 −2 0 1 1
matrices 0, I are upper as well as lower triangular matrices.
1.3. MATRIX OPERATIONS 11
4. An m × n matrix A = [aij ] is said to have an uppertriangular form if aij = 0 for all

a11 a12 · · · a1n
 
0 0 1
  " #
 0 a22 · · · a2n   0 1 0

 1 2 0 0 1
i > j. For example, the matrices  . ,  and
  
 .. .. .. .. 
 . . .  
 0 0 2 
 0 0 0 1 1
0 0 · · · ann 0 0 0
have upper triangular forms.
5. For 1 ≤ i ≤ n, define ei = In [:, i], a matrix of order n × 1. Then the column matrices
e1 , . . . , en are called the standard unit vectors or the standard basis of Mn,1 (C) or
Cn . The dependence of n is omitted as it is understood
  from the context. For example,
" # 1
1
if e1 ∈ C then, e1 =
2 and if e1 ∈ C then e1 = 0
3
 
.

0
0
1.3 Matrix Operations

As a first operation, we define ‘transpose’ and/or ‘conjugate transpose’ of a matrix. This allows
us to interchange the ideas related with the rows of a matrix with the columns of a matrix and
vice-versa. It’s use also helps us in looking at geometrical ideas that are useful in applications.
T
1.3.1 Transpose and Conjugate Transpose of Matrices

AF
Definition 1.3.1. Let A = [aij ] ∈ Mm,n (C). Then

DR
1. the transpose of A, denoted AT , is an n × m matrix with (AT )ij = aji , for all i, j.
2. the conjugate transpose of A, denoted A∗ , is an n × m matrix with (A∗ )ij = aji (the
complex-conjugate of aji ), for all i, j.
" # " # " #
1 4+i 1 0 ∗ 1 0
If A = T
then A = and A = . Note that A∗ 6= AT .
0 1−i 4+i 1−i 4−i 1+i
" #
1 h i
Note that if x = is a column vector then xT = 1 2 and x∗ are row vectors.
2
Theorem 1.3.2. For any matrix A, (A∗ )∗ = A and (AT )T = A.
Proof. Let A = [aij ], A∗ = [bij ] and (A∗ )∗ = [cij ]. Clearly, the order of A and (A∗ )∗ is the
same. Also, by definition cij = bji = aij = aij for all i, j.
1.3.2 Sum and Scalar Multiplication of Matrices
Definition 1.3.3. Let A = [aij ], B = [bij ] ∈ Mm,n (C) and k ∈ C.

1. . Then the sum of A and B, denoted A + B, is defined to be the matrix C = [cij ] ∈
Mm,n (C) with cij = aij + bij for all i, j.
2. Then, the product of k ∈ C with A, denoted kA, equals kA = [kaij ] = [aij k] = Ak.
" # " #
1 4 5 1 −4 6
Example 1.3.4. If A = ,B= then
0 1 2 1 1 7
" # " #
2 0 11 5 20 25
A+B = and 5A = .
1 2 9 0 5 10
Theorem 1.3.5. Let A, B, C ∈ Mm,n (C) and let k, ` ∈ C. Then
1. A + B = B + A (commutativity).
2. (A + B) + C = A + (B + C) (associativity).
3. k(À) = (k`)A.
4. (k + `)A = kA + À.
Proof. (1). Let A = [aij ] and B = [bij ]. Then by definition
A + B = [aij ] + [bij ] = [aij + bij ] = [bij + aij ] = [bij ] + [aij ] = B + A
as complex numbers commute. The other parts are left for the reader.
Definition 1.3.6. Let A ∈ Mm,n (C). Then
1. the matrix 0m×n satisfying A + 0 = 0 + A = A is called the additive identity.

T
2. the matrix B with A + B = 0 is called the additive inverse of A, denoted −A = (−1)A.

AF
Exercise 1.3.7. 1. Find non zero, non-identity matrices A satisfying

DR
(a) A∗ = A (such matrices are called Hermitian matrices).

(b) A∗ = −A (such matrices are called skew-Hermitian matrices).
2. Suppose A = [aij ], B = [bij ] ∈ Mm,n (C).

(a) If A + B = 0 then show that B = (−1)A = [−aij ].
(b) If A + B = A then show that B = 0.
3. Let A ∈ Mn (C). Then there exists matrices B and C such that A = B +C, where B T = B
(Symmetric matrix) and C T = −C (skew-symmetric matrix).
 
1 + i −1 " #
  2 3 −1 ∗ ∗
4. Let A = 
 2 3 and B = 1 1 − i 2 . Compute A + B and B + A .
i 1
5. Write the 3 × 3 matrices A = [aij ] satisfying

(a) aij = 1 if i 6= j and 2 otherwise.
(b) aij = 1 if | i − j | ≤ 1 and 0 otherwise.
(c) aij = i + j.
(d) aij = 2i+j .
1.3.3 Multiplication of Matrices
We now come to the most important operation between matrices, called the matrix multipli-
cation. We define it as follows.
Definition 1.3.8. Let A = [aij ] ∈ Mm,n (C) and B = [bij ] ∈ Mn,r (C). Then, the product of A
and B, denoted AB, is a matrix C = [cij ] ∈ Mm,r (C) such that for 1 ≤ i ≤ m, 1 ≤ j ≤ r
 
b1j
n
 
 b2j  X
cij = A[i, :]B[:, j] = [ai1 , ai2 , . . . , ain ] .  = ai1 b1j + ai2 b2j + · · · + ain bnj = aik bkj .
 
 .. 
  k=1
bnj
Thus, AB is defined if and only if the number of columns of A = the number of rows of
B. The way matrix product is defined seems quite complicated. Most of you have already seen
it. But, we will find other ways (3 more ways) to understand this matrix multiplication. These
will be quite useful at different stages in our study. So, we need to spend enough time on it.
 
1 −1 " #
3 4 5
T
 
Example 1.3.9. Let A = 
2 0  and B = −1 0 1 .

AF
0 1
DR
1. Entry-wise Method: (AB)  11 = 1 · 3 +

 (−1) · (−1) = 3 + 1 = 4. Similarly, compute the
4 4 4
 
rest and verify that AB = 
 6 8 10 .

−1 0 1
2. Row Method: Note that A[1, :] is a 1 × 2 matrix and B is a 2 × 3 matrix and hence
A[1, :]B is a 1 × 3 matrix. So, matrix multiplication is defined. Thus,
" #
h i 3 4 5 h i h i h i
A[1, :]B = 1 −1 = 1 · 3 4 5 + (−1) · −1 0 1 = 4 4 4
−1 0 1
" #
h i 3 4 5 h i h i h i
A[2, :]B = 2 0 = 2 · 3 4 5 + 0 · −1 0 1 = 6 8 10
−1 0 1
" #
h i 3 4 5 h i h i h i
A[3, :]B = 0 1 = 0 · 3 4 5 + 1 · −1 0 1 = −1 0 1 .
−1 0 1
       
A[1, :] A[1, :] A[1, :]B 4 4 4
       
Hence, if A = 
A[2, :] then AB = A[2, :]B = A[2, :]B  =  6 8 10.
      
A[3, :] A[3, :] A[3, :]B −1 0 1
3. Column Method: Note that A is a 3 × 2 matrix and B[:, 1] is a 2 × 1 matrix and hence
A(B[:, 1]) is a 3 × 1 matrix. So, matrix multiplication is defined. Thus,

       
1 −1 " # 1 −1 4
  3      
A · B[:, 1] = 2 0 
  = 2 · 3 +  0  · −1 =  6 
    
−1 
0 1 0 1 −1
       
1 −1 " # 1 −1 4
  4      
A · B[:, 2] = 2 0 
  = 2 · 4 +  0  · 0 = 8
    
0 
0 1 0 1 0
       
1 −1 " # 1 −1 4
  5      
A · B[:, 1] = 2 0 
  = 2 · 5 +  0  · 1 = 10
    
1 
0 1 0 1 1
h i
Thus, if B = B[:, 1] B[:, 2] B[:, 3] then
 
h i h i  4 4 4

AB = A B[:, 1 B[:, 2] B[:, 3] = A · B[:, 1 A · B[:, 2] A · B[:, 3] = 
6 .
8 10
−1 0 1
" #
h i B[1, :]
4. Matrix Method: We also have if A = A[:, 1] A[:, 2] and B = then A[:, 1]
T
B[2, :]
AF
is a 3 × 1 matrix and B[1, :] is a 1 × 3 matrix. Thus, the matrix product A[:, 1] B[1, :] is
DR
defined and is a 3 × 3 matrix. Hence,

   
1
 h i −1h i
A[:, 1]B[1, :] + A[:, 2]B[2, :] = 2 3 4 5 + 
  0
 
 −1 0 1
0 1
     
3 4 5 1 0 −1 4 4 4
     
=
6 8 10  +  0 0 0  =  6 8 10.
    
0 0 0 −1 0 1 −1 0 1
Remark 1.3.10. Let A ∈ Mm,n (C) and B ∈ Mn,p (C). Then the product AB is defined and
observe the following:
1. AB corresponds to operating (combining certain multiples of rows) on the rows of B. This

is called the row method for calculating the matrix product. Here,
(AB)[i, :] = A[i, :]B = ai1 B[1, :] + · · · + ain B[n, :], for 1 ≤ i ≤ m.
2. AB also corresponds to operating (combining certain multiples of columns) on the columns

of A. This is called the column method for calculating the matrix product. Here,
(AB)[:, j] = AB[:, j] = A[:, 1]b1j + · · · + A[:, n]bnj , for 1 ≤ j ≤ p.

 
A[1, :]
 .  h i
3. Write A =  ..  and B = B[:, 1] · · · B[:, p] then
 
A[m, :]
 
A[1, :]B[:, 1] A[1, :]B[:, 2] ··· A[1, :]B[:, p]
 
 A[2, :]B[:, 1] A[2, :]B[:, 2] · · · A[2, :]B[:, p] 
AB =  .
 
.. .. ..

 . . · · · . 

A[m, :]B[:, 1] A[m, :]B[:, 2] · · · A[m, :]B[:, p]

B[1, :]
h i  . 
4. Write A = A[:, 1] · · · A[:, n] and B =  . 
 . . Then
B[n, :]
AB = A[:, 1]B[1, :] + A[:, 2]B[2, :] + · · · + A[:, n]B[n, :].
5. If m 6= p then the product BA is NOT defined.

 
1
 
6. Let m = p. Here BA and AB can still be different. For example, if A = 2

 and B =
3
T
 
AF
h i −1 2 3
 
−1 2 3 then AB = −2 4 6 whereas BA = −1 + 4 + 9 = 12. As matrices, they

DR
−3 6 9
look quite different but it will be shown during the study of eigenvalues and eigenvectors
that they have similar structure.
7. If m = n = p, then the orders

" of AB#and BA are" same.
# Even then AB
" may NOT
# equal
1 1 1 1 2 2
BA. For example, if A = and B = then AB = whereas
−1 −1 1 1 −2 −2
" #
0 0
BA = . Thus, AB 6= BA and hence
0 0
(A + B)2 = A2 + AB + BA + B 2 6= A2 + B 2 + 2AB.
" # " #
2 1 3 3
Whereas if C = then BC = CB = = 3A 6= A = CA. Note that cancella-
1 2 3 3
tion laws don’t hold.
Definition 1.3.11. Two square matrices A and B are said to commute if AB = BA.
Theorem 1.3.12. Let A ∈ Mm,n (C), B ∈ Mn,p (C) and C ∈ Mp,q (C).
1. Then (AB)C = A(BC), i.e., the matrix multiplication is associative.
2. For any k ∈ C, (kA)B = k(AB) = A(kB).

3. Then A(B + C) = AB + AC, i.e., multiplication distributes over addition.
4. If A ∈ Mn (C) then AIn = In A = A.

p
P n
P
Proof. (1). Verify that (BC)kj = bk` c`j and (AB)i` = aik bk` . Therefore,
`=1 k=1
n
X n
X p
X p
n X
X
A(BC) ij = aik BC kj
= aik bk` c`j = aik bk` c`j
k=1 k=1 `=1 k=1 `=1
Xn Xp p
X Xn X T

= aik bk` c`j = aik bk` c`j = AB c
i` `j
= (AB)C ij
.
k=1 `=1 `=1 k=1 `=1
Using a similar argument, the next part follows. The other parts are left for the reader.
Exercise 1.3.13. 1. Let A ∈ Mn (C) and e1 , . . . , en ∈ Mn,1 (C) (see Definition 5). Then
(a) Ae1 = A[:, 1], . . . , Aen = A[:, n].
(b) eT1 A = e∗1 A = A[1, :], . . . , eTn A = e∗n A = A[n, :].
2. Let L1 , L2 ∈ Mn (C) be lower triangular matrices. If D ∈ Mn (C) is a diagonal matrix then
(a) L1 L2 is a lower triangular matrix.

(b) DL1 and L1 D are lower triangular matrices.
T
The same holds for upper triangular matrices.

AF
3. Let A ∈ Mm,n (C) and B ∈ Mn,p (C).

DR
(a) Prove that (AB)∗ = B ∗ A∗ .

(b) If A[1, :] = 0T then (AB)[1, :] = 0T .
(c) If B[:, 1] = 0 then (AB)[:, 1] = 0.
(d) If A[i, :] = A[j, :] for some i and j then (AB)[i, :] = (AB)[j, :].
(e) If B[:, i] = B[:, j] for some i and j then (AB)[:, i] = (AB)[:, j].
4. Construct matrices A and B that satisfy the following statements.
(a) The product AB is defined but BA is not defined.

(b) The products AB and BA are defined but they have different orders.
(c) The products AB and BA are defined, they have the same order but AB 6= BA.
(d) Construct a 2 × 2 matrix satisfying A2 = A.
 
" # 0 1 1
0 1
formula for An and B n and prove it?
 
(e) Let A = and B = 0 0 1 . Guess a
0 0  
0 0 0
   
" # 1 1 1 1 1 1
1 1     2
(f ) Let A = , B = 0 1 1 and C = 1 1 1. Is it true that A −2A+I = 0?
  
0 1
0 0 1 1 1 1
3 2 2
What is B − 3B + 3B − I? Is C = 3C?
5. Let A ∈ Mm,n (C). If Ax = 0 for all x ∈ Mn,1 (C) then A = 0, the zero matrix.
6. Let A, B ∈ Mm,n (C). If Ax = Bx, for all x ∈ Mn,1 (C) then prove that A = B.
   
x1 y1
 .  . n n
.. , y =  ..  ∈ Mn,1 (C). Then y∗ x = ∗x = |xi |2 ,
P P
7. Let x = 
    y i x i , x
i=1 i=1
xn yn
 
  |x1 |2 x1 x2 · · · x1 xn
x1 y1 x1 y2 · · · x1 yn
 x2 x1 |x2 |2 · · · x2 xn 
 
.. .. .. 
xy∗ =  ∗
 
.  and xx =  .
 
 . . ···  ... .
.. . .. .
.. 
xn y1 xn y2 · · · xn yn
 
xn x1 xn x2 · · · |xn | 2
8. Let A be an upper triangular matrix. If A∗ A = AA∗ then prove that A is a diagonal

matrix. The same holds for lower triangular matrix.
9. Let A be a 3 × 3 upper triangular matrix with diagonal entries a, b, c. Then
(A − aI3 )(A − bI3 )(A − cI3 ) = 0.
Note that (A − aI3 )[:, 1] = 0. So, if A[:, 1] = 0 then B[1, :] doesn’t play any role in AB.
T
AF
DR
10. Let A and B be two m × n matrices. Then, prove that (A + B)∗ = A∗ + B ∗ .
11. Find A, B, C ∈ M2 (C) such that AB = AC but B 6= C (Cancellation laws don’t

hold).
 
0 1 0
2 3 3 3 2
 
0 0 1. Compute A and A . Is A = I? Determine aA + bA + cA .
12. Let A =  
1 0 0
1.3.4 Inverse of a Matrix
Definition 1.3.14. Let A ∈ Mn (C). Then
1. B ∈ Mn (C) is said to be a left inverse of A if BA = In .
2. C ∈ Mn (C) is called a right inverse of A if AC = In .
3. A is invertible (has an inverse) if there exists B ∈ Mn (C) such that AB = BA = In .
Lemma 1.3.15. Let A ∈ Mn (C). If there exist B, C ∈ Mn (C) such that AB = In and CA = In
then B = C, i.e., If A has a left inverse and a right inverse then they are equal.
Proof. Note that C = CIn = C(AB) = (CA)B = In B = B.

Remark 1.3.16. Lemma 1.3.15 implies that whenever A is invertible, the inverse is unique.
Thus, we denote the inverse of A by A−1 . That is, AA−1 = A−1 A = I.
Theorem 1.3.17. Let A and B be two invertible matrices. Then,
1. (A−1 )−1 = A.
2. (AB)−1 = B −1 A−1 .
3. (A∗ )−1 = (A−1 )∗ .
Proof. (1). Let B = A−1 . Then AB = BA = I. Thus, by definition, B is invertible and

B −1 = A. Or equivalently, (A−1 )−1 = A.
(2). By associativity (AB)(B −1 A−1 ) = A(BB −1 )A−1 = I = (B −1 A−1 )(AB).
(3). As AA−1 = A−1 A = I, we get (AA−1 )∗ = (A−1 A)∗ = I ∗ . Or equivalently, (A−1 )∗ A∗ =
A∗ (A−1 )∗ = I. Thus, by definition (A∗ )−1 = (A−1 )∗ .
We will again come back to the study of invertible matrices in Sections 2.4 and 2.8.
Exercise 1.3.18. 1. If A is an invertible matrix then (A−1 )r = A−r , for all r ∈ N.
2. If A1 , . . . , Ar are invertible matrices then B = A1 A2 · · · Ar is also invertible.

T
AF
" # " #
cos(θ) sin(θ) cos(θ) − sin(θ)
3. Find the inverse of and .
DR
sin(θ) − cos(θ) sin(θ) cos(θ)
4. Let A ∈ Mn (C) be an invertible matrix. Then
(a) A[i, :] 6= 0T , for any i.
(b) A[:, j] 6= 0, for any j.
(c) A[i, :] 6= A[j, :], for any i and j.
(d) A[:, i] 6= A[:, j], for any i and j.
(e) A[3, :] 6= αA[1, :] + βA[2, :], for any α, β ∈ C, whenever n ≥ 3.
(f ) A[:, 3] 6= αA[:, 1] + βA[:, 2], for any α, β ∈ C, whenever n ≥ 3.
" #
1 2
5. Determine A that satisfies (I + 3A)−1 = .
2 1
1 2
6. Let A be an invertible matrix satisfying A3 + A − 2I = 0. Then A−1 =

A +I .
2
7. Let A = [aij ] be an invertible matrix and B = [pi−j aij ], for some p ∈ C, p 6= 0. Then
B −1 = [pi−j (A−1 )ij ].
1.4. SOME MORE SPECIAL MATRICES 19
1.4 Some More Special Matrices

Definition 1.4.1. 1. For 1 ≤ k ≤ m and 1 ≤ ` ≤ n, define ek` ∈ Mm,n (C) by
(
1, if (k, `) = (i, j)
(ek` )ij =
0, otherwise.
Then, the matrices ek` for 1 ≤ k ≤ m and 1 ≤ ` ≤ n are called the standard basis
elements for Mm,n (C).
" " # # " # " #
1 0 0
1 h i 0 1 0 1 h i
So, if ek` ∈ M2,3 (C) then e11 = = 1 0 0 , e12 = = 0 1 0
0 0 0 0 0 0 0 0
" # " #
0 0 0 0 h i
and e22 = = 0 1 0 .
0 1 0 1
In particular, if eij ∈ Mn (C) then eij = ei eTj = ei e∗j , for 1 ≤ i, j ≤ n.
2. Let A ∈ Mn (R). Then

" #
1 3
(a) A is called symmetric if AT = A. For example, A = .
3 2
" #
0 3
(b) A is called skew-symmetric if AT = −A. For example, A = .
−3 0
T
" #
1 1 1
AF
(c) A is called orthogonal if AAT = AT A = I. For example, A = √ .

2 1 −1
DR
(d) A is said to be a permutation matrix if A has exactly one non-zero entry, "namely
#
0 1
1, in each row and column. For example, In for each positive integer n, ,
1 0
     
0 1 0 0 0 1 0 1 0
     
0 0 1, 0 1 0 and 1 0 0 are permutation matrices. Verify that per-
     
1 0 0 1 0 0 0 0 1
mutation matrices are Orthogonal matrices.
3. Let A ∈ Mn (C). Then

" #
1 i
(a) A is called normal if A∗ A = AA∗ . For example, is a normal matrix.
i 1
" #
1 1+i
(b) A is called Hermitian if A∗ = A. For example, A = .
1−i 2
" #
0 1+i
(c) A is called skew-Hermitian if A∗ = −A. For example, A = .
−1 + i 0
" #
∗ ∗ 1 1+i 1
(d) A is called unitary if AA = A A = I. For example, A = √ .
3 −1 1 − i
Verify that Hermitian, skew-Hermitian and Unitary matrices are normal matrices.
4. A vector u ∈ Mn,1 (C) such that u∗ u = 1 is called a unit vector.

" #
1 0
5. A matrix A is called idempotent if A2 = A. For example, A = is idempotent.
1 0
6. An idempotent matrix which is also Hermitian is called a projection matrix. For example,
if u ∈ Mn,1 (C) is a unit vector then A = uu∗ is a Hermitian, idempotent matrix. Thus A
is a projection matrix.
In particular, if u ∈ Mn,1 (R) is a unit vector then A = uuT . Then verify that uT (x−Ax) =
uT x − uT Ax = uT x − uT (uuT )x = 0 (as uT u = 1), for any x ∈ R3 . Thus, with respect
to the dot product in R3 , Ax is the foot of the perpendicular from the point x on the
1
vector u. In particular, if u = √ [1, 2, −1]T and A = uuT . Then, for any vector
6
x = [x1 , x2 , x3 ]T ∈ M3,1 (R),
x1 + 2x2 − x3 x1 + 2x2 − x3
Ax = (uuT )x = u(uT x) = √ u= [1, 2, −1]T .
6 6
7. Fix a unit vector u ∈ Mn,1 (R) and let A = 2uuT − In . Then, verify that A ∈ Mn (R) and
Ay = 2(uT y)u − y, for all y ∈ Rn . This matrix is called the reflection matrix about the
line, say `, containing the points 0 and u. This matrix fixes each point on the line ` and
send the vector v, which is orthogonal to u, to −v.
8. Let A ∈ Mn (C). Then, A is said to be nilpotent if there exists a positive integer n

such that An = 0. The least positive integer k for which Ak = 0 is called the order of
T
AF
nilpotency. For example, if A = [aij ] ∈ Mn (C) with aij equal to 1 if i − j = 1 and 0,

otherwise then An = 0 and A` 6= 0 for 1 ≤ ` ≤ n − 1.
DR
Exercise 1.4.2. 1. Consider the matrices eij ∈ Mn (C) for 1 ≤ i, j, ≤ n. Is e12 e11 = e11 e12 ?
What about e12 e22 and e22 e12 ?
2. Let {u1 , u2 , u3 } be three vectors in R3 such that u∗i ui = 1, for 1 ≤ i ≤ 3, and u∗i uj = 0
whenever i 6= j. Prove the following.
(a) If U = [u1 u2 u3 ] then U ∗ U = I. What about U U ∗ = u1 u∗1 + u2 u∗2 + u3 u∗3 ?

(b) If A = ui u∗i , for 1 ≤ i ≤ 3 then A2 = A. Is A Hermitian? Is A a projection matrix?
(c) If A = ui u∗i + uj u∗j , for i 6= j then A2 = A. Is A a projection matrix?
3. Let A be an n × n upper triangular matrix. If A is also an orthogonal matrix then A is a

diagonal matrix with diagonal entries ±1.
4. Prove that in M5 (R), there are infinitely many orthogonal matrices of which only finitely
many are diagonal (in fact, there number is just 32).
5. Prove that permutation matrices are real orthogonal.
6. Let A, B ∈ Mn (C) be two unitary matrices. Then both AB and BA are unitary matrices.
7. Let A ∈ Mn (C) be a Hermitian matrix.

(a) Then the diagonal entries of A are necessarily real numbers.
1.5. SUBMATRIX OF A MATRIX 21
(b) Then, for any x ∈ Mn,1 (C). x∗ Ax is a real number.

(c) For each B ∈ Mn (C) the matrix B ∗ AB is Hermitian.
(d) Further, if A2 = 0 then A = 0.
8. Let A ∈ Mn (C). If x∗ Ax ∈ R for every x ∈ Mn,1 (C) then A is a Hermitian matrix. [Hint:
Use ej , ej + ek and ej + iek of Mn,1 (C) for x.]
9. Let A and B be Hermitian matrices. Then AB is Hermitian if and only if AB = BA.
10. Let A ∈ Mn (C) be a skew-Hermitian matrix. Then prove that

(a) the diagonal entries of A are either zero or purely imaginary.
(b) for each B ∈ Mn (C) prove that B ∗ AB is a skew-Hermitian matrix.
(c) Then, for any x ∈ Mn,1 (C), x∗ Ax is either 0 or purely imaginary.
11. Let A ∈ Mn (C). Then A = S1 +S2 , where S1 = 21 (A+A∗ ) is Hermitian and S2 = 21 (A−A∗ )
is skew-Hermitian.
12. Let A, B be skew-Hermitian matrices with AB = BA. Is the matrix AB Hermitian or

skew-Hermitian?
13. Let A be a nilpotent matrix. Then prove that I + A is invertible.

   
1 0 0 1 0 0
T
AF
   
14. Let A = 0 cos θ − sin θ and B = 0 cos θ , for θ ∈ [−π, π). Are they
sin θ
 
0 sin θ cos θ 0 sin θ − cos θ
DR
orthogonal?
1.5 Submatrix of a Matrix

Definition 1.5.1. For k ∈ N, let [k] = {1, . . . , k}. Also, let A ∈ Mm×n (C).
1. Then, a matrix obtained by deleting some of the rows and/or columns of A is said to be
a submatrix of A.
2. If S ⊆ [m] and T ⊆ [n] then by A(S|T) , we denote the submatrix obtained from A by
deleting the rows with indices in S and columns with indices in T . By A[S, T ], we mean
A(S c |T c ), where S c = [m] \ S and T c = [n] \ T . Whenever, S or T consist of a single
element, then we just write the element. If S = [m], then A[S, T ] = A[:, T ] and if T = [n]
then A[S, T ] = A[S, :] which matches with our notation in Definition 1.2.1.
3. If m = n, the submatrix A[S, S] is called a principal submatrix of A.
" # " #
1 4 5 1 5
Example 1.5.2. 1. Let A = . Then, A[{1, 2}, {1, 3}] = A[:, {1, 3}] = ,
0 1 2 0 2
" #
1
A[1, 1] = [1], A[2, 3] = [2], A[{1, 2}, 1] = A[:, 1] = , A[1, {1, 3}] = [1 5] and A are a few
0
" # " #
1 4 1 4
sub-matrices of A. But the matrices and are not sub-matrices of A.
1 0 0 2
 
1 2 3 " #
 
, S = {1, 3} and T = {2, 3}. Then, A[S, S] = 1 3 h i
2. Let A =  5 6 7 , A(S | S) = 6 ,
  9 7
9 8 7
" #
6 7 h i
A[T, T ] = and A(T | T ) = 1 are principal sub-matrices of A.
8 7
Let A ∈ Mn,m (C) and B ∈ Mm,p (C). Then the product AB" is
# defined. Suppose r < m.
H
Then A and B can be decomposed as A = [P Q] and B = , where P ∈ Mn,r (C) and
K
H ∈ Mr,p (C) so that AB = P H + QK. This is proved next.
Theorem 1.5.3. Let the matrices A, B, P, H, Q and K be defined as above. Then
AB = P H + QK.
Proof. Verify that the matrix products P H and QK are valid. Further, their sum is defined
as P H, QK ∈ Mn,p (C). Now, let P = [Pij ], Q = [Qij ], H = [Hij ], and K = [Kij ]. Then, for
1 ≤ i ≤ n and 1 ≤ j ≤ p, we have
m
X r
X m
X r
X m
X
(AB)ij = aik bkj = aik bkj + aik bkj = Pik Hkj + Qik Kkj
k=1 k=1 k=r+1 k=1 k=r+1
T
= (P H)ij + (QK)ij = (P H + QK)ij .

AF
Thus, the required result follows.

DR
Remark 1.5.4. Theorem 1.5.3 is very useful due to the following reasons:
1. The matrices P, Q, H and K can be further partitioned so as to form blocks that are either
identity or zero or have certain nice properties. So, such partitions are useful during
different matrix operations.
" # Examples of such partitions
" # appear throughout the notes. For
Ir 0 h i Q1
example, let A = , P = P1 P2 and Q = . Then, verify that P AQ = P1 Q1 .
0 0 Q2
This is similar to the understanding that
" #" #
i a
11 a12 y1
h
x1 x2 = x1 a11 y1 + x1 a12 y2 + x2 a21 y1 + x2 a22 y2 .
a21 a22 y2
2. Suppose one wants to prove a result for a square matrix A. If we want to prove it using
induction then we can prove it for the 1 × 1 matrix (the initial step of induction). Then
assume the result to hold for all k × k sub-matrices
" # A or just the first k × k principal
of
B x
sub-matrix of A. At the next step write A = T , where B is a k × k matrix. Then
x a
the result holds for B and then one can proceed to prove it for A.
Exercise 1.5.5. 1. Complete the proofs of Theorems 1.3.5 and 1.3.12.

" # " # " # " #
x1 y1 cos α − sin α cos(2θ) sin(2θ)
2. Let x = ,y= ,A= and B = .
x2 y2 sin α cos α sin(2θ) − cos(2θ)
1.5. SUBMATRIX OF A MATRIX 23
(a) Then y = Ax gives the counter-clockwise rotation through an angle α.

(b) Then y = Bx gives the reflection about the line y = tan(θ)x.
(c) Let α = θ and compute y = (AB)x and y = (BA)x. Do they correspond to reflec-
tion? If yes, then about which line(s)?
(d) Further, if y = Cx gives the counter-clockwise rotation through β and y = Dx gives
the reflections about the line y = tan(δ) x. Then prove that
i. AC = CA and y = (AC)x gives the counter-clockwise rotation through α + β.
ii. y = (BD)x and y = (DB)x give rotations. Which angles do they represent?
3. Let A ∈ Mn (C). If AB = BA for all B ∈ Mn (C) then A is a scalar matrix, i.e., A = αI

for some α ∈ C (use the matrices eij in Definition 1.4.1.1).
4. For An×n = [aij ], the trace of A, denoted tr(A), is defined by tr(A) = a11 + a22 + · · · + ann .
" # " #
3 2 4 −3
(a) Compute tr(A) for A = and A = .
2 2 −5 1
" # " # " # " #
1 1 1 1
(b) Let A be a matrix with A =2 and A =3 . Determine tr(A)?
2 2 −2 −2
(c) Let A and B be two square matrices of the same order. Then
T
i. tr(A + B) = tr(A) + tr(B).

AF
ii. tr(AB) = tr(BA).

DR
(d) Does there exist matrices A, B ∈ Mn (C) such that AB − BA = cI, for some c 6= 0?
5. Let J ∈ Mn (R) be a matrix having each entry 1.
(a) Verify that J = 11T , where 1 is a column vector having all entries 1.
(b) Verify that J 2 = nJ.
(c) Also, for any α1 , α2 , β1 , β2 ∈ R, verify that there exist α3 , β3 ∈ R such that
(α1 In + β1 J) · (α2 In + β2 J) = α3 In + β3 J.
(d) Let α, β ∈ R such that α 6= 0 and α + nβ 6= 0. Now, define A = αIn + βJ. Then,
use the above to prove that A is invertible.
6. Suppose the matrices B and C are invertible and the involved partitioned products are
defined, then verify that that
" #−1 " #
A B 0 C −1
= .
C 0 B −1 −B −1 AC −1
" #
A11 x
7. Let A = , where A11 ∈ Mn (C) is invertible and c ∈ C.
y∗ c
(a) If p = c − y∗ A−1
11 x is non zero, then verify that
" # " #
−1 −1
A 0 1 A x h i
A−1 = 11
+ 11
y∗ A−1
11 −1 .
0 0 p −1
   
0 −1 2 0 −1 2
   
(b) Use the above to find the inverse of 
 1 1 4 
 and  3
 1 4 
.
−2 1 1 −2 5 −3
8. Let x ∈ Mn,1 (R) be a unit vector (recall the reflection matrix).
(a) Define A = In − 2xxT . Prove that A is symmetric and A2 = I. The matrix A is

commonly known as the Householder matrix.
(b) Let α 6= 1 be a real number and define A = In − αxxT . Prove that A is symmetric
and invertible. [The inverse is also of the form In + βxxT , for some β.]
9. Let A ∈ Mn (R) be an invertible matrix and let x, y ∈ Mn,1 (R). Also, let β ∈ R such that
α = 1 + βyT A−1 x 6= 0. Then, verify the famous Shermon-Morrison formula
β −1 T −1
(A + βxyT )−1 = A−1 − A xy A .
α
This formula gives the information about the inverse when an invertible matrix is modified
T
by a rank (see Definition 2.5.1) one matrix.

AF
10. Let A ∈ Mm,n (C). Then, a matrix G ∈ Mn,m (C) is called a generalized inverse (for
DR
short, g-inverse) of A if AGA = A.

" #
1 − 2α
(a) For example, a generalized inverse of the matrix A = [1, 2] is a matrix G = ,
α
for all α ∈ R. So, for a fixed singular matrix A, there are infinitely many g-inverses.
(b) A generalized inverse G is called a pseudo inverse or a Moore-Penrose inverse
if GAG = G and the matrices AG and GA are symmetric. Thus, verify that AG and
GA are both idempotent matrices. It can also be shown that the pseudo inverse of a
2
matrix is unique. Check that for α = the matrix G is a pseudo inverse of A.
5
(c) It turns out that among all the g-inverses, the inverse with the least euclidean norm
is the pseudo inverse.
1.6 Summary
In this chapter, we started with the definition of a matrix and came across lots of examples.
We recall these examples as they will be used in later chapters to relate different ideas:
1. The zero matrix of size m × n, denoted 0m×n or 0.
2. The identity matrix of size n × n, denoted In or I.
3. Triangular matrices.
1.6. SUMMARY 25
4. Hermitian/Symmetric matrices.
5. Skew-Hermitian/skew-symmetric matrices.
6. Unitary/Orthogonal matrices.
7. Idempotent matrices.
8. Nilpotent matrices.
We also learnt product of two matrices. Even though it seemed complicated, it basically
tells that multiplying by a matrix on the
1. left of A is same as operating on (playing with) the rows of A.
2. right of A is same as operating on (playing with) the columns of A.
The matrix multiplication is not commutative. We also defined the inverse of a matrix. Further,
there were exercises that informs us that the rows and columns of invertible matrices cannot
have certain properties.
T
AF
DR
T
AF
DR
Chapter 2
System of Linear Equations
2.1 Introduction
We start this section with our understanding of the system of linear equations.
Example 2.1.1. Let us look at some examples of linear systems.
1. Suppose a, b ∈ R. Consider the system ax = b in the variable x. If
(a) a 6= 0 then the system has a unique solution x = ab .

(b) a = 0 and
T
AF
i. b 6= 0 then the system has no solution.

ii. b = 0 then the system has infinite number of solutions, namely all x ∈ R.
DR
2. Recall that the linear system ax + by = c for (a, b) 6= (0, 0), in the variables x and y,
represents a line in R2 . So, let us consider the points of intersection of the two lines
a1 x + b1 y = c1 , a2 x + b2 y = c2 , (2.1.1)
where a1 , a2 , b1 , b2 , c1 , c2 ∈ R with (a1 , b1 ), (a2 , b2 ) 6= (0, 0) (see Figure 2.1 for illustration
of different cases).
❵✶
❵✶
❵✷ ❵✶ ✄☎❞ ❵✷
✝ ❵✷
◆♦ ❙♦❧ t✐♦♥ ■♥☞♥✐t✂ ◆ ♠❜✂✁ ♦❢ ❙♦❧ t✐♦♥s ❯♥✐✞ ✂ ❙♦❧ t✐♦♥✿ ■♥t✂✁s✂❝t✐♥❣ ▲✐♥✂s
P❛✐✁ ♦❢ P❛✁❛❧❧✂❧ ❧✐♥✂s ❈♦✐♥❝✐✆✂♥t ▲✐♥✂s ✟ ✿ P♦✐♥t ♦❢ ■♥t✂✁s✂❝t✐♦♥
Figure 2.1: Examples in 2 dimension.
(a) Unique
" # Solution
" # (a1 b2 − a2 b1 6= 0): The linear system x − y = 3 and 2x + 3y = 11
x 4
has = as the unique solution.
y 1
27
28 CHAPTER 2. SYSTEM OF LINEAR EQUATIONS
(b) No Solution (a1 b2 − a2 b1 = 0 but a1 c2 − a2 c1 6= 0): The linear system x + 2y = 1

and 2x + 4y = 3 represent a pair of parallel lines which have no point of intersection.
(c) Infinite Number of Solutions (a1 b2 − a2 b1 = 0 and a1 c2 − a2 c1 = 0): The linear
system "x #+ 2y"= 1 and # 2x
"+ # 4y =
" 2 #represent the same line. So, the solution set
x 1 − 2y 1 −2
equals = = +y with y arbitrary. Observe that the vector
y y 0 1
" #
1
i. corresponds to the solution x = 1, y = 0 of the given system.
0
" #
−2
ii. gives x = −2, y = 1 as the solution of x + 2y = 0, 2x + 4y = 0.
1
(d) If the linear system ax + by = c has
i. (a, b) = (0, 0) and c 6= 0 then ax + by = c has no solution.
ii. (a, b, c) = (0, 0, 0) then ax+by = c has infinite number of solutions, namely
whole of R2 .
Let us now look at different interpretations of the solution concept.
" # 2.1.2. Observe the following of the linear system in Example 2.1.1.2a.
Example
4
1. corresponds to the point of intersection of the corresponding two lines.
1
T
" #
1 −1
2. Using matrix multiplication, the given system equals Ax = b, where A = ,
AF
2 3
DR
" # " # " #" # " #

x 3 1 3 1 3 4
x= and b = . So, the solution is x = A−1 b = 5 = .
y 11 −2 1 11 1
" # " # " # " # " # " #
1 −1 3 1 −1 3
3. Re-writing Ax = b as x+ y = gives us 4 +1 = . This
2 3 11 2 3 11
corresponds to addition of vectors in the Euclidean plane.
Thus, there are three ways of looking at the linear system Ax = b, where, as the name
suggests, one of the ways is looking at the point of intersection of planes, the other is the vector
sum approach and the third is the matrix multiplication approach. We will see that all the
three approaches are fundamental to the understanding of linear algebra.
Definition 2.1.3. A system of m linear equations in n variables x1 , x2 , . . . , xn is a set of

equations of the form
a11 x1 + a12 x2 + · · · + a1n xn = b1

a21 x1 + a22 x2 + · · · + a2n xn = b2
.. ..
. . (2.1.2)
am1 x1 + am2 x2 + · · · + amn xn = bm
where for 1 ≤ i ≤ m and 1 ≤ j ≤ n; aij , bi ∈ R. The linear system (2.1.2) is called homoge-
neous if b1 = 0 = b2 = · · · = bm and non-homogeneous, otherwise.
2.1. INTRODUCTION 29
 
a11 a12 ··· a1n    
  x1 b1
 a21 a22 · · · a2n   .   . 
Definition 2.1.4. Let A =  . ,x= .   . 
 .  and b =  . . Then, Equa-
 
 .. .. .. .. 
. . . 
xn bm
 
am1 am2 · · · amn
tion (2.1.2) can be re-written as Ax = b, where A is called the coefficient matrix and the
block matrix [A b] is called the augmented matrix .
In the above definition, note the following.

1. the i-th row of the augmented matrix, namely, ([A b])[i, :], corresponds to the i-th linear
equation.
2. the j-th column of the augmented matrix, namely, ([A b])[:, j], corresponds to the j-th
unknown/variable whenever 1 ≤ j ≤ n and
3. the (n + 1)-th column, namely ([A b])[:, n + 1], corresponds to the vector b.
Definition 2.1.5. Consider a linear system Ax = b. Then

1. a solution of Ax = b is a vector y such that the matrix product Ay indeed equals b.
2. the set of all solutions is called the solution set of the system.
3. this linear system is called consistent if it admits a solution and is called inconsistent
if it admits no solution.
T
     
AF
1 1 1 1  0 

 

   
For example, Ax = b, with A = 1 4 2 and b = 0 has −1 as the solution set.
DR
    

4 1 1 1  2 
 
" # " # (" #)
1 1 2 1
Similarly, A = and b = has as the solution set. Further, they are consistent
1 2 3 1
systems. Whereas, the system x + y = 2, 2x + 2y = 3 is inconsistent (has no solution).
Definition 2.1.6. For the linear system Ax = b the corresponding linear homogeneous system
Ax = 0 is called the associated homogeneous system.
The readers are advised to supply the proof of the next remark.
Remark 2.1.7. Consider the linear system Ax = b with two distinct solutions, say u and v.
1. Then xh = u − v is a non-zero solution of the associated homogeneous system Ax = 0.
2. Thus, any two distinct solutions of Ax = b differs by a solution of the associated homoge-
neous system Ax = 0, i.e., {x0 + xh } is the solution set of Ax = b with x0 as a particular
solution and xh , a solution of the associated homogeneous system Ax = 0.
3. Equivalently, A (αu + (1 − α)v) = αAu + (1 − α)Av = αb + (1 − α)b = b. Thus, the line

joining the two points u and v is also a solution of the system Ax = b.
4. Now, consider the associated homogeneous linear system Ax = 0.

(a) Then, x = 0, the zero vector, is always a solution, called the trivial solution.
(b) Let w 6= 0 be a solution of Ax = 0. Then w is called a non-trivial solution. Thus

y = cw is also a solution for all c ∈ R. So, a non-trivial solution implies the solution
set has infinite number of elements.
k
P
(c) Let w1 , . . . , wk be solutions of Ax = 0. Then, ai wi is also a solution of Ax = 0,
i=1
for each choice of ai ∈ R, 1 ≤ i ≤ k.
" # " #
1 1 1
Example 2.1.8. Let A = . Then x = is a non-trivial solution of Ax = 0.
1 1 −1
Exercise 2.1.9. 1. Consider a system of 2 equations in 3 variables. If this system is con-
sistent then how many solutions does it have?
2. Give a linear system of 3 equations in 2 variables such that the system is inconsistent
whereas it has 2 equations which form a consistent system.
3. Give a linear system of 4 equations in 3 variables such that the system is inconsistent
whereas it has three equations which form a consistent system.
4. Let Ax = b be a system of m equations in n variables, where A ∈ Mm,n (R).

(a) Can the system, Ax = b have exactly two distinct solutions for any choice of m and
n? Give reasons for your answer.
(b) Can the system Ax = b have only a finitely many (greater than 1) solutions for any
T
choice of m and n? Give reasons for your answer.

AF
DR
2.2 Row-Reduced Echelon Form (RREF)

A system of linear equations can be solved by people differently. But, the final solution set
remains the same. In this section, we use a systematic way to solve any linear system which is
popularly known as the Guass - Jordan. We start with Gauss Elimination Method.
2.2.1 Gauss Elimination Method
To proceed with the understanding of the solution set of a system of linear equations, we start
with the definition of a pivot.
Definition 2.2.1. Let A be a non-zero matrix. Then, in each non-zero row of A, the left most
non-zero entry is called a pivot/leading entry. The column containing the pivot is called a
pivotal column.
If aij is a pivot then we denote it by aij . For example, the entries a12 and a23 are pivots
 
0 3 4 2
 
in A =   0 0 0 0. Thus, columns 2 and 3 are pivotal columns.

0 0 2 1
Definition 2.2.2. A matrix is in row echelon form (REF) (staircase/ ladder like)
1. if the zero rows are at the bottom;

2.2. ROW-REDUCED ECHELON FORM (RREF) 31
2. if the pivot of the (i + 1)-th row, if it exists, comes to the right of the pivot of the i-th
row.
3. if the entries below the pivot in a pivotal column are 0.

Example 2.2.3. 1. The following matrices are in echelon form.

    1 2 0 5  
0 2 4 2 1 1 0 2 3   1 0 0
    0 2 0 6  
0 0 1 1,  0 0 0 3
  4 , 
   and 0 1 0.
 0 0 0 1  
0 0 0 0 0 0 0 0 1 0 0 1
 
0 0 0 0
2. The
 following matrices
 are not in echelon 
form (determine the rule(s) that fail).
0 1 4 2 1 1 0 2 3
   
0
 0 0 0 and  0
  0 0 0 1.
0 0 1 1 0 0 0 1 4
We now start with solving two systems of linear equations. The idea is to manipulate the
rows of the augmented matrix in place of the linear equations themselves. Since, multiplying
a matrix on the left corresponds to row operations, we left multiply by certain matrices to
the augmented matrix so that the final matrix is in row echelon form (REF). The process of
obtaining the REF of a matrix is called the Gauss Elimination method. The readers should
carefully look at the matrices being multiplied on the left in the examples given below.
T
Example 2.2.4. 1. Solve the linear system y + z = 2, 2x + 3z = 5,x + y + z =3.

AF
0 1 1 2
 
DR
Solution: Let B0 = [A b], the augmented matrix. Then, B0 = 2 0 3


. We now
5
1 1 1 3
systematically proceed to get the solution.
(a) Interchange 1-st and 2-nd equations (interchange B0 [1, :] and B0 [2, :] to get B1 ).
   
2x + 3z = 5 0 1 0 2 0 3 5
   
y+z =2 B1 = 1 0 0B0 =  0 1 1 2 .
  
x+y+z =3 0 0 1 1 1 1 3
1
(b) In the new system, replace 3-rd equation by 3-rd equation minus times the 1-st
2
1
equation (replace B1 [3, :] by B1 [3, :] − B1 [1, :] to get B2 ).
2
   
2x + 3z = 5 1 0 0 2 0 3 5
   
y+z =2 B2 =   0 1 0 B1 =  0
 1 1 2 .
1 1
y − 2z = 2 −1/2 0 1 0 1 −1/2 1/2
(c) In the new system, replace 3-rd equation by 3-rd equation minus 2-nd equation
(replace B2 [3, :] by B2 [3, :] − B2 [2, :] to get B3 ).
   
2x + 3z =5 1 0 0 2 0 3 5
   
y+z =2 B3 =  0 1 0B2 = 0
 1 1 2  .
3 3 0 0 -3/2 −3/2
−2z = −2 0 −1 1
Observe that the matrix B3 is in REF. Using the last row of B3 , we get z = 1. Using
this and the second row of B3 gives y = 1. Finally, the first row gives x = 1. Hence,
the solution set of Ax = b is {[x, y, z]T | [x, y, z] = [1, 1, 1]}, a unique solution. The
method of finding the values of the unknowns y and x, using the 2-nd and 1-st row of B3
and the value of z is called back substitution.
2. Solve the linear system x + y +
 z = 4, 2x +
 3z = 5, y + z = 3.
1 1 1 4
 
Solution: Let B0 = [A b] = 2 0 3 5

 be the augmented matrix. Then
0 1 1 3
(a) The given system looks like (correspond to the augment matrix B0 ).
 
x+y+z =4 1 1 1 4
 
2x + 3z = 5  2 0 3 5 .
B0 =  
y+z =3 0 1 1 3
(b) In the new system, replace 2-nd equation by 2-nd equation minus 2 times the 1-st
equation (replace B0 [2, :] by B0 [2, :] − 2 · B0 [1, :] to get B1 ).
   
x+y+z = 4 1 0 0 1 1 1 4
   
−2y + z = −3 B1 = −2 1 0B0 =  0
   -2 .
1 −3
T
y+z = 3 0 0 1 0 1 1 3
AF
DR
(c) In the new system, replace 3-rd equation by 3-rd equation plus 1/2 times the 2-nd
equation (replace B1 [3, :] by B1 [3, :] + 1/2 · B1 [2, :] to get B2 ).
   
x+y+z = 4 1 0 0 1 1 1 4
−2y + z = −3
   
B2 = 0 1 0B1 =  0
   -2 1 −3 
.
3 3
z= 0 1/2 1 0 0 3/2 3/2
2 2
Observe that the matrix B2 is in REF. Verify that the solution set is {[x, y, z]T | [x, y, z] =
[1, 2, 1]}, again a unique solution.
In both the Examples, observe the following.

1. Each operation on the linear system corresponds to a similar operation on the rows of the
augmented matrix.
2. At each stage, the new augmented matrix was obtained by left multiplication by a matrix,
say E. Note that E is obtained by changing exactly one row of the identity matrix. The
readers should find the relationship between the matrix E and the row operations that
have been made on the augmented matrix.
3. Thus, the task of understanding the solution set of a linear system reduces to understand-
ing the form of the matrix E.
4. Note that E corresponds to a row operation made on the identity matrix I3 .
5. Also, for each matrix E note that we have a matrix F , again a variant of I3 such that
EF = I3 = F E.
We use the above ideas to define elementary row operations and the corresponding elemen-
tary matrices in the next subsection.
2.2.2 Elementary Row Operations
Definition 2.2.5. Let A ∈ Mm,n (C). Then, the elementary row operations are
1. Eij : Interchange the i-th and j-th rows, namely, interchange A[i, :] and A[j, :].
2. Ek (c) for c 6= 0: Multiply the k-th row by c, namely, multiply A[k, :] by c.
3. Eij (c) for c 6= 0: Replace the i-th row by i-th row plus c-times the j-th row, namely,
replace A[i, :] by A[i, :] + cA[j, :].
Definition 2.2.6. A matrix E ∈ Mm (R) is called an elementary matrix if it is obtained by

applying exactly one elementary row operation to the identity matrix Im .
For better understanding we give the elementary matrices for m = 3.

Notation Operations on Elementary Row Opera- Elementary
Equations tions: A is 3 × n matrix Matrix
 
T
1 0 0
AF
 
E2 (c), c 6= 0; Multiply the 2-th A[2, :] ← cA[2, :] 0 c 0
 
DR
Ek (c), c 6= 0 row by c 0 0 1
 
1 0 0
 
E21 (c), c 6= 0; Replace 2-th row by A[2, :] ← A[2, :] + cA[1, :]  c 1 0
 
Eij (c), c 6= 0 2-nd row plus c-times 0 0 1
1-st row
 
1 0 0
 
E23 ; Eij Interchange 2-nd and Interchange A[2, :] and A[3, :] 
0 0 1

3-rd rows 0 1 0
     
1 0 0 1 −5 0 0 0 1
     
Example 2.2.7. Verify that E2 (5) = 0 5
 0, E12 (−5) = 0 1 0 and E13 = 0 1 0
   
0 0 1 0 0 1 1 0 0
are elementary matrices.
Exercise 2.2.8. 1. Which of the following matrices are elementary?
           
1
2 0 1 0 0 1 −1 0 1 0 0 0 0 1 0 0 1
  2         
0 1 0 ,  0 1 0 , 0 1 0 , 5 1 0 , 0 1 0 , 1 0 0 .
     
    
0 0 1 0 0 1 0 0 1 0 0 1 1 0 0 0 1 0
" #
2 1
2. Find some elementary matrices E1 , . . . , Ek such that Ek · · · E1 = I2 .
1 2
We now give the elementary matrices for general n.
Example 2.2.9. Let e1 , . . . , en be the standard unit vectors of Mn,1 (R). Then, using eTi ej =
0 = eTj ei and eTi ei = 1 = eTj ej , verify that each elementary matrix is invertible.
1. Ek (c) = In + (c − 1)ek eTk for c 6= 0. Verify that
Ek (c)Ek (1/c) = In + (c − 1)ek eTk In + (1/c − 1)ek eTk = In = Ek (1/c)Ek (c).

2. Eij (c) = In + c ei eTj for c 6= 0. Verify that
Eij (c)Eij (−c) = In + c ei eTj In − c ei eTj = In = Eij (−c)Eij (c).

3. Eij = In − ei eTi − ej eTj + ei eTj + ej eTi . Verify that
Eij Eij = In − ei eTi − ej eTj + ei eTj + ej eTi In − ei eTi − ej eTj + ei eTj + ej eTi = In .

We now show that the above elementary matrices correspond to respective row operations.
Remark 2.2.10. Let A ∈ Mm,n (R).

1. For c 6= 0, Ek (c)A corresponds to the replacement of A[k, :] by cA[k, :].
Using eTk ek = 1 and A[k, :] = eTk A, we get
T
(Ek (c)A)[k, :] = eTk (Ek (c)A) = eTk Im + (c − 1)ek eTk A = eTk + (c − 1)eTk (ek eTk ) A

AF
= eTk + (c − 1)eTk A = ceTk A = cA[k, :].

DR
A similar argument with eTi ek = 0, for i 6= k, gives (Ek (c)A)[i, :] = A[i, :], for i 6= k.
2. For c 6= 0, Eij (c)A corresponds to the replacement of A[i, :] by A[i, :] + cA[j, :].
Using eTi ei = 1 and A[i, :] = eTi A, we get
(Eij (c)A)[i, :] = eTi (Eij (c)A) = eTi Im + c ei eTj A = eTi + c eTi (ei eTj ) A

= eTi + c eTj A = A[i, :] + cA[j, :].

A similar argument with eTk ei = 0, for k 6= i, gives (Eij (c)A)[k, :] = A[k, :], for k 6= i.
3. Eij A corresponds to interchange of A[i, :] and A[j, :].
Using eTi ei = 1, eTi ej = 0 and A[i, :] = eTi A, we get
(Eij A)[i, :] = eTi (Eij A) = eTi Im − ei eTi − ej eTj + ei eTj + ej eTi A

= eTi − eTi − 0T + eTj + 0T A = eTj A = A[j, :]

Similarly, using eTj ej = 1, eTj ei = 0 and A[j, :] = eTj A show that (Eij A)[j, :] = A[i, :].
Further, using eTk ei = 0 = eTk ej , for k 6= i, j show that (Eij A)[k, :] = A[k, :].
Definition 2.2.11. Two matrices A and B are said to be row equivalent if one can be
obtained from the other by a finite number of elementary row operations. Or equivalently,
there exists elementary matrices E1 , . . . , Ek such that B = E1 · · · Ek A.
Definition 2.2.12. The linear systems Ax = b and Cx = d are said to be row equivalent if
their respective augmented matrices, [A b] and [C d], are row equivalent.
Thus, note that the linear systems at each step in Example 2.2.4 are row equivalent to each
other. We now prove that the solution set of two row equivalent linear systems are same.
Theorem 2.2.13. Let Ax = b and Cx = d be two row equivalent linear systems. Then they
have the same solution set.
Proof. Let E1 , . . . , Ek be the elementary matrices such that E1 · · · Ek [A b] = [C d]. Put

E = E1 · · · Ek . Then
EA = C, E b = d, A = E −1 C and b = E −1 d. (2.2.3)
Now assume that Ay = b holds. Then, by Equation (2.2.3)
C y = EA y = E b = d. (2.2.4)
On the other hand if C z = d holds then using Equation (2.2.3), we have
A z = E −1 C z = E −1 d = b. (2.2.5)
Therefore, using Equations (2.2.4) and (2.2.5) the required result follows.
T
The following result is a particular case of Theorem 2.2.13.

AF
Corollary 2.2.14. Let A and B be two row equivalent matrices. Then, the systems Ax = 0
DR
and Bx = 0 have the same solution set.

   
1 0 0 1 0 a
   
Example 2.2.15. Are the matrices A =  0 1 0 and B = 0 1 b row equivalent?
  
0 0 1 0 0 0
 
a
 
Solution: No, as 
 b  is a solution of Bx = 0 but it isn’t a solution of Ax = 0.

−1
The following exercise shows that every square matrix is row equivalent to an upper trian-
gular matrix. We will come back to this idea again in the chapter titled “Advanced Topics”.
Exercise 2.2.16. 1. Let A = [aij ] ∈ Mn (R). Then there exists an orthogonal matrix U
such that U A is upper triangular. The proof uses the following ideas.
(a) If A[1, :] = 0 then proceed to the next column. Else, A[:, 1] 6= 0.
(b) If A[:, 1] = αe1 , for some α ∈ R, α 6= 0, proceed to the next column. Else, either
a11 = 0 or a11 6= 0.
(c) If a11 = 0 then left multiply A with E1i (an orthogonal matrix) so that the (1, 1)
entry of B = E1i A is non-zero. Hence, without loss of generality, let a11 6= 0.
(d) Let [w1 , . . . , wn ]T = w ∈ Rn with w1 6= 0. Then use the Householder matrix H such
that Hw = w1 e1 , i.e., find x ∈ Rn such that (In − 2xxT )w = w1 e1 .
" #
w1 ∗
(e) So, Part 1d gives an orthogonal matrix H1 with H1 A = .
0 A1
(f ) Use induction to get H2 ∈ Mn−1 (R) satisfying H2 A1 = T1 , an upper triangular
matrix. " # " #
1 0T w1 ∗
(g) Define H = H1 . Then H is an orthogonal matrix and HA = , an
0 H2 0 T1
upper triangular matrix.
2. Let A ∈ Mn (R) such that tr(A) = 0. Then prove that there exists a non-singular matrix
S such that SAS −1 = B with B = [bij ] and bii = 0, for 1 ≤ i ≤ n.
2.3 Initial Results on LU Decomposition

Consider the linear system Ax = b, where A ∈ Mn (R) and b ∈ Mn (R) are known matrices and
x ∈ Mn (R) is the unknown matrix. Recall that in Example 2.2.4 we used the Gauss Elimination
method to get the REF of the augmented matrix. If the REF was [C d] then C was an upper
triangular matrix. The upper triangular form of C was used during back substitution.
The decomposition of a square matrix A as LU , where L is a lower triangular matrix and U is
an upper triangular matrix plays an important role in numerical linear algebra. To get a better
understanding of the LU -decomposition (LU -factorization), we recall a few observations
T
and give a few examples.

AF
1. Observe that solving the system Ax = b is quite easy whenever A is a triangular matrix.
DR
2. As L is a triangular matrix, the linear system Ly = b can be easily solved. Let y0 be a

solution of Ly = b.
3. Now consider the linear system U z = y0 . As U is again triangular, this system can again
be easily solved. Let u0 be a solution of U z = y0 . Then, z0 is a solution of the system
Ax = b as
Az0 = (LU )z0 = L(U z0 ) = Ly0 = b.
Hence, we observe that solving the system Ax = b reduces to solving two easier linear systems,
namely Ly = b and U z = y, where y is obtained as a solution of Ly = b.
To give the LU -decomposition for a square matrix A, we need to know the determinant of A,
namely det(A), and its properties. Since, we haven’t yet studied it, we just give the idea of the
LU -decomposition. For the general case, the readers should see the chapter titled “Advanced
Topics”. Let us start with a few examples.
" #
0 1
Example 2.3.1. 1. Let A = . Then A cannot be decomposed into LU .
1 0
" #" #
a 0 e f
For if, A = LU = then the numbers a, b, c, e, f, g ∈ R satisfy
b c 0 g
ae = 0, af = 1, be = 1 and bf + cg = 0.
But, ae = 0 implies either a = 0 or e = 0, contradicting either af = 1 or be = 1.

2.3. INITIAL RESULTS ON LU DECOMPOSITION 37
" # " # " #
1 1 0 1
2. Let > 0 and A = . Then, A = LU , where L = 1 and U = 1
. Thus,
1 0 1 0 −
comparing it with the previous example, we see that the LU -decomposition is highly
unstable.
" # " # " #
1 2 1 0 1 2
3. Let A = . Then, for any choice of α 6= 0, L = and U = gives
2 2 2 α 0 − α2
A = LU . Check that if we restrict ourselves with the condition that the diagonal entries
of L are all 1 then the decomposition is unique.
 
1 1 1
 
4. Let A =  2 0 3. Then, using ideas in Example 2.2.4.2 verify that A = LU , where
01 1
   
1 0 0 1 1 1
   
L = 2
 1 0 and U = 0 −2 1 .
 
0 −1/2 1 0 0 3/2
5. Recall that in Example 2.2.4.2, we had pivots pivots at each stage. Whereas, in Exam-
ple 2.2.4.1, we had to interchange the first and second row to get a pivot. So, it is not
possible to write A = LU .
6. Finally, using A 
= LU , the system
 Ax  = b reduces to LU x = b. Here,
 the solution of
4 4 1
T
     
Ly = b, for b = 5 equals y =  −3 . This, in turn implies x = 2 as the solution of
AF
    
3 3/2 1
DR
both U x = y or Ax = b.
So, to proceed further, let A ∈ Mn (R). Then, recall that for any S ⊆ {1, 2, . . . , n}, A[S, S]
denotes the principal submatrix of A corresponding to the elements of S (see Page 21). Then,
we assume that det(A[S, S]) 6= 0, for every S = {1, 2, . . . , i}, 1 ≤ i ≤ n.
We need to show that there exists an invertible lower triangular matrix L such that LA is
an invertible upper triangular matrix. The proof uses the following ideas.
" #
a11 A12
1. By assumption A[1, 1] = a11 6= 0. Write A = , where A22 is a (n − 1) × (n − 1).
A21 A22
" #
1 0T −1
2. Let L1 = , where x = A21 . Then L1 is a lower triangular matrix and
x In−1 a11
" #" # " # " #
1 0T a11 A12 a11 A12 a11 A12
L1 A = = = .
x In−1 A21 A22 a11 x + A21 xA12 + A22 0 xA12 + A22
3. Note that (2, 2)-th entry of L1 A equals the (1, 1)-th entry of xA12 + A22 . This equals
   
a21
 −1  . h i a11 a22 − a12 a21 A[{1, 2}, {1, 2}]
 .  a12 · · · a1n  + (A22 )11 = = 6= 0.
 a11  . 

 a11 a11
an1
11
" #
a11 ∗
4. Thus, L1 is an invertible lower triangular matrix with L1 A = and (A1 )11 6= 0.
0 A1
Hence, det(A) = a11 det(A1 ) and det(A1 [S, S]) 6= 0, for all S ⊆ {1, 2, . . . , n − 1} as
(a) the determinant of a lower triangular matrix equals product of diagonal entries and
(b) if A and B are two n × n matrices then det(AB) = det(A) · det(B).
5. Now, using induction, we get L2 ∈ Mn−1 (R), an invertible lower triangular matrix, with
1’s on the diagonal such that L2 A1 = T1 , an invertible upper triangular matrix.
" # " #
1 0 T a ∗
11
6. Define Le= L1 . Then, verify that LA
e = , is an upper triangular matrix
0 L2 0 T1
with Le as an invertible lower triangular matrix.
−1
7. Defining L = L e , we see that L is a lower triangular matrix (inverse of a lower trian-
" #
a11 ∗
gular matrix is lower triangular) with A = LU and U = , an upper triangular
0 T1
invertible matrix.
As mentioned above, we will again come back to this at a later stage.
2.4 Row-Reduced Echelon Form (RREF)

T
We now proceed to understand the row-reduced echelon form (RREF) of a matrix. This un-
AF
derstanding will be used to define the row-rank of a matrix in the next section. In subsequent
DR
sections and chapters, RREF is used to obtain very important results.
Definition 2.4.1. A matrix C is said to be in row-reduced echelon form (RREF)
1. if C is already in REF,
2. if the pivot of each non-zero row is 1,
3. if every other entry in each pivotal column is zero.
A matrix in RREF is also called a row-reduced echelon matrix.

Example 2.4.2. 1. The following matrices
 are in RREF.
    1 0 0 5  
0 1 0 −2 0 1 3 0   1 1 0 0 0
0 0 0 1 ,  0 1 0 6
      
0 0 1 1 ,  0
 and  0 0 0 1 0.
   0 1 2 
0 0 0 0 0 0 0 0 0 0 0 0 1
 
0 0 0 0
Note that if we look at the pivotal rows and columns then I2 is present in the first two
matrices and I3 is there in the next two. Also, the subscripts 2 and 3, respectively, in I2
and I3 correspond to the number of pivots.
2. The
 following matrices
  are not inRREF
 (determinethe rule(s) that fail).
0 3 3 0 0 1 3 0 0 1 3 1
     
, 0 ,
0 0 0 1  0 0 0 0 0 0 1 .
   
0 0 0 0 0 0 0 1 0 0 0 0
Even though, we have two pivots in examples 1 and 3, the matrix "I2 doesn’t
# appear as a
3 0
submatrix in pivotal rows and columns. In the first one, we have as a submatrix
0 1
" #
1 1
and in the third the corresponding submatrix is .
0 1
We now give another examples to indicate its application to the theory of the system of
linear equations.
Example 2.4.3. Consider a linear system Ax = b, where A ∈ M3 (C) and A[:, 1] 6= 0. If

[C d] = RREF([A b]) then what are the possible choices for [C d] and what are its
implication?
Solution: Since there are 3 rows, the number of pivots can be at most 3. So, let us verify that
there are 7 different choices for [C d] = RREF([A b]).
1. There are exactly 3 pivots. These pivots can be in either the columns {1, 2, 3}, {1, 2, 4}
and {1, 3, 4} as we have assumed A[:, 1] 6= 0. The corresponding cases are given below.
 
1 0 0 d1
 
(a) Pivots in the columns 1, 2, 3 ⇒ [C d] =  0 1 0 d2 . Here, Ax = b is consistent.

0 0 1 d3
   
T
x d
   1
AF
The unique solution equals  y  = d2 .

  
DR
z d3
   
1 0 α 0 1 α 0 0
   
(b) Pivots in the columns 1, 2, 4 or 1, 3, 4 ⇒ [C d] equals 
0 1 β 0  or 0 0 1

.
0
0 0 0 1 0 0 0 1
Here, Ax = b is inconsistent for any choice of α, β as there is a row of [0 0 0 1]. This
corresponds to solving 0 · x + 0 · y + 0 · z = 1, an equation which has no solution.
2. There are exactly 2 pivots. These pivots can be in either the columns {1, 2}, {1, 3} or
{1, 4} as we have assumed A[:, 1] 6= 0. The corresponding cases are given below.
   
1 0 α d1 1 α 0 d1
   
(a) Pivots in the columns 1, 2 or 1, 3 ⇒ [C d] equals 
0 1 β d2
 or 0 0 1 d2 .
  
0 0 0 0 0 0 0 0
Here, for the first matrix, the solution set equals
       
x d1 − αz d1 −α
       
y  = d2 − βz  = d2  + z −β ,
       
z z d3 1
where z is arbitrary. Here, z is called the “Free variable” as z can be assigned any
value and x and y are called “Basic Variables” and they can be written in terms of
the free variable z and constant.
 
1 α β 0
 
(b) Pivots in the columns 1, 4 ⇒ [C d] =  0 0 0 1  which has a row of [0 0 0 1].

0 0 0 0
This corresponds to solving 0 · x + 0 · y + 0 · z = 1, an equation which has no solution.
 
1 α β d1
 
3. There is exactly one pivot. In this case [C d] =   0 0 0 0 . Here, Ax = b is consis-

0 0 0 0
tent and has infinite number of solutions for every choice of α, β as RREF([A b])
has no row of the form [0 0 0 1].
2.4.1 The Gauss-Jordan Elimination Method
So, having seen the application of the RREF to the augmented matrix, let us proceed with the
algorithm, commonly known as the Gauss-Jordan Elimination (GJE), which helps us compute
the RREF.
1. Input: A ∈ Mm,n (R).
2. Output: a matrix B in RREF such that A is row equivalent to B.
3. Step 1: Put ‘Region’ = A.
4. Step 2: If all entries in the Region are 0, STOP. Else, in the Region, find the leftmost
T
nonzero column and find its topmost nonzero entry. Suppose this nonzero entry is aij = c
AF
(say). Box it. This is a pivot.

DR
5. Step 3: Interchange the row containing the pivot with the top row of the region. Also,
make the pivot entry 1 by dividing this top row by c. Use this pivot to make other entries
in the pivotal column as 0.
6. Step 4: Put Region = the submatrix below and to the right of the current pivot. Now,
go to step 2.
Important: The process will stop, as we can get at most min{m, n} pivots.
 
0 2 3 7
 
1 1 1 1
Example 2.4.4. Apply GJE to  1 3 4 8

 
0 0 0 1
1. Region = A as A 6= 0.
   
1 1 1 1 1 1 1 1
   
 0 2 3 7 0 2 3 7
 1 3 4 8. Also, E31 (−1)E12 A =  0
2. Then, E12 A =     = B (say).
   2 3 7 
0 0 0 1 0 0 0 1
 
  1 1 1 1
2 3 7
1 32 27 
 
 6= 0. Then, E2 ( 1 )B =  0
  
3. Now, Region =  2 3 7 2 0
 = C(say). Then,
  2 3 7
0 0 1
 
0 0 0 1
 
−1 −5
1 0 2 2

0 3 7 
1 2 2 

E12 (−1)E32 (−2)C = 
0 = D(say).
 0 0
0 
0 0 0 1
 
1 0 −1 2
−5
2
" #  3 7 
0 0 0 1 2 2 

4. Now, Region = . Then, E34 D =  . Now, multiply on the left
0 1 0
 0 0 1 
0 0 0 0
 
1 0 − 12 0

0 3 
5 −7 1 2 0
by E13 ( 2 ) and E23 ( 2 ) to get 
 , a matrix in RREF. Thus, A is row
 0 0 0 1 

0 0 0 0
 
1 0 − 12 0

0 3 
1 2 0
equivalent to F , where F = RREF(A) =   .
 0 0 0 1 

0 0 0 0
5. Note that we have multiplied A on the left by the elementary matrices, E12 , E31 (−1),
E2 (1/2), E32 −2, E12 (−1), E34 , E23 (−7/2), E13 (5/2), i.e.,
E13 (5/2)E23 (−7/2)E34 E12 (−1)E32 (−2)E2 (1/2)E31 (−1)E12 A = F = RREF(A).

T
AF
6. Or equivalently, we have an invertible matrix P such that P A = F = RREF(A), where

DR
P = E13 (5/2)E23 (−7/2)E34 E12 (−1)E32 (−2)E2 (1/2)E31 (−1)E12 .

     
0 0 1 0 1 1 3 0 −1 1
     
Exercise 2.4.5. Let A = 
 1 0 3  , B = 0 0 1 3 and C = −2 0 3 . Determine
    
3 0 7 1 1 0 0 −5 1 0
their RREF.
2.4.2 Results on RREF
The proof of the next result is beyond the scope of this book and hence is omitted.
Theorem 2.4.6. Let A and B be two row equivalent matrices in RREF. Then A = B.
As an immediate corollary, we obtain the following important result.
Corollary 2.4.7. The RREF of a matrix A is unique.
Proof. Suppose there exists a matrix A having B and C as RREFs. As the RREFs are obtained
by left multiplication of elementary matrices, there exist elementary matrices E1 , . . . , Ek and
F1 , . . . , F` such that B = E1 · · · Ek A and C = F1 · · · F` A. Thus,
B = E1 · · · Ek A = E1 · · · Ek (F1 · · · F` )−1 C = E1 · · · Ek F`−1 · · · F2−1 F1−1 C.
As inverse of an elementary matrix is an elementary matrix, B and C are are row equivalent.
As B and C are in RREF, using Theorem 2.4.6, B = C.
Remark 2.4.8. Let A ∈ Mm,n (R).

1. Then, the uniqueness of RREF implies that RREF(A) is independent of the choice of the
row operations used to get the final matrix which is in RREF.
2. Let B = EA, for some elementary matrix E. Then, RREF(A) = RREF(B).
3. Then RREF(A) = P A, for some invertible matrix P .
4. Let F = RREF(A) and B = [A[:, 1], . . . , A[:, s]], for some s ≤ n. Then,
RREF(B) = [F [:, 1], . . . , F [:, s]].
Proof. By Remark 2.4.8.3, there exist an invertible matrix P , such that
F = P A = [P A[:, 1], . . . , P A[:, n]] = [F [:, 1], . . . , F [:, n]].
But, P B = [P A[:, 1], . . . , P A[:, s]] = [F [:, 1], . . . , F [:, s]]. As F is in RREF, its first s
columns are also in RREF. Thus, by Corollary 2.4.7, RREF(P B) = [F [:, 1], . . . , F [:, s]].
Now, a repeated application of Remark 2.4.8.2 implies RREF(B) = [F [:, 1], . . . , F [:, s]].
Proposition 2.4.9. Let A ∈ Mn (R). Then, A is invertible if and only if RREF(A) = In , i.e.,
every invertible matrix is a product of elementary matrices.
T
Proof. If RREF(A) = In then there exist elementary matrices E1 , . . . , Ek such that In =

AF
RREF(A) = E1 · · · Ek A. As elementary matrices are invertible and product of invertible ma-

DR
trices are invertible, we get A = (E1 · · · Ek )−1 ⇔ A−1 = E1 · · · Ek .

Now, let A be invertible with B = RREF(A) = E1 · · · Ek A, for some elementary matrices
E1 , . . . , Ek . As A and Ei ’s are invertible, the matrix B is invertible. Hence, B doesn’t have any
zero row. Thus, all the n rows of B have pivots. Therefore, B has n pivotal columns. As B
has exactly n columns, each column is a pivotal column and hence B = In . Thus, the required
result follows.
Exercise 2.4.10. 1. Let A ∈ Mm,n (R). Then RREF(SA) = RREF(A), for any invertible
matrix S ∈ Mn (R).
h i
2. Let A ∈ Mn (R) be an invertible matrix. Then, for any matrix B, define C = A B and
" # " #
A h i In
D= . Then, RREF(C) = In A−1 B and RREF(D) = .
B 0
2.4.3 Computing the Inverse using GJE
Recall that if A ∈ Mn (C) is invertible then there exists a matrix B such that AB = In = BA.
So, we want to find a B such that
h i h i h i
e1 · · · en = In = AB = A B[:, 1] · · · B[:, n] = AB[:, 1] · · · AB[:, n] .
h i
So, if B = B[:, 1] · · · B[:, n] is the matrix of unknowns then we need to solve n-system
of linear equations AB[:, 1] = e1 , . . ., AB[:, n] = en . Thus, we have n-augmented matrices
2.5. RANK OF A MATRIX 43
[A | e1 ], . . ., [A | en ]. So, in place of solving the n-augmented matrices separately, the idea of

GJE is to consider the augmented matrix
h i h i
A e1 e2 · · · en = A In .
h i h i
Thus, if E is an invertible matrix such that E A In = In B then EA = In and E = B.
Hence, invertibility of E implies AE = In and hence, B = E = A−1 . This idea together with
Remark 2.4.8.4 is used to compute A−1 whenever it exists.
Let A ∈ Mn (R). Compute RREF(C), where C = [A I In ]. Then RREF(C) = [RREF(A) B].
Now, either RREF(A) = In or RREF(A) 6= In . Thus, if RREF(A) = In then we must have
B = A−1 . Else, A is not invertible. We show this with an example.
 
0 0 1
 
Example 2.4.11. Use GJE to find the inverse of A =  0 1 1.

1 1 1
 
0 0 1 1 0 0
 
Solution: Applying GJE to [A | I3 ] = 
 0 1 1 0 1 0  gives

1 1 1 0 0 1
   
1 1 1 0 0 1 1 1 0 −1 0 1
E13   E13 (−1),E23 (−2)  
[A | I3 ] →  0 1 1 0 1 0 → 0 1 0 −1 1 0
T
  
AF
0 0 1 1 0 0 0 0 1 1 0 0
 
DR
1 0 0 0 −1 1
E12 (−1)  
→  0 1 0 −1 1 .
0
0 0 1 1 0 0
 
0 −1 1
Thus, A−1 = 
 
−1 1 .
0
1 0 0
   
1 2 3 1 3 3
   
Exercise 2.4.12. Use GJE to compute the inverseA = 
1 3 2 and B = 2 3 2.
  
2 4 7 3 5 4
2.5 Rank of a Matrix

Recall that the RREF of a matrix is unique. So, we use RREF(A) to define the rank of a A.
Definition 2.5.1. Let A ∈ Mm,n (C). Then, the rank of A, denoted Rank(A), is the number
of pivots in the RREF(A).
Note that Rank(A) is defined using the number of pivots in RREF (A). These pivots
were obtained using the row operations. The question arises, what if we had applied column
operations? That is, what happens when we multiply by invertible matrices on the right?
Will the pivots using column operations remain the same or change? This question cannot be
answered at this stage. Using the ideas in vector spaces, we can show that the number of pivots
do not change and hence, we just use the word Rank(A).
We now illustrate the calculation of the rank by giving a few examples.
Example 2.5.2. Determine the rank of the following matrices.
1. Rank(In ) = n and Rank(0) = 0.
2. Let A = diag(d1 , . . . , dn ). Then, Rank(A) equals the number of non-zero di ’s.

" # " #
1 2 1 2
3. Let A = . Then RREF(A) = ⇒ Rank(A) = 1.
2 4 0 0
" # " #
1 2 3 1 2 3
4. Let A = . Then RREF(A) = ⇒ Rank(A) = 1.
2 4 6 0 0 0
   
1 0 0 1 0 0
   
5. Let A =  0 1 0 . Then RREF(A) = 0 1 0 ⇒ Rank(A) = 2.
   
2 2 0 0 0 0
" # " #
1 2 −2 −2
6. Let A = and B = . Then, Rank(A) = Rank(B) = 1. Also, verify that
2 4 1 1
T
AF
" #
−6 −12
AB = 0 and BA = . So, Rank(AB) = 0 6= 1 = Rank(BA). Observe that A
3 6
DR
and B are not invertible. So, the rank can either remain the same or reduce.
 
1 2 1 1 1
 
7. Let A = 
2 3 1 2 2 . Then, Rank(A) = 2 as it’s REF has two pivots.

1 1 0 1 1
Remark 2.5.3. Before proceeding further, for A, B ∈ Mm,n (C), we observe the following.
1. If A and B are row-equivalent then Rank(A) = Rank(B).
2. The number of pivots in the RREF(A) equals the number of pivots in REF of A. Hence,
one needs to compute only the REF to determine the rank.
Exercise 2.5.4. Let A ∈ Mm,n (R).

1. Then Rank(A) ≤ min{m, n}.
" #
A C
2. If B = then Rank(B) = Rank([A C]).
0 0
" #
A 0
3. If B = then Rank(B) = Rank(A)
0 0
4. If A = P B, for some invertible matrix P then Rank(A) = Rank(B).
" #
A11 A12 h i h i
5. If A = then Rank(A) ≤ Rank A11 A12 + Rank A21 A22 .
A21 A22
We now have the following result.
Corollary 2.5.5. Let A ∈ Mm,n (R) and B ∈ Mn,q (R). Then, Rank(AB) ≤ Rank(A).
In particular, if B ∈ Mn (R) is invertible then Rank(AB) = Rank(A).
Proof. Let Rank(A) = r." Then,

# there exists an
" invertible
# # P and A1 ∈ Mr,n (R) such
" matrix
A1 A1 A1 B
that P A = RREF(A) = . Then, P AB = B= . So, using Remark 2.5.3 and
0 0 0
Exercise 2.5.4.2
" #!
A1 B
Rank(AB) = Rank(P AB) = Rank = Rank(A1 B) ≤ r = Rank(A). (2.5.1)
0
In particular, if B is invertible then, using Equation (2.5.1), we get
Rank(A) = Rank(ABB −1 ) ≤ Rank(AB)
and hence the required result follows.
Proposition 2.5.6. Let A ∈ Mn (C) be an invertible matrix and let S be any subset of {1, 2, . . . , n}.
Then Rank(A[S, :]) = |S| and Rank(A[:, S]) = |S|.
Proof. Without loss of generality, let S = {1, . . . , r} and S c = {r + 1, . . . , n}. Write A1 = A[:, S]
T
and A2 = A[:, S c ]. Since A is invertible, RREF(A) = In . Hence, by Remark 2.4.8.3, there exists
AF
an invertible matrix P such that P A = In . Thus,

DR
" #
h i h i Ir 0
P A1 P A2 = P A1 A2 = P A = In = .
0 In−r
" # " #
Ir 0
Thus, P A1 = and P A2 = . So, using Corollary 2.5.5, Rank(A1 ) = r.
0 In−r
For the second part, let B1 = A[S, :], B2 = A[S c , :] and let Rank(B1 ) = t < s. Then, by
Exercise 2, there exists an s × s invertible matrix Q and a matrix C in RREF, of size t × n and
having exactly t pivots, such that
" #
C
QB1 = RREF(B1 ) = . (2.5.2)
0
As t < s, QB1 has at least one zero row. Then

 
" # " #" # " # C
Q 0 Q 0 B1 QB1  
A= = =
 0 .

0 In−r 0 In−r B2 B2
B2
" #
Q 0
As and A are invertible, their product is invertible. But, their product has a zero
0 In−r
row, a contradiction. Thus, Rank(B1 ) = s.
Let us also look at the following example to understand the prove of the next theorem.
   
1 2 3 4 1 2 0 0
   
Example 2.5.7. Let A = 2 4 5
 1. Then RREF(A) = 0 0 1
 
. Thus, we have
0
1 2 4 7 0 0 0 1
an invertible matrix P such that P A = RREF(A). Note that Rank(A)= 3 and
h hence
i I3 is
a submatrix of RREF(A). So, we need to find Q ∈ M4 (R) such that P AQ = I3 0 . Now,
consider the following column operations.
 
1 0 0 2
 
1. Let B = 0 1 0 0. Then B = RREF(A) E23 E34 .

0 0 1 0
 
  1 0 0 −2
1 0 0 2   h
T
 0 1 0 0 i
2. B(E41 (−2)) = BE14 (−2) = 0 1 0 0
 
0
 = I3 0 .
0 1 0
0 0 1 0
 
0 0 0 1
h i
3. Thus, define Q = E23 E34 E14 (−2) to get P AQ = I3 0 .
Theorem 2.5.8. Let A ∈ Mm,n (R). If Rank(A) = r then, there exist invertible matrices P and
Q such that
" #
Ir 0
P AQ= .
T
0 0
AF
Proof. Let C = RREF(A). Then, by Remark 2.4.8.3 there exists as invertible matrix P such
DR
that C = P A. Note that C has r pivots and they appear in columns, say i1 < i2 < · · · < ir .
Now, let Q1 = E1i1 E2i2 · · · Erir . As Ejij ’s are elementary matrices that interchange the
" #
Ir B
columns of C, one has D = CQ1 = , where B ∈ Mr,n−r (R).
0 0
" #
Ir −B
Now, let Q2 = and Q = Q1 Q2 . Then Q is invertible and
0 In−r
" #" # " #
Ir B Ir −B Ir 0
P AQ = CQ = CQ1 Q2 = DQ2 = = .
0 0 0 In−r 0 0

Since we are multiplying by invertible matrices on the right, the idea of the above theorem
cannot be used for solving the system of linear equations. But, this idea can be used to
understand the properties of the given matrix, such as ideas related to rank-factorization, row-
space, column space and so on which have not yet been defined.
As a corollary of Theorem 2.5.8, we now give the rank-factorization of a matrix A.
Corollary 2.5.9. Let A ∈ Mm,n (R). If Rank(A) = r then there exist matrices B ∈ Mm,r (R)
r
and C ∈ Mr,n (R) such that Rank(B) = Rank(C) = r and A = BC. Furthermore, A = xi yiT ,
P
i=1
for some xi ∈ Rm and yi ∈ Rn .
" #
Ir 0
Proof. By Theorem 2.5.8, there exist invertible matrices P and Q such that P A Q = .
0 0
" # " #
I r 0 C
Or equivalently, A = P −1 Q−1 . Decompose P −1 = [B D] and Q−1 = such that
0 0 F
B ∈ Mm,r (R) and C ∈ Mr,n (R). Then Rank(B) = Rank(C) = r (see Proposition 2.5.6) and
" #" # " #
Ir 0 C h i C
A = [B D] = B 0 = BC.
0 0 F F
 
y1T
h i  .  r
.. . Then A = BC = xi yiT .
P
Furthermore, assume that B = x1 · · · xr and C = 
 
i=1
yrT
Proposition 2.5.10. Let A, B ∈ Mm,n (R). Then, prove that Rank(A + B) ≤ Rank(A) +
k
xi yiT , for some xi , yi ∈ R, for 1 ≤ i ≤ k, then Rank(A) ≤ k.
P
Rank(B). In particular, if A =
i=1
Proof. Let Rank(A) = r. Then, # exists an invertible matrix P and a matrix A1 ∈ Mr,n (R)
" there
A1
such that P A = RREF(A) = . Then,
0
" # " # " #
T
A1 B1 A1 + B 1
P (A + B) = P A + P B = + = .
AF
0 B2 B2
DR
Now using Corollary 2.5.5, we have
Rank(A + B) = Rank(P (A + B)) ≤ r + Rank(B2 ) ≤ r + Rank(B) = Rank(A) + Rank(B).
Thus, the required result follows. The other part follows, as Rank(xi yiT ) = 1, for 1 ≤ i ≤ k.
Exercise 2.5.11. 1. Let A ∈ Mm,n (R) be a matrix of rank 1. Then prove that A = xyT ,
for non-zero vectors x ∈ Rm and y ∈ Rn .
2. Let A ∈ Mm (R). If Rank(A) = 1 then prove that A2 = αA, for some scalar α.
" # " #
2 4 8 1 0 0
3. Let A = and B = .
1 3 2 0 1 0
h i
(a) Find P and Q such that B = P AQ. Thus, A = P −1 I2 0 Q−1 .
" #
I2
(b) Define G = Q T P . Then, verify that AGA = A. Hence, G is a g-inverse of A.
x
" #
I2
(c) In particular, if b = 0 then G = Q T P . In this case, verify that GAG = G,
0
T T
(AG) = AG and (GA) = GA. Hence, this G is the pseudo-inverse of A.
" #
1 2 3
4. Let A = .
2 1 1
 
a α
 
(a) Find a matrix G such that AG = I2 . Hint: Let G = 
 b β . Now, use AG = I2 to

c γ
get the solution space and proceed.
(b) What can you say about the number of such matrices? Give reasons for your answer.
(c) Does the choice of G in part (a) also satisfies (AG)T = AG and (GA)T = GA? Give
reasons for your answer.
(d) Does there exist a matrix C such that CA = I3 ? Give reasons for your answer.
(e) Could you have used the ideas from Exercise 2.5.11.3 to get your answers?
5. Let A ∈ Mm,n (R) with Rank(A) = r. Then," using

# Theorem 2.5.8, we can find invertible
Ir 0
matrices P and Q such that P A Q = . Choose arbitrary matrices U, V and W
0 0
" # " #
Ir U Ir U
such that the matrix is an n × m matrix. Define G = Q P . Then, prove
V W V W
that G is a g-inverse of A.
6. Let A ∈ Mm,n (R) with Rank(A) = r. Then, using Corollary 2.5.9 there exist matrices B ∈
Mm,r (R) and C ∈ Mr,n (R) such that Rank(B) = Rank(C) = r and A = BC. Thus, B T B
and CC T are invertible matrices. Now, define G1 = (B T B)−1 B T and G2 = C T (CC T )−1 .
T
Then, prove that G = G2 G1 is a pseudo-inverse of A.

AF
DR
2.6 Solution set of a Linear System

Having learnt RREF of a matrix and the properties of the rank function, let us go back to the
system of linear equations and apply those ideas to have a better understanding of the system
of linear equations. So, let us consider the linear system Ax = b. Then, using Remark 2.4.8.4
Rank([A b]) ≥ Rank(A). Further, the augmented matrix has exactly one extra column. Hence,
either
Rank([A b]) = Rank(Augmented Matrix) = Rank(Coefficient matrix) + 1 = Rank(A) + 1

or
Rank([A b]) = Rank(Augmented Matrix) = Rank(Coefficient matrix) = Rank(A).
In the first case, there is a pivot in the (n + 1)-th column of the augmented matrix [A b].
Thus, the column corresponding to b has a pivot. This implies b 6= 0. This implies that the
row corresponding to this pivot in RREF([A b]) has all entries before this pivot as 0. Thus,
in RREF([A b]) this pivotal row equals [0 0 · · · 0 1]. But, this corresponds to the equation
0 · x1 + 0 · x2 + · · · + 0 · xn = 1. This implies that the Ax = b has no solution whenever
b 6= 0 and Rank(Augmented Matrix) > Rank(Coefficient matrix).
We now define the words “basic variables” and “free variables”.

2.6. SOLUTION SET OF A LINEAR SYSTEM 49
Definition 2.6.1. Consider the linear system Ax = b. If RREF([A b]) = [C d]. Then,
the variables corresponding to the pivotal columns of C are called the basic variables and the
variables that correspond to non-pivotal columns are called free variables.
To understand the second case, we look at the homogeneous system Ax = 0.
Example 2.6.2. Consider a linear system Ax = 0. Suppose RREF(A) = C, where

 
1 0 2 −1 0 0 2
 
0 1 1 3 0 0 5
 
C=0 0 0 0 1 0 −1 .
 
 
0
 0 0 0 0 1 −4 
0 0 0 0 0 0 0
Then to get the solution set, observe that C has 4 pivotal columns, namely, the columns 1, 2, 5
and 6. Thus, x1 , x2 , x5 and x6 are basic variables. Therefore, the remaining variables x3 , x4 and
x7 are free variables. Hence, the solution set is given by
         
x1 −2x3 + x4 − 2x7 −2 1 −2
         
x2  −x3 − 3x4 − 5x7  −1 −3 −5
         
x3   x3 1 0 0
         

         
x4  =  x 4
 = x 3
 0  + x 4
 1  + x 7
 0 ,
T
         
AF
4x7 0 0 4

         
x5   
         
x  
 6  4 − x7
  0   0  −1
DR
      
x7 x7 0 0 1
where x3 , x4 and x7 are arbitrary. Let

 
−2 1 −2
 
−1 −3 −5
 
1 0 0
 
h i  

u1 u2 u3 =  0
 1 0.
0 0 4
 
 
0
 0 −1
0 0 1
Then, for 1 ≤ i ≤ 3, Cui = 0 ⇒

h Aui = 0.i Further, as x3 , x4 and x7 are the free variables,
observe that the submatrix of u1 u2 u3 corresponding to the 3-rd, 4-th and 7-th rows
equals I3 .
Let us now summarize the above ideas and examples.
Theorem 2.6.3. Let Ax = b be a linear system in n variables with RREF([A b]) = [C d].
1. Then the system Ax = b is inconsistent if Rank([A b]) > Rank(A).
2. Then the system Ax = b is consistent if Rank([A b]) = Rank(A).

(a) Further, Ax = b has a unique solution if Rank(A) = n.

(b) Further, Ax = b has infinite number of solutions if Rank(A) < n. In this
case, there exist vectors x0 , u1 , . . . , un−r ∈ Rn with Ax0 = b and Aui = 0, for
1 ≤ i ≤ n − r. Furthermore, the solution set is given by
{x0 + k1 u1 + k2 u2 + · · · + kn−r un−r | ki ∈ R, 1 ≤ i ≤ n − r}.
Proof. Part 1: As Rank([A b]) > Rank(A), by Remark 2.4.8.4 ([C d])[r + 1, :] = [0T 1]. Note
that this row corresponds to the linear equation
0 · x1 + 0 · x2 + · · · + 0 · xn = 1
which clearly has no solution. Thus, by Theorem 2.2.13, Ax = b is inconsistent.

Part 2: As Rank([A b]) = Rank(A), by Remark 2.4.8.4, [C d] doesn’t have a row of the
form [0T 1]. Further, the number of pivots in [C d] and that in C is same, namely, r pivots.
Part 2a: As Rank(A) = r = n, there are no free variables. Hence, xi = di , for 1 ≤ i ≤ n,
is the unique solution.
Part 2b: As Rank(A) = r < n. Suppose the pivots appear in columns i1 , . . . , ir with
1 ≤ i1 < · · · < ir ≤ n. Thus, the variables xij , for 1 ≤ j ≤ r, are basic variables and the
remaining n − r variables, say xt1 , . . . , xtn−r , are free variables with t1 < · · · < tn−r . Since C is
in RREF, in terms of the free variables and basic variables, the `-th row of [C d], for 1 ≤ ` ≤ r,
corresponds to the equation (writing basic variables in terms of a constant and free variables)
T
n−r n−r
AF
X X
x i` + c`tk xtk = d` ⇔ xi` = d` − c`tk xtk .
k=1 k=1
DR
Thus, the system Cx = d is consistent. Hence, by Theorem 2.2.13 the system Ax = b is

consistent and the solution set of the system Ax = b and Cx = d are the same. Therefore, the
solution set of the system Cx = d (or equivalently Ax = b) is given by
 n−r
P 
d − c x
 1 k=1 1tk tk 
         
x i1 d1 c1t1 c1t2 c1tn−r
 .   ..   .. 
    .   .   . 
 .    .   .   . 
 .    .  .
  .   .   . 
 
 x ir   n−r         
  d −
P
c x  dr 
   crt1 
  crt2 
 crt 
n−r 
 r rtk tk
  
 
 xt1  =  k=1 = 0  + xt1  1  + xt2  0  + · · · + xtn−r  0 . (2.6.3)
       
 
xt1
           
 x    0
    0   1   0 
 t2        
 .  
 .   xt2   .. 
    .   .   . 
. . . 
 .   ..  .  .   .   . 
    
.

xtn−r 0 0 0 1
 
xtn−r
     
d1 c1t1 c1tn−r
.  .   . 
.  .   . 
.  .   . 
     
dr  crt  crt 
   1  n−r 
Define x0 =  0  and u1 =  1 , . . . , un−r =  0 . Then, it can be easily verified
     
     
0  0   0 
     
.  .   . 
.  .   . 
.  .   . 
0 0 1
2.6. SOLUTION SET OF A LINEAR SYSTEM 51
that Ax0 = b and, for 1 ≤ i ≤ n − r, Aui = 0. Also, by Equation (2.6.3) the solution set has
indeed the required form, where ki corresponds to the free variable xti . As there is at least one
free variable the system has infinite number of solutions.
Thus, note that the solution set of Ax = b depends on the rank of the coefficient matrix, the
rank of the augmented matrix and the number of unknowns. In some sense, it is independent
of the choice of m.
Exercise 2.6.4. Consider the linear system given below. Use GJE to find the RREF of its
augmented matrix and use it to find the solution.
x +y −2 u +v = 2
z +u +2 v = 3
v +w = 3
v +2 w = 5
Let A ∈ Mm,n (R). Then, Rank(A) ≤ m. Thus, using Theorem 2.6.3 the next result follows.
Corollary 2.6.5. Let A ∈ Mm,n (R). If Rank(A) = r < n then the homogeneous system Ax = 0
has at least one non-trivial solution.
Remark 2.6.6. Let A ∈ Mm,n (R). Then, Theorem 2.6.3 implies that Ax = b is consistent
if and only if Rank(A) = Rank([A b]). Further, the the vectors ui ’s associated with the free
T
variables in Equation (2.6.3) are solutions of the associated homogeneous system Ax = 0.

AF
DR
We end this subsection with some applications.
Example 2.6.7. 1. Determine the equation of the circle passing through the points (−1, 4), (0, 1)
and (1, 4).
Solution: The equation a(x2 + y 2 ) + bx + cy + d = 0, for a, b, c, d ∈ R, represents a circle.
Since this curve passes through the given points, we get a homogeneous system having 3
equations in4 unknowns, namely
 
  a
(−1)2 + 42 −1 4 1  
  b 
 (0)2 + 12 0 1 1   = 0.
  c 
12 + 4 2 1 4 1
 
d
3
Solving this system, we get [a, b, c, d] = [ 13 d, 0, − 16
13 d, d]. Hence, choosing d = 13, the
required circle is given by 3(x2 + y 2 ) − 16y + 13 = 0.
2. Determine the equation of the plane that contains the points (1, 1, 1), (1, 3, 2) and (2, −1, 2).
Solution: The general equation of a plane in space is given by ax + by + cz + d = 0,
where a, b, c and d are unknowns. Since this plane passes through the 3 given points, we
get a homogeneous system in 3 equations and 4 variables. So, it has a non-trivial solution,
namely [a, b, c, d] = [− 43 d, − d3 , − 32 d, d]. Hence, choosing d = 3, the required plane is given
by −4x − y + 2z + 3 = 0.
 
2 3 4
 
3. Let A = 0 −1 0 . Then, find a non-trivial solution of Ax = 2x. Does there exist a

0 −3 4
nonzero vector y ∈ R3 such that Ay = 4y?
Solution: Solving for Ax = 2x is equivalentto solving (A − 2I)x = 0. The augmented
0 3 4 0
. Verify that xT = [1, 0, 0] is a nonzero
 
matrix of this system equals 
0 −3 0 0 
0 −3 2 0
solution.
  other part, the augmented matrix for solving (A − 4I)y = 0 equals
For the
−2 3 4 0
 0 −5 0 0. Thus, verify that yT = [2, 0, 1] is a nonzero solution.
 
 
0 −3 0 0
Exercise 2.6.8. 1. Let A ∈ Mn (R). If A2 x = 0 has a non trivial solution then show that
Ax = 0 also has a non trivial solution.
2. Let u = (1, 1, −2)T and v = (−1, 2, 3)T . Find condition on x, y and z such that the system
cu + dv = (x, y, z)T in the unknowns c and d is consistent.
3. Find condition(s) on x, y, z so that the systems given below (in the unknowns a, b and c)
is consistent?
T
AF
(a) a + 2b − 3c = x, 2a + 6b − 11c = y, a − 2b + 7c = z.
DR
(b) a + b + 5c = x, a + 3c = y, 2a − b + 4c = z.
4. For what values of c and k, the following systems have i) no solution, ii) a unique
solution and iii) infinite number of solutions.
(a) x + y + z = 3, x + 2y + cz = 4, 2x + 3y + 2cz = k.
(b) x + y + z = 3, x + y + 2cz = 7, x + 2y + 3cz = k.
(c) x + y + 2z = 3, x + 2y + cz = 5, x + 2y + 4z = k.
(d) x + 2y + 3z = 4, 2x + 5y + 5z = 6, 2x + (c2 − 6)z = c + 20.
(e) x + y + z = 3, 2x + 5y + 4z = c, 3x + (c2 − 8)z = 12.
5. Consider the linear system Ax = b in m equations and 3 unknowns. Then, for each of
the given solution set, determine the possible choices of m? Further, for each choice of
m, determine a choice of A and b.
(a) (1, 1, 1)T is the only solution.

(b) {c(1, 2, 1)T |c ∈ R} as the solution set.
(c) {(1, 1, 1)T + c(1, 2, 1)T |c ∈ R} as the solution set.
(d) {c(1, 2, 1)T + d(2, 2, −1)T |c, d ∈ R} as the solution set.
(e) {(1, 1, 1)T + c(1, 2, 1)T + d(2, 2, −1)T |c, d ∈ R} as the solution set.
2.7. SQUARE MATRICES AND LINEAR SYSTEMS 53
2.7 Square Matrices and Linear Systems

In this section, we apply our ideas to the linear system Ax = b, where the coefficient matrix is
square. We start with proving a few equivalent conditions that relate different ideas.
Theorem 2.7.1. Let A ∈ Mn (R). Then, the following statements are equivalent.
1. A is invertible.
2. RREF(A) = In .
3. A is a product of elementary matrices.
4. Rank(A) = n.
5. The homogeneous system Ax = 0 has only the trivial solution.
6. The system Ax = b has a unique solution for every b.
7. The system Ax = b is consistent for every b.
Proof. 1 ⇔ 2 and 2 ⇔ 3 See Proposition 2.4.9.

2⇔4 By definition. For the converse, Rank(A) = n ⇒ A has n pivots and A has n
columns. So, all columns are pivots. Thus, RREF(A) = In .
As A is invertible A−1 A = In . So, if x0 is any solution of the homogeneous
T
1 =⇒ 5
AF
system Ax = 0 then
DR
x0 = In · x0 = (A−1 A)x0 = A−1 (Ax0 ) = A−1 0 = 0.
5 =⇒ 1 Ax = 0 has only the trivial solution implies that there are no free variables. So,
all the unknowns are basic variables. So, each column is a pivotal column. Thus, RREF(A) = In .
1⇒6 Note that x0 = A−1 b is the unique solution of Ax = b.
6⇒7 A unique solution implies that is at least one solution. So, nothing to show.
7 =⇒ 1 Given assumption implies that for 1 ≤ i ≤ n, the linear system Ax = ei has a
solution, say ui . Define B = [u1 , u2 , . . . , un ]. Then
AB = A[u1 , u2 , . . . , un ] = [Au1 , Au2 , . . . , Aun ] = [e1 , e2 , . . . , en ] = In .
Now, consider the linear homogeneous system Bx = 0. Then AB = In implies that
x0 = In x0 = (AB)x0 = A(Bx0 ) = A0 = 0.
Thus, the homogeneous system Bx = 0 has a only the trivial solution. Hence, using Part 5, B
is invertible. As AB = In and B is invertible, we get BA = In . Thus AB = In = BA. Thus, A
is invertible as well.
We now give an immediate application of Theorem 2.7.1 without proof.
Theorem 2.7.2. The following two statements cannot hold together for A ∈ Mn (R).
1. The system Ax = b has a solution for every b.

2. The system Ax = 0 has a non-trivial solution.
As an immediate consequence of Theorem 2.7.1, the readers should prove that one needs to
compute either the left or the right inverse to prove invertibility of A ∈ Mn (R).
Corollary 2.7.3. Let A ∈ Mn (R). Then the following holds.
1. If there exists C such that CA = In then A−1 exists.
2. If there exists B such that AB = In then A−1 exists.
Corollary 2.7.4. (Theorem of the Alternative) The following two statements cannot hold
together for A ∈ Mn (C) and b ∈ Rn .
1. The system Ax = b has a solution.
2. The system yT A = 0T , yT b 6= 0 has a solution.
Proof. Observe that if x0 is a solution of Ax = b and y0 is a solution of yT A = 0T then

y0T b = y0T (Ax0 ) = (y0T A)x0 = 0T x0 = 0.
Note that one of the requirement in the last corollary is yT b 6= 0. Thus, we want non-zero
vectors x0 and y0 in Rn such that they are solutions of Ax = b and yT A = 0T , respectively,
T
with the added condition that y0 and b are not orthogonal or perpendicular (their dot product
AF
is not zero).
DR
Exercise 2.7.5. 1. Give the proof of Theorem 2.7.2 and Corollary 2.7.3.
2. Let A ∈ Mn,m (R) and B ∈ Mm,n (R). Either use Theorem 2.7.1.5 or multiply the matrices
to verify the following statementes.
(a) Then, prove that I − BA is invertible if and only if I − AB is invertible.

(b) If I − AB is invertible then, prove that (I − BA)−1 = I + B(I − AB)−1 A.
(c) If I − AB is invertible then, prove that (I − BA)−1 B = B(I − AB)−1 .
(d) If A, B and A + B are invertible then, prove that (A−1 + B −1 )−1 = A(A + B)−1 B.
3. Let bT = [1, 2, −1, −2]. Suppose A is a 4 × 4 matrix such that the linear system Ax = b
has no solution. Mark each of the statements given below as true or false?
(a) The homogeneous system Ax = 0 has only the trivial solution.

(b) The matrix A is invertible.
(c) Let cT = [−1, −2, 1, 2]. Then, the system Ax = c has no solution.
(d) Let B = RREF(A). Then,
i. B[4, :] = [0, 0, 0, 0].
ii. B[4, :] = [0, 0, 0, 1].
2.8. DETERMINANT 55
2.8 Determinant
 
1 2 3 " #
  1 2
Recall the notations used in Section 1.5 on Page 21 . If A =  1 3 2 then A(1 | 2) = 2 7

2 4 7
and A({1, 2} | {1, 3}) = [4]. The actual definition of the determinant requires an understanding
of group theory. So, we will just give an inductive definition which will help us to compute
the determinant and a few results. The advanced students can find the main definition of the
determinant in Appendix 9.2.22, where it is proved that the definition given below corresponds
to the expansion of determinant along the first row.
Definition 2.8.1. Let A be a square matrix of order n. Then, the determinant of A, denoted
det(A) (or | A | ) is defined by

 a,
 if A = [a] (corresponds to n = 1),
det(A) = n
(−1)1+j a1j det A(1 | j) ,
P

 otherwise.
j=1
Example 2.8.2. 1. Let A = [−2]. Then, det(A) = | A | = −2.

" #
a b
2. Let A = . Then, det(A) = a det(A(1 | 1)) − b det(A(1 | 2)) = ad − bc.
c d
" #
1 2
T
(a) If A = then det(A) = 1 · 5 − 2 · 0 = 5.

0 5
AF
" #
2 1
DR
(b) If B = then det(B) = 2 · 0 − 1 · 5 = −5.

5 0
Observe that B is obtained from A by interchanging the columns. This has resulted in
the value of the determinant getting multiplied by −1. So, if we think of the columns of
the matrix as vectors in R2 then, the sign of determinant gets related with the orientation
of the parallelogram formed using the two column vectors.
3. Let A = [aij ] be a 3 × 3 matrix. Then,
det(A) = | A | = a11 det(A(1 | 1)) − a12 det(A(1 | 2)) + a13 det(A(1 | 3))

a
22 a23
a
21 a23
a
21 a22

= a11 − a12 + a 13
a32 a33 a31 a33 a31 a32
= a11 (a22 a33 − a23 a32 ) − a12 (a21 a33 − a31 a23 ) + a13 (a21 a32 − a31 a22 ).
 
1 2 3
3 1

2 1

2 3
 
2 3 1, det(A) = 1 · 2 2 − 2 · 1 2 + 3 · 1 2 = 4 − 2(3) + 3(1) = 1.
For A =  
1 2 2
Exercise
 2.8.3.
 Find the determinant
 of the following matrices.
1 2 7 8 3 0 0 1  
    1 a a2
0 4 3 2 0 2 0 5  
i) 
0 0 2 3 ii) 6 −7 1 0 iii) 1
    b b2 
.
1 c c2
   
0 0 0 5 3 2 0 6
Definition 2.8.4. A matrix A is said to be a singular if det(A) = 0 and is called non-

singular if det(A) 6= 0.
It turns out that the determinant of a matrix equals the volume of the parallelepiped formed
using the columns of the matrix. With this understanding, the singularity of A gets related with
the dimension in which we are looking at the parallelepiped. For, example, the length makes
sense in one-dimension but it doesn’t make sense to talk of area (which is a two-dimensional
idea) of a line segment. Similarly, it makes sense to talk of volume of a cube but it doesn’t make
sense to talk of the volume of a square or rectangle or parallelogram which are two-dimensional
objects.
We now state a few properties of the determinant function. For proof, see Appendix 9.3.
Theorem 2.8.5. Let A be an n × n matrix.
1. det(In ) = 1.
2. If A is a triangular matrix with d1 , . . . , dn on the diagonal then det(A) = d1 · · · dn .
3. If B = Eij A, for 1 ≤ i 6= j ≤ n, then det(B) = − det(A). In particular, det(Eij ) = −1
4. If B = Ei (c)A, for c 6= 0, 1 ≤ i ≤ n, then det(B) = c det(A). In particular, det(Ei (c)) = c.

T
AF
5. If B = Eij (c)A, for c 6= 0 and 1 ≤ i 6= j ≤ n, then det(B) = det(A). In particular,

det(Eij (c)) = 1.
DR
6. If A[i, :] = 0T , for 1 ≤ i, j ≤ n then det(A) = 0.
7. If A[i, :] = A[j, :] for 1 ≤ i 6= j ≤ n then det(A) = 0.

 
2 2 6 1 1 3 1 1 3

1
  E1 ( 2 ) E21 (−1)E31 (−1)
Example 2.8.6. Let A = 1 3 2. Then A → 1
  3 2 → 0
2 −1.
1 1 2 1 1 2 0 0 −1

Thus, using Theorem 2.8.5, det(A) = 2 · (1 · 2 · (−1)) = −4, where the first 2 appears from the
1
elementary matrix E1 ( ).
2
Exercise 2.8.7. Prove the following without computing the determinant (use Theorem 2.8.5).
h i
1. Let A = u v 2u + 3v , where u, v ∈ R3 . Then, det(A) = 0.
   
a b c a e x2 a + xe + h
   
2
2. Let A =  e f g . If x 6= 0 and B =  b f x b + xf + j  then det(A) = det(B).
  
h j ` c g x2 c + xg + `

3 1 2

Hence, conclude that 3 divides 4 7 1 .
1 4 −2

2.8. DETERMINANT 57
Remark 2.8.8. Theorem 2.8.5.3 implies that the determinant can be calculated by expanding
along any row. Hence, the readers are advised to verify that
n
X
det(A) = (−1)k+j akj det(A(k | j)), for 1 ≤ k ≤ n.
j=1
Example
Using Remark 2.8.8, one has
2.8.9.
2 2 6 1
2 2 1 2 2 6

0 0 2 1 2+3
2+4

0 1 = (−1) · 2 · 0 1 0 + (−1)
· 0 1 2 = −2 · 1 + (−8) = −10.

2 0
1 2 1 1 2 1

1 2 1 1
2.8.1 Inverse of a Matrix
Definition 2.8.10. Let A ∈ Mn (R). Then, the cofactor matrix, denoted Cof(A), is an Mn (R)
matrix with Cof(A) = [Cij ], where
Cij = (−1)i+j det (A(i | j)) , for 1 ≤ i, j ≤ n.

T
AF
And, the Adjugate (classical Adjoint) of A, denoted Adj(A), equals CofT (A).
DR
 
1 2 3
 
2 3 1. Then,

1 2 4
 
C C21 C31
T
 11 
Adj(A) = Cof (A) = 
C12 C22 C32 

C13 C23 C33
 
(−1)1+1 det(A(1|1)) (−1)2+1 det(A(2|1)) (−1)3+1 det(A(3|1))
 
=  1+2 det(A(1|2)) (−1)2+2 det(A(2|2)) (−1)3+2 det(A(3|2))
(−1) 
(−1)1+3 det(A(1|3)) (−1)2+3 det(A(2|3)) (−1)3+3 det(A(3|3))
 
10 −2 −7
 
= 
−7 1 5 .

1 0 −1
   
−1 0 0 det(A) 0 0
   
Now, verify that AAdj(A) = 
 0 −1 0 = 0
  det(A) 0  = Adj(A)A.

0 0 −1 0 0 det(A)
 
x − 1 −2 −3
 
Consider xI3 − A = 
 −2 x − 3 −1 . Then,

−1 −2 x − 4
   
C C21 C31 x2 − 7x + 10 2x − 2 3x − 7
 11   
Adj(xI − A) = C12 C22
 C32  =  2x − 7 x 2 − 5x + 1 x + 5 
  
C13 C23 C33 x+1 2x x2 − 4x − 1
 
−7 2 3
2 2
 
= x I + x 2
 −5 1   + Adj(A) = x I + Bx + C(say).
1 2 −4
Hence, we observe that Adj(xI − A) = x2 I + Bx + C is a polynomial in x with coefficients as

matrices. Also, note that (xI − A)Adj(xI − A) = (x3 − 8x2 + 10x − det(A))I3 . Thus
(xI − A)(x2 I + Bx + C) = (x3 − 8x2 + 10x − det(A))I3 .
That is, we have obtained a matrix identity. Hence, replacing x by A makes sense. But, then
the LHS is 0. So, for the RHS to be zero, we must have A3 − 8A2 + 10A − det(A)I = 0 (this
equality is famously known as the Cayley-Hamilton Theorem).
The next result relates adjugate matrix with the inverse, in case det(A) 6= 0.
T
AF
Theorem 2.8.12. Let A ∈ Mn (R).

n n
DR
aij (−1)i+j det(A(i|j)) = det(A), for 1 ≤ i ≤ n.

P P
1. Then aij Cij =
j=1 j=1
n n
aij (−1)i+j det(A(`|j)) = 0, for i 6= `.
P P
2. Then aij C`j =
j=1 j=1
3. Thus A(Adj(A)) = det(A)In . Hence,

1
whenever det(A) 6= 0 one has A−1 = Adj(A). (2.8.1)
det(A)
Proof. Part 1: It follows directly from Remark 2.8.8 and the definition of the cofactor.
Part 2: Fix positive integers i, ` with 1 ≤ i 6= ` ≤ n. Suppose that the i-th and `-th rows of
B are equal to the i-th row of A and B[t, :] = A[t, :], for t 6= i, `. Since two rows of B are equal,
det(B) = 0. Now, let us expand the determinant of B along the `-th row. We see that
n
X
(−1)`+j b`j det B(` | j)

0 = det(B) = (2.8.2)
j=1
n
X
(−1)`+j aij det B(` | j)

= (bij = b`j = aij for all j)
j=1
Xn n
X
= (−1)`+j aij det A(` | j) = aij C`j . (2.8.3)
j=1 j=1
This completes the proof of Part 2.

2.8. DETERMINANT 59
Part 3: Using the first two parts, observe that

n n
(
0, if i 6= j,
X X

A Adj(A) = aik Adj(A) kj = aik Cjk =
ij k=1 k=1
det(A), if i = j.

1
Thus, A(Adj(A)) = det(A)In . Therefore, if det(A) 6= 0 then A det(A) Adj(A) = In . Hence,
1
by Proposition 2.4.9, A−1 = det(A) Adj(A).
   
1 −1 0 −1 1 −1
   
Example 2.8.13. For A = 0 1 1 , Adj(A) =  1 1 −1 and det(A) = −2. Thus,
 
1 2 1 −1 −3 1
 
1/2 −1/2 1/2
−1
 
by Theorem 2.8.12.3, A = 
−1/2 −1/2 1/2 .

1/2 3/2 −1/2
1
Let A be a non-singular matrix. Then, by Theorem 2.8.12.3, A−1 = det(A) Adj(A). Thus

A Adj(A) = Adj(A) A = det(A) In and this completes the proof of the next result
Corollary 2.8.14. Let A be a non-singular matrix. Then,

n
(
X det(A), if j = k,
Cik aij =
0, if j 6= k.
T
i=1
AF
The next result gives another equivalent condition for a square matrix to be invertible.
DR
Theorem 2.8.15. A square matrix A is non-singular if and only if A is invertible.
1
Proof. Let A be non-singular. Then, det(A) 6= 0 and hence A−1 = det(A) Adj(A).
Now, let us assume that A is invertible. Then, using Theorem 2.7.1, A = E1 · · · Ek , a product
of elementary matrices. Thus, a repeated application of Parts 3, 4 and 5 of Theorem 2.8.5 gives
det(A) 6= 0.
2.8.2 Results on the Determinant
The next result relates the determinant of a matrix with the determinant of its transpose. Thus,
the determinant can be computed by expanding along any column as well.
Theorem 2.8.16. Let A ∈ Mn (R). Then det(A) = det(AT ). Further, det(A∗ ) = det(A).
Proof. If A is singular then, by Theorem 2.8.15, A is not invertible. So, AT is also not invertible
and hence by Theorem 2.8.15, det(AT ) = 0 = det(A).
Now, let A be a non-singular and let AT = B. Then, by definition,
n
X n
X
det(AT ) = det(B) = (−1)1+j b1j det B(1 | j) = (−1)1+j aj1 det A(j | 1)

j=1 j=1
n
X
= aj1 Cj1 = det(A)
j=1
using Corollary 2.8.14. Further, using induction and the first part, one has
n
X
det(A∗ ) = det((A)T ) = det(A) = (−1)1+j a1j det A(1 | j)

j=1
n
X
= (−1)1+j a1j det A(1 | j) = det(A)
j=1
Hence, the required result follows.

The next result relates the determinant of product of two matrices with their determinants.
Theorem 2.8.17. Let A and B be square matrices of order n. Then,
det(AB) = det(A) · det(B) = det(BA).
Proof. Case 1: Let A be non-singular. Then, by Theorem 2.8.12.3, A is invertible and by

Theorem 2.7.1, A = E1 · · · Ek , a product of elementary matrices. Thus, a repeated application
of Parts 3, 4 and 5 of Theorem 2.8.5 and an inductive argument gives the desired result as
det(AB) = det(E1 · · · Ek B) = det(E1 ) det(E2 · · · Ek B)

= det(E1 ) det(E2 · · · Ek ) det(B) = det(E1 E2 · · · Ek ) det(B)
= det(A) det(B).
T
AF
Case 2: Let A be singular. Then, by Theorem 2.8.15 A is"not#invertible. So, "by Proposi-
#
C1 C1
DR
tion 2.4.9 there exists an invertible matrix P such that P A = . So A = P −1 . As P

0 0
is invertible, using Part 1, we have
" #! ! " #! " #!
−1 C1 −1 C1 B −1 C1 B
det(AB) = det P B = det P = det(P ) · det
0 0 0
= det(P ) · 0 = 0 = 0 · det(B) = det(A) det(B).
Thus, the proof of the theorem is complete.

We now give an application of Theorem 2.8.17.
Example 2.8.18. Let A ∈ Mn (R) be an orthogonal matrix then, by definition, AAT = I.

Thus, by Theorems 2.8.17 and 2.8.16
1 = det(I) = det(AAT ) = det(A) det(AT ) = det(A) det(A) = (det(A))2 .

" #
a b
Hence det A = ±1. In particular, if A = ∈ M2 (R) then the following holds.
c d
" #
T a2 + b2 ac + bd
1. I = AA = .
ac + bd c2 + d2
2. Thus a2 + b2 = 1 and hence there exists θ ∈ [−π, π) such that a = cos θ and b = sin θ.
3. Further, ac + bd = 0 and c2 + d2 = 1 ⇒ c = sin θ, d = − cos θ or c = − sin θ, d = cos θ.
2.8. DETERMINANT 61
" # " #
cos θ sin θ cos θ sin θ
4. Thus A = or A = .
sin θ − cos θ − sin θ cos θ
" #
cos θ sin θ
5. For A = , det(A) = −1. Then A represents a reflection about the line
sin θ − cos θ
y = mx. Determine m? (see Exercise 2.2b).
" #
cos θ sin θ
6. For A = , det(A) = 1. Then A represents a rotation through the angle θ.
− sin θ cos θ
Is the rotation clockwise or counter-clockwise (see Exercise 2.2a)?
Exercise 2.8.19. 1. Let A be a square matrix. Then, prove that A is invertible ⇔ AT is

invertible ⇔ AT A is invertible ⇔ AAT is invertible.
2. Let A and B be two matrices having positive entries and of orders 1 × n and n × 1,
respectively. Which of BA or AB is invertible? Give reasons.
2.8.3 Cramer’s Rule
Consider the linear system Ax = b. Then, using Theorems 2.7.1 and 2.8.15, we conclude
that Ax = b has a unique solution for every b if and only if det(A) 6= 0. The next theorem,
commonly known as the Cramer’s rule gives a direct method of finding the solution of the
linear system Ax = b when det(A) 6= 0.
T
AF
Theorem 2.8.20. Let A be an n × n non-singular matrix. Then, the unique solution of the
DR
linear system Ax = b with the unknown vector xT = [x1 , . . . , xn ] is given by
det(Aj )
xj = , for j = 1, 2, . . . , n,
det(A)
where Aj is the matrix obtained from A by replacing the j-th column of A, namely A[:, j], by b.
Proof. Since det(A) 6= 0, A is invertible. Thus A−1 [A | b] = [I | A−1 b]. Let d = A−1 b. Then
Ax = b has the unique solution xj = dj , for 1 ≤ j ≤ n. Thus,

−1 −1
A Aj = A A[:, 1], . . . , A[:, j − 1], b, A[:, j + 1], . . . , A[:, n]

−1 −1 −1 −1 −1
= A A[:, 1], . . . , A A[:, j − 1], A b, A A[:, j + 1], . . . , A A[:, n]
= [e1 , . . . , ej−1 , d, ej+1 , . . . , en ].
Thus, det(A−1 Aj ) = dj , for 1 ≤ j ≤ n. Also,
dj det(A−1 Aj ) det(A−1 Aj ) det(A−1 ) det(Aj ) det(Aj )

dj = = = −1
= −1
= .
1 det(I) det(A A) det(A ) det(A) det(A)
det(Aj )
Hence, xj = and the required result follows.
det(A)
   
1 2 3 1
   
Example 2.8.21. Solve Ax = b using Cramer’s rule, where A = 2 3 1 and b = 1
  
.
1 2 2 1
T
Solution: Check that det(A) = 1 and x = [−1, 1, 0] as

1 2 3 1 1 3 1 2 1

x1 = 1 3 1 = −1, x2 = 2 1 1 = 1, and x3 = 2 3 1 = 0.
1 2 2 1 1 2 1 2 1

2.9 Miscellaneous Exercises

Exercise 2.9.1. 1. Determine the determinant of an orthogonal matrix.
2. Let A be a unitary matrix then what can you say about | det(A) |?
3. Let A ∈ Mn (R). Prove that the following statements are equivalent:
(a) A is not invertible.

(b) Rank(A) 6= n.
(c) det(A) = 0.
(d) A is not row-equivalent to In .
T
AF
(e) The homogeneous system Ax = 0 has a non-trivial solution.

DR
(f ) The system Ax = b is either inconsistent or it has an infinite number of solutions.

(g) A is not a product of elementary matrices.
4. Let A ∈ M2n+1 (R) be a skew-symmetric matrix. Then det(A) = 0.
5. If A is a Hermitian matrix then det A is a real number.
6. Let A ∈ Mn (R). Then A is invertible if and only if Adj(A) is invertible.
7. Let A and B be invertible matrices. Prove that Adj(AB) = Adj(B)Adj(A).

" #
A B
8. Let A be an n × n invertible matrix and let P = . Then, show that Rank(P ) = n
C D
if and only if D = CA−1 B.
9. Let A be a 2 × 2 matrix with tr(A) = 0 and det(A) = 0. Then, A is a nilpotent matrix.
10. Determine necessary and sufficient condition for a triangular matrix to be invertible.
11. Let A and B be two non-singular matrices. Are the matrices A + B and A − B non-
singular? Justify your answer.
12. For what value(s) of λ does the following systems have non-trivial solutions? Also, for
each value of λ, determine a non-trivial solution.
2.10. SUMMARY 63
(a) (λ − 2)x + y = 0, x + (λ + 2)y = 0.

(b) λx + 3y = 0, (λ + 6)y = 0.
13. Let a1 , . . . , an ∈ R and define A = [aij ]n×n with aij = aj−1

i . Prove that det(A) =
Q
(aj − ai ). This matrix is usually called the van der monde matrix.
1≤i<j≤n
14. Let A = [aij ] ∈ Mn (R) with aij = max{i, j}. Prove that det A = (−1)n−1 n.
15. Let p ∈ R, p 6= 0. Let A = [aij ], B = [bij ] ∈ Mn (R) with bij = pi−j aij , for 1 ≤ i, j ≤ n.
Then, compute det(B) in terms of det(A).
16. The position of an element aij of a determinant is called even or odd according as i + j is
even or odd. Prove that if all the entries in
(a) odd positions are multiplied with −1 then the value of determinant doesn’t change.
(b) even positions are multiplied with −1 then the value of determinant
i. does not change if the matrix is of even order.
ii. is multiplied by −1 if the matrix is of odd order.
2.10 Summary
T
AF
In this chapter, we started with a system of m linear equations in n variables and formally
DR
wrote it as Ax = b and in turn to the augmented matrix [A | b]. Then, the basic operations on
equations led to multiplication by elementary matrices on the right of [A | b]. These elementary
matrices are invertible and applying the GJE on a matrix A, resulted in getting the RREF of
A. We used the pivots in RREF matrix to define the rank of a matrix. So, if Rank(A) = r and
Rank([A | b]) = ra
1. then, r < ra implied the linear system Ax = b is inconsistent.
2. then, r = ra implied the linear system Ax = b is consistent. Further,
(a) if r = n then the system Ax = b has a unique solution.

(b) if r < n then the system Ax = b has an infinite number of solutions.
We have also seen that the following conditions are equivalent for A ∈ Mn (R).
1. A is invertible.
3. The row reduced echelon form of A is I.

7. Rank(A) = n.
8. det(A) 6= 0.
So, overall we have learnt to solve the following type of problems:
1. Solving the linear system Ax = b. This idea will lead to the question “is the vector b a
linear combination of the columns of A”?
2. Solving the linear system Ax = 0. This will lead to the question “are the columns of A
linearly independent/dependent”? In particular, we will see that
(a) if Ax = 0 has a unique solution then the columns of A are linear independent.
(b) if Ax = 0 has a non-trivial solution then the columns of A are linearly dependent.
T
AF
DR
Chapter 3
Vector Spaces
In this chapter, we will mainly be concerned with finite dimensional vector spaces over R or C.
Please note that the real and complex numbers have the property that any pair of elements can
be added, subtracted or multiplied. Also, division is allowed by a non-zero element. Such sets in
mathematics are called field. So, Q, R and C are examples of field and they have infinite number
of elements. But, in mathematics, we do have fields that have only finitely many elements. For
example, consider the set Z5 = {0, 1, 2, 3, 4}. In Z5 , we define addition and multiplication,
respectively, as
+ 0 1 2 3 4 · 0 1 2 3 4
T
AF
0 0 1 2 3 4 0 0 0 0 0 0
1 1 2 3 4 0 1 0 1 2 3 4
DR
and .
2 2 3 4 0 1 2 0 2 4 1 3
3 3 4 0 1 2 3 0 3 1 4 2
4 4 0 1 2 3 4 0 4 3 2 1
Then, we see that the elements of Z5 can be added, subtracted and multiplied. Note that 4
behaves as −1 and 3 behaves as −2. Thus, 1 behaves as −4 and 2 behaves as −3. Also, we see
that in this multiplication 2 · 3 = 1 and 4 · 4 = 1. Hence,
1. the division by 2 is similar to multiplying by 3,
2. the division by 3 is similar to multiplying by 2, and
3. the division by 4 is similar to multiplying by 4.
Thus, Z5 indeed behaves like a field. So, in this chapter, F will represent a field.
3.1 Vector Spaces: Definition and Examples

Let us recall that the vectors in R2 and R3 satisfy the following properties:
1. Vector Addition: To every pair u, v ∈ R3 there corresponds a unique element u+v ∈ R3

(called the addition of vectors) such that
(a) u + v = v + u (Commutative law).
65
66 CHAPTER 3. VECTOR SPACES
(b) (u + v) + w = u + (v + w) (Associative law).

(c) R3 has a unique element, denoted 0, called the zero vector that satisfies u + 0 = u,
for every u ∈ R3 (called the additive identity).
(d) for every u ∈ R3 there is an element w ∈ R3 that satisfies u + w = 0.
2. Scalar Multiplication: For each u ∈ R3 and α ∈ R, there corresponds a unique element

α · u ∈ R3 (called the scalar multiplication) such that
(a) α · (β · u) = (α · β) · u for every α, β ∈ R and u ∈ R3 .

(b) 1 · u = u for every u ∈ R3 , where 1 ∈ R.
3. Distributive Laws: relating vector addition with scalar multiplication

For any α, β ∈ R and u, v ∈ R3 , the following distributive laws hold:
(a) α · (u + v) = (α · u) + (α · v).
(b) (α + β) · u = (α · u) + (β · u).
So, we want the above properties to hold for any collection of vectors. Thus, formally, we have
the following definition.
Definition 3.1.1. A vector space V over F, denoted V(F) or in short V (if the field F is clear
from the context), is a non-empty set, in which one can define vector addition, scalar multipli-
T
AF
cation. Further, with these definitions, the properties of vector addition, scalar multiplication
and distributive laws (see items 1, 2 and 3 above) are satisfied.
DR
Remark 3.1.2. 1. The elements of F are called scalars.

2. The elements of V are called vectors.
3. We denote the zero element of F by 0, whereas the zero element of V will be denoted by 0.
4. Observe that Condition 1d implies that for every u ∈ V, the vector w ∈ V such that
u + w = 0 holds, is unique. For if, w1 , w2 ∈ V with u + wi = 0, for i = 1, 2 then by
commutativity of vector addition, we see that
w1 = w1 + 0 = w1 + (u + w2 ) = (w1 + u) + w2 = 0 + w2 = w2 .
Hence, we represent this unique vector by −u and call it the additive inverse.
5. If V is a vector space over R then V is called a real vector space.
6. If V is a vector space over C then V is called a complex vector space.
7. In general, a vector space over R or C is called a linear space.
Some interesting consequences of Definition 3.1.1 is stated next. Intuitively, they seem
obvious. The proof are given for better understanding of the given conditions.
Theorem 3.1.3. Let V be a vector space over F. Then,
1. u + v = u implies v = 0.
3.1. VECTOR SPACES: DEFINITION AND EXAMPLES 67
2. α · u = 0 if and only if either u = 0 or α = 0.
3. (−1) · u = −u, for every u ∈ V.
Proof. Part 1: By Condition 1d and Remark 3.1.2.4, for each u ∈ V there exists −u ∈ V such
that −u + u = 0. Hence u + v = u implies
0 = −u + u = −u + (u + v) = (−u + u) + v = 0 + v = v.
Part 2: As 0 = 0 + 0, using Condition 3, α · 0 = α · (0 + 0) = (α · 0) + (α · 0). Thus, using

Part 1, α · 0 = 0 for any α ∈ F. Similarly, 0 · u = (0 + 0) · u = (0 · u) + (0 · u) implies 0 · u = 0,
for any u ∈ V.
Now suppose α · u = 0. If α = 0 then the proof is over. So, assume that α 6= 0, α ∈ F. Then
(α)−1 ∈ F and using, 1 · u = u for every vector u ∈ V (see Condition 2.2b), we have
0 = (α)−1 · 0 = (α)−1 · (α · u) = ((α)−1 · α) · u = 1 · u = u.
Thus, if α 6= 0 and α · u = 0 then u = 0.

Part 3: As 0 = 0 · u = (1 + (−1))u = u + (−1) · u, one has (−1) · u = −u.
Example 3.1.4. The readers are advised to justify the statements given below.
1. Let V = {0}. Then, V is a real as well as a complex vector space.
2. Let A ∈ Mm,n (F) and define V = {x ∈ Mn,1 (F) : Ax = 0}. Then, by Theorem 2.1.7, V
T
AF
satisfies:
DR
(a) 0 ∈ V as A0 = 0.
(b) if x ∈ V then αx ∈ V, for all α ∈ F. In particular, for α = −1, −x ∈ V.
(c) if x, y ∈ V then, for any α, β ∈ F, αx + βy ∈ V.
Thus, V is a vector space over F.
3. Consider R with the usual addition and multiplication. Then R forms a real vector space.
4. Let Rn = {(a1 , . . . , an )T | ai ∈ R, 1 ≤ i ≤ n}. For u = (a1 , . . . , an )T , v = (b1 , . . . , bn )T ∈

V and α ∈ R, define
u + v = (a1 + b1 , . . . , an + bn )T and α · u = (αa1 , . . . , αan )T
(called component-wise operations). Then, V is a real vector space. The vector

space Rn is called the real vector space of n-tuples.
√
Recall that the symbol i represents the complex number −1.
5. Let Cn = {(z1 , . . . , zn )T | zi ∈ C, 1 ≤ i ≤ n}. For z = (z1 , . . . , zn ), w = (w1 , . . . , wn )T ∈

Cn and α ∈ F, define component-wise vector sum and scalar multiplication. Then, verify
that Cn forms a vector space over C (called the complex vector space) as well as over R
(called the real vector space). Unless specified otherwise, Cn will be considered a complex
vector space.
Remark 3.1.5. If F = C then i(1, 0, . . . , 0)T = (i, 0, . . . , 0)T is allowed. Whereas, if

F = R then i(1, 0, . . . , 0)T doesn’t make sense as i 6∈ R.
6. Fix m, n ∈ N and let Mm,n (C) = {Am×n = [aij ] | aij ∈ C}. Then, with usual addition
and scalar multiplication of matrices, Mm,n (C) is a complex vector space. If m = n, the
vector space Mm,n (C) is denoted by Mn (C).
7. Let S be a non-empty set and let RS = {f | f is a function from S to R}. For f, g ∈ RS

and α ∈ R, define (f + αg)(x) = f (x) + αg(x), for all x ∈ S. Then, RS is a real vector
space. In particular, for S = N, observe that RN consists of all real sequences and forms
a real vector space.
8. Fix a, b ∈ R with a < b and let C([a, b], R) = {f : [a, b] → R | f is continuous}. Then,
C([a, b], R) with (f + αg)(x) = f (x) + αg(x), for all x ∈ [a, b], is a real vector space.
9. Let C(R, R) = {f : R → R | f is continuous}. Then, C(R, R) is a real vector space, where

(f + αg)(x) = f (x) + αg(x), for all x ∈ R.
10. Fix a < b ∈ R and let C 2 ((a, b), R) = {f : (a, b) → R | f 00 is continuous}. Then,
C 2 ((a, b), R) with (f + αg)(x) = f (x) + αg(x), for all x ∈ (a, b), is a real vector space.
11. Let R[x] = {a0 + a1 x + · · · + an xn | ai ∈ R, for 0 ≤ i ≤ n}. Now, let p(x), q(x) ∈ R[x].
T
AF
Then, we can choose m such that p(x) = a0 + a1 x + · · · + am xm and q(x) = b0 + b1 x +

· · · + bm xm , where some of the ai ’s or bj ’s may be zero. Then, we define
DR
p(x) + q(x) = (a0 + b0 ) + (a1 + b1 )x + · · · + (am + bm )xm
and αp(x) = (αa0 ) + (αa1 )x + · · · + (αam )xm , for α ∈ R. With these operations
“component-wise addition and multiplication”, it can be easily verified that R[x] forms a
real vector space.
12. Fix n ∈ N and let R[x; n] = {p(x) ∈ R[x] | p(x) has degree ≤ n}. Then, with component-
wise addition and multiplication, the set R[x; n] forms a real vector space.
13. Let V and W be vector spaces over F, with operations (+, •) and (⊕, ), respectively. Let
V × W = {(v, w) | v ∈ V, w ∈ W}. Then, V × W forms a vector space over F, if for every
(v1 , w1 ), (v2 , w2 ) ∈ V × W and α ∈ R, we define
(v1 , w1 ) ⊕0 (v2 , w2 ) = (v1 + v2 , w1 ⊕ w2 ), and

α ◦ (v1 , w1 ) = (α • v1 , α w1 ).
v1 +v2 and w1 ⊕w2 on the right hand side mean vector addition in V and W, respectively.
Similarly, α • v1 and α w1 correspond to scalar multiplication in V and W, respectively.
Note that R2 is similar to R × R, where the operations are the same in both spaces.
14. Let Q be the set of scalars. Then,

3.1. VECTOR SPACES: DEFINITION AND EXAMPLES 69
(a) R is a vector space over Q. In this space, all the irrational numbers are vectors but
not scalars.
√
(b) V = {a + b 2 : a, b ∈ Q} is a vector space.
√ √ √
(c) V = {a + b 2 + c 3 + d 6 : a, b, c, d ∈ Q} is a vector space.
√
(d) V = {a + b −3 : a, b ∈ Q} is a vector space.
15. Let R+ = {x ∈ R | x > 0}. Then,

(a) R+ is not a vector space under usual operations of addition and scalar multiplication.
(b) R+ is a real vector space with 1 as the additive identity if we define
u ⊕ v = u · v and α u = uα , for all u, v ∈ R+ and α ∈ R.
16. For any α ∈ R and x = (x1 , x2 )T , y = (y1 , y2 )T ∈ R2 , define
x ⊕ y = (x1 + y1 + 1, x2 + y2 − 3)T and α x = (αx1 + α − 1, αx2 − 3α + 3)T .
Then, R2 is a real vector space with (−1, 3)T as the additive identity.
17. Recall the field Z5 = {0, 1, 2, 3, 4} given on the first page of this chapter. Then, V =
{(a, b) | a, b ∈ Z5 } is a vector space over Z5 having 25 elements/vectors.
From now on, we will use ‘u + v’ for ‘u ⊕ v’ and ‘αu or α · u’ for ‘α u’.
T
AF
Exercise 3.1.6. 1. Verify that the vectors spaces mentioned in Example 3.1.4 do satisfy all
DR
the conditions for vector spaces.
2. Does R with x ⊕ y = x − y and α x = −αx, for all x, y, α ∈ R form a vector space?
3. Let V = R2 . For x = (x1 , x2 )T , y = (y1 , y2 )T ∈ R2 and α ∈ R, define

(a) (x1 , y1 )T ⊕ (x2 , y2 )T = (x1 + x2 , 0)T and α (x1 , y1 )T = (αx1 , 0)T .
(b) x + y = (x1 + y1 , x2 + y2 )T and αx = (αx1 , 0)T .
Then, does V form a vector space under any of the two operations?
3.1.1 Vector Subspace
Definition 3.1.7. Let V be a vector space over F. Then, a non-empty subset W of V is called
a subspace of V if W is also a vector space with vector addition and scalar multiplication in
W coming from that in V (compute the vector addition and scalar multiplication in V and then
the computed vector should be an element of W).
Example 3.1.8.
1. The vector space R[x; n] is a subspace of R[x].
2. Is V = {xp(x) | p(x) ∈ R[x]} a subspace of R[x]?
3. Let V be a vector space. Then V and {0} are subspaces, called trivial subspaces.
4. The real vector space R has no non-trivial subspace. To check this, let V 6= {0} be a
vector subspace of R. Then, there exists x ∈ R, x 6= 0 such that x ∈ V. Now, using scalar
multiplication, we see that {αx | α ∈ R} ⊆ V. As, x 6= 0, the set {αx | α ∈ R} = R. This
in turn implies that V = R.
5. W = {x ∈ R3 | [1, 2, −1]x = 0} is a subspace. It represents a plane in R3 containing 0.

( " # )
1 1 1
6. W = x ∈ R3 | x = 0 is a subspace. What does it represent?
1 −1 −1
7. Verify that W = {(x, 0)T ∈ R2 | x ∈ R} is a subspace of R2 .
8. Is the set of sequences converging to 0 a subspace of the set of all bounded sequences?
9. Let V be the vector space of Example 3.1.4.16. Then,

(a) S = {(x, 0)T | x ∈ R} is not a subspace of V as (x, 0)T ⊕(y, 0)T = (x+y+1, −3)T 6∈ S.
(b) Verify that W = {(x, 3)T | x ∈ R} is a subspace of V.
10. The vector space R+ defined in Example 3.1.4.15 is not a subspace of R.
Let V(F) be a vector space and W ⊆ V, W 6= ∅. We now prove a result which implies that
to check W to be a subspace, we need to verify only one condition.
T
Theorem 3.1.9. Let V(F) be a vector space and W ⊆ V, W 6= ∅. Then, W is a subspace of V

AF
if and only if αu + βv ∈ W whenever α, β ∈ F and u, v ∈ W.

DR
Proof. Let W be a subspace of V and let u, v ∈ W. As W is a subspace, the scalar multiplication

and vector addition gives elements of W itself. Hence, for every α, β ∈ F, αu, βv ∈ W and
αu + βv ∈ W.
Now, we assume that αu + βv ∈ W, whenever α, β ∈ F and u, v ∈ W. To show, W is a
subspace of V:
1. Taking α = 0 and β = 0 ⇒ 0 ∈ W. So, W is non-empty.
2. Taking α = 1 and β = 1, we see that u + v ∈ W, for every u, v ∈ W.
3. Taking β = 0, we see that αu ∈ W, for every α ∈ F and u ∈ W. Hence, using Theo-

rem 3.1.3.3, −u = (−1)u ∈ W as well.
4. The commutative and associative laws of vector addition hold as they hold in V.
5. The conditions related with scalar multiplication and the distributive laws also hold as
they hold in V.
Exercise 3.1.10. 1. Prove that a line in R2 is a subspace if and only if it passes through
origin.
2. Prove that {(x, y, z)T ∈ R3 | ax + by + cz = d} is a subspace of R3 if and only if d = 0.

3.2. LINEAR COMBINATION AND LINEAR SPAN 71
3. Does the set V given below form a subspace? Give reasons for your answer.
(a) Let V = {(x, y, z)T | x + y + z = 1}.

(b) Let V = {(x, y)T ∈ R2 | x · y = 0}.
(c) Let V = {(x, y)T ∈ R2 | x = y 2 }.
(d) Let V = {(x, y)T ∈ R2 | x, y ≥ 0}.
4. Determine all the subspaces of R and R2 .
5. Fix n ∈ N. In the examples given below, is W a subspace of Mn (R), where
(a) W = {A ∈ Mn (R) | A is upper triangular}?
(b) W = {A ∈ Mn (R) | A is symmetric}?
(c) W = {A ∈ Mn (R) | A is skew-symmetric}?
(d) W = {A ∈ Mn (R) | A is a diagonal matrix}?
(e) W = {A ∈ Mn (R) | trace(A) = 0}?
(f ) W = {A ∈ Mn (R) | AT = 2A}?
6. Fix n ∈ N. Then, is W = {A = [aij ] ∈ Mn (R | a11 + a22 = 0} a subspace of the complex

vector space Mn (C)? What if Mn (C) is a real vector space?
T
AF
7. Is W = {f ∈ C([−1, 1]) | f (−1/2) = 0, f (1/2) = 0} a subspace of C([−1, 1])?

DR
8. Are all the sets given below subspaces of R[x]?
(a) W = {f (x) ∈ R[x] | deg(f (x)) = 3}.
(b) W = {f (x) ∈ R[x] | xg(x) for some g(x) ∈ R[x]}.
9. Among the following, determine the subspaces of the complex vector space Cn ?
(a) {(z1 , z2 , . . . , zn )T | z1 is real }.
(b) {(z1 , z2 , . . . , zn )T | z1 + z2 = z3 }.
(c) {(z1 , z2 , . . . , zn )T | | z1 |=| z2 |}.
10. Prove that G = {A ∈ Mn (R) | det(A) = 0} is not subspaces of Mn (R).
3.2 Linear Combination and Linear Span
Let us recollect that system Ax = b was either consistent (has a solution) or inconsistent (no
solution). It turns out that the system Ax = b is consistent leads to the idea that the vector b
is a linear combination of the columns of A. Let us try to understand them using examples.
Example 3.2.1.
           
1 1 2 2 1 1 2
           
1. Let A = 1 2 and b = 3. Then, 3 = 1 + 2
        
. Thus,
3 is a linear
 
1 3 4 4 1 3 4
     
1 1 10

 
  
combination of the vectors in S =  1, 2 . Similarly, the vector
   
16 is a linear
 

 1

3 

22
     
10 1 1 " #
      4
combination of the vectors in S as 
16 = 41 + 62 = A 6 .
    
22 1 3
   
2 1 1 2
   
2. Let b = 
3. Then, the system Ax = b has no solution as REF ([A b]) = 0 1 1.
  
5 0 0 1
Formally, we have the following definition.
Definition 3.2.2. Let V be a vector space over F and let S = {u1 , . . . , un } ⊆ V. Then, a
vector u ∈ V is called a linear combination of elements of S if we can find α1 , . . . , αn ∈ F
such that
n
X
u = α 1 u1 + · · · + α n un = α i ui .
T
i=1
AF
n
αi ui , where α1 , . . . , αn ∈ F, is said to be a linear
P
Or equivalently, any vector of the form
DR
i=1
combination of the elements of S.
Thus, the system Ax = b has a solution ⇒ b is a linear combination of the columns of A.

Or equivalently, b is a linear combination means the system Ax = b has a solution. So, recall
that when we were solving a system of linear equations, we looked at the point of intersections
of lines or plane etc. But, here it leads us to the study of whether a given vector is a linear
combination of a given set S or not? Or in the language of matrices, is b a linear combination
of columns of the matrix A or not?
Example 3.2.3.
1. (3, 4, 5) is not a linear combination of (1, 1, 1) and (1, 2, 1) as the linear system (3, 4, 5) =
a(1, 1, 1) + b(1, 2, 1), in the unknowns a and b has no solution.
2. Is (4, 5, 5) a linear combination

 of(1, 0, 0), (2,1,0) and (3, 3, 1)?
1 2 3 4
   
Solution: Define A =  0 1 3 and b = 5. Then, does the system Ax = b has a
  
0 0 1 5
T
solution? Verify that x = [9, −10, 5] is a solution.
Exercise 3.2.4. 1. Let x ∈ R3 . Prove that xT is a linear combination of (1, 0, 0), (2, 1, 0)
and (3, 3, 1).
2. Find condition(s) on x, y, z ∈ R such that

(a) (x, y, z) is a linear combination of (1, 2, 3), (−1, 1, 4) and (3, 3, 2).
(b) (x, y, z) is a linear combination of (1, 2, 1), (1, 0, −1) and (1, 1, 0).
(c) (x, y, z) is a linear combination of (1, 1, 1), (1, 1, 0) and (1, −1, 0).
3.2.1 Linear Span
Let V be a vector space over F and S a subset of V. We now look at ‘linear span’ of a collection
of vectors. So, here we ask “what is the largest collection of vectors that can be obtained as
linear combination of vectors from S”? Or equivalently, what is the smallest subspace of V that
contains S? We first look at an example for clarity.
Example 3.2.5. Let S = {(1, 0, 0), (1, 2, 0)} ⊆ R3 . We want the largest possible subspace
of R3 which contains vectors of the form α(1, 0, 0), β(1, 2, 0) and α(1, 0, 0) + β(1, 2, 0) for all
possible choices of α, β ∈ R. Note that
1. `1 = {α(1, 0, 0) : α ∈ R} gives the X-axis.
2. `2 = {β(1, 2, 0) : β ∈ R} gives the line passing through (0, 0, 0) and (1, 2, 0).
So, we want the largest subspace of R3 that contains vectors which are formed as sum of
any two points on the two lines `1 and `2 . Or the smallest subspace of R3 that contains S? We
T
give the definition next.

AF
Definition 3.2.6. Let V be a vector space over F and S ⊆ V.

DR
1. Then, the linear span of S, denoted LS(S), is defined as
LS(S) = {α1 u1 + · · · + αn un | αi ∈ F, ui ∈ S, for 1 ≤ i ≤ n}.
That is, LS(S) is the set of all possible linear combinations of finitely many vectors of S.
If S is an empty set, we define LS(S) = {0}.
2. V is said to be finite dimensional if there exists a finite set S such that V = LS(S).
3. If there does not exist any finite subset S of V such that V = LS(S) then V is called
infinite dimensional.
Example 3.2.7. For the set S given below, determine LS(S).
1. S = {(1, 0)T , (0, 1)T } ⊆ R2 .

Solution: LS(S) = {a(1, 0)T + b(0, 1)T | a, b ∈ R} = {(a, b)T | a, b ∈ R} = R2 . Thus,
R2 is finite dimensional.
2. S = {(1, 1, 1)T , (2, 1, 3)T }. What does LS(S) represent in R3 ?

Solution: LS(S) = {a(1, 1, 1)T + b(2, 1, 3)T | a, b ∈ R} = {(a + 2b, a + b, a + 3b)T | a, b ∈
R}. Note that LS(S) represents a plane passing through the points (0, 0, 0)T , (1, 1, 1)T
and (2, 1, 3)T . To get he equation of the plane, we proceed as follows:
Find conditions on x, y and z such that (a + 2b, a + b, a + 3b) = (x, y, z). Or equivalently,
find conditions on x, y and z such that a + 2b = x, a + b = y and

 a + 3b = z has asolution
1 0 2y − x
for all a, b ∈ R. The RREF of the augmented matrix equals 
 
0 1 x−y  . Thus,
0 0 z + y − 2x
the required condition on x, y and z is given by z + y − 2x = 0. Hence,
LS(S) = {a(1, 1, 1)T + b(2, 1, 3)T | a, b ∈ R} = {(x, y, z)T ∈ R3 | 2x − y − z = 0}.
Verify that if T = S ∪ {(1, 1, 0)T } then LS(T ) = R3 . Hence, R3 is finite dimensional. In

general, for every fixed n ∈ N, Rn is finite dimensional as Rn = LS ({e1 , . . . , en }).
3. S = {1 + 2x + 3x2 , 1 + x + 2x2 , 1 + 2x + x3 }.
Solution: To understand LS(S), we need to find condition(s) on α, β, γ, δ such that the
linear system
a(1 + 2x + 3x2 ) + b(1 + x + 2x2 ) + c(1 + 2x + x3 ) = α + βx + γx2 + δx3
in the unknowns a, b, c is always consistent. An application of GJE method gives

α + β − γ − 3δ = 0 as the required condition. Thus,
LS(S) = {α + βx + γx2 + δx3 ∈ R[x] | α + β − γ − 3δ = 0}.
Note that, for every fixed n ∈ N, R[x; n] is finite dimensional as R[x; n] = LS ({1, x, . . . , xn }).
T
AF
    


  0 1 1 0 1 2 


DR
4. S = I3 , 1 1 2, 1 0 2 ⊆ M3 (R).
 
   
 
1 2 0 2 2 4 

 
Solution: To get the equation, we need to find conditions on aij ’s such that the system
   
α β+γ β + 2γ a11 a12 a13
   
 = a21 a22 a23 ,
β+γ α + β 2β + 2γ   

β + 2γ 2β + 2γ α + 2γ a31 a32 a33
in the unknowns α, β, γ is always consistent. Now, verify that the required condition
equals
a22 + a33 − a13
LS(S) = {A = [aij ] ∈ M3 (R) | A = AT , a11 = ,
2
a22 − a33 + 3a13 a22 − a33 + 3a13

a12 = , a23 = .
4 2
In general, for each fixed m, n ∈ N, the vector space Mm,n (R) is finite dimensional as
Mm,n (R) = LS ({eij : 1 ≤ i ≤ m, 1 ≤ j ≤ n}).
5. C[x] is not finite dimensional as the degree of a polynomial can be any large positive
integer. Indeed, verify that C[x] = LS({1, x, x2 , . . . , xn , . . .}).
6. The vector space R over Q is infinite dimensional.
Exercise 3.2.8. Determine the equation of the geometrical object represented by LS(S).
1. S = {π} ⊆ R.
2. S = {(x, y)T : x, y < 0} ⊆ R2 .
3. S = {(x, y)T : either x 6= 0 or y 6= 0} ⊆ R2 .
4. S = {(1, 0, 1)T , (0, 1, 0)T , (2, 0, 2)T } ⊆ R3 . Give two examples of vectors u, v different
from the given set such that LS(S) = LS(u, v).
5. S = {(x, y, z)T : x, y, z > 0} ⊆ R3 .

     


 0 1 0 0 0 1 0 1 1 


0 1, −1 0 0 ⊆ M3 (R).
   
6. S = −1 0 1,  0
     
 
 0 −1 0

−1 −1 0 −1 0 0 

7. S = {(1, 2, 3, 4)T , (−1, 1, 4, 5)T , (3, 3, 2, 3)T } ⊆ R4 .
8. S = {1 + 2x + 3x2 , −1 + x + 4x2 , 3 + 3x + 2x2 } ⊆ C[x; 2].
9. S = {1, x, x2 , . . .} ⊆ C[x].
Lemma 3.2.9. Let V be a vector space over F with S ⊆ V. Then LS(S) is a subspace of V.
Proof. By definition, 0 ∈ LS(S). So, LS(S) is non-empty. Let u, v ∈ LS(S). To show,

au + bv ∈ LS(S) for all a, b ∈ F. As u, v ∈ LS(S), there exist n ∈ N, vectors wi ∈ S and
T
scalars αi , βi ∈ F such that u = α1 w1 + · · · + αn wn and v = β1 w1 + · · · + βn wn . Hence,

AF
DR
au + bv = (aα1 + bβ1 )w1 + · · · + (aαn + bβn )wn ∈ LS(S)
as aαi + bβi ∈ F for 1 ≤ i ≤ n. Thus, by Theorem 3.1.9, LS(S) is a vector subspace.
Exercise 3.2.10. Let V be a vector space over F and W ⊆ V.

1. Then LS(W ) = W if and only if W is a subspace of V.
2. If W is a subspace of V and S ⊆ W then LS(S) is a subspace of W as well.
Theorem 3.2.11. Let V be a vector space over F and S ⊆ V. Then LS(S) is the smallest
subspace of V containing S.
Proof. For every u ∈ S, u = 1 · u ∈ LS(S). Thus, S ⊆ LS(S). Need to show that LS(S) is the
smallest subspace of V containing S. So, let W be any subspace of V containing S. Then, by
Exercise 3.2.10, LS(S) ⊆ W and hence the result follows.
Definition 3.2.12. Let V be a vector space over F and S, T be two subsets of V. Then, the
sum of S and T , denoted S + T equals {s + t| s ∈ S, t ∈ T }.
Example 3.2.13.
1. If V = R, S = {0, 1, 2, 3, 4, 5, 6} and T = {5, 10, 15} then S + T = {5, 6, . . . , 21}.
(" #) (" #) (" #)
1 −1 0
2. If V = R2 , S = and T = then S + T = .
1 1 2
(" #) " #! (" # " # )
1 −1 1 −1
3. If V = R2 , S= and T = LS then S + T = +c |c∈R .
1 1 1 1
Exercise 3.2.14. Let P and Q be two non-trivial, distinct subspaces of R2 . Then P + Q = R2 .
We leave the proof of the next result for readers.
Lemma 3.2.15. Let P and Q be two subspaces of a vector space V over F. Then P + Q is a
subspace of V. Furthermore, P + Q is the smallest subspace of V containing both P and Q.
Exercise 3.2.16. 1. Let a ∈ R2 , a 6= 0. Then {x ∈ R2 | aT x = 0} is a non-trivial subspace

of R2 . Geometrically, what does this set represent in R2 ?
2. Find all subspaces of R3 .

(" # ) (" # )
a b a 0
3. Let U = | a, b ∈ R and W = | a, d ∈ R be subspaces of M2 (R).
−b 0 0 d
Determine U ∩ W. Is M2 (R) = U + W?
4. Let W and U be two subspaces of a vector space V over F.
(a) Prove that W ∩ U is a subspace of V.

(b) Give examples of W and U such that W ∪ U is not a subspace of V.
(c) Determine conditions on W and U such that W ∪ U a subspace of V?
T
AF
(d) Prove that LS(W ∪ U) = W + U.

DR
5. Let S = {x1 , x2 , x3 , x4 }, where x1 = (1, 0, 0)T , x2 = (1, 1, 0)T , x3 = (1, 2, 0)T and x4 =
(1, 1, 1)T . Then, determine all xi such that LS(S) = LS(S \ {xi }).
6. Let W = LS((1, 0, 0)T , (1, 1, 0)T ) and U = LS((1, 1, 1)T ). Prove that W + U = R3 and
W ∩ U = {0}. If v ∈ R3 , determine w ∈ W and u ∈ U such that v = w + u. Is it
necessary that w and u are unique?
7. Let W = LS((1, −1, 0), (1, 1, 0)) and U = LS((1, 1, 1), (1, 2, 1)). Prove that W + U = R3
and W ∩ U 6= {0}. Find v ∈ R3 such that v = w + u, for 2 different choices of w ∈ W
and u ∈ U. Thus, the choice of vectors w and u is not unique.
8. Let S = {(1, 1, 1, 1)T , (1, −1, 1, 2)T , (1, 1, −1, 1)T } ⊆ R4 . Does (1, 1, 2, 1)T ∈ LS(S)? Fur-
thermore, determine conditions on x, y, z and u such that (x, y, z, u)T ∈ LS(S).
3.3 Linear Independence

Let us now go back to homogeneous system Ax = 0. Here, we saw that this system has either a
non-trivial solution or only the trivial solution. The idea of a non-trivial solution leads to linear
dependence of vectors and the idea of only the trivial solution leads to linear independence. We
look at a few examples for better understanding.
3.3. LINEAR INDEPENDENCE 77
Example 3.3.1.
 
1 1
 
1. Let A =  1 2 . Then Ax = 0 has only the trivial solution. So, we say that the columns

1 3
   
1 1

 

of A are linearly independent. Thus, the set S =  1, 2 , consisting of columns of
   

 1

3 

A, is linearly independent.
   
1 1 2 1 1 2
   
2. Let A =  1 2 3. As REF (A) = 0 1 1, Ax = 0 has only the trivial solution.
  
1 3 5 0 0 1
     
1 1 2

 

Hence, the set S =  ,  ,
1 2 3, consisting of columns of A, is linearly independent.
 

 1

3 5 

   
1 1 2 1 1 2
   
3. Let A = 1 2 3. As REF (A) = 0 1
  
, Ax = 0 has a non-trivial solution. Hence,
1
1 3 4 0 0 0
     
1 1 2

 

T
the set S =  ,  ,
1 2 3, consisting
  of columns of A, is linearly dependent.
AF

 1

3 4 

DR
Formally, we have the following definition.
Definition 3.3.2. Let S = {u1 , . . . , um } be a non-empty subset of a vector space V over F.

Then, S is said to be linearly independent if the linear system
α1 u1 + α2 u2 + · · · + αm um = 0, (3.3.1)
in the unknowns αi ’s, 1 ≤ i ≤ m, has only the trivial solution. If Equation (3.3.1) has a
non-trivial solution then S is said to be linearly dependent. If S has infinitely many vectors
then S is said to be linearly independent if for every finite subset T of S, T is linearly
independent.
Observe that we are solving a linear system over F. Hence, whether a set is linearly inde-
pendent or linearly dependent depends on the set of scalars.
Example 3.3.3.
1. Consider C2 as a vector space over R. Let S = {(1, 2)T , (i, 2i)T }. Then, the linear system
a · (1, 2)T + b · (i, 2i)T = (0, 0)T , in the unknowns a, b ∈ R has only the trivial solution,
namely a = b = 0. So, S is a linear independent subset of the vector space C2 over R.
2. Consider C2 as a vector space over C. Then S = {(1, 2)T , (i, 2i)T } is a linear dependent
subset of the vector space C2 over C as a = −i and b = 1 is a non-trivial solution.
3. Let V be the vector space of all real valued continuous functions with domain [−π, π].
Then V is a vector space over R. Question: What can you say about the linear indepen-
dence or dependence of the set S = {1, sin(x), cos(x)}?
Solution: For all x ∈ [−π, π], consider the system
 
h ia
1 sin(x) cos(x)   b  = 0 ⇔ a · 1 + b · sin(x) + c · cos(x) = 0, (3.3.2)

c
in the unknowns a, b and c. Even though we seem to have only one linear system, we we
can obtain the following two linear systems (the first using differentiation and the second
π
using evaluation at 0, and π of the domain).
2
 
a + b sin x + c cos x =0 
  a+c =0


0 · a + b cos x − c sin x = 0 or a+b =0
 
0 · a − b sin x − c cos x = 0 a−c =0

 

Clearly, the above systems has only the trivial solution. Hence, S is linearly independent.
4. Let A ∈ Mm,n (C). If Rank(A) < m then, the rows of A are linearly dependent. " #
C
Solution: As Rank(A) < m, there exists an invertible matrix P such that P A = .
0
T
m
AF
Thus, 0T = (P A)[m, :] =
P
pmi A[i, :]. As P is invertible, at least one pmi 6= 0. Thus, the
i=1
required result follows.
DR
5. Let A ∈ Mm,n (C). If Rank(A) < n then, the columns of A are linearly dependent.
Solution: As Rank(A) < n the system Ax = 0 has a non-trivial solution.
6. Let S = {0}. Is S linearly independent?
Solution: Let u = 0. So, consider the system αu = 0. This has a non-trivial solution
α = 1 as 1 · 0 = 0.
(" # " #) " #
0 1 0 1
7. Let S = , . Then Ax = 0 corresponds to A = . This has a non-trivial
0 2 0 2
" #
1
solution x = . Hence, S is linearly dependent.
0
(" #)
1
8. Let S = . Is S linearly independent?
2
" #
1
Solution: Let u = . Then the system αu = 0 has only the trivial solution. Hence S
2
is linearly independent.
So, we observe that 0, the zero-vector cannot belong to any linearly independent set. Fur-
ther, a set consisting of a single non-zero vector is linearly independent.
Exercise 3.3.4. 1. Show that S = {(1, 2, 3)T , (−2, 1, 1)T , (8, 6, 10)T } ⊆ R3 is linearly de-
pendent.
2. Let A ∈ Mn (R). Suppose x, y ∈ Rn \ {0} such that Ax = 3x and Ay = 2y. Then, prove
that x and y are linearly independent.
 
2 1 3
. Determine x, y, z ∈ R3 \ {0} such that Ax = 6x, Ay = 2y and
 
3. Let A = 4 −1 3 
3 −2 5
Az = −2z. Use the vectors x, y and z obtained above to prove the following.
(a) A2 v = 4v, where v = cy + dz for any c, d ∈ R.

(b) The set {x, y, z} is linearly independent.
(c) Let P = [x, y, z] be a 3 × 3 matrix. Then, P is invertible.
 
6 0 0
 
(d) Let D = 
0 2 0 . Then AP = P D.
0 0 −2
3.3.1 Basic Results on Linear Independence
The reader is expected to supply the proof of the next proposition.
Proposition 3.3.5. Let V be a vector space over F.
1. Then, 0, the zero-vector, cannot belong to a linearly independent set.

T
AF
2. Then, every subset of a linearly independent set in V is also linearly independent.

DR
3. Then, a set containing a linearly dependent set of V is also linearly dependent.
We now prove a couple of results which will be very useful in the next section.
Proposition 3.3.6. Let S be a linearly independent subset of a vector space V over F. If

T1 , T2 are two subsets of S such that T1 ∩ T2 = ∅ then, LS(T1 ) ∩ LS(T2 ) = {0}. That is, if
v ∈ LS(T1 ) ∩ LS(T2 ) then v = 0.
Proof. Let v ∈ LS(T1 )∩LS(T2 ). Then, there exist vectors u1 , . . . , uk ∈ T1 , w1 , . . . , w` ∈ T2 and

k
P P̀
scalars αi ’s and βj ’s (not all zero) such that v = αi ui and v = βj wj . Thus, we see that
i=1 j=1
k
P P̀
αi ui + (−βj )wj = 0. As the scalars αi ’s and βj ’s are not all zero, we see that a non-trivial
i=1 j=1
linear combination of some vectors in T1 ∪ T2 ⊆ S is 0. This contradicts the assumption that S
is a linearly independent subset of V. Hence, each of α’s and βj ’s is zero. That is v = 0.
Lemma 3.3.7. Let S be a linearly independent subset of a vector space V over F. Then, each
v ∈ LS(S) is a unique linear combination of vectors from S.
Proof. Suppose there exists v ∈ LS(S) with v ∈ LS(T1 ), LS(T2 ) with T1 , T2 ⊆ S. Let T1 =
{v1 , . . . , vk } and T2 = {w1 , . . . , w` }, for some vi ’s and wj ’s in S. Define T = T1 ∪ T2 . Then,
T is a subset of S. Hence, using Proposition 3.3.5, the set T is linearly independent. Let T =
{u1 , . . . , up }. Then, there exist αi ’s and βj ’s in F, not all zero, such that v = α1 u1 + · · · + αp up
as well as v = β1 u1 + · · · + βp up . Equating the two expressions for v gives
(α1 − β1 )u1 + · · · + (αp − βp )up = 0. (3.3.3)
As T is a linearly independent subset of V, the system c1 v1 + · · · + cp vp = 0, in the variables

c1 , . . . , cp , has only the trivial solution. Thus, in Equation (3.3.3), αi − βi = 0, for 1 ≤ i ≤ p.
Thus, for 1 ≤ i ≤ p, αi = βi and the required result follows.
Theorem 3.3.8. Let S = {u1 , . . . , uk } be a non-empty subset of a vector space V over F. If

Z ⊆ LS(S) having more than k vectors then, Z is a linearly dependent subset in V.
Proof. Let Z = {w1 , . . . , wm }. As wi ∈ LS(S), there exist aij ∈ F such that
wi = ai1 u1 + · · · + aik uk , for 1 ≤ i ≤ m.
So,       
w1 a11 u1 + · · · + a1k uk a11 · · · a1k u1
 .   .   . . .  . 
 . = .. = . .. ..  .. .
 .     .  
wm am1 u1 + · · · + amk uk am1 · · · amk uk
As m > k, the homogeneous system AT x = 0 has a non-trivial solution, say y 6= 0, i.e.,
T
AT y = 0 ⇔ yT A = 0T . Thus,
AF
        
w1 u1 u1 u1
DR
 .    . 
yT  T A .  = (yT A) .  = 0T  ..  = 0T .
.
   
. 
 . =y   .   .   . 
wm uk uk uk
As y 6= 0, a non-trivial linear combination of vectors in Z is 0. Thus, the set Z is linearly

dependent subset of V.
Corollary 3.3.9. Fix n ∈ N. Then, any subset S of Rn with | S | ≥ n + 1 is linearly dependent.
Proof. Observe that Rn = LS({e1 , . . . , en }), where ei = In [:, i], is the i-th column of In . Hence,
using Theorem 3.3.8, the required result follows.
Theorem 3.3.10. Let S be a linearly independent subset of a vector space V over F. Then, for
any v ∈ V the set S ∪ {v} is linearly dependent if and only if v ∈ LS(S).
Proof. Let us assume that S ∪ {v} is linearly dependent. Then, there exist vi ’s in S such that
the linear system
α1 v1 + · · · + αp vp + αp+1 v = 0 (3.3.4)
in the variables αi ’s has a non-trivial solution, say αi = ci , for 1 ≤ i ≤ p + 1. We claim that

cp+1 6= 0.
For, if cp+1 = 0 then, Equation (3.3.4) has a non-trivial solution corresponds to having a
non-trivial solution of the linear system α1 v1 + · · · + αp vp = 0 in the variables α1 , . . . , αp . This
contradicts Proposition 3.3.5.2 as {v1 , . . . , vp } ⊆ S, a linearly independent set. Thus, cp+1 6= 0

and we get
1 ci
v=− (c1 v1 + · · · + cp vp ) ∈ LS(v1 , . . . , vp ) as − ∈ F, for 1 ≤ i ≤ p.
cp+1 cp+1
Now, assume that v ∈ LS(S). Then, there exists vi ∈ S and ci ∈ F, not all zero, such that
p
P
v= ci vi . Thus, the linear system α1 v1 + · · · + αp vp + αp+1 v = 0 in the variables αi ’s has a
i=1
non-trivial solution [c1 , . . . , cp , −1]. Hence, S ∪ {v} is linearly dependent.
We now state a very important corollary of Theorem 3.3.10 without proof. This result can
also be used as an alternative definition of linear independence and dependence.
Corollary 3.3.11. Let V be a vector space over F and let S be a subset of V containing a
non-zero vector u1 .
1. If S is linearly dependent then, there exists k such that LS(u1 , . . . , uk ) = LS(u1 , . . . , uk−1 ).
Or equivalently, if S is a linearly dependent set then there exists a vector uk , for k ≥ 2,
which is a linear combination of the previous vectors.
2. If S linearly independent then, v ∈ V \ LS(S) if and only if S ∪ {v} is also a linearly

independent subset of V.
3. If S is linearly independent then, LS(S) = V if and only if each proper superset of S is

linearly dependent.
T
AF
As an application, we have the following result about finite dimensional vector spaces. We
leave the proof for the reader as it directly follows from Corollary 3.3.11 and the idea that an
DR
algorithm has to finally stop if it has finite number of steps to implement.

Theorem 3.3.12. Let V is a finite dimensional vector space over F.
1. If S is a finite subset of V such that LS(S) = V then we can find a subset T of S such
that T is linearly independent and LS(T ) = V.
2. Let T be a linearly independent subset of V. Then, we can find a superset S of T such
that S is linearly independent and LS(S) = V.
Exercise 3.3.13.
1. Prove Corollary 3.3.11.
2. Let V and W be subspaces of Rn such that V + W = Rn and V ∩ W = {0}. Prove that
each u ∈ Rn is uniquely expressible as u = v + w, where v ∈ V and w ∈ W.
3. Let W be a subspace of a vector space V over F. For u, v ∈ V \ W, define K = LS(W, u)
and M = LS(W, v). Then, prove that v ∈ K if and only if u ∈ M .
4. Suppose V is a vector space over R as well as over C. Then, prove that {u1 , . . . , uk }
is a linearly independent subset of V over C if and only if {u1 , . . . , uk , iu1 , . . . , iuk } is a
linearly independent subset of V over R.
5. Is the set {1, x, x2 , . . .} a linearly independent subset of the vector space C[x] over C?
6. Is the set {eij | 1 ≤ i ≤ m, 1 ≤ j ≤ n} a linearly independent subset of the vector space
Mm,n (C) over C (see Definition 1.4.1.1)?
3.3.2 Application to Matrices
In this subsection, we use the understanding of vector spaces to relate the rank of a matrix
with linear independence and dependence of rows and columns of a matrix. We start with our
understanding of the RREF.
Theorem 3.3.14. Let A ∈ Mm,n (C) with Rank(A) = r. Then,

1. there exist r rows of A that are linearly independent.
2. every collection of (r + 1) rows of A are linearly dependent.
3. there exist r columns of A that are linearly independent.
4. every collection of (r + 1) columns of A are linearly dependent.
Proof. As Rank(A) = r, there exist an invertible

" # matrix P and an r × n matrix B having r
B
pivots such that P A = RREF(A) = R = . As B is in RREF, the matrix Ir is a submatrix
0
of B. Hence, the rows of B are linearly independent. Thus, we have shown that the pivotal
rows of R are linearly independent. These pivotal rows would have come from certain initial
rows, say i1 , . . . , ir ,"of #A. Thus, the rows "{A[i
#1 , :], . . . , A[i"r , :}]
# is a linearly independent set.
B B B
Further, P A = implies, A = P −1 = [P1 P2 ] = P1 B, for some matrix P1 Thus,
0 0 0
T
every row of A is a linear combination of the r-rows of B. Hence, using Theorem 3.3.8 any
AF
collection of (r + 1) rows of A are linearly dependent.

Let B[:, i1 ], . . . , B[:, ir ] be the pivotal columns of B. Then, they are linearly independent
DR
due to pivotal 1’s. As B = RREF(A), there exists an invertible matrix P such that B = P A.
Then, the corresponding columns of A satisfy
[A[:, i1 ], . . . , A[:, ir ]] = [P −1 B[:, i1 ], . . . , P −1 B[:, ir ]] = P −1 [B[:, i1 ], . . . , B[:, ir ]].

   
x1 x1
. .
As P is invertible, the systems [A[:, i1 ], . . . , A[:, ir ]] . .
 .  = 0 and [B[:, i1 ], . . . , B[:, ir ]] .  = 0
xr xr
are row-equivalent. Thus, they have the same solution set. Hence, {A[:, i1 ], . . . , A[:, ir ]} is
linearly independent if and only if {B[:, i1 ], . . . , B[:, ir ]} is linear independent. Thus, the required
result follows.
We consider an example for clarity of the above result.

   
1 1 1 0 1 0 −1 0

1 0 −1
  
1 with RREF(A) = B = 0
 1 2 0
Example 3.3.15. Let A =  2 1 0
.
 1 0
 0 0 1
1 1 1 2 0 0 0 0
1. Then, B[:, 3] = −B[:, 1] + 2B[:, 2]. Thus, A[:, 3] = −A[:, 1] + 2A[:, 2].
2. As the 1-st, 2-nd and 4-th columns of B are linearly independent, the set consisting of
corresponding columns {A[:, 1], A[:, 2], A[:, 4]} is linearly independent.
3.4. BASIS OF A VECTOR SPACE 83
3. Also, note that during the application of GJE, the 3-rd and 4-th rows were interchanged.
Hence, the rows A[1, :], A[2, :] and A[4, :] are linearly independent.
As an immediate corollary of Theorem 3.3.14 one has the following result.
Corollary 3.3.16. The following statements are equivalent for A ∈ Mn (C).

1. A is invertible.
2. The columns of A are linearly independent.
3. The rows of A are linearly independent.
Exercise 3.3.17. 1. Let S1h= {u1 , . . . , uni} and

h S2 = {w1 ,i. . . , wn } be subsets of a complex
vector space V. Also, let w1 · · · wn = u1 · · · un A for some matrix A ∈ Mn (C).
(a) If A = [aij ] is invertible then S1 is a linearly independent if and only if S2 is linearly

independent.
(b) If S2 is linearly independent then prove that A is invertible. Further, in this case,
the set S1 is necessarily linearly independent.
2. Let S = {u1 , . . . , un } ⊆ Cn and T = {Au1 , . . . , Aun }, for some matrix A ∈ Mn (C).

(a) If S is linearly dependent then prove that T is linear dependent.
T
AF
(b) If S is linearly independent then prove that T is linearly independent for every in-
vertible matrix A.
DR
(c) If T is linearly independent then S is linearly independent. Further, in this case, the
matrix A is necessarily invertible.
3.4 Basis of a Vector Space

Definition 3.4.1. Let S be a subset of a set T . Then, S is said to be a maximal subset of
T having property P if
1. S has property P and
2. no proper superset of S in T has property P .
Example 3.4.2. Let T = {2, 3, 4, 7, 8, 10, 12, 13, 14, 15}. Then, a maximal subset of T of
consecutive integers is S = {2, 3, 4}. Other maximal subsets are {7, 8}, {10} and {12, 13, 14, 15}.
Note that {12, 13} is not maximal. Why?
Definition 3.4.3. Let V be a vector space over F. Then, S is called a maximal linearly
independent subset of V if
1. S is linearly independent and
2. no proper superset of S in V is linearly independent.
Example 3.4.4.
1. In R3 , the set S = {e1 , e2 } is linearly independent but not maximal as S ∪ {(1, 1, 1)T } is
a linearly independent set containing S.
2. In R3 , S = {(1, 0, 0)T , (1, 1, 0)T , (1, 1, −1)T } is a maximal linearly independent set as S is
linearly independent and any collection of 4 or more vectors from R3 is linearly dependent
(see Corollary 3.3.9).
3. Let S = {v1 , . . . , vk } ⊆ Rn . Now, form the matrix A = [v1 , . . . , vk ] and let B =
RREF(A). Then, using Theorem 3.3.14, we see that if B[:, i1 ], . . . , B[:, ir ] are the piv-
otal columns of B then {vi1 , . . . , vir } is a maximal linearly independent subset of S.
4. Is the set {1, x, x2 , . . .} a maximal linearly independent subset of C[x] over C?
5. Is the set {eij | 1 ≤ i ≤ m, 1 ≤ j ≤ n} a maximal linearly independent subset of Mm,n (C)
over C?
Theorem 3.4.5. Let V be a vector space over F and S a linearly independent set in V. Then,
S is maximal linearly independent if and only if LS(S) = V.
Proof. Let v ∈ V. As S is linearly independent, using Corollary 3.3.11.2, the set S ∪ {v} is
linearly independent if and only if v ∈ V \ LS(S). Thus, the required result follows.
Let V = LS(S) for some set S with | S | = k. Then, using Theorem 3.3.8, we see that if
T ⊆ V is linearly independent then | T | ≤ k. Hence, a maximal linearly independent subset
T
of V can have at most k vectors. Thus, we arrive at the following important result.
AF
Theorem 3.4.6. Let V be a vector space over F and let S and T be two finite maximal linearly
DR
independent subsets of V. Then | S | = | T | .
Proof. By Theorem 3.4.5, S and T are maximal linearly independent if and only if LS(S) =
V = LS(T ). Now, use the previous paragraph to get the required result.
Let V be a finite dimensional vector space. Then, by Theorem 3.4.6, the number of vectors
in any two maximal linearly independent set is the same. We use this number to now define
the dimension of a vector space.
Definition 3.4.7. Let V be a finite dimensional vector space over F. Then, the number of
vectors in any maximal linearly independent set is called the dimension of V, denoted dim(V).
By convention, dim({0}) = 0.
Example 3.4.8.
1. As {1} is a maximal linearly independent subset of R, dim(R) = 1.
2. As {e1 , . . . , en } is a maximal linearly independent subset in Rn , dim(Rn ) = n.
3. As {e1 , . . . , en } is a maximal linearly independent subset in Cn over C, dim(Cn ) = n.
4. Using Exercise 3.3.13.4, {e1 , . . . , en , ie1 , . . . , ien } is a maximal linearly independent subset
in Cn over R. Thus, as a real vector space, dim(Cn ) = 2n.
5. As {eij | 1 ≤ i ≤ m, 1 ≤ j ≤ n} is a maximal linearly independent subset of Mm,n (C) over
C, dim(Mm,n (C)) = mn.
Definition 3.4.9. Let V be a finite dimensional vector space over F. Then, a maximal linearly
independent subset of V is called a basis of V. The vectors in a basis are called basis vectors.
By convention, a basis of {0} is the empty set.
Thus, using Theorem 3.3.12 we see that every finite dimensional vector space has a basis.
Remark 3.4.10 (Standard Basis). The readers should verify the statements given below.
1. All the maximal linearly independent set given in Example 3.4.8 form the standard basis
of the respective vector space.
2. {1, x, x2 , . . .} is the standard basis of R[x] over R.
3. Fix a positive integer n. Then {1, x, x2 , . . . , xn } is the standard basis of R[x; n] over R.
4. Let V = {A ∈ Mn (R) | A = AT }. Then, V is a vector space over R with standard basis

{eii , eij + eji | 1 ≤ i < j ≤ n}.
5. Let V = {A ∈ Mn (R) | AT = −A}. Then, V is a vector space over R with standard basis
{eij − eji | 1 ≤ i < j ≤ n}.
Definition 3.4.11. Let V be a vector space over F. Then, a subset S of V is called minimal
spanning if LS(S) = V and no proper subset of S spans V.
T
AF
Example 3.4.12.
DR
1. Note that {−2} is a basis and a minimal spanning subset in R.
2. Let u1 , u2 , u3 ∈ R2 . Then {u1 , u2 , u3 } can neither be a basis nor a minimal spanning

subset of R2 .
3. Let V = {(x, y, 0)T | x, y ∈ R} ⊆ R3 . Then, B = {(1, 0, 0)T , (1, 3, 0)T } is a basis of V.
4. Let V = {(x, y, z)T ∈ R3 | x + y − z = 0} ⊆ R3 . As each element (x, y, z)T ∈ V satisfies

x + y − z = 0. Or equivalently z = x + y, we see that
(x, y, z) = (x, y, x + y) = (x, 0, x) + (0, y, y) = x(1, 0, 1) + y(0, 1, 1).
Hence, {(1, 0, 1)T , (0, 1, 1)T } forms a basis of V.
5. Let S = {a1 , . . . , an }. Then, RS is a real vector space (see Example 3.1.4.7). For 1 ≤ i ≤ n,
define the functions (
1 if j = i
ei (aj ) = .
0 otherwise
Then, prove that B = {e1 , . . . , en } is a linearly independent subset of RS over R. Is it a

basis of RS over R? What can you say if S is a countable set?
6. Let S = {v1 , . . . , vk } ⊆ Rn . Define A = [v1 , . . . , vk ]. Then, using Example 3.4.4.3,

we see that dim(LS(S)) = Rank(A). Further, using Theorem 3.3.14, the columns of A
corresponding to the pivotal columns in RREF(A) form a basis of LS(S).
3.4.1 Main Results associated with Bases
Theorem 3.4.13. Let V be a non-zero vector space over F. Then, the following statements are
equivalent.
1. B is a basis (maximal linearly independent subset) of V.
2. B is linearly independent and spans V.
3. B is a minimal spanning set in V.
Proof. 1 ⇒ 2 By definition, every basis is a maximal linearly independent subset of V.

Thus, using Corollary 3.3.11.2, we see that B spans V.
2 ⇒ 3 Let S be a linearly independent set that spans V. As S is linearly independent,
/ LS (S − {x}). Hence LS (S − {x}) $ LS(S) = V.
for any x ∈ S, x ∈
3 ⇒ 1 If B is linearly dependent then using Corollary 3.3.11.1, B is not minimal
spanning. A contradiction. Hence, B is linearly independent.
We now need to show that B is a maximal linearly independent set. Since LS(B) = V, for
any x ∈ V \ B, using Corollary 3.3.11.2, the set B ∪ {x} is linearly dependent. That is, every
proper superset of B is linearly dependent. Hence, the required result follows.
Now, using Lemma 3.3.7, we get the following result.
Remark 3.4.14. Let B be a basis of a vector space V over F. Then, for each v ∈ V, there exist
n
T
unique ui ∈ B and unique αi ∈ F, for 1 ≤ i ≤ n, such that v =

P
αi ui .
AF
i=1
DR
The next result is generally known as “every linearly independent set can be extended to
form a basis of a finite dimensional vector space”. Also, recall Theorem 3.3.12.
Theorem 3.4.15. Let V be a vector space over F with dim(V) = n. If S is a linearly independent
subset of V then there exists a basis T of V such that S ⊆ T .
Proof. If LS(S) = V, done. Else, choose u1 ∈ V \ LS(S). Thus, by Corollary 3.3.11.2, the set
S ∪{u1 } is linearly independent. We repeat this process till we get n vectors in T as dim(V) = n.
By Theorem 3.4.13, this T is indeed a required basis.
3.4.2 Constructing a Basis of a Finite Dimensional Vector Space
We end this section with an algorithm which is based on the proof of the previous theorem.
Step 1: Let v1 ∈ V with v1 6= 0. Then, {v1 } is linearly independent.
Step 2: If V = LS(v1 ), we have got a basis of V. Else, pick v2 ∈ V \ LS(v1 ). Then, by

Corollary 3.3.11.2, {v1 , v2 } is linearly independent.
Step i: Either V = LS(v1 , . . . , vi ) or LS(v1 , . . . , vi ) $ V. In the first case, {v1 , . . . , vi } is

a basis of V. Else, pick vi+1 ∈ V \ LS(v1 , . . . , vi ). Then, by Corollary 3.3.11.2, the set
{v1 , . . . , vi+1 } is linearly independent.
This process will finally end as V is a finite dimensional vector space.
Exercise 3.4.16. 1. Let B = {u1 , . . . , un } be a basis of a vector space V over F. Then,

n
P
does the condition αi ui = 0 in αi ’s imply that αi = 0, for 1 ≤ i ≤ n?
i=1
2. Let S = {v1 , . . . , vp } be a subset of a vector space V over F. Suppose LS(S) = V but S

is not a linearly independent set. Then, does this imply that each v ∈ V is expressible
in more than one way as a linear combination of vectors from S? Is it possible to get a
subset T of S such that T is a basis of V over F? Give reasons for your answer.
3. Let V be a vector space of dimension n and let S be a subset of V having n vectors .
(a) If S is linearly independent then prove that S forms a basis of V.

(b) If LS(S) = V then prove that S forms a basis of V.
4. Let {v1 , . . . , vn } be a basis of Cn . Then, prove that the two matrices B = [v1 , . . . , vn ] and
 
v1T
 . 
C= . 
 .  are invertible.
vnT
5. Let W1 and W2 be two subspaces of a finite dimensional vector space V such that W1 ⊆ W2 .
Then, prove that W1 = W2 if and only if dim(W1 ) = dim(W2 ).
6. Let W1 be a subspace of a finite dimensional vector space V over F. Then, prove that
T
there exists a subspace W2 of V such that

AF
W1 ∩ W2 = {0}, W1 + W2 = V and dim(W2 ) = dim(V) − dim(W1 ).

DR
Also, prove that for each v ∈ V there exist unique vectors w1 ∈ W1 and w2 ∈ W2 with
v = w1 + w2 . The subspace W2 is called the complementary subspace of W1 in V and
we write V = W1 ⊕ W2 .
7. Let V be a finite dimensional vector space over F. If W1 and W2 are two subspaces of V
such that W1 ∩W2 = {0} and dim(W1 )+dim(W2 ) = dim(V) then prove that W1 +W2 = V.
8. Consider the vector space C([−π, π]) over R. For each n ∈ N, define en (x) = sin(nx).
Then, prove that S = {en | n ∈ N} is linearly independent. [Hint: Need to show that every
finite subset of S is linearly independent. So, on the contrary assume that there exists ` ∈ N and
functions ek1 , . . . , ek` such that α1 ek1 + · · · + α` ek` = 0, for some αt 6= 0 with 1 ≤ t ≤ `. But,
the above system is equivalent to looking at α1 sin(k1 x) + · · · + α` sin(k` x) = 0 for all x ∈ [−π, π].
Now in the integral
Z π Z π
sin(mx) (α1 sin(k1 x) + · · · + α` sin(k` x)) dx = sin(mx)0 dx = 0
−π −π
replace m with ki ’s to show that αi = 0, for all i, 1 ≤ i ≤ `. This gives the required contradiction.]
9. Is the set {1, sin(x), cos(x), sin(2x), cos(2x), sin(3x), cos(3x), . . .} a linearly subset of the
vector space C([−π, π], R) over R?
10. Find a basis of R3 containing the vector (1, 1, −2)T and (1, 2, −1)T .
11. Determine a basis and dimension of W = {(x, y, z, w)T ∈ R4 | x + y − z + w = 0}.
12. Find a basis of V = {(x, y, z, u) ∈ R4 | x − y − z = 0, x + z − u = 0}.

 
1 0 1 1 0
. Find a basis of V = {x ∈ R5 | Ax = 0}.
 
13. Let A = 
 0 1 2 3 0
0 0 0 0 1
14. Let uT = (1, 1, −2), vT = (−1, 2, 3) and wT = (1, 10, 1). Find a basis of LS(u, v, w).
Determine a geometrical representation of LS(u, v, w).
15. Is the set W = {p(x) ∈ R[x; 4] | p(−1) = p(1) = 0} a subspace of R[x; 4]? If yes, find its
dimension.
3.5 Fundamental Subspaces Associated with a Matrix

In this section, we will study results that are intrinsic to the understanding of linear algebra
from the point of view of matrices. For the sake of clarity, we will also restrict our attention to
matrices with real entries. So, we start with defining the four fundamental subspaces associated
with a matrix.
T
AF
Definition 3.5.1. Let A ∈ Mm,n (R). Then, we define the four fundamental subspaces associ-
DR
ated with A as
1. Col(A) = {Ax | x ∈ Rn } is a subspace of Rm , called the Column space, and is the
linear span of the columns of A.
2. Row(A) = Col(AT ) = {AT x | x ∈ Rm } is a subspace of Rn , called the row space of A
and is the linear span of the rows of A.
3. Null(A) = {x ∈ Rn | Ax = 0}, called the Null space of A.
4. Null(AT ) = {x ∈ Rm | AT x = 0}, also called the left-null space.
Exercise 3.5.2. Let A ∈ Mm,n (R). Then prove that

1. Null(A) and Row(A) are subspaces of Rn .
2. Null(AT ) and Col(A) are subspaces of Rm .
Example 3.5.3.
 
1 1 1 −2
 
1. Compute the fundamental subspaces for A = 
1 2 −1 .
1 
1 −2 7 −11
Solution: Verify the following
(a) Row(A) = {(x, y, z, u)T ∈ R4 | 3x − 2y = z, 5x − 3y + u = 0}.

(b) Col(A) = {(x, y, z)T ∈ R3 | 4x − 3y − z = 0}.
3.5. FUNDAMENTAL SUBSPACES ASSOCIATED WITH A MATRIX 89
(c) Null(A) = {(x, y, z, u)T ∈ R4 | x + 3z − 5u = 0, y − 2z + 3u = 0}.

(d) Null(AT ) = {(x, y, z)T ∈ R3 | x + 4z = 0, y − 3z = 0}.
 
1 1 0 −1
 
2. Let A = 
1 −1 1 2 . Then, verify that

2 0 1 1
(a) Col(A) = {x = (x1 , x2 , x3 )T ∈ R3 | x1 + x2 − x3 = 0}.

(b) Row(A) = {x = (x1 , x2 , x3 , x4 )T ∈ R4 | x1 − x2 − 2x3 = 0, x1 − 3x2 − 2x4 = 0}.
(c) Null(A) = LS({(1, −1, −2, 0)T , (1, −3, 0, −2)T }).
(d) Null(AT ) = LS((1, 1, −1)T ).
Remark 3.5.4. Let A ∈ Mm,n (R). Then, in Example 3.5.3, observe that the direction ratios
of normal vectors of Col(A) matches with vector in Null(AT ). Similarly, the direction ratios
of normal vectors of Row(A) matches with vectors in Null(A). Are these true in the general
setting?
Exercise 3.5.5. 1. For the matrices given below, determine the four fundamental spaces.
Further,
 find the dimensions
 of all 
the vector subspacesso obtained.
1 2 1 3 2 2 4 0 6
 
 and B = −1 0 −2 5 .
 
0 2 2 2 4
T
 
A=
AF
2
 −2 4 0 8  
−3 −5 1 −4
 
4 2 5 6 10 −1 −1 1 2
DR
2. Let A = [X Y ]. Then, determine the condition under which Col(X) = Col(Y ).
The next result is a re-writing of the results on system of linear equations. The readers are
advised to provide the proof for clarity.
Lemma 3.5.6. Let A ∈ Mm×n (C) and let E be an elementary matrix. If

1. B = EA then Null(A) = Null(B), Row(A) = Row(B). Thus, the dimensions of the
corresponding spaces are equal.
2. B = AE then Null(AT ) = Null(B T ), Col(A) = Col(B). Thus, the dimensions of the

corresponding spaces are equal.
Let W1 and W1 be two subspaces of a vector space V over F. Then, recall that (see
Exercise 3.2.16.4d) W1 + W2 = {u + v | u ∈ W1 , v ∈ W2 } = LS(W1 ∪ W2 ) is the smallest
subspace of V containing both W1 and W2 . We now state a result similar to a result in Venn
diagram that states | A | + | B | = | A ∪ B | + | A ∩ B |, whenever the sets A and B are
finite (for a proof, see Appendix 9.4.1).
Theorem 3.5.7. Let V be a finite dimensional vector space over F. If W1 and W2 are two
subspaces of V then
dim(W1 ) + dim(W2 ) = dim(W1 + W2 ) + dim(W1 ∩ W2 ). (3.5.1)

For better understanding, we give an example for finite subsets of Rn . The example uses
Theorem 3.3.14 to obtain bases of LS(S), for different choices S. The readers are advised to
see Example 3.3.14 before proceeding further.
Example 3.5.8. Let V = {(v, w, x, y, z)T ∈ R5 | v + x + z = 3y} and W = {(v, w, x, y, z)T ∈

R5 | w − x = z, v = y}. Find bases of V and W containing a basis of V ∩ W.
Solution: One can first find a basis of V ∩ W and then heuristically add a few vectors to get
bases for V and W, separately.
Alternatively, First find bases of V, W and V ∩ W, say BV , BW and B. Now, consider
S = B ∪ BV . This set is linearly dependent. So, obtain a linearly independent subset of S that
contains all the elements of B. Similarly, do for T = B ∪ BW .
So, we first find a basis of V ∩ W. Note that (v, w, x, y, z)T ∈ V ∩ W if v, w, x, y and z satisfy
v + x − 3y + z = 0, w − x − z = 0 and v = y. The solution of the system is given by
(v, w, x, y, z)T = (y, 2y, x, y, 2y − x)T = y(1, 2, 0, 1, 2)T + x(0, 0, 1, 0, −1)T .
Thus, B = {(1, 2, 0, 1, 2)T , (0, 0, 1, 0, −1)T } is a basis of V ∩ W. Similarly, a basis of V is

given by C = {(−1, 0, 1, 0, 0)T , (0, 1, 0, 0, 0)T , (3, 0, 0, 1, 0)T , (−1, 0, 0, 0, 1)T } and that of W is
given by D = {(1, 0, 0, 1, 0)T , (0, 1, 1, 0, 0)T , (0, 1, 0, 0, 1)T }. To find the required basis form a
matrix whose rows are the vectors in B, C and D (see below) and apply row operations other
than Eij . Then, after a few row operations, we get
T
AF
  
1 2 0 1 2 1 2 0 1 2    
1 2 0 1 2 1 2 0 1 2
DR
0 0 1 0 −1 0 0 1 0 −1
   
   

−1
   0 0 1 0 −1

0
 0 1 0 −1
0 1 0 0
 0 1 0 0 0  
→ 1 0 0 1 0  → 0 1 0 0 1 .
      
 
0 1 0 0 0
 0 0 0 1 3    
   0 1 1 0 0 0 0 0 0 0
3 0 0 1 0 0 0 0 0 0
       
0 1 0 0 1 0 0 0 0 0
−1 0 0 0 1 0 0 0 0 0
Thus, a required basis of V is {(1, 2, 0, 1, 2)T , (0, 0, 1, 0, −1)T , (0, 1, 0, 0, 0)T , (0, 0, 0, 1, 3)T }. Sim-
ilarly, a required basis of W is {(1, 2, 0, 1, 2)T , (0, 0, 1, 0, −1)T , (0, 1, 0, 0, 1)T }.
Exercise 3.5.9. 1. Give an example to show that if A and B are equivalent then Col(A)
need not equal Col(B).
2. Let V = {(x, y, z, w)T ∈ R4 | x + y − z + w = 0, x + y + z + w = 0, x + 2y = 0} and

W = {(x, y, z, w)T ∈ R4 | x − y − z + w = 0, x + 2y − w = 0} be two subspaces of R4 .
Think of a method to find bases and dimensions of V, W, V ∩ W and V + W.
3. Let W1 and W2 be two subspaces of a vector space V. If dim(W1 ) + dim(W2 ) > dim(V),
then prove that dim(W1 ∩ W2 ) ≥ 1.
3.6 Fundamental Theorem of Linear Algebra and Applications

We start with proving the rank-nullity theorem and give some of it’s consequences.
3.6. FUNDAMENTAL THEOREM OF LINEAR ALGEBRA AND APPLICATIONS 91
Theorem 3.6.1 (Rank-Nullity Theorem). Let A ∈ Mm×n (C). Then,
dim(Col(A)) + dim(Null(A)) = n. (3.6.2)
Proof. Let dim(Null(A)) = r ≤ n and let B = {u1 , . . . , ur } be a basis of Null(A). Since B is

a linearly independent set in Rn , extend it to get C = {u1 , . . . , un } as a basis of Rn . Then,
Col(A) = LS(AB) = LS(Au1 , . . . , Aun )

= LS(0, . . . , 0, Aur+1 , . . . , Aun ) = LS(Aur+1 , . . . , Aun ).
So, D = {Aur+1 , . . . , Aun } spans Col(A). We further need to show that D is linearly indepen-
dent. So, consider the homogeneous linear system given below in the unknowns α1 , . . . , αn−r .
α1 Aur+1 + · · · + αn−r Aun = 0 ⇔ A(α1 ur+1 + · · · + αn−r un ) = 0 (3.6.3)
Thus, α1 ur+1 + · · · + αn−r un ∈ Null(A) = LS(B). Therefore, there exist scalars βi , 1 ≤ i ≤ r,

n−r
P Pr
such that αi ur+i = βj uj . Or equivalently,
i=1 j=1
β1 u1 + · · · + βr ur − α1 ur+1 − · · · − αn−r un = 0. (3.6.4)
Equation (3.6.4) is a linear system in vectors from C with αi ’s and βj ’s as unknowns. As C is a

T
linearly independent set, the only solution of Equation (3.6.4) is

AF
αi = 0, for 1 ≤ i ≤ n − r and βj = 0, for 1 ≤ j ≤ r.

DR
In other words, we have shown that the only solution of Equation (3.6.3) is the trivial solution.
Hence, {Aur+1 , . . . , Aun } is a basis of Col(A). Thus, the required result follows.
Theorem 3.6.1 is part of what is known as the fundamental theorem of linear algebra (see
Theorem 3.6.5). The following are some of the consequences of the rank-nullity theorem. The
proofs are left as an exercise for the reader.
Exercise 3.6.2. 1. Let A ∈ Mm,n (R).
(a) If n > m then the system Ax = 0 has infinitely many solutions,

(b) If n < m then there exists b ∈ Rm \ {0} such that Ax = b is inconsistent.
2. The following statements are equivalent for an m × n matrix A.
(a) Rank (A) = k.

(b) There exist a set of k rows of A that are linearly independent.
(c) There exist a set of k columns of A that are linearly independent.
(d) dim(Col(A)) = k.
(e) There exists a k × k submatrix B of A with det(B) 6= 0. Further, the determinant of
every (k + 1) × (k + 1) submatrix of A is zero.
(f ) There exists a linearly independent subset {b1 , . . . , bk } of Rm such that the system
Ax = bi , for 1 ≤ i ≤ k, is consistent.
(g) dim(Null(A)) = n − k.
3. Let A ∈ Mn (R) and define a function f : Rn → Rn by f (x) = Ax. Then, the following
statements are equivalent.
(a) f is one-one.
(b) f is onto.
(c) f is invertible.
" #
1 −1
4. Let A = . Then, verify that Null(A) = Col(A). Can such examples exist in Rn
1 −1
for n odd? What about n even? Further, verify that R2 6= Null(A) + Col(A). Does it
contradict the rank-nullity theorem?
5. Determine a 2 × 2 matrix A of rank 1 such that R2 = Null(A) + Col(A).
We end this section by proving the fundamental theorem of linear algebra. We start with
the following result.
Lemma 3.6.3. Let A ∈ Mm,n (R). Then, Null(A) = Null(AT A).

T
Proof. Clearly, Null(A) ⊆ Null(AT A) as Ax = 0 implies (AT A)x = AT (Ax) = 0.

AF
So, let x ∈ Null(AT A). Then, (AT A)x = 0 implies (Ax)T (Ax) = xT AT Ax = xT 0 = 0.
DR
Thus, Ax = 0 and the required result follows.
Let u, v ∈ Rn . Then u is said to be orthogonal to v if uT v = 0 (dot product of vectors in

Rn ). Further, for S ⊆ Rn , the orthogonal complement of S, denoted S ⊥ , is defined as
S ⊥ = {x ∈ Rn : xT s = 0 for all s ∈ S}.
The readers are required to prove the following lemma.
Lemma 3.6.4. Consider the vector space Rn . Then, for S ⊆ Rn prove that
1. S ⊥ is a subspace of Rn .
2. S ⊥ = (LS(S))⊥ .
3. (S ⊥ )⊥ = S ⊥ if and only if S is a subspace of Rn .
4. Let W be a subspace of Rn . Then, there exists a subspace V of Rn such that
(a) Rn = W ⊕ V. Or equivalently, W and V are complementary subspaces.
(b) vT u = 0, for every u ∈ W and v ∈ V. This, further implies that W and V are also
orthogonal to each other. Such spaces are called orthogonal complements.
Theorem 3.6.5 (Fundamental Theorem of Linear Algebra). Let A ∈ Mm,n (R). Then,
1. dim(Null(A)) + dim(Col(A)) = n.
3.6. FUNDAMENTAL THEOREM OF LINEAR ALGEBRA AND APPLICATIONS 93
⊥ ⊥
2. Null(A) = Col(AT ) and Null(AT ) = Col(A) .
3. dim(Col(A)) = dim(Col(AT )). Or equivalently, Row-rank(A) = Column-rank(A).
Proof. Part 1: Proved in Theorem 3.6.1.

Part 2: To show: Null(A) ⊆ Col(AT )⊥ . Equivalently, need to show that for each
x ∈ Null(A) and u ∈ Col(AT ), uT x = 0. As u ∈ Col(AT ) there exists y ∈ Rm such that
u = AT y. Further, x ∈ Null(A) implies Ax = 0. Thus, we see that
T
uT x = AT y x = (yT A)x = yT (Ax) = yT 0 = 0.
Hence, Null(A) ⊆ Col(AT )⊥ .

We now show that Col(AT )⊥ ⊆ Null(A). Let z ∈ Col(AT )⊥ ⊆ Rn . Then, for every
y ∈ Rm , AT y ∈ Col(AT ) and hence (AT y)T z = 0. In particular, for y = Az ∈ Rm , we have
0 = (AT y)T z = yT Az = yT y ⇔ y = 0.
Thus Az = 0 and z ∈ Null(A). This completes the proof of the first equality in Part 2. A
similar argument gives the second equality.
Part 3: Note that, using the rank-nullity theorem we have
⊥
dim(Col(A)) = n − dim(Null(A)) = n − dim Col(AT ) = n − n − dim Col(AT ) .

T
Thus, dim(Col(A)) = dim Col(AT ) .

AF
Hence the proof of the fundamental theorem is complete.

DR
Remark 3.6.6. Let A ∈ Mm,n (R). Then, Theorem 3.6.5.2 implies the following:
⊥
1. Null(A) = Col(AT ) . This is just stating the usual fact that if x ∈ Null(A) then
Ax = 0. Hence, the dot product of every row of A with x equals 0.
2. Rn = Null(A) ⊕ Col(AT ). Further, Null(A) is orthogonal complement of Col(AT ).
3. Rm = Null(AT ) ⊕ Col(A). Further, Null(AT ) is orthogonal complement of Col(A).
As an implication of last two parts of Theorem 3.6.5, we show the existence of an invertible
function f : Col(AT ) → Col(A).
Corollary 3.6.7. Let A ∈ Mm,n (R). Then, the function f : Col(AT ) → Col(A) defined by
f (x) = Ax is invertible.
Proof. Let us first show that f is one-one. So, let x, y ∈ Col(AT ) such that f (x) = f (y).
Hence, Ax = Ay. Thus x − y ∈ Null(A) = (Col(AT ))⊥ (by Theorem 3.6.5.2). Therefore,
x − y ∈ (Col(AT ))⊥ ∩ Col(AT ) = {0}. Thus x = y and hence f is one-one.
We now show that f is onto. So, let z ∈ Col(A). To find y ∈ Col(AT ) such that f (y) = z.
As z ∈ Col(A) there exists w ∈ Rn with z = Aw. But Null(A) and Col(AT ) are
complementary subspaces and hence, there exists unique vectors, w1 ∈ Null(A) and w2 ∈
Col(AT ), such that w = w1 + w2 . Thus, z = Aw implies
z = Aw = A(w1 + w2 ) = Aw1 + Aw2 = 0 + Aw2 = Aw2 = f (w2 ),

AX
text
AX
AAtul
feelin
y T
AF
for w2 ∈ Col(AT . Thus, the required result follows.

DR
The readers should look at Example 3.5.3 and Remark 3.5.4. We give one more example.
 
1 1 0
 
Example 3.6.8. Let A =  2 1 1. Then, verify that

3 2 1
1. {(0, 1, 1)T , (1, 1, 2)T } is a basis of Col(A).
2. {(1, 1, −1)T } is a basis of Null(AT ).
3. Null(AT ) = (Col(A))⊥ .
For more information related with the fundamental theorem of linear algebra the interested
readers are advised to see the article “The Fundamental Theorem of Linear Algebra, Gilbert
Strang, The American Mathematical Monthly, Vol. 100, No. 9, Nov., 1993, pp. 848 - 855.” The
diagram 3.6 has been taken from the above paper. It also explains Corollary 3.6.7.
Exercise 3.6.9. 1. Find subspaces W1 6= {0} and W2 6= {0} in R3 such that they are
orthogonal but they are not orthogonal complement of each other.
2. Let A ∈ Mm,n (R). Prove that Col(AT ) = Col(AT A). Thus, Rank(A) = n if and only if
Rank(AT A) = n. [ Hint: Use the rank-nullity theorem and/ or Lemma 3.6.3]
3. Let A ∈ Mm,n (R). Then, for every
(a) x ∈ Rn , x = u + v, where u ∈ Col(AT ) and v ∈ Null(A) are unique.
(b) y ∈ Rm , y = w + z, where w ∈ Col(A) and z ∈ Null(AT ) are unique.
3.7. SUMMARY 95
4. Let A, B ∈ Mn (R) such that A is idempotent and AB = BA = 0. Then prove that

Col(A + B) = Col(A) + Col(B).
5. Let A ∈ Mm,n (R). Then, a matrix G ∈ Mn,m (R) is a g-inverse of A if and only if for any
b ∈ Col(A), the vector y = Gb is a solution of the system Ay = b.
6. Let A ∈ Mm,n (R). If G ∈ Mn,m (R) is a g-inverse of A then, for any b ∈ Col(A) the
solution set of the system Ay = b is given by Gb + (I − GA)z, for any arbitrary vector z.
3.7 Summary
In this chapter, we defined vector spaces over F. The set F was either R or C. To define a vector
space, we start with a non-empty set V of vectors and F the set of scalars. We also needed to
do the following:
1. first define vector addition and scalar multiplication and
2. then verify the conditions in Definition 3.1.1.
If all conditions in Definition 3.1.1 are satisfied then V is a vector space over F. If W was a
non-empty subset of a vector space V over F then for W to be a space, we only need to check
whether the vector addition and scalar multiplication inherited from that in V hold in W.
We then learnt linear combination of vectors and the linear span of vectors. It was also shown
T
that the linear span of a subset S of a vector space V is the smallest subspace of V containing
AF
S. Also, to check whether a given vector v is a linear combination of u1 , . . . , un , we needed to

DR
solve the linear system c1 u1 + · · · + cn un = v in the variables c1 , . . . , cn . Or equivalently, the

system Ax = b, where in some sense A[:, i] = ui , 1 ≤ i ≤ n, xT = [c1 , . . . , cn ] and b = v. It
was also shown that the geometrical representation of the linear span of S = {u1 , . . . , un } is
equivalent to finding conditions in the entries of b such that Ax = b was always consistent.
Then, we learnt linear independence and dependence. A set S = {u1 , . . . , un } is linearly
independent set in the vector space V over F if the homogeneous system Ax = 0 has only the
trivial solution in F. Else S is linearly dependent, whereas before the columns of A correspond
to the vectors ui ’s.
We then talked about the maximal linearly independent set (coming from the homogeneous
system) and the minimal spanning set (coming from the non-homogeneous system) and culmi-
nating in the notion of the basis of a finite dimensional vector space V over F. The following
important results were proved.
1. A linearly independent set can be extended to form a basis of V.
2. Any two bases of V have the same number of elements.
This number was defined as the dimension of V, denoted dim(V).

Now let A ∈ Mn (R). Then, combining a few results from the previous chapter, we have the
following equivalent conditions.
1. A is invertible.
3. RREF(A) = In .
7. Rank(A) = n.
8. det(A) 6= 0.
9. Col(AT ) = Row(A) = Rn .
10. Rows of A form a basis of Rn .
11. Col(A) = Rn .
12. Columns of A form a basis of Rn .
13. Null(A) = {0}.

T
AF
DR
Chapter 4
Inner Product Spaces
4.1 Definition and Basic Properties

Recall the dot product in R2 and R3 . It satisfies the following properties.
1. u · (a v + w) = a v · u + u · w, for all u, v, w ∈ R3 and a ∈ R.
2. u · v = v · u, for all u, v ∈ R3 .
3. u · u ≥ 0 for all u ∈ R3 . Further, equality holds if and only if u = 0.
The dot product helped us to compute the length of vectors and talk of perpendicularity of
T
vectors. This enabled us to rephrase geometrical problems in R2 and R3 in the language of

AF
vectors. We now generalize the idea of dot product to achieve similar goal for a general vector
DR
space over R or C.
Definition 4.1.1. Let V be a vector space over F. An inner product over V, denoted by
h , i, is a map from V × V to F satisfying
1. ha u + v, wi = a hu, wi + hv, wi, for all u, v, w ∈ V and a ∈ F,
2. hu, vi = hv, ui, the complex conjugate of hu, vi, for all u, v ∈ V and
3. hu, ui ≥ 0 for all u ∈ V. Furthermore, equality holds if and only if u = 0.
Remark 4.1.2. Using the definition of inner product, we immediately observe that
1. hu, 0i = hu, 0 + 0i = hu, 0i + hu, 0i. Thus, hu, 0i = 0, for all u ∈ V.
2. hv, α wi = hα w, vi = α hw, vi = α hv, wi, for all α ∈ F and v, w ∈ V.
3. If hu, vi = 0 for all v ∈ V then in particular hu, ui = 0. Hence u = 0.
Definition 4.1.3. Let V be a vector space with an inner product h , i. Then, (V, h , i) is called
an inner product space (in short, ips).
Example 4.1.4. Examples 1 to 4 that appear below are called the standard inner product
or the dot product. Whenever an inner product is not clearly mentioned, it will be assumed
to be the standard inner product.
97
98 CHAPTER 4. INNER PRODUCT SPACES
1. For u = (u1 , . . . , un )T , v = (v1 , . . . , vn )T ∈ Rn define hu, vi = u1 v1 + · · · + un vn = vT u.

Then, h , i is indeed an inner product and hence Rn , h , i is an ips.

2. For u = (u1 , . . . , un )T , v = (v1 , . . . , vn )T ∈ Cn define hu, vi = u1 v1 + · · · + un vn = v∗ u.

Then, Cn , h , i is an ips.

3. For A, B ∈ Mn (R), define hA, Bi = tr(BT A). Then,
hA + B, Ci = tr CT (A + B) = tr(CT A) + tr(CT B) = hA, Ci + hB, Ci and

hA, Bi = tr(BT A) = tr( (BT A)T ) = tr(AT B) = hB, Ai.

n n n
If A = [aij ] then hA, Ai = tr(AT A) = (AT A)ii = a2ij and therefore,
P P P
aij aij =
i=1 i,j=1 i,j=1
hA, Ai > 0 for all non-zero matrix A.
R1
4. Consider the real vector space C[−1, 1] and define hf, gi = f (x)g(x)dx. Then,
−1
R1
(a) hf , f i = | f (x) |2 dx ≥ 0 as | f (x) |2 ≥ 0. Further, hf , f i = 0 if and only if f ≡ 0 as
−1
f is continuous.
R1 R1 R1
(b) hg, f i = g(x)f (x)dx = g(x)f (x)dx = f (x)g(x)dx = hf , gi.
−1 −1 −1
R1 R1
(c) hf + g, hi = (f + g)(x)h(x)dx = [f (x)h(x) + g(x)h(x)]dx = hf , hi + hg, hi.
T
−1 −1
AF
R1 R1
(d) hαf , gi = (αf (x))g(x)dx = α f (x)g(x)dx = αhf , gi.
DR
−1 −1
" #
4 −1
5. For x = (x1 , x2 )T , y = (y1 , y2 )T ∈ R2 and A = , define hx, yi = yT Ax. Then,
−1 2
h , i is an inner product as hx, xi = (x1 − x2 )2 + 3x21 + x22 .
" #
a b
6. Fix A = with a, c > 0 and ac > b2 . Then, hx, yi = yT Ax is an inner product on
b c
h i2
R2 as hx, xi = ax21 + 2bx1 x2 + cx22 = a x1 + bxa2 + a1 ac − b2 x22 .

7. Verify that for x = (x1 , x2 , x3 )T , y = (y1 , y2 , y3 )T ∈ R3 , hx, yi = 10x1 y1 + 3x1 y2 + 3x2 y1 +

2x2 y2 + x2 y3 + x3 y2 + x3 y3 defines an inner product.
Exercise 4.1.5. For x = (x1 , x2 )T , y = (y1 , y2 )T ∈ R2 , we define three maps that satisfy at
least one condition out of the three conditions for an inner product. Determine the condition
which is not satisfied. Give reasons for your answer.
1. hx, yi = x1 y1 .
2. hx, yi = x21 + y12 + x22 + y22 .
3. hx, yi = x1 y13 + x2 y23 .
As hu, ui > 0, for all u 6= 0, we use inner product to define the length/ norm of a vector.
4.2. CAUCHY-SCHWARTZ INEQUALITY 99
Definition 4.1.6. Let V be an inner product space over F. Then, for any vector u ∈ V, we
p
define the length (norm) of u, denoted kuk = hu, ui, the positive square root. A vector of
u
norm 1 is called a unit vector. Thus, is called the unit vector in the direction of u.
kuk
1. Let V be an ips and u ∈ V. Then, for any scalar α, kαuk = α · kuk.

Example 4.1.7.
√ √
2. Let u = (1, −1, 2, −3)T ∈ R4 . Then, kuk = 1 + 1 + 4 + 9 = 15. Thus, √1 u and
15
− √115 u are unit vectors in the direction of u.
Exercise 4.1.8. 1. Let u = (−1, 1, 2, 3, 7)T ∈ R5 . Find all α ∈ R such that kαuk = 1.
2. Let u = (−1, 1, 2, 3, 7)T ∈ C5 . Find all α ∈ C such that kαuk = 1.
3. Let u = (1, 2)T , v = (2, −1)T ∈ R2 . Then, does there "exist an

# inner product in R such
2
a b
that kuk = 1, kvk = 1 and hu, vi = 0? [Hint: Let A = and define hx, yi = yT Ax.
b c
Use given conditions to get a linear system of 3 equations in the variables a, b, c.]
4. Prove that under the standard inner product in Mm,n (R),

m
X n
X
kAk2 = tr(AT A) = kA[k, :]T k2 = kA[:, `]k2 , for all A ∈ Mm,n (R).
k=1 `=1
4.2 Cauchy-Schwartz Inequality

T
AF
A very useful and a fundamental inequality, commonly called the Cauchy-Schwartz inequality,
DR
is a generalization of |x · y| ≤ kxk · kyk, and is proved next.
Theorem 4.2.1 (Cauchy- Schwartz inequality). Let V be an inner product space over F. Then,
for any u, v ∈ V
| hu, vi | ≤ kuk kvk. (4.2.1)
Moreover, equality holds in Inequality
(4.2.1) if and only if u and v are linearly dependent. In

u u
particular, if u 6= 0 then v = v, .
kuk kuk
Proof. If u = 0 then Inequality (4.2.1) holds. Hence, let u 6= 0. Then, by Definition 4.1.1.3,
hv, ui
hλu + v, λu + vi ≥ 0 for all λ ∈ F and v ∈ V. In particular, for λ = − , we have
kuk2
0 ≤ hλu + v, λu + vi = λλkuk2 + λhu, vi + λhv, ui + kvk2
hv, ui hv, ui 2 hv, ui hv, ui 2 2 | hv, ui |2
= kuk − hu, vi − hv, ui + kvk = kvk − .
kuk2 kuk2 kuk2 kuk2 kuk2
Or, in other words | hv, ui |2 ≤ kuk2 kvk2 and the proof of the inequality is over.
Now, note that equality holds in Inequality (4.2.1) if and only if hλu + v, λu + vi = 0, or
equivalently, λu + v = 0. Hence, u and v are linearly dependent. Moreover,
0 = h0, ui = hλu + v, ui = λhu, ui + hv, ui

hv, ui

u u
implies that v = −λu = 2
u = v, .
kuk kuk kuk
As an immediate corollary, the following hold.
Corollary 4.2.2. Prove the following results.

n 2 n n
2 2 2 yi , for all x, y ∈ Rn .
P P 2 P 2
1. xi yi = |x · y| ≤ kxk · kyk = xi
i=1 i=1 i=1
2. kAxk ≤ kAk · kxk, for all A ∈ Mm,n (R) and x ∈ Rn .
Proof. We will just prove the second part. Note that

m
X m
X n
X
2 2 2
kAxk = | (Ax)k | = | (A[k, :] x) | = |hx, A[k, :]T i|2
k=1 k=1 k=1
m
X m
X
≤ kxk2 · kA[k, :]T k2 = kxk2 kA[k, :]T k2 = kxk2 kAk2
k=1 k=1

1 1
Exercise 4.2.3. 1. Let a, b ∈ R with a, b > 0. Then, prove that (a + b) ≥ 4. In +
n n a b
P 1
general, for 1 ≤ i ≤ n, let ai ∈ R with ai > 0. Then ≥ n2 .
P
ai
i=1 i=1 ai
2. Prove that | z1 + · · · + zn | ≤ n( | z1 |2 + · · · + | zn |2 ), for z1 , . . . , zn ∈ C. When does

p
the equality hold?
3. Let V be an ips. If u, v ∈ V with kuk = 1, kvk = 1 and hu, vi = 1 then prove that u = αv
for some α ∈ F. Is α = 1?
T
AF
Let V be a real vector space. Then, for u, v ∈ V, the Cauchy-Schwartz inequality implies
hu, vi
that −1 ≤ ≤ 1. This together with the properties of the cosine function is used to
DR
kuk kvk
define the angle between two vectors in a real inner product space.
Definition 4.2.4. Let V be a real vector space. If θ ∈ [0, π] is the angle between u, v ∈ V \ {0}
then we define
hu, vi
cos θ = .
kuk kvk
Example 4.2.5. 1. Take (1, 0)T , (1, 1)T ∈ R2 . Then cos θ = √1 . So θ = π/4.
2
2. Take (1, 1, 0)T , (1, 1, 1)T ∈ R3 . Then, angle between them, say β = cos−1 √2 .
6
3. Angle depends on the IP. Take hx, yi = 2x1 y1 + x1 y2 + x2 y1 + x2 y2 on R2 . Then, angle

between (1, 0)T , (1, 1)T ∈ R2 equals cos−1 √3 .
10
4. As hx, yi = hy, xi for any real vector space, the angle between x and y is same as the
angle between y and x.
a
b
A B
c
Figure 4.1: Triangle with vertices A, B and C

4.3. NORMED LINEAR SPACE 101
We will now prove that if A, B and C are the vertices of a triangle (see Figure 4.1) and a, b
b2 + c2 − a2
and c, respectively, are the lengths of the corresponding sides then cos(A) = . This
2bc
in turn implies that the angle between vectors has been rightly defined.
Lemma 4.2.6. Let A, B and C be the vertices of a triangle (see Figure 4.1) with corresponding
side lengths a, b and c, respectively, in a real inner product space V then
b2 + c2 − a2
cos(A) = .
2bc
Proof. Let 0, u and v be the coordinates of the vertices A, B and C, respectively, of the triangle
ABC. Then, AB ~ = u, AC
~ = v and BC ~ = v − u. Thus, we need to prove that
kvk2 + kuk2 − kv − uk2

cos(A) = ⇔ kvk2 + kuk2 − kv − uk2 = 2 kvk kuk cos(A).
2kvkkuk
Now, by definition kv−uk2 = kvk2 +kuk2 −2hv, ui and hence kvk2 +kuk2 −kv−uk2 = 2 hu, vi.
As hv, ui = kvk kuk cos(A), the required result follows.
Exercise 4.2.7. Let x, y ∈ Rn then prove that
1. hx, yi = 0 ⇐⇒ kx − yk2 = kxk2 + kyk2 (Pythagoras Theorem).

Solution: Use kx − yk2 = kxk2 + kyk2 − 2hx, yi to get the required result follows.
T
2. kxk = kyk ⇐⇒ hx + y, x − yi = 0 (x and y form adjacent sides of a rhombus as the

AF
diagonals x + y and x − y are orthogonal).

DR
Solution: Use hx + y, x − yi = kxk2 − kyk2 to get the required result follows.
3. 4hx, yi = kx + yk2 − kx − yk2 (polarization identity in Rn ).

Solution: Just expand the right hand side to get the required result follows.
4. kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 (parallelogram law: the sum of squares of the
lengths of the diagonals of a parallelogram equals twice the sum of squares of the lengths
of its sides).
Solution: Just expand the left hand side to get the required result follows.
4.3 Normed Linear Space

In the last two sections, we have used the idea of inner product to define the norm/ length of
a vector. This idea was used to get the Cauchy-Schwartz inequality, the basic back ground for
defining the angle between two vectors. The question arises ‘does every norm come from an
inner product’. To understand it, we first state the properties that a norm must enjoy. We only
look at linear spaces which are vector spaces over R or C.
Definition 4.3.1. Let V be a linear space.

1. Then, a norm on V is a function f (x) = kxk from V to R such that
(a) kxk ≥ 0 for all x ∈ V and if kxk = 0 then x = 0.
(b) kαxk = | α | kxk for all α ∈ F and x ∈ V.

(c) kx + yk ≤ kxk + kyk for all x, y ∈ V (triangle inequality).
2. A linear space with a norm on it is called a normed linear space (nls).

1. Let V be an ips. Is it true that f (x) = hx, xi is a norm?
p
Remark 4.3.2.
Proof. Yes, f (x) indeed defines an inner product. The readers should verify the first two
conditions. For the third condition, using the Cauchy-Schwartz inequality, we get
f (x + y)2 = hx + y, x + yi = hx, xi + hx, yi + hy, xi + hy, yi

≤ kxk2 + kxk · kyk + kxk · kyk + kyk2 = (f (x) + f (y))2 .
p
Thus, f (x) = hx, xi is a norm, called the norm induced by the inner product h·, ·i.
2. If k · k is a norm in V then d(x, y) = kx − yk, for x, y ∈ V, defines a distance function.
Proof. To see this, note that
(a) d(x, x) = 0, for each x ∈ V.
(b) using the triangle inequality, for any z ∈ V, we have
d(x, y) = kx−yk = k (x − z)−(y − z) k ≤ k (x − z) k+k (y − z) k = d(x, z)+d(z, y).

Theorem 4.3.3. Let V be a normed linear space and x, y ∈ V. Then kxk − kyk ≤ kx − yk.

Proof. As kxk = kx − y + yk ≤ kx − yk + kyk one has kxk − kyk ≤ kx − yk. Similarly, one
obtains kyk − kxk ≤ ky − xk = kx − yk. Combining the two, the required result follows.
T
1. Let V be a complex ips. Then,

AF
Exercise 4.3.4.
4hx, yi = kx + yk2 − kx − yk2 + ikx + iyk2 − ikx − iyk2 (Polarization Identity).

DR
2. Consider the complex vector space Cn . If x, y ∈ Cn then prove that
(a) If x 6= 0 then kx + ixk2 = kxk2 + kixk2 , even though hx, ixi =

6 0.
(b) hx, yi = 0 whenever kx + yk2 = kxk2 + kyk2 and kx + iyk2 = kxk2 + kiyk2 .
3. Let A ∈ Mn (C) satisfy kAxk ≤ kxk for all x ∈ Cn . Then, prove that if α ∈ C with
| α | > 1 then A − αI is invertible.
The next result is stated without proof as the proof is beyond the scope of this book.
Theorem 4.3.5. Let k · k be a norm on a normed linear space V. Then the norm k · k is induced
by some inner product if and only if k · k satisfies the parallelogram law:
kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 .
We now define a norm which doesn’t come from an inner product.

Example 4.3.6. For x = (x1 , x2 )T ∈ R2 , we define kxk = |x1 | + |x2 |. Verify that kxk is
indeed a norm. But, for x = e1 and y = e2 , 2(kxk2 + kyk2 ) = 4 whereas
kx + yk2 + kx − yk2 = k(1, 1)T k2 + k(1, −1)T k2 = (|1| + |1|)2 + (|1| + | − 1|)2 = 8.
So the parallelogram law fails. Thus, kxk is not induced by any inner product in R2 .
Exercise 4.3.7. Does there exist an inner product in R2 such that kxk = max{|x1 |, |x2 |}?
4.4. ORTHOGONALITY IN INNER PRODUCT SPACE 103
4.4 Orthogonality in Inner Product Space
We come back to the study of Inner product spaces the topic which is a building block for most
of the applications. To start with, we give the definition of orthogonality of two vectors.
Definition 4.4.1. Let V be an inner product space over F. Then,
1. the vectors u, v ∈ V are called orthogonal/perpendicular if hu, vi = 0.
2. Let S ⊆ V. Then, the orthogonal complement of S in V, denoted S ⊥ , equals
S ⊥ = {v ∈ V : hv, wi = 0, for all w ∈ S}.
Example 4.4.2. 1. 0 is orthogonal to every vector as h0, xi = 0 for all x ∈ V.
2. If V is a vector space over R or C then 0 is the only vector that is orthogonal to itself.
3. Let V = R.
(a) If S = {0} then, S ⊥ = R.
(b) If S = R then, S ⊥ = {0}.
(c) Let S be any subset of R containing a non-zero real number. Then S ⊥ = {0}.
T
AF
4. Let u = (1, 2)T . What is u⊥ in R2 ?

DR
Solution: By definition, u⊥ = {(x, y)T ∈ R2 | x + 2y = 0}. Thus, u⊥ is the Null(u).

Note that u⊥ = LS (2, −1)T . Further, observe that for any vector x ∈ R2 ,

2x1 − x2

u u x1 + 2x2
x = hx, ui 2
+ x − hx, ui 2
= (1, 2)T + (2, −1)T
kuk kuk 5 5
is a decomposition of x into two vectors, one parallel to u and the other parallel to u⊥ .
5. Fix u = (1, 1, 1, 1)T , v = (1, 1, −1, 0)T ∈ R4 . Determine z, w ∈ R4 such that u = z + w

with the condition that z is parallel to v and w is orthogonal to v.
Solution: As z is parallel to v, z = kv = (k, k, −k, 0)T , for some k ∈ R. Since w is
orthogonal to v the vector w = (a, b, c, d)T satisfies a + b − c = 0. Thus, c = a + b and
(1, 1, 1, 1)T = u = z + w = (k, k, −k, 0)T + (a, b, a + b, d)T .
Comparing the corresponding coordinates, gives the linear system d = 1, a + k = 1,

b + k = 1 and a + b − k = 1 in the unknowns a, b, d and k. Thus, solving for a, b, d and k
1 1
gives z = (1, 1, −1, 0)T and w = (2, 2, 4, 3)T .
3 3
6. Apollonius’ Identity: Let the length of the sides of a triangle be a, b, c ∈ R and that of
the median be d ∈ R. If the median is drawn on the side with length a then prove that
a 2
b2 + c2 = 2 d2 + .
2
7. Let P = (1, 1, 1)T , Q = (2, 1, 3)T and R = (−1, 1, 2)T be three vertices of a triangle in R3 .
Compute the angle between the sides P Q and P R.
Solution: Method 1: Note that P~Q = (2, 1, 3)T − (1, 1, 1)T = (1, 0, 2)T , P~R =
(−2, 0, 1)T and RQ~ = (−3, 0, −1)T . As hP~Q, P~Ri = 0, the angle between the sides
π
P Q and P R is .
2
√ √ √
Method 2: kP Qk = 5, kP Rk = 5 and kQRk = 10. As kQRk2 = kP Qk2 + kP Rk2 ,
π
by Pythagoras theorem, the angle between the sides P Q and P R is .
2
Exercise 4.4.3. 1. Let V be an ips.
(a) If S ⊆ V then S ⊥ is a subspace of V and S ⊥ = (LS(S))⊥ .
(b) Furthermore, if V is finite dimensional then S ⊥ and LS(S) are complementary.
Thus, V = LS(S) ⊕ S ⊥ . Equivalently, hu, wi = 0, for all u ∈ LS(S) and w ∈ S ⊥ .
2. Find v, w ∈ R3 such that v, w and (1, −1, −2)T are mutually orthogonal.
3. Let W = {(x, y, z, w)T ∈ R4 : x + y + z − w = 0}. Find a basis of W⊥ .
4. Determine W⊥ , where W = {A ∈ Mn (R) | AT = A}.
5. Consider R3 with the standard inner product. Find

(a) S ⊥ for S = {(1, 1, 1)T , (0, 1, −1)T }.
T
AF
(b) k such that cos−1 (hu, vi) = π/3, where u = (1, −1, 1)T and v = (1, k, 1)T .
(c) vectors v, w ∈ R3 such that v, w and u = (1, 1, 1)T are mutually orthogonal.
DR
6. Consider R3 with the standard inner product. Find the plane containing
(a) (1, 1 − 1) with (a, b, c) 6= 0 as the normal vector.
(b) (2, −2, 1)T and perpendicular to the line ` = {(t − 1, 3t + 2, t + 1) : t ∈ R}.
(c) the lines (1, 2, −2) + t(1, 1, 0) and (1, 2, −2) + t(0, 1, 2).
(d) (1, 1, 2)T and orthogonal to the line `{(2 + t, 3, 1 − t) : t ∈ R}.
7. Let P = (3, 0, 2)T , Q = (1, 2, −1)T and R = (2, −1, 1)T be three points in R3 . Then,
(a) find the area of the triangle with vertices P, Q and R.
(b) find the area of the parallelogram built on vectors P~Q and QR.
~
(c) find a non-zero vector orthogonal to the plane of the above triangle.
√
(d) find all vectors x orthogonal to P~Q and QR~ with kxk = 2.
(e) the volume of the parallelepiped built on vectors P~Q and QR
~ and x, where x is one
of the vectors found in Part (d). Do you think the volume would be different if you
choose the other vector x?
8. Let p1 be a plane containing the point A = (1, 2, 3)T and the vector (2, −1, 1)T as its
normal. Then,
(a) find the equation of the plane p2 that is parallel to p1 and contains (−1, 2, −3)T .
(b) calculate the distance between the planes p1 and p2 .
9. In the parallelogram ABCD, ABkDC and ADkBC and A = (−2, 1, 3)T , B = (−1, 2, 2)T
and C = (−3, 1, 5)T . Find the
(a) coordinates of the point D,

(b) cosine of the angle BCD.
(c) area of the triangle ABC
(d) volume of the parallelepiped determined by AB, AD and (0, 0, −7)T .
4.4.1 Properties of Orthonormal Vectors
We start with the definition of an orthonormal set.
Definition 4.4.4. Let V be an ips. Then, a non-empty set S = {v1 , . . . , vn } ⊆ V is called an

orthogonal set if vi and vj are mutually orthogonal, for 1 ≤ i 6= j ≤ n, i.e.,
hui , uj i = 0, for 1 ≤ i < j ≤ n.
Further, if kvi k = 1, for 1 ≤ i ≤ n, Then S is called an orthonormal set. If S is also a

basis of V then S is called an orthonormal basis of V.
T
Example 4.4.5. 1. A few orthonormal sets in R2 are

AF
1 1 1 1
DR
(1, 0)T , (0, 1)T , √ (1, 1)T , √ (1, −1)T and √ (2, 1)T , √ (1, −2)T .

2 2 5 5
2. Let S = {e1 , . . . , en } be the standard basis of Rn . Then, S is an orthonormal set as
(a) kei k = 1, for 1 ≤ i ≤ n.

(b) hei , ej i = 0, for 1 ≤ i 6= j ≤ n.
h iT h iT h iT
1 1 1 1 1 2 1 1
3. The set √ , −√ , √
3 3 3
, 0, √2 , √2 , √6 , √6 , − √6 is an orthonormal set in R3 .
Rπ
4. Recall that hf (x), g(x)i = f (x)g(x)dx defines the standard inner product in C[−π, π].
−π
Consider S = {1} ∪ {em | m ≥ 1} ∪ {fn | n ≥ 1}, where 1(x) = 1, em (x) = cos(mx) and
fn (x) = sin(nx), for all m, n ≥ 1 and for all x ∈ [−π, π]. Then,
(a) S is a linearly independent set.
(b) k1k2 = 2π, kem k2 = π and kfn k2 = π.
(c) the functions in S are orthogonal.

1 1 1
Hence, √ ∪ √ em | m ≥ 1 ∪ √ fn | n ≥ 1 is an orthonormal set in C[−π, π].
2π π π
We now prove the most important initial result of this section.
Theorem 4.4.6. Let S = {u1 , . . . , un } be an orthonormal subset of an ips V(F).

1. Then S is a linearly independent subset of V.

n
αi ui , for some αi ’s in F. Then,
P
2. Suppose v ∈ LS(S) with v =
i=1
(a) αi = hv, ui i.
n n
(b) kvk2 = k αi ui k2 = | αi |2 .
P P
i=1 i=1
n
3. Let z ∈ V and y =
P
hz, ui iui .
i=1
(a) Then z = y + (z − y) with hz − y, yi = 0, i.e., z − y ∈ LS(S)⊥ .
(b) Pythagoras Theorem: kzk2 = kyk2 + kz − yk2 .
(c) Thus, y is the nearest vector in LS(S). That is, if w ∈ LS(S) with w 6= y then
kz − wk2 = kz − y + y − wk2 = kz − yk2 + ky − wk2 > kz − yk2 .
4. Let dim(V) = n. Then hv, ui i = 0 for all i = 1, 2, . . . , n if and only if v = 0.
Proof. Part 1: Consider the linear system c1 u1 + · · · + cn un = 0 in the unknowns c1 , . . . , cn . As

h0, ui = 0 and huj , ui i = 0, for all j 6= i, we have
n
X
0 = h0, ui i = hc1 u1 + · · · + cn un , ui i = cj huj , ui i = ci hui , ui i = ci .
j=1
T
AF
Hence ci = 0, for 1 ≤ i ≤ n. Thus, the above linear system has only the trivial solution. So,
the set S is linearly independent.
DR
n
P n
P
Part 2: Note that hv, ui i = h αj uj , ui i = αj huj , ui i = αi hui , ui i = αi . This completes
j=1 j=1
Sub-part (a). For Sub-part (b), we have
n
* n n
+ n
* n
+
X X X X X
2 2
kvk = k α i ui k = αi ui , αi ui = αi ui , αj uj
i=1 i=1 i=1 i=1 j=1
n
X n
X n
X n
X
= αi αj hui , uj i = αi αi hui , ui i = | αi |2 .
i=1 j=1 i=1 i=1
Part 3: Note that for 1 ≤ i ≤ n,

* n +
X
hz − y, ui i = hz, ui i − hy, ui i = hz, ui i − hz, uj iuj , ui
j=1
n
X
= hz, ui i − hz, uj ihuj , ui i = hz, ui i − hz, ui i = 0.
j=1
So, z − y ∈ LS(S)⊥ . As y ∈ LS(S), hz − y, yi = 0 and
kzk2 = ky + (z − y)k2 = kyk2 + kz − yk2 ≥ kyk2 .
Further, w, y ∈ LS(S) implies w − y ∈ LS(S). Hence hz − y, w − yi = 0 and
kz − wk2 = kz − y + y − wk2 = kz − yk2 + ky − wk2 > kz − yk2 .

Part 4: Follows directly using Part 2b as {u1 , . . . , un } is a basis of V.
A rephrasing of Theorem 4.4.6.2b gives a generalization of the Pythagoras theorem, popu-

larly known as the Parseval’s formula. The proof is left as an exercise for the reader.
Theorem 4.4.7. Let V be an with an orthonormal basis {v1 , · · · , vn }. Then

n
X
hx, yi = hx, vi ihy, vi i, for each x, y ∈ V.
i=1
n
Furthermore, if x = y then kxk2 = | hx, vi i |2 (generalizing the Pythagoras Theorem).
P
i=1
We have another corollary of Theorem 4.4.6 which talks about an orthogonal set.
Theorem 4.4.8 (Bessel’s Inequality). Let V be an ips with {v1 , · · · , vn } as an orthogonal set.
n | hz, v i |2 n hz, v i
k k
Then, for each z ∈ V, 2 . Equality holds if and only if z =
P P
2
≤ kzk v .
2 k
k=1 kvk k k=1 kvk k
vk
Proof. For 1 ≤ k ≤ n, define uk = and use Theorem 4.4.6.4 to get the required result.
kvk k

Remark 4.4.9. Using Theorem 4.4.6, we see that if B = v1 , . . . , vn is an ordered orthonormal
basis of an ips V then
 
hu, v1 i
T
 . 
AF
. 
 . ,
[u]B =  for each u ∈ V.
DR
hu, vn i
Thus, to get the coordinates of a vector with respect to an orthonormal ordered basis, we just
need to compute the inner product with basis vectors.
To proceed further with the applications of the above ideas, we pose a question for better
understanding.
Example 4.4.10. Which point on the plane P is closest to the point, say Q?
P lane − P
0 y
Solution: Let y be the foot of the perpendicular from Q on P . Thus, by Pythagoras

Theorem (see Theorem 4.4.6.3c), y is unique. So, the question arises: how do we find y?
−→ →
− −→
Note that yQ gives a normal vector of the plane P . Hence, Q = y + yQ. So, need to
→
−
decompose Q into two vectors such that one of them lies on the plane P and the other is
orthogonal to the plane.
Thus, we see that given u, v ∈ V \ {0}, we need to find two vectors, say y and z, such that
y is parallel to u and z is perpendicular to u. Thus, y = u cos(θ) and z = u sin(θ), where θ is
the angle between u and v.
R v P
⃗ =v− ⟨v,u⟩ ⃗ =
OQ ⟨v,u⟩
u
OR ∥u∥2
u ∥u∥2
u
Q
O θ
Figure 4.2: Decomposition of vector v
u
We do this as follows (see Figure 4.2). Let û = be the unit vector in the direction
kuk
~
of u. Then, using trigonometry, cos(θ) = kOQk ~
. Hence kOQk = kOP~ k cos(θ). Now using
~ k
kOP
~ hv,ui hv,ui
Definition 4.2.4, kOQk = kvk kvk kuk = kuk , where the absolute value is taken as the
length/norm is a positive quantity. Thus,

~ = kOQk
~ û = u u
OQ v, .
kuk kuk

Hence, y = OQ ~ = v, u u
and z = v − v,
u u
. In literature, the vector y = OQ ~
kuk kuk kuk kuk
is called the orthogonal projection of v on u, denoted Proju (v). Thus,

hv, ui

u u ~
Proju (v) = v, and kProju (v)k = kOQk = . (4.4.2)
kuk kuk kuk
T
AF

Moreover, the distance of u from the point P equals kORk~ = kP~Qk = u
v − hv, kuk u
i kuk .
DR
Example 4.4.11. 1. Determine the foot of the perpendicular from the point (1, 2, 3) on the
XY -plane.
Solution: Verify that the required point is (1, 2, 0)?
2. Determine the foot of the perpendicular from the point Q = (1, 2, 3, 4) on the plane
generated by (1, 1, 0, 0), (1, 0, 1, 0) and (0, 1, 1, 1).
Answer: (x, y, z, w) lies on the plane x−y −z +2w = 0 ⇔ h(1, −1, −1, 2), (x, y, z, w)i = 0.
So, the required point equals
1 1
(1, 2, 3, 4) − h(1, 2, 3, 4), √ (1, −1, −1, 2)i √ (1, −1, −1, 2)
7 7
4 1
= (1, 2, 3, 4) − (1, −1, −1, 2) = (3, 18, 25, 20).
7 7
3. Determine the projection of v = (1, 1, 1, 1)T on u = (1, 1, −1, 0)T .

u 1 T
Solution: By Equation (4.4.2), we have Projv (u) = hv, ui = 3 (1, 1, −1, 0) and
kuk2
w = (1, 1, 1, 1)T − Projv (u) = 31 (2, 2, 4, 3)T is orthogonal to u.
4. Let u = (1, 1, 1, 1)T , v = (1, 1, −1, 0)T , w = (1, 1, 0, −1)T ∈ R4 . Write v = v1 + v2 , where
v1 is parallel to u and v2 is orthogonal to u. Also, write w = w1 + w2 + w3 such that
w1 is parallel to u, w2 is parallel to v2 and w3 is orthogonal to both u and v2 .
Solution: Note that
4.5. GRAM-SCHMIDT ORTHONORMALIZATION PROCESS 109
u 1 1 T is parallel to u.
(a) v1 = Proju (v) = hv, ui kuk2 = 4 u = 4 (1, 1, 1, 1)
(b) v2 = v − 14 u = 41 (3, 3, −5, −1)T is orthogonal to u.
Note that Proju (w) is parallel to u and Projv2 (w) is parallel to v2 . Hence, we have
u 1 1 T is parallel to u,
(a) w1 = Proju (w) = hw, ui kuk2 = 4 u = 4 (1, 1, 1, 1)
(b) w2 = Projv2 (w) = hw, v2 i kvv22k2 = 7

44 (3, 3, −5, −1)
T is parallel to v2 and
3 T
(c) w3 = w − w1 − w2 = 11 (1, 1, 2, −4) is orthogonal to both u and v2 .
4.5 Gram-Schmidt Orthonormalization Process

In view of the importance of Theorem 4.4.6, we inquire into the question of extracting an
orthonormal basis from a given basis. The process of extracting an orthonormal basis from a
finite linearly independent set is called the Gram-Schmidt Orthonormalization process.
We first consider a few examples. Note that Theorem 4.4.6 also gives us an algorithm for doing
so, i.e., from the given vector subtract all the orthogonal projections/components. If the new
vector is nonzero then this vector is orthogonal to the previous ones. The proof follows directly
from Theorem 4.4.6 but we give it again for the sake of completeness.
Theorem 4.5.1 (Gram-Schmidt Orthogonalization Process). Let V be an ips. If {v1 , . . . , vn }

T
is a set of linearly independent vectors in V then there exists an orthonormal set {w1 , . . . , wn }
AF
in V. Furthermore, LS(w1 , . . . , wi ) = LS(v1 , . . . , vi ), for 1 ≤ i ≤ n.

DR
Proof. Note that for orthonormality, we need kwi k = 1, for 1 ≤ i ≤ n and hwi , wj i = 0, for
1 ≤ i 6= j ≤ n. Also, by Corollary 3.3.11.2, vi ∈
/ LS(v1 , . . . , vi−1 ), for 2 ≤ i ≤ n, as {v1 , . . . , vn }
is a linearly independent set. We are now ready to prove the result by induction.
v1
Step 1: Define w1 = then LS(v1 ) = LS(w1 ).
kv1 k
u2
Step 2: Define u2 = v2 − hv2 , w1 iw1 . Then, u2 6= 0 as v2 6∈ LS(v1 ). So, let w2 = .
ku2 k
Note that {w1 , w2 } is orthonormal and LS(w1 , w2 ) = LS(v1 , v2 ).
Step 3: For induction, assume that we have obtained an orthonormal set {w1 , . . . , wk−1 } such
that LS(v1 , . . . , vk−1 ) = LS(w1 , . . . , wk−1 ). Now, note that
k−1
P k−1
P
uk = v k − hvk , wi iwi = vk − Projwi (vk ) 6= 0 as vk ∈
/ LS(v1 , . . . , vk−1 ). So, let us put
i=1 i=1
uk
wk = . Then, {w1 , . . . , wk } is orthonormal as kwk k = 1 and
kuk k
k−1
X k−1
X
kuk khwk , w1 i = huk , w1 i = hvk − hvk , wi iwi , w1 i = hvk , w1 i − h hvk , wi iwi , w1 i
i=1 i=1
k−1
X
= hvk , w1 i − hvk , wi ihwi , w1 i = hvk , w1 i − hvk , w1 i = 0.
i=1
Similarly, hwk , wi i = 0, for 2 ≤ i ≤ k − 1. Clearly, wk = uk /kuk k ∈ LS(w1 , . . . , wk−1 , vk ). So,

wk ∈ LS(v1 , . . . , vk ).
k−1
P
As vk = kuk kwk + hvk , wi iwi , we get vk ∈ LS(w1 , . . . , wk ). Hence, by the principle of
i=1
mathematical induction LS(w1 , . . . , wk ) = LS(v1 , . . . , vk ) and the required result follows.
We now illustrate the Gram-Schmidt process with a few examples.
Example 4.5.2. 1. Let S = {(1, −1, 1, 1), (1, 0, 1, 0), (0, 1, 0, 1)} ⊆ R4 . Find an orthonormal
set T such that LS(S) = LS(T ).
Solution: As we just require LS(S) = LS(T ), we can order the vectors as per our
convenience. So, let v1 = (1, 0, 1, 0)T , v2 = (0, 1, 0, 1)T and v3 = (1, −1, 1, 1)T . Then,
w1 = √1 (1, 0, 1, 0)T . As hv2 , w1 i = 0, we get w2 = √1 (0, 1, 0, 1)T . For the third vector,
2 2
let u3 = v3 − hv3 , w1 iw1 − hv3 , w2 iw2 = (0, −1, 0, 1)T . Thus, w3 = √12 (0, −1, 0, 1)T .
h iT h iT h iT h iT
2. Let S = v1 = 2 0 0 , v2 = 3 , v3 = 1 3 , v4 = 1 1 1 . Find
2 2 0 2 2 0
an orthonormal set T such that LS(S) = LS(T ).
h iT
Solution: Take w1 = kvv11 k = 1 0 0 = e1 . For the second vector, consider u2 =
h iT h iT
3 u2
v2 − 2 w1 = 0 2 0 . So, put w2 = ku2 k = 0 1 0 = e2 .
2
hv3 , wi iwi = (0, 0, 0)T . So, v3 ∈ LS((w1 , w2 )). Or
P
For the third vector, let u3 = v3 −
i=1
equivalently, the set {v1 , v2 , v3 } is linearly dependent.
2
T
P
So, for again computing the third vector, define u4 = v4 − hv4 , wi iwi . Then, u4 =
AF
i=1
v4 − w1 − w2 = e3 . So w4 = e3 . Hence, T = {w1 , w2 , w4 } = {e1 , e2 , e3 }.
DR
3. Find an orthonormal set in R3 containing (1, 2, 1)T .

Solution: Let (x, y, z)T ∈ R3 with (1, 2, 1), (x, y, z) = 0. Thus,

(x, y, z) = (−2y − z, y, z) = y(−2, 1, 0) + z(−1, 0, 1).
Observe that (−2, 1, 0) and (−1, 0, 1) are orthogonal to (1, 2, 1) but are themselves not
orthogonal.
Method 1: Apply Gram-Schmidt process to { √16 (1, 2, 1)T , (−2, 1, 0)T , (−1, 0, 1)T } ⊆ R3 .
Method 2: Valid only in R3 using the cross product of two vectors.
−1
In either case, verify that { √16 (1, 2, 1), √ 5
(2, −1, 0), √−1
30
(1, 2, −5)} is the required set.
We now state the following result without proof.
Corollary 4.5.3. Let V 6= {0} be a finite dimensional ips. Then

1. V has an orthonormal basis.
2. any linearly independent set S in V can be extended to form an orthonormal basis of V.
Remark 4.5.4. Let S = {v1 , . . . , vn } =

6 {0} be a non-empty subset of a finite dimensional
vector space V. Then, we observe the following.
1. If S is linearly independent then we obtain an orthonormal basis {w1 , . . . , wn } of LS(S).
4.6. QR DECOMPOSITION 111
2. If S is linearly dependent then as in Example 4.5.2.2, there will be stages at which the
vector uk = 0. Thus, we will obtain an orthonormal basis {w1 , . . . , wm } of LS(S), but
note that m < n.
3. a re-arrangement of elements of S then we may obtain another orthonormal basis of
LS(v1 , . . . , vn ). But, observe that the size of the two bases will be the same.
Exercise 4.5.5. 1. Let (V, h , i) be an n-dimensional ips. If u ∈ V with kuk = 1 then give
reasons for the following statements.
(a) Let S ⊥ = {v ∈ V | hv, ui = 0}. Then, dim(S ⊥ ) = n − 1.

(b) Let 0 6= β ∈ F. Then S = {v ∈ V : hv, ui = β} is not a subspace of V.
(c) Let v ∈ V. Then v = v0 + hv, uiu for a vector v0 ∈ S ⊥ . Thus V = LS(u, S ⊥ ).
2. Let V be an ips with B = {v1 , . . . , vn } as a basis. Then, prove that B is orthonormal if

n
and only if for each x ∈ V, x = hx, vi ivi . [Hint: Since B is a basis, each x ∈ V has a
P
i=1
unique linear combination in terms of vi ’s.]
3. Let S be a subset of V having 101 elements. Suppose that the application of the Gram-
Schmidt process yields u5 = 0. Does it imply that LS(v1 , . . . , v5 ) = LS(v1 , . . . , v4 )? Give
reasons for your answer.
k
4. Let B = {v1 , . . . , vn } be an orthonormal set in Rn . For 1 ≤ k ≤ n, define Ak = vi viT .
P
i=1
T
Then prove that ATk = Ak and A2k = Ak . Thus, Ak ’s are projection matrices. Further,
AF
show that Rank(Ak ) = k.

DR
5. Determine an orthonormal basis of R4 containing (1, −2, 1, 3)T and (2, 1, −3, 1)T .
6. Let x ∈ Rn with kxk = 1.
(a) Then prove that {x} can be extended to form an orthonormal basis of Rn .
(b) Let the extended basis be {x,x2 , . . . , xn } and B = [e 1 , . . . , en ] the standard ordered
basis of Rn . Prove that A = [x]B , [x2 ]B , . . . , [xn ]B is an orthogonal matrix.
7. Let v, w ∈ Rn , n ≥ 1 with kuk = kwk = 1. Prove that there exists an orthogonal matrix
P such that P v = w. Prove also that A can be chosen such that det(P ) = 1.
4.6 QR Decomposition
In this section, we study the QR-decomposition of a matrix A ∈ Mn (R). The decomposition
is obtained by applying the Gram-Schmidt Orthogonalization process to the columns of the
matrix A. Thus, the set {A[:, 1], . . . , A[:, n]} of the columns of A are taken as the collection of
vectors {v1 , . . . , vn }.
If Rank(A) = n then the columns of A are linearly independent and the application of
the
h Gram-Schmidt
i process gives us vectors {w1 , . . . , wn } ⊆ Rn such that the matrix Q =
w1 · · · wn is an orthogonal matrix. Further the condition
LS(A[:, 1], . . . , A[:, k]) = LS(w1 , . . . , wk ), for 1 ≤ k ≤ n,

in the Gram-Schmidt process implies that A[:, k] = LS(w1 , . . . , wk ), for 1 ≤ k ≤ n. Hence,

 
α1k
 . 
there exist αjk ∈ R, 1 ≤ j ≤ k, such that A[:, k] = [w1 , . . . , wk ] . 
 . . Thus A = QR, where
αkk
 
α11 α12 · · · α1n
 
h i  0 α22 · · · α2n 
Q = w1 · · · wn and R =  . .
 
 .. .. .. .. 
 . . . 

0 0 · · · αnn
This decomposition is stated next.
Theorem 4.6.1 (QR Decomposition). Let A ∈ Mn (R) be a matrix with Rank(A) = n. Then,
there exist matrices Q and R such that Q is orthogonal and R is upper triangular with A = QR.
Furthermore, the diagonal entries of R can be chosen to be positive. Also, in this case, the
decomposition is unique.
Proof. The argument before the statement of the theorem gives us A = QR, with
1. Q being an orthogonal matrix (see Exercise 5.8.8.5) and
2. R being an upper triangular matrix.
Thus, this completes the proof of the first part. Note that
T
AF
1. αii 6= 0, for 1 ≤ i ≤ n, as A[:, 1] 6= 0 and A[:, i] ∈

/ LS(w1 , . . . , wi−1 ) as A has full column
rank.
DR
2. if αii < 0, for some i, 1 ≤ i ≤ n then we can replace vi in Q by −vi to get new matrices
Q and R with the added condition that the diagonal entries of R are positive.
Uniqueness: Suppose Q1 R1 = Q2 R2 for some orthogonal matrices Qi ’s and upper triangular

matrices Ri ’s with positive diagonal entries. As Qi ’s and Ri ’s are invertible, we get Q−1
2 Q1 =
R2 R1−1 . As product of upper triangular matrices is also upper triangular (see Exercise 2) the
matrix R2 R1−1 is an upper triangular matrix. Similarly, Q−1
2 Q1 is an orthogonal matrix.
So, the matrix R2 R1−1 is an orthogonal upper triangular matrix. Hence R2 R1−1 = In . So,
R2 = R1 and therefore Q2 = Q1 .
Remark 4.6.2. Note that in the proof of Theorem 4.6.1, we just used the idea that A[:, i] ∈
LS(w1 , . . . , wi ) to get the scalars αji , for 1 ≤ j ≤ i. As {w1 , . . . , wi } is an orthonormal set
αji = hA[:, i], wj i, for 1 ≤ j ≤ i.
So, it is quite easy to compute the entries of the upper triangular matrix R.
Now, let A be an m × n matrix with Rank(A) = r. Then, by Remark 4.5.4, we obtain an

orthonormal set {w1 , . . . , wr } ⊆ Rn such that
LS(A[:, 1], . . . , A[:, j]) = LS(w1 , . . . , wi ), for 1 ≤ i ≤ j ≤ n.
Hence, proceeding on the lines of the above theorem, one has the following result.
4.6. QR DECOMPOSITION 113
Theorem 4.6.3 (Generalized QR Decomposition). Let A be an m×n matrix with Rank(A) = r.

Then A = QR, where
1. Q = [w1 , . . . , wr ] is an m × r matrix with QT Q = Ir ,
2. LS(A[:, 1], . . . , A[:, j]) = LS(v1 , . . . , vi ), for 1 ≤ i ≤ j ≤ n and
3. R is an r × n matrix with Rank(R) = r.

We look at a few examples to understand it better.
 
1 0 1 2

0 1 −1 1

Example 4.6.4. 1. Let A = 
1 0 1 1. Find an orthogonal matrix Q and an upper

 
0 1 1 1
triangular matrix R such that A = QR.
Solution: From Example 4.5.2, we know that w1 = √1 (1, 0, 1, 0)T , w2 = √1 (0, 1, 0, 1)T
2 2
and w3 = √1 (0, −1, 0, 1)T . We now compute w4 . If v4 = (2, 1, 1, 1)T then
2
1
u4 = v4 − hv4 , w1 iw1 − hv4 , w2 iw2 − hv4 , w3 iw3 = (1, 0, −1, 0)T .
2
1 T
Thus, w4 = √2 (−1, 0, 1, 0) . Hence, we see that A = QR with
  √ √
√1 √1 2 − √32

2
0 0 2
2 0

1 −1
 √ √ 
 0 √2 √2 0 

   0 2 0 − 2
Q = w1 , . . . , w4 =  1 and R = √ .
T

−1 

√ 0 0 √  0 0 2 0 
AF
 2 2  
1 1 1
0 √ √ 0 0 0 0 √
2
2 2
DR
 
1 1 1 0

−1 0 −2 1

2. Let A =   . Find a 4 × 3 matrix Q satisfying QT Q = I3 and an upper
 1 1 1 0 

1 0 2 1
triangular matrix R such that A = QR.
Solution: Let us apply the Gram-Schmidt orthonormalization process to the columns of
A. As v1 = (1, −1, 1, 1)T , we get w1 = 21 v1 . Let v2 = (1, 0, 1, 0)T . Then,
1
u2 = v2 − hv2 , w1 iw1 = (1, 0, 1, 0)T − w1 = (1, 1, 1, −1)T .
2
Hence, w2 = 12 (1, 1, 1, −1)T . Let v3 = (1, −2, 1, 2)T . Then,
u3 = v3 − hv3 , w1 iw1 − hv3 , w2 iw2 = v3 − 3w1 + w2 = 0.
So, we again take v3 = (0, 1, 0, 1)T . Then,
u3 = v3 − hv3 , w1 iw1 − hv3 , w2 iw2 = v3 − 0w1 − 0w2 = v3 .
So, w3 = √1 (0, 1, 0, 1)T . Hence,

2
1 1
 
2 2 0  
 −1 1 2 1 3 0
√1 

 2 2 2  and R = 
 
Q = [v1 , v2 , v3 ] = 
 1 1
0 1 −1 0  .
 2 2 0   √ 
1 −1 1
0 0 0 2
√
2 2 2
The readers are advised to check the following:
(a) Rank(A) = 3,
(b) A = QR with QT Q = I3 , and
(c) R is a 3 × 4 upper triangular matrix with Rank(R) = 3.
Remark 4.6.5. Let A ∈ Mm,n (R) with Rank(A) = n .
1. Then AT A is invertible (see Exercise 3.6.9.3).
2. By Theorem 4.6.3, there exist matrices Q ∈ Mm,n (R) and R ∈ Mn,n (R) such that A = QR.
3. Further, the columns of Q form an orthonormal set and hence QT Q = In .
4. Furthermore, Rank(R) = n as Rank(A) = n. Thus R is invertible. Hence RT R is

invertible and (RT R)−1 = R−1 (RT )−1 .
5. So, if Q = [w1 , . . . , wn ] then
A(AT A)−1 AT = QR(RT R)−1 RT QT = (QR)(R−1 (RT )−1 )RT QT = QQT .
 
w1T
 .  n
6. Thus P = A(AT A)−1 AT = QQT = [w1 , . . . , wn ] ..  = wi wiT is the projection
P
 
i=1
T
wnT
AF
matrix that projects on Col(A) (see Exercise 4.5.5.4).

DR
4.7 Summary
In the previous chapter, we learnt that if V is vector space over F with dim(V) = n then V
basically looks like Fn . Also, any subspace of Fn is either Col(A) or Null(A) or both, for some
matrix A with entries from F.
So, we started this chapter with inner product, a generalization of the dot product in R3
or Rn . We used the inner product to define the length/norm of a vector. The norm has the
property that “the norm of a vector is zero if and only if the vector itself is the zero vector”. We
then proved the Cauchy-Schwartz Inequality which helped us in defining the angle between two
vector. Thus, one can talk of geometrical problems in Rn and proved some geometrical results.
We then independently defined the notion of a norm in Rn and showed that a norm is
induced by an inner product if and only if the norm satisfies the parallelogram law (sum of
squares of the diagonal equals twice the sum of square of the two non-parallel sides).
The next subsection dealt with the fundamental theorem of linear algebra where we showed
that if A ∈ Mm,n (C) then
1. dim(Null(A)) + dim(Col(A)) = n.
⊥ ⊥
2. Null(A) = Col(A∗ ) and Null(A∗ ) = Col(A) .
3. dim(Col(A)) = dim(Col(A∗ )).

4.7. SUMMARY 115
We then saw that having an orthonormal basis is an asset as determining the coordinates
of a vector boils down to computing the inner product.
So, the question arises, how do we compute an orthonormal basis? This is where we came
across the Gram-Schmidt Orthonormalization process. This algorithm helps us to determine
an orthonormal basis of LS(S) for any finite subset S of a vector space. This also lead to the
QR-decomposition of a matrix.
T
AF
DR
T
AF
DR
Chapter 5
Linear Transformations
5.1 Definitions and Basic Properties

Recall that understanding functions, their domain, co-domain and their properties, such as
one-one, onto etc. played an important role. So, in this chapter, we study functions over vector
spaces that preserve the operations of vector addition and scalar multiplication.
Definition 5.1.1. Let V and W be vector spaces over F with vector operations +, · in V and
⊕, in W. A function (map) f : V → W is called a linear transformation if for all α ∈ F
and u, v ∈ V the function f satisfies
T
AF
f (α · u) = α f (u) and f (u + v) = f (u) ⊕ f (v). (5.1.1)

DR
By L(V, W), we denote the set of all linear transformations from V to W. In particular, if
W = V then the linear transformation f is called a linear operator and the corresponding set
of linear operators is denoted by L(V).
Even though, in the definition above, we have differentiated between the vector addition
and scalar multiplication for domain and co-domain, we will not differentiate them in the book
unless necessary.
Equation (5.1.1) just states that the two operations, namely, taking the image (apply f ) and
doing ‘vector space operations (vector addition and scalar multiplication) commute, i.e., first
apply vector operations (u + v or αv) and then look at their images f (u + v) or f (αv)) is same
as first computing the images (f (u), f (v)) and then compute vector operations (f (u) + f (v)
and αf (v)). Or equivalently, we look at only those functions which preserve vector operations.
Definition 5.1.2. Let g, h ∈ L(V, W). Then g and h are said to be equal if g(x) = h(x), for
all x ∈ V.
We now give examples of linear transformations.
Example 5.1.3. 1. Let V be a vector space. Then, the maps Id, 0 ∈ L(V), where
(a) Id(v) = v, for all v ∈ V, is commonly called the identity operator.
(b) 0(v) = 0, for all v ∈ V, is commonly called the zero operator.
117
118 CHAPTER 5. LINEAR TRANSFORMATIONS
2. Let V and W be vector spaces over F. Then, 0 ∈ L(V, W), where 0(v) = 0, for all v ∈ V,
is commonly called the zero transformation.
3. The map f (x) = 7x, for all x ∈ R, is an element of L(R) as
f (ax) = 7(ax) = a(7x) = af (x) and f (x + y) = 7(x + y) = 7x + 7y = f (x) + f (y).
4. Let V, W and Z be vector spaces over F. Then, for any T ∈ L(V, W) and S ∈ L(W, Z),
the map S ◦ T ∈ L(V, Z), defined by (S ◦ T )(v) = S T (v) for all v ∈ V, is called the

composition of maps. Observe that for each u, v ∈ V and α, β ∈ R,

(S ◦ T )(αv + βu) = S T (αv + βu) = S αT (v) + βf (u)

= αS T (v) + βS T (u) = α(S ◦ T )(v) + β(S ◦ T )(u).
Hence S ◦ T , in short ST , is an element of L(V, Z).
5. Fix a ∈ Rn and define f (x) = aT x, for all x ∈ Rn . Then f ∈ L(Rn , R). In particular, if
x = [x1 , . . . , xn ]T then, for all x ∈ Rn ,
n
xi = 1T x is a linear transformation.
P
(a) f (x) =
i=1
(b) fi (x) = xi = eTi x is a linear transformation, for 1 ≤ i ≤ n.
T
6. Define f : R2 → R3 by f (x, y)T = (x + y, 2x − y, x + 3y)T . Then f ∈ L(R2 , R3 ). Here

AF
f (e1 ) = (1, 2, 1)T and f (e2 ) = (1, −1, 3)T .

DR
7. Fix A ∈ Mm×n (C). Define fA (x) = Ax, for every x ∈ Cn . Then, fA ∈ L(Cn , Cm ). Thus,
for each A ∈ Mm,n (C), there exists a linear transformation fA ∈ L(Cn , Cm ).
8. Define f : Rn+1 → R[x; n] by f (a1 , . . . , an+1 )T = a1 + a2 x + · · · + an+1 xn , for each

(a1 , . . . , an+1 ) ∈ Rn+1 . Then f is a linear transformation.
9. Fix A ∈ Mn (C). Now, define fA : Mn (C) → Mn (C) and gA : Mn (C) → C by
fA (B) = AB and gA (B) = tr(AB), for every B ∈ Mn (C).
Then fA and gA are both linear transformations.
10. Is the map T : R[x; n] → R[x; n + 1] defined by T (f (x)) = xf (x), for all f (x) ∈ R[x; n] a
linear transformation?
d
Rx
11. The maps T, S : R[x] → R[x] defined by T (f (x)) = dx f (x) and S(f (x)) = f (t)dt, for all
0
f (x) ∈ R[x] are linear transformations. Is it true that T S = Id? What about ST ?
12. Recall the vector space RN in Example 3.1.4.7. Now, define maps T, S : RN → RN
by T ({a1 , a2 , . . .}) = {0, a1 , a2 , . . .} and S({a1 , a2 , . . .}) = {a2 , a3 , . . .}. Then, T and S,
commonly called the shift operators, are linear operators with exactly one of ST or T S
as the Id map.
5.1. DEFINITIONS AND BASIC PROPERTIES 119
13. Recall the vector space C(R, R) (see Example 3.1.4.9). Define T : C(R, R) → C(R, R) by
Rx Rx
T (f )(x) = f (t)dt. For example, T (sin)(x) = sin(t)dt = 1−cos(x), for all x ∈ R. Then,
0 0
verify that T is a linear transformation.
Remark 5.1.4. Let A ∈ Mn (C) and define TA : Cn → Cn by TA (x) = Ax, for every x ∈ Cn .
Then, verify that TAk (x) = (TA ◦ TA ◦ · · · ◦ TA )(x) = Ak x, for any positive integer k.
| {z }
k times
Exercise 5.1.5. Fix A ∈ Mn (C). Then, do the following maps define linear transformations?
1. Define f, g : Mn (C) → Mn (C) by f (B) = A∗ B and g(B) = BA, for every B ∈ Mn (C).
2. Define h, t : Mn (C) → C by h(B) = tr(A∗ B) and t(B) = tr(BA), for every B ∈ Mn (C).
We now prove that any linear transformation sends the zero vector to a zero vector.
Proposition 5.1.6. Let T ∈ L(V, W). Suppose that 0V is the zero vector in V and 0W is the
zero vector of W. Then T (0V ) = 0W .
Proof. Since 0V = 0V + 0V , we get T (0V ) = T (0V + 0V ) = T (0V ) + T (0V ). As T (0V ) ∈ W,
0W + T (0V ) = T (0V ) = T (0V ) + T (0V ).

T
Hence T (0V ) = 0W .
AF
From now on 0 will be used as the zero vector of the domain and co-domain. We now
DR
consider a few more examples for better understanding.

Example 5.1.7. 1. Does there exist a linear transformation T : V → W such that T (v) 6= 0,
for all v ∈ V?
Solution: No, as T (0) = 0 (see Proposition 5.1.6).
2. Does there exist a linear transformation T : R → R such that T (x) = x2 , for all x ∈ R?
Solution: No, as T (ax) = (ax)2 = a2 x2 = a2 T (x) 6= aT (x), unless a = 0, 1.
3. Does there exist a linear transformation T : R → R such that T (5) = 10 and T (10) = 5?
Solution: No, as T (10) = T (5 + 5) = T (5) + t(5) = 10 + 10 = 20 6= 5.
4. Does there exist a linear transformation f : R2 → R2 such that f ((x, y)T ) = (x + y, 2)T ?
Solution: No, as f (0) 6= 0.
5. Does there exist a linear transformation f : R2 → R2 such that f ((x, y)T ) = (x + y, xy)T ?
Solution: No, as f ((2, 2)T ) = (4, 4)T 6= 2(2, 1)T = 2f ((1, 1)T ).
6. Define a map T : C → C by T (z) = z, the complex conjugate of z. Is T a linear operator
over the real vector space R?
Solution: Yes, as for any α ∈ R, T (αz) = αz = αz = αT (z).
We now define the range space.
Definition 5.1.8. Let f ∈ L(V, W). Then the range/ image of f , denoted Rng(f ) or Im(f ),
is given by Rng(f ) = {f (x) : x ∈ V}.
As an exercise, show that Rng(f ) is a subspace of W. The next result, which is a very
important result, states that a linear transformation is known if we know its image on a basis
of the domain space.
Lemma 5.1.9. Let V and W be vector spaces over F with B = {v1 , v2 , . . .} as a basis of V. If
f ∈ L(V, W) then T is determined if we know the set {f (v1 ), f (v2 ), . . .}, ı.e., if we know the
image of f on the basis vectors of V, or equivalently, Rng(f ) = LS(f (x)|x ∈ B).
Proof. Let B be a basis of V over F. Then, for each v ∈ V, there exist vectors u1 , . . . , uk in B
k
and scalars c1 , . . . , ck ∈ F such that v =
P
ci ui . Thus
i=1
k k k
!
X X X
T (v) = f ci ui = f (ci ui ) = ci T (ui ).
i=1 i=1 i=1
Or equivalently, whenever
   
c1 c
. i 1 
 .. 
h
v = [u1 , . . . , uk ] .
 .  then f (v) = f (u1 ) · · · f (uk )  . . (5.1.2)
ck ck
Thus, the image of f on v just depends on where the basis vectors are mapped. Equation 5.1.2
T
also shows that Rng(f ) = {f (x) : x ∈ V} = LS(f (x)|x ∈ B).

AF
DR
Example 5.1.10. Determine Rng(T ) of the following linear transformations.

1. f ∈ L(R3 , R4 ), where f ((x, y, z)T ) = (x − y + z, y − z, x, 2x − 5y + 5z)T .
Solution: Consider the standard basis {e1 , e2 , e3 } of R3 . Then
Rng(f ) = LS(f (e1 ), T (e2 ), T (e3 )) = LS (1, 0, 1, 2)T , (−1, 1, 0, −5)T , (1, −1, 0, 5)T

= LS (1, 0, 1, 2)T , (1, −1, 0, 5)T = {λ(1, 0, 1, 2)T + β(1, −1, 0, 5)T | λ, β ∈ R}

= {(λ + β, −β, λ, 2λ + 5β) : λ, β ∈ R}

= {(x, y, z, w)T ∈ R4 | x + y − z = 0, 5y − 2z + w = 0},
2. Let B ∈ M2 (R). Now, define a map T : M2 (R) → M2 (R) by T (A) = BA − AB, for all
A ∈ M2 (R). Determine Rng(T ) and Null(T ).
Solution: Recall that {eij |1 ≤ i, j ≤ 2} is a basis of M2 (R). So,
(a) if B = cI2 then Rng(T ) = {0}.
" # " # " # " #
1 2 0 −2 −2 −3 2 0
(b) if B = then T (e11 ) = , T (e12 ) = , T (e21 ) = and
2 4 2 0 0 2 3 −2
" # " # " # " #!
0 2 0 2 2 3 −2 0
T (e22 ) = . Thus, Rng(T ) = LS , , .
−2 0 −2 0 0 −2 −3 2
" # " # " # " #!
1 2 0 2 2 2 −2 0
(c) for B = , verify that Rng(T ) = LS , , .
2 3 −2 0 0 −2 −2 2
Recall that by Example 5.1.3.5, for each a ∈ Rn , the map T (x) = aT x, for each x ∈ Rn , is
a linear transformation from Rn to R. We now show that these are the only ones.
Corollary 5.1.11. [Reisz Representation Theorem] Let T ∈ L(Rn , R). Then, there exists
a ∈ Rn such that T (x) = aT x.
Proof. By Lemma 5.1.9, T is known if we know the image of T on {e1 , . . . , en }, the standard
basis of Rn . So, for 1 ≤ i ≤ n, let T (ei ) = ai , for some ai ∈ R. Now define a = [a1 , . . . , an ]T
and x = [x1 , . . . , xn ]T ∈ Rn . Then, for all x ∈ Rn ,
n n n
!
X X X
T (x) = T xi ei = xi T (ei ) = xi ai = aT x.
i=1 i=1 i=1
Example 5.1.12. In each of the examples given below, state whether a linear transformation
exists or not. If yes, give at least one linear transformation. If not, then give the condition due
to which a linear transformation doesn’t exist.
1. Can we construct a linear transformation T : R2 → R2 such that T ((1, 1)T ) = (e, 2)T and
T ((2, 1)T ) = (5, 4)T ?
Solution: The first thing that we need to answer is “is the set {(1, 1), (2, 1)} linearly
independent”? The answer is ‘Yes’. So, we can construct it. So, how do we do it?
T
AF
" # " # " # " #" #

x 1 2 1 2 α
We now need to write any vector =α +β = as by definition of
DR
y 1 1 1 1 β
linear transformation
" #! " # " #! " #! " #! " # " #
x 1 2 1 2 e 5
T = T α +β = αT + βT =α +β
y 1 1 1 1 2 4
" #" # " # " #!−1 " #
e 5 α e 5 1 2 x
= =
2 4 β 2 4 1 1 y
" #" #" # " #
e 5 −1 2 x (5 − e)x + (2e − 5)y
= = .
2 4 1 −1 y 2x
2. T : R2 → R2 such that T ((1, 1)T ) = (1, 2)T and T ((1, −1)T ) = (5, 10)T ?" #
1 1
Solution: Yes, as the set {(1, 1), (1, −1)} is a basis of R2 . Write B = . Then,
1 −1
" #! " #! " #!!
x x x
T = T (BB −1 ) = T B B −1
y y y
" " #! " #!# "  #−1 " #
1 1 1 1 x
= T , T  
1 −1 1 −1 y
" # " #−1 " # " #" # " #
x+y
1 5 1 1 x 1 5 2 3x − 2y
=  = = .
2 10 1 −1 y 2 10 x−y 2 6x − 4y
3. T : R2 → R2 such that T ((1, 1)T ) = (1, 2)T and T ((5, 5)T ) = (5, 11)T ?
Solution: Note that the set {(1, 1), (5, 5)} is linearly dependent. Further, (5, 11)T =
T ((5, 5)T ) = 5T ((1, 1)T )5(1, 2)T = (5, 10)T gives us a contradiction. Hence, there is no
such linear transformation.
4. Does there exist a linear transformation T : R3 → R2 with T (1, 1, 1) = (1, 2), T (1, 2, 3) =
(4, 3) and T (2, 3, 4) = (7, 8)?
Solution: Here, the set {(1, 1, 1), (1, 2, 3), (2, 3, 4)} is linearly dependent and (2, 3, 4) =
(1, 1, 1) + (1, 2, 3). So, we need T ((2, 3, 4)) = T ((1, 1, 1) + (1, 2, 3)) = T ((1, 1, 1)) +
T ((1, 2, 3)) = (1, 2) + (4, 3) = (5, 5). But, we are given T (2, 3, 4) = (7, 8), a contradiction.
So, such a linear transformation doesn’t exist.
5. T : R2 → R2 such that T ((1, 1)T ) = (1, 2)T and T ((5, 5)T ) = (5, 10)T ?
Solution: Yes, as (5, 10)T = T ((5, 5)T ) = 5T ((1, 1)T ) = 5(1, 2)T = (5, 10)T .
# such linear transformation, note that {(1, 1) , (1, 0) } is a basis of R .

To construct T T 2
" one
1 1
Let B = and pick v ∈ R2 . Now define T ((1, 0)T ) = v = (v1 , v2 )T . Then, as in the
1 0
" # " #" # " #
x 0 1 x y
previous example, note that B −1 = = and hence
y 1 −1 y x−y
" #! " " #! " #!# " # " #" # " #
x 1 1 −1 x 1 v1 y 1
T = T ,T B = =y + (x − y)v.
T
y 1 0 y 2 v2 x−y 2
AF
6. Does there exist a linear transformation T : R3 → R2 with T (1, 1, 1) = (1, 2), T (1, 2, 3) =
DR
(4, 3) and T (2, 3, 4) = (5, 5)?

Solution: As (2, 3, 4) = (1, 1, 1) + (1, 2, 3) and T ((2, 3, 4)) = T ((1, 1, 1) + (1, 2, 3)) =
T ((1, 1, 1)) + T ((1, 2, 3)), such a linear transformation exists. To get the linear transfor-
mation, get a basis, namely {1, 1, 1), (1, 2, 3), e1 }, of R3 . Note that this basis contains
(1, 1, 1) and (1, 2, 3). Now, define T (e1 ) as any vector of R2 . This give us a linear trans-
formation satisfying the given condition.
7. T : R2 → R2 such that Null(T ) = {x ∈ R2 | T (x) = 0} = LS{(1, π)T }?
Solution: Yes. Take {(1, π)T , u} as a basis of R2 and define T ((1, π)T ) = 0 and T (u) = u.
(" #)
1 2
8. T : M2 (R) → R such that Null(T ) = {x ∈ M2 (R) | T (x) = 0} = LS
3 ?
2 1
(" # )
1 2
Solution: Yes. Take , e11 , e12 , e21 as a basis of M2 (R) and define
2 1
" #!
1 2
T (e11 ) = e1 , T (e12 ) = e2 , T (e21 ) = e3 and T = 0.
2 1
Exercise 5.1.13. 1. Use matrices to construct linear operators T, S : R3 → R3 that satisfy:
(a) T 6= 0, T ◦ T = T 2 6= 0, T ◦ T ◦ T = T 3 = 0.
(b) T 6= 0, S 6= 0, S ◦ T = ST 6= 0, T ◦ S = T S = 0.
(c) S ◦ S = S 2 = T 2 = T ◦ T, S 6= T .
(d) T ◦ T = T 2 = Id, T 6= Id.
2. Fix a positive integer p and let T : Rn → Rn be a linear operator with T k 6= 0 for

1 ≤ k ≤ p and T p+1 = 0. Then prove that there exists a vector x ∈ Rn such that the set
{x, T (x), . . . , T p (x)} is linearly independent.
3. Fix x0 ∈ Rn with x0 6= 0. Now, define T ∈ L(Rn , Rm ) by T (x0 ) = y0 , for some y0 ∈ Rm .

Define T −1 (y0 ) = {x ∈ Rn : T (x) = y0 }. Then prove that x ∈ T −1 (y0 ) if and only if
x − x0 ∈ T −1 (0). Further, T −1 (y0 ) is a subspace of Rn if and only if y0 = 0.
4. Let V and W be vector spaces over F. If {v1 , . . . , vn } is a basis of V and {w1 , . . . , wn } ⊆ W

then prove that there exists a unique T ∈ L(V, W) with T (vi ) = wi , for i = 1, . . . , n.
5. Let V be a vector space and let a ∈ V. Then the map Ta : V → V defined by Ta (x) = x+a,
for all x ∈ V is called the translation map. Prove that Ta ∈ L(V) if and only if a = 0.
6. Prove that there exists infinitely many linear transformations T : R3 → R2 such that
T ((1, −1, 1)T ) = (1, 2)T and T ((−1, 1, 2)T ) = (1, 0)T ?
7. Does there exist a linear transformation T : R3 → R2 such that

(a) T ((1, 0, 1)T ) = (1, 2)T , T ((0, 1, 1)T ) = (1, 0)T and T ((1, 1, 1)T ) = (2, 3)T ?
T
(b) T ((1, 0, 1)T ) = (1, 2)T , T ((0, 1, 1)T ) = (1, 0)T and T ((1, 1, 2)T ) = (2, 3)T ?
AF
DR
8. Find T ∈ L(R3 ) for which Rng(T ) = LS (1, 2, 0)T , (0, 1, 1)T , (1, 3, 1)T .

9. Let T : R3 → R3 be defined by T ((x, y, z)T ) = (2x + 3y + 4z, x + y + z, x + y + 3z)T . Find

the value of k for which there exists a vector x ∈ R3 such that T (x) = (9, 3, k)T .
10. Let T : R3 → R3 be defined by T ((x, y, z)T ) = (2x − 2y + 2z, −2x + 5y + 2z, x + y + 4z)T .
Find x ∈ R3 such that T (x) = (1, 1, −1)T .
11. Let T : R3 → R3 be defined by T ((x, y, z)T ) = (2x + y + 3z, 4x − y + 3z, 3x − 2y + 5z)T .

Determine x, y, z ∈ R3 \ {0} such that T (x) = 6x, T (y) = 2y and T (z) = −2z. Is the set
{x, y, z} linearly independent?
12. Let T : R3 → R3 be defined by T ((x, y, z)T ) = (2x + 3y + 4z, −y, −3y + 4z)T . Determine
x, y, z ∈ R3 \ {0} such that T (x) = 2x, T (y) = 4y and T (z) = −z. Is the set {x, y, z}
linearly independent?
13. Does there exist a linear transformation T : R3 → Rn such that T ((1, 1, −2)T ) = x,
T ((−1, 2, 3)T ) = y and T ((1, 10, 1)T ) = z
(a) with z = x + y?
(b) with z = cx + dy, for some choice of c, d ∈ R?
14. For each matrix A given below, define T ∈ L(R2 ) by T (x) = Ax. What do these linear
operators signify geometrically?
( "√ # " # " √ # " # " #)
cos 2π 2π

1 3 −1 1 1 −1 1 1 − 3 0 −1 3 − sin 3
(a) A ∈ √ ,√ , √ , , .
2 1 2 sin 2π 2π

3 2 1 1 3 1 1 0 3 cos 3
( " # " # " # " #)
1 1 −1 1 1 2 0 0 1 0
(b) A ∈ , , , .
2 −1 1 5 2 4 0 1 0 0
( "√ # " # " √ # " #)
cos 2π 2π

1 3 1 1 1 1 1 1 3 3 sin 3
(c) A ∈ √ ,√ , √ , .
2 1 2 1 −1 2 sin 2π 2π

− 3 3 −1 3 − cos 3
15. Consider the space C3 over C. If f ∈ L(C3 ) with f (x) = x, f (y) = (1 + i)y and f (z) =
(2 + 3i)z, for x, y, z ∈ C3 \ {0} then prove that {x, y, z} forms a basis of C3 .
5.2 Rank-Nullity Theorem

Recall that for any f ∈ L(V, W), Rng(f ) = {f (v)|v ∈ V} (see Definition 5.1.8). Now, in line
with the ideas in Theorem 3.6.1, we define the null-space or the kernel of a linear transformation.
At this stage, the readers are advised to recall Section 3.6 for clarity and similarity with the
results in this section.
Definition 5.2.1. Let f ∈ L(V, W). Then the null space of f , denoted Null(f ) or Ker(f ),
is given by Null(f ) = {v ∈ V|f (v) = 0}. In most linear algebra books, it is also called the
kernel of f and written Ker(f ). Further, if V is finite dimensional then one writes
T
dim(Rng(T )) = Rank(T ) and dim(Null(T )) = Nullity(T ).

AF
Example 5.2.2. 1. Define f ∈ L(R3 , R4 ) by f ((x, y, z)T ) = (x−y+z, y−z, x, 2x−5y+5z)T .

DR
Then, by definition,
Null(f ) = {(x, y, z)T ∈ R3 : f ((x, y, z)T ) = 0}

= {(x, y, z)T ∈ R3 : (x − y + z, y − z, x, 2x − 5y + 5z)T = 0}
= {(x, y, z)T ∈ R3 : x − y + z = 0, y − z = 0, x = 0, 2x − 5y + 5z = 0}
= {(x, y, z)T ∈ R3 : y − z = 0, x = 0}
= {(0, z, z)T ∈ R3 : z ∈ R} = LS((0, 1, 1)T )
2. Fix B ∈ M2 (R). Now, define T : M2 (R) → M2 (R) by T (A) = BA−AB, for all A ∈ M2 (R).
Solution: Then A ∈ Null(T ) if and only if A commutes with B. In particular,
{I, B, B 2 , . . .} ⊆ Null(T ). For example, if B = αI, for some α then Null(T ) = M2 (R).
Exercise 5.2.3. 1. Let T ∈ L(V, W). Then Null(T ) is a subspace of V. Furthermore, if

V is finite dimensional then dim(Null(T )), dim(Rng(T )) ≤ dim(V).
2. Define T ∈ L(R2 , R4 ) by T ((x, y)T ) = (x+y, x−y, 2x+y, 3x−4y)T . Determine Null(T ).
3. Describe Null(D) and Rng(D), where D ∈ L(R[x; n]) is defined by (D(f ))(x) = f 0 (x),
the differentiation with respect to x. Note that Rng(D) ⊆ R[x; n − 1].
4. Define T ∈ L(R[x]) by (T (f ))(x) = xf (x), for all f (x) ∈ L(R[x]). What can you say
about Null(T ) and Rng(T )?
5.2. RANK-NULLITY THEOREM 125
5. Define T ∈ L(R3 ) by T (e1 ) = e1 + e3 , T (e2 ) = e2 + e3 and T (e3 ) = −e3 . Then
(a) determine T ((x, y, z)T ), for x, y, z ∈ R.

(b) determine Null(T ) and Rng(T ).
(c) is it true that T 2 = Id?
We now prove a result which is similar to Exercise 3.3.17.2.
Theorem 5.2.4. Let V and W be vector spaces over F and let T ∈ L(V, W).
1. If S ⊆ V is linearly dependent then T (S) = {T (v) | v ∈ V} is linearly dependent.
2. Suppose S ⊆ V such that T (S) is linearly independent then S is linearly independent.
Proof. Part 1 : As S is linearly dependent, there exist k ∈ N and vi ∈ S, for 1 ≤ i ≤ k, such that
k
xi vi = 0, in the unknowns xi ’s, has a non-trivial solution, say xi = ai ∈ F, 1 ≤ i ≤
P
the system
i=1
k
P Pk
k. Thus ai vi = 0. Then ai ’s also give a non-trivial solution to the system yi T (vi ) = 0,
i=1 i=1
k k k

P P P
where yi ’s are unknown, as ai T (vi ) = T (ai vi ) = T ai vi = T (0) = 0. Hence the
i=1 i=1 i=1
Part 2 : On the contrary assume that S is linearly dependent. Then by Part 1, T (S) is
linearly dependent, a contradiction to the given assumption that T (S) is linearly independent.
T
We now prove the rank-nullity Theorem. The proof of this result is similar to the proof of
AF
Theorem 3.6.1. We give it again for the sake of completeness.

DR
Theorem 5.2.5 (Rank-Nullity Theorem). Let V and W be vector spaces over F. If dim(V) is
finite and T ∈ L(V, W) then
Rank(T ) + Nullity(T ) = dim(Rng(T )) + dim(Null(T )) = dim(V).
Proof. Let dim(V) = n. As Null(T ) ⊆ V, let dim(Null(T )) = k ≤ n. Let B = {v1 , . . . , vk }

be a basis of Null(T ). We extend it to form a basis C = {v1 , . . . , vk , vk+1 , . . . , vn } of V. As
T (v) = 0, for all v ∈ B,
Rng(T ) = LS(T (v1 ), . . . , T (vk ), T (vk+1 ), . . . , T (vn )) = LS(T (vk+1 ), . . . , T (vn )).
We claim that {T (vk+1 ), . . . , T (vn )} is linearly independent subset of W.

n−k
P
So, consider the system ai T (vk+i ) = 0 in the unknown a1 , . . . , an−k . Note that
i=1
n−k n−k n−k

!
X X X
ai T (vk+i ) = 0 ⇔ T ai vk+i =0⇔ ai vk+i ∈ Null(T ).
i=1 i=1 i=1
n−k k
Hence, there exists b1 , . . . , bk ∈ F such that
P P
ai vk+i = bj vj . This gives a new system
i=1 j=1
n−k
X k
X
ai vk+i + (−bj )vj = 0,
i=1 j=1
in the unknowns ai ’s and bj ’s. As C is linearly independent, the new system has only the trivial
n−k
solution, namely [a1 , . . . , ak , −b1 , . . . , −b` ]T = 0. Hence, the system
P
ai T (vk+i ) = 0 has only
i=1
the trivial solution. Thus, the set {T (vk+1 ), . . . , T (vn )} is linearly independent subset of W. It
also spans Rng(T ) and hence is a basis of Rng(T ). Therefore,
dim(Rng(T )) + dim(Null(T )) = k + (n − k) = n = dim(V).
Thus, we have proved the required result.

As an immediate corollary, we have the following result. The proof is left for the reader.
Corollary 5.2.6. Let V and W be finite dimensional vector spaces over F and let T ∈ L(V, W).
If dim(V) = dim(W) then the following statements are equivalent.
1. T is one-one.
2. Ker(T ) = {0}.
3. T is onto.
4. dim(Rng(T )) = dim(W) = dim(V).
Exercise 5.2.7. 1. Prove Corollary 5.2.6.
2. Let V and W be finite dimensional vector spaces over F. If T ∈ L(V, W) then

T
AF
(a) T cannot be onto if dim(V) < dim(W).

DR
(b) T cannot be one-one if dim(V) > dim(W).
3. Let A ∈ Mn (R) with A2 = A. Define T ∈ L(Rn ) by T (v) = Av for all v ∈ Rn . Then

prove that
(a) T 2 = T , or equivalently, (T (Id − T ))(x) = 0, for all x ∈ Rn .
(b) Null(T ) ∩ Rng(T ) = {0}.
(c) Rn = Rng(T ) + Null(T ).
4. Let z1 , z2 , . . . , zk be k distinct complex numbers. Define T ∈ L(C[x; n], Ck ) by T P (z) =

T
P (z1 ), . . . , P (zk ) , for all P (z) ∈ C[x; n]. Determine Rank(T ).
5.3 Algebra of Linear Transformations
We start with the following definition.
Definition 5.3.1. Let V, W be vector spaces over F and let S, T ∈ L(V, W). Then, we define
the point-wise
1. sum of S and T , denoted S + T , by (S + T )(v) = S(v) + T (v), for all v ∈ V.
2. scalar multiplication, denoted c T for c ∈ F, by (c T )(v) = c (T (v)), for all v ∈ V.

5.3. ALGEBRA OF LINEAR TRANSFORMATIONS 127
( " # " #)
1 0
To understand the next result, consider L(R2 , R3 )
and let B = v1 = , v2 =
0 1
      


 1 0 0 
 
0, w2 = 1, w3 = 0 be bases of R and R , respectively. Now, for
2 3
   
and C = w1 =      

0 0 1 

 
1 ≤ i ≤ 2, 1 ≤ j ≤ 3, define elements of L(R2 , R3 ) by

(
wj , if k = i
fji (vk ) =
0, 6 i.
if k =
Then verify that the above maps correspond to the following collection of matrices?
           
1 0 0 1 0 0 0 0 0 0 0 0
           
f11 = 0 0, f12 = 0 0, f21 = 1 0, f22 = 0 1, f31 = 0 0, f32 = 0 0.
          
0 0 0 0 0 0 0 0 1 0 0 1
Theorem 5.3.2. Let V and W be vector spaces over F. Then L(V, W) is a vector space over
F. Furthermore, if dim V = n and dim W = m, then dim L(V, W) = mn.
Proof. It can be easily verified that under point-wise addition and scalar multiplication, defined
above, L(V, W) is indeed a vector space over F. We now prove the other part. So, let us
assume that B = {v1 , . . . , vn } and C = {w1 , . . . , wm } are bases of V and W, respectively. For
T
1 ≤ i ≤ n, 1 ≤ j ≤ m, we define the functions fji on the basis vectors of V by

AF
(
DR
wj , if k = i
fji (vk ) =
0, if k 6= i.
n
For other vectors of V, we extend the definition by linearity, i.e., if v =
P
αs vs then
s=1
n n
!
X X
fji (v) = fji αs vs = αs fji (vs ) = αi fji (vi ) = αi wj . (5.3.1)
s=1 s=1
Thus fji ∈ L(V, W). We now show that {fji |1 ≤ i ≤ n, 1 ≤ j ≤ m} is a basis of L(V, W).
As a first step, we show that fji ’s are linearly independent. So, consider the linear system
n P
P m
cji fji = 0, in the unknowns cji ’s, for 1 ≤ i ≤ n, 1 ≤ j ≤ m. Using the point-wise addition
i=1 j=1
and scalar multiplication, we get
 
X n Xm n X
X m m
X
0 = 0(vk ) =  cji fji  (vk ) = cji fji (vk ) = cjk wj .
i=1 j=1 i=1 j=1 j=1
But, the set {w1 , . . . , wm } is linearly independent. Hence the only solution equals cjk = 0, for
1 ≤ j ≤ m. Now, as we vary vk from v1 to vn , we see that cji = 0, for 1 ≤ j ≤ m and 1 ≤ i ≤ n.
Thus, we have proved the linear independence of {fji |1 ≤ i ≤ n, 1 ≤ j ≤ m}.
Now, let us prove that LS ({fji |1 ≤ i ≤ n, 1 ≤ j ≤ m}) = L(V, W). So, let f ∈ L(V, W).
m
Then, for 1 ≤ s ≤ n, f (vs ) ∈ W and hence there exists βts ’s such that f (vs ) =
P
βts wt . So, if
t=1
n
αs vs ∈ V then, using Equation (5.3.1), we get
P
v=
s=1
n n n m n X
m
! !
X X X X X
f (v) = f α s vs = αs f (vs ) = αs βts wt = βts (αs wt )
s=1 s=1 s=1 t=1 s=1 t=1
n m n X
m
!
XX X
= βts fts (v) = βts fts (v).
s=1 t=1 s=1 t=1
n P
m
Since the above is true for every v ∈ V, we get f =
P
βts fts . Thus, we conclude that
s=1 t=1
f ∈ LS ({fji |1 ≤ i ≤ n, 1 ≤ j ≤ m}). Hence, LS ({fji |1 ≤ i ≤ n, 1 ≤ j ≤ m}) = L(V, W) and
thus the required result follows.
We now give a corollary of the rank-nullity theorem.
Corollary 5.3.3. Let V be a vector space over F with dim(V) = n. If S, T ∈ L(V) then
1. Nullity(T ) + Nullity(S) ≥ Nullity(ST ) ≥ max{Nullity(T ), Nullity(S)}.
2. min{Rank(S), Rank(T )} ≥ Rank(ST ) ≥ n − Rank(S) − Rank(T ).
Proof. The prove of Part 2 is omitted as it directly follows from Part 1 and Theorem 5.2.5.
Part 1 - Second Inequality: Suppose v ∈ Ker(T ). Then
T
(ST )(v) = S(T (v)) = S(0) = 0

AF
DR
implies Ker(T ) ⊆ Ker(ST ). Thus Nullity(T ) ≤ Nullity(ST ).

By Theorem 5.2.5, Nullity(S) ≤ Nullity(ST ) ⇔ Rank(S) ≥ Rank(ST ). This holds as
Rng(T ) ⊆ V implies Rng(ST ) = S(Rng(T )) ⊆ S(V) = Rng(S).
Part 1 - First Inequality: Let {v1 , . . . , vk } be a basis of Null(T ). Then {v1 , . . . , vk } ⊆
Null(ST ). So, let us extend it to get a basis {v1 , . . . , vk , u1 , . . . , u` } of Null(ST ).
Now, proceeding as in the proof of the rank-nullity theorem, implies that {T (u1 ), . . . , T (u` )}
is a linearly independent subset of Null(S). Hence, Nullity(S) ≥ ` and therefore, we get
Nullity(ST ) = k + ` ≤ Nullity(T ) + Nullity(S).
Before proceeding further, recall the following definition about a function.
Definition 5.3.4. Let f : S → T be any function. Then
1. a function g : T → S is called a left inverse of f if (g ◦ f )(x) = x, for all x ∈ S. That is,

g ◦ f = Id, the identity function on S.
2. a function h : T → S is called a right inverse of f if (f ◦ h)(y) = y, for all y ∈ T . That

is, f ◦ h = Id, the identity function on T .
3. f is said to be invertible if it has a right inverse and a left inverse.
Remark 5.3.5. Let f : S → T be invertible. Then, it can be easily shown that any right inverse
and any left inverse are the same. Thus, the inverse function is unique and is denoted by f −1 .
It is well known that f is invertible if and only if f is both one-one and onto.
5.3. ALGEBRA OF LINEAR TRANSFORMATIONS 129
Lemma 5.3.6. Let V and W be vector spaces over F and let T ∈ L(V, W). If T is one-one and
onto then, the map T −1 : W → V is also a linear transformation. The map T −1 is called the
inverse linear transform of T and is defined by T −1 (w) = v whenever T (v) = w.
Proof. Part 1: As T is one-one and onto, by Theorem 5.2.5, dim(V) = dim(W). So, by
Corollary 5.2.6, for each w ∈ W there exists a unique v ∈ V such that T (v) = w. Thus, one
defines T −1 (w) = v.
We need to show that T −1 (α1 w1 + α2 w2 ) = α1 T −1 (w1 ) + α2 T −1 (w2 ), for all α1 , α2 ∈ F
and w1 , w2 ∈ W. Note that by previous paragraph, there exist unique vectors v1 , v2 ∈ V such
that T −1 (w1 ) = v1 and T −1 (w2 ) = v2 . Or equivalently, T (v1 ) = w1 and T (v2 ) = w2 . So,
T (α1 v1 + α2 v2 ) = α1 w1 + α2 w2 , for all α1 , α2 ∈ F. Hence, for all α1 , α2 ∈ F, we get
T −1 (α1 w1 + α2 w2 ) = α1 v1 + α2 v2 = α1 T −1 (w1 ) + α2 T −1 (w2 ).
Example 5.3.7. 1. Let T : R2 → R2 be given by (x, y) (x + y, x − y). Then, verify that

x+y x−y
T −1 is given by 2 , 2 .
n
2. Let T ∈ L(Rn , R[x; n − 1]) be given by (a1 , . . . , an ) ai xi−1 , for (a1 , . . . , an ) ∈ Rn .
P
i=1
n n
Then, T −1 maps ai xi−1 ai xi−1 ∈ R[x; n − 1].
P P
(a1 , . . . , an ), for each polynomial
i=1 i=1
T
Verify that T −1 ∈ L(R[x; n − 1], Rn ).

AF
Definition 5.3.8. Let V and W be vector spaces over F and let T ∈ L(V, W). Then, T is said
DR
to be singular if {0} $ Ker(T ), i.e., Ker(T ) contains a non-zero vector. If Ker(T ) = {0}
then, T is called non-singular.
 
" #! x
x
Example 5.3.9. Let T ∈ L(R2 , R3 ) be defined by T
 
=
y . Then, verify that T is

y
0
non-singular. Is T invertible?
We now prove a result that relates non-singularity with linear independence.
Theorem 5.3.10. Let V and W be vector spaces over F and let T ∈ L(V, W). Then the
following statements are equivalent.
1. T is one-one.
2. T is non-singular.
3. Whenever S ⊆ V is linearly independent then T (S) is necessarily linearly independent.
Proof. 1⇒2 On the contrary, let T be singular. Then, there exists v 6= 0 such that T (v) =
0 = T (0). This implies that T is not one-one, a contradiction.
2⇒3 Let S ⊆ V be linearly independent. Let if possible T (S) be linearly dependent.
k
Then, there exists v1 , . . . , vk ∈ S and α = (α1 , . . . , αk )T 6= 0 such that
P
αi T (vi ) = 0.
i=1
k k

P P
Thus, T αi vi = 0. But T is non-singular and hence we get αi vi = 0 with α 6= 0, a
i=1 i=1
contradiction to S being a linearly independent set.
3⇒1 Suppose that T is not one-one. Then, there exists x, y ∈ V such that x 6= y but
T (x) = T (y). Thus, we have obtained S = {x − y}, a linearly independent subset of V with
T (S) = {0}, a linearly dependent set. A contradiction to our assumption. Thus, the required
result follows.
Definition 5.3.11. Let V and W be vector spaces over F and let T ∈ L(V, W). Then, T is
said to be an isomorphism if T is one-one and onto. The vector spaces V and W are said to
∼ W, if there is an isomorphism from V to W.
be isomorphic, denoted V =
We now give a formal proof of the statement that every finite dimensional vector space V
over F looks like Fn , where n = dim(V).
Theorem 5.3.12. Let V be an n-dimensional vector space over F. Then V ∼

= Fn .
vn } be abasis of V and {e1 , . . . , en }, the standard basis of Fn . Define

Proof. Let {v1 , . . . ,
n n n
T ∈ L(V, Fn ) by T vi ei ∈ V. Then, it is easy to
P P P
αi vi = αi ei , whenever v =
i=1 i=1 i=1
observe that T is one-one and onto. Hence, T is an isomorphism.
T
As a direct application using the countability argument, one obtains the following result
AF
DR
Corollary 5.3.13. The vector space R over Q is not finite dimensional. Similarly, the vector
space C over Q is not finite dimensional.
We now summarize the different definitions related with a linear operator on a finite dimen-
sional vector space. The proof basically uses the rank-nullity theorem and they appear in some
form in previous results. Hence, we leave the proof for the reader.
Theorem 5.3.14. Let V be a finite dimensional vector space over F with dim V = n. Then the
following statements are equivalent for T ∈ L(V).
1. T is one-one.
2. Ker(T ) = {0}.
3. Rank(T ) = n.
4. T is onto.
5. T is an isomorphism.
6. If {v1 , . . . , vn } is a basis for V then so is {T (v1 ), . . . , T (vn )}.
7. T is non-singular.
8. T is invertible.
5.4. ORDERED BASES 131
5.4 Ordered Bases
Let V be a vector space of dimension n over F. Then Theorem 5.3.12 implies that V is isomorphic
to Fn . So, one should be able to visualize the elements of V as an n-tuple. Further, our problem
may require us to look at a subspace W of V whose dimension is very small as compared to the
dimension of V (this is generally encountered when we work with sparse matrices or whenever
we do computational work). It may also be possible that a basis of W may not look like a
standard basis of Fn , where the coefficient of ei gave the i-th component of the vector. We start
with the following example. Note that we will be using ‘small brackets’ in place of ‘braces’ to
represent a basis.
Example 5.4.1.
1. Let f (x) =1 −x2 ∈ R[x; 2]. If B = (1, x, x2 ) be a basis of R[x; 2] then, f (x) =
h i 1 
2
1 x x   0 .

−1
2. Let V = {(u, v, w, x, y)T ∈ R5 | w − x = u, v = y, u + v + x = 3y}. Then, verify that

B = (−1, 0, 0, 1, 0)T , (2, 1, 2, 0, 1)T = (u1 , u2 ), say, can be taken as a basis of V. So, even

though V is a subspace of R5 , we just need two scalars" α, # β to understand any vector in

T
" #
3 5
AF
V. For example, (7, 5, 10, 3, 5)T = 3u1 + 5u2 = [u1 , u2 ] = [u2 , u1 ] .

5 3
DR
So, from Example 5.4.1 we conclude the following: Let V be a vector space of dimension n
n
over F. If we fix a basis, say, B = (u1 , u2 , . . . , un ) of V and if v ∈ V with v =
P
α i ui ⇒
i=1
   
α1 α2
   
 α2   α1 
v = [u1 , u2 , . . . , un ] .  = [u2 , u1 , . . . , un ] . 
   
 ..   .. 
   
αn αn
Note the change in the first two components of the column vectors which are elements of Fn .
So, a change in the position of the vectors ui ’s gives a change in the column vector. Hence,
if we fix the order of the basis vectors ui ’s then with respect to this order all vectors can be
thought of as elements of Fn . We use the above discussion to define an ordered basis.
Definition 5.4.2. Let W be a vector space over F with a basis B = {u1 , . . . , um }. Then, an
ordered basis for W is a basis B together with a one-to-one correspondence between B and
{1, 2, . . . , m}. Since there is an order among the elements of B, we write B = (u1 , . . . , um ). The
matrix B = [u1 , . . . , um ] containing the basis vectors of Wm and is called the basis matrix.
Example 5.4.3. Note that for Example 5.4.1.1 [1, x, x2 ] is a basis matrix, whereas for Exam-
ple 5.4.1.2, [u1 , u2 ] and [u2 , u1 ] are basis matrices.
Definition 5.4.4. Let B = [v1 , . . . , vm ] be the basis matrix corresponding to an ordered basis
B = (v1 , . . . , vm ) of W. Since B is a basis of W, for each v ∈ W, there exist βi , 1 ≤ i ≤ m,
   
β1 β1
m  .   . 
.   . , denoted [v]B , is called the coordinate
P
such that v = βi vi = B 
 . . The vector  . 
i=1
βm βm
vector of v with respect to B. Thus,
 
v1
 . 
v = B[v]B = [v1 , . . . , vm ][v]B , or equivalently, v = [v]TB  . 
 . . (5.4.1)
vm
The expressions in Equation (5.4.1) are generally viewed as a symbolic expressions.
Example 5.4.5. Consider Example 5.4.1. Then for

 
1
1. f (x) = 1 − x2 ∈ R[x; 2] with B = (1, x, x2 ) as an ordered basis of R[x; 2] ⇒ (x)]B = 
 
 0 .

−1
2. (7, 5, 10, 3, 5) ∈ V = {(u, v, w, x, y)T ∈ R5 | w − x = u, v = y, u + v + x = 3y}" with #

3
B = (−1, 0, 0, 1, 0)T , (2, 1, 2, 0, 1)T as an ordered basis of V ⇒ [(7, 5, 10, 3, 5)]B =

.
5
T
AF
Remark 5.4.6. Let V be a vector space over F of dimension n. Suppose B = (v1 , . . . , vn ) is

an ordered basis of V.
DR
1. Then [αv + w]B = α[v]B + [w]B , for all α ∈ F and v, w ∈ V.

2. So, once we have fixed B, we can think of each element of V as a vector in Fn . Therefore,
if S = {w1 , . . . , wm } ⊆ V then in place of working with S, we can work with its coordinates
or equivalently with S 0 = {[w1 ]B , . . . , [wm ]B }. Hence,
(a) S is linearly independent if and only if S 0 is linearly independent in Fn .
(b) S is linearly dependent if and only if S 0 is linearly dependent in Fn .
(c) a vector v ∈ LS(S) if and only if [v]B ∈ LS(S 0 ).
h ordered basis C i= (w1 , . . . , wn ) of V, we can work with the n × n matrix

(d) for any
C = [w1 ]B · · · [wm ]B . Note that C is invertible.
(e) the symbolic expression v = B[v]B can also be thought of as a matrix equation. Thus
v = B[v]B ⇔ [v]B = B −1 v for every v ∈ V. (5.4.2)

 
1 2 3
2 1 3 ∈ Mm,n (R). If B = (e11 , e12 , e13 , . . . , e33 )
 
Example 5.4.7. Consider the matrix A =  
3 1 4
h i
is an ordered basis of M3 (R) then [A]B = 1 2 3 2 1 3 3 1 4 .
T
Thus, a little thought implies that Mm,n (R) can be mapped to Rmn with respect to the
ordered basis B = (e11 , . . . , e1n , e21 , . . . , e2n , . . . , em1 , . . . , emn ) of Mm,n (R).
5.4. ORDERED BASES 133
The next definition relates the coordinates of a vector with respect to two distinct ordered
bases. This allows us to move from one ordered basis to another ordered basis.
Definition 5.4.8. Let V be a vector space over F with dim(V) = n. Let A = [v1 , . . . , vn ] and
B = [u1 , . . . , un ] be basis matrices corresponding to the ordered bases A and B, respectively, of
V. Thus, continuing with the symbolic expression in Equation (5.4.1), we have
A = [v1 , . . . , vn ] = [B[v1 ]B , . . . , B[vn ]B ] = B [[v1 ]B , . . . , [vn ]B ] = B[A]B , (5.4.3)
where [A]B = [[v1 ]B , . . . , [vn ]B ], is called the matrix of A with respect to the ordered basis
B or the change of basis matrix from A to B.
We now summarize the ideas related with ordered bases. This also helps us to understand
the nomenclature ‘change of basis matrix’ for the matrix [A]B .
Theorem 5.4.9. Let V be a vector space over F with dim(V) = n. Further, let A = (v1 , . . . , vn )
and B = (u1 , . . . , un ) be two ordered bases of V.
1. Then the matrix [A]B is invertible. Further, Equation (5.4.2) gives [A]B = B −1 A.
2. Similarly, the matrix [B]A is invertible and [B]A = A−1 B.
3. Moreover, [x]B = [A]B [x]A , for all x ∈ V, i.e., [A]B takes coordinate vector of x with
respect to A to the coordinate vector of x with respect to B.
T
AF
4. Similarly, [x]A = [B]A [x]B , for all x ∈ V.

DR
5. Furthermore ([A]B )−1 = [B]A .
Proof. Part 1: Note that using Equation (5.4.3), we see that the matrix [A]B takes a linearly
independent set to another linearly independent set. Hence, by Exercise 3.3.17, the matrix [A]B
is invertible, which proves Part 1. A similar argument gives Part 2.
Part 3: Using Equation (5.4.2), [x]B = B −1 x = B −1 (AA−1 )x = (B −1 A)(A−1 x) = [A]B [x]A ,
for all x ∈ V. A similar argument gives Part 4 and clearly Part 5.
Example 5.4.10.
1. Let V = Rn , A = [v1 , . . . , vn ] and B = (e1 , . . . , en ) be the standard ordered basis. Then
A = [v1 , . . . , vn ] = [[v1 ]B , . . . , [vn ]B ] = [A]B .
2. Suppose A = (1, 0, 0)T , (1, 1, 0)T , (1, 1, 1)T and B = (1, 1, 1)T , (1, −1, 1)T , (1, 1, 0)T are

two ordered bases of R3 . Then, we verify the statements in the previous result.
   −1    
x 1 1 1 x x−y
       
y  = 0 1 1 y  =  y − z .
(a) Using Equation (5.4.2),        
z 0 0 1 z z
A
   −1       
x 1 1 1 x −1 1 2 x −x + y + 2z
      1   1  
(b) Similarly, 
y  = 1 −1 1 y  = 2  1 −1 0 y  = 2 
         x−y .

z 1 1 0 z 2 0 −2 z 2x − 2z
B
   
−1/2 0 1 0 2 0
   
(c) [A]B = 
 1/2 0 0 , [B]A = 0 −2 1 and [A]B [B]A = I3 .
  
1 1 0 1 1 0
Exercise 5.4.11. In R3 , let A = (1, 2, 0)T , (1, 3, 2)T , (0, 1, 3)T be an ordered basis.

1. If B = (1, 2, 1)T , (0, 1, 2)T , (1, 4, 6)T is another ordered basis of R3 . Then, determine

[A]B , [B]A and verify that [A]B [B]A = I3 .

 
2 1 3
 
2. Determine the ordered basis C such that [A]C = 
3 1 2.

0 0 1
5.5 Matrix of a linear transformation

In Example 5.1.3.7, we saw that for each A ∈ Mm×n (R) there exists a linear transformation
T ∈ L(Rn , Rm ) given by T (x) = Ax, for all x ∈ Rn . In this section, we prove that if V
and W are finite dimensional vector spaces over F with ordered bases A = (v1 , . . . , vn ) and
B = (w1 , . . . , wm ), respectively, then any T ∈ L(V, W) corresponds to an m × n matrix.
To understand it let A = [v1 , . . . , vn ] and B = [w1 , . . . , wm ] be the basis matrix of A and
B, respectively. Thus, for any T ∈ L(V, W) and v ∈ V, using the symbolic expression in
Equation (5.4.1), we see that T (v) = B[T(v)]B and v = A[v]A . Hence, for any x ∈ V
T
h i
AF
B[T(x)]B = T (x) = T ([v1 , . . . , vn ][x]A ) = T (v1 ) · · · T (vn ) [x]A

h i h i
DR
= B[T (v1 )]B · · · B[T (vn )]B [x]A = B [T (v1 )]B · · · [T (vn )]B [x]A .
As we can think of B as an invertible matrix (see Equation (5.4.2)), we get
[T(x)]B = [[T (v1 )]B , . . . , [T (vn )]B ] [x]A , for each x ∈ V.

h i
Note that the matrix [T (v1 )]B · · · [T (vn )]B , denoted T [A, B], is an m × n matrix and is
unique with respect to the ordered bases A and B as the i-th column equals [T (vi )]B , for the
i-th vector vi ∈ A, 1 ≤ i ≤ n. So, we immediately have the following definition and result.
Definition 5.5.1. Let A = (v1 , . . . , vn ) and B = (w1 , . . . , wm ) be ordered bases of V and W,

respectively. If T ∈ L(V, W) then the matrix T [A, B] is called the coordinate matrix of T or
the matrix of the linear transformation T with respect to the bases A and B, respectively.
When there is no mention of bases, we take it to be the standard ordered bases and denote
the corresponding matrix by [T ]. Also, note that for each x ∈ V, the matrix T [A, B][x]A is
the coordinate vector of T (x) with respect to the ordered basis B of the co-domain. Thus,
the matrix T [A, B] takes coordinate vector of the domain points to the coordinate vector of its
images. The above discussion is stated as the next result.
Theorem 5.5.2. Let A = (v1 , . . . , vn ) and B = (w1 , . . . , wm ) be ordered bases of V and W,

respectively. If T ∈ L(V, W) then there exists a matrix S ∈ Mm×n (F) with
h i
S = T [A, B] = [T (v1 )]B · · · [T (vn )]B and [T (x)]B = S [x]A , for all x ∈ V.
5.5. MATRIX OF A LINEAR TRANSFORMATION 135
𝒜 ℬ
[x]𝒜 [𝑇 (x)]ℬ = 𝑇 [𝒜, ℬ] [x]𝒜

‖ ‖
x 𝑇 (x)
𝕍 𝕎
Figure 5.1: Matrix of the Linear Transformation
See Figure 5.1 for clarity on which basis occurs at which place.
Remark 5.5.3. Let V and W be vector spaces over F with ordered bases A1 = (v1 , . . . , vn )
and B1 = (w1 , . . . , wm ), respectively. Also, for α ∈ F with α 6= 0, let A2 = (αv1 , . . . , αvn ) and
B2 = (αw1 , . . . , αwm ) be another set of ordered bases of V and W, respectively. Then, for any
T ∈ L(V, W)
h i h i
T [A2 , B2 ] = [T (αv1 )]B2 · · · [T (αvn )]B2 = [T (v1 )]B1 · · · [T (vn )]B1 = T [A1 , B1 ].
Thus, the same matrix can be the matrix representation of T for two different pairs of bases.
T
AF
We now give a few examples to understand the above discussion and Theorem 5.5.2.
DR
Q = (0, 1)
Q′ = (− sin θ, cos θ)
P ′ = (x′ , y ′ )
θ P ′ = (cos θ, sin θ)
θ P = (x, y)
P = (1, 0)
θ α
O O
1
Figure 5.2: Counter-clockwise Rotation by an angle θ
Example 5.5.4. 1. Let T ∈ L(R2 ) represent a counter-clockwise rotation by an angle θ, for

some θ ∈ [0, 2π]. Then, using the right figure in Figure 5.2, we see that x = OP cos α
and y = OP sin α. Thus, verify that
" # " # " # " #" #
x0 OP 0 cos(α + θ) OP cos α cos θ − sin α sin θ cos θ − sin θ x
= = = .
y0 OP 0 sin(α + θ) OP sin α cos θ + cos α sin θ sin θ cos θ y
Or equivalently, using the left figure in Figure 5.2 we see that the matrix in the standard
ordered basis of R2 equals
" #
h i cos θ − sin θ
[T ] = T (e1 ), T (e2 ) = . (5.5.1)
sin θ cos θ
2. Let T ∈ L(R2 ) with T ((x, y)T ) = (x + y, x − y)T .

" #
h i 1 1
(a) Then [T ] = [T (e1 )] [T (e2 )] = .
1 −1
" # " #!
1 1
(b) On the image space take the ordered basis as B = , . Then
0 1
"" # " # # " #
h i 1 1 0 2
[T ] = [T (e1 )]B [T (e2 )]B = = .
1 −1 1 −1
B B " # " #!
−1 3
(c) In the above, let the ordered basis of the domain space be A = , . Then
1 1
"" " ## " " ## # "" # " # # " #
−1 3 0 4 2 2
T [A, B] = T T = = .
1 1 −2 2 −2 2
B B B B
3. Let A = (e1 , e2 ) and B = (e1 + e2 , e1 − e2 ) be two ordered bases of R2 . Then Compute

T [A, A] and T [B, B], where T ((x, y)T ) = (x + y, x − 2y)T .
Solution:
" Note
# that the bases matrices for the two ordered
" #bases are A = Id2 and
1 1 1 1 1
B= , respectively. So, A−1 = Id2 and B −1 = . Thus,
1 −1 2 1 −1
"" " #!# " " #!# # "" # " # # " #
1 0 1 1 1 1
T [A, A] = T , T = , = and
0 1 1 −2 1 −2
" #!#A " A
# "" A # " #A # "
T
"" " #!# #

1 3
1 1 2 0
AF
2 2
T [B, B] = T , T = , = 3 3
1 −1 −1 3 2 −2
DR
B B B B
" # " # " # " #
2 2 0 −1 0
as = B −1 and =B .
−1 −1 3 3
B B
4. Let T ∈ L(R3 , R2 ) be defined by T ((x, y, z)T ) = (x + y − z, x + z)T . Determine [T ].

Solution: By definition
"" # " # " ## " #
1 1 −1 1 1 −1
[T ] = [[T (e1 )], [T (e2 )], [T (e3 )]] = , , = .
1 0 1 1 0 1
5. Define T ∈ L(C3 ) by T (x) = x, for all x ∈ C3 . Note that T is the Id map. De-

termine the coordinate matrix with respect to the ordered basis A = e1 , e2 , e3 and

B = (1, 0, 0), (1, 1, 0), (1, 1, 1) .
Solution: By definition, verify that
        
1 0 0 1 −1 0
        
T [A, B] = [[T (e1 )]B , [T (e2 )]B , [T (e3 )]B ] = 
0 , 1 , 0
       = 0 1 −1
  
0 0 1 0 0 1
B B B
and         
1 1 1 1 1 1
        
T [B, A] = 
0 , 1 , 1
       = 0 1 1 .
  
0 0 1 0 0 1
A A A
5.5. MATRIX OF A LINEAR TRANSFORMATION 137
Thus, verify that T [B, A]−1 = T [A, B] and T [A, A] = T [B, B] = I3 as the given map is
indeed the identity map.
We now give a remark which relates the above ideas with respect to matrix multiplication.
Remark 5.5.5. 1. Fix S ∈ Mn (C) and define T ∈ L(Cn ) by T (x) = Sx, for all x ∈ Cn . If
A is the standard basis of Cn then [T ] = S as
[T ][:, i] = [T (ei )]A = [S(ei )]A = [S[:, i]]A = S[:, i], for 1 ≤ i ≤ n.
2. Fix S ∈ Mm,n (C) and define T ∈ L(Cn , Cm ) by T (x) = Sx, for all x ∈ Cn . Let A and B
be the standard ordered bases of Cn and Cm , respectively. Then T [A, B] = S as
(T [A, B])[:, i] = [T (ei )]B = [Sei ]B = [S[:, i]]B = S[:, i], for 1 ≤ i ≤ n.
3. Fix S ∈ Mn (C) and define T ∈ L(Cn ) by T (x) = Sx, for all x ∈ Cn . Let A = (v1 , . . . , vn )
and B = (u1 , . . . , un ) be two ordered bases of Cn with respective basis matrices A and B.
Then
h i h i
T [A, B] = [T (v1 )]B · · · [T (v1 )]B = B −1 T (v1 ) · · · B −1 T (v1 )
h i h i
= B −1 Sv1 · · · B −1 Sv1 = B −1 S v1 · · · vn = B −1 SA.
T
In particular, if A = B then T [A, A] = A−1 SA. Thus, if S = In then

AF
(a) T = Id and Id[A, A] = In .

DR
(b) Id[A, B] = B −1 A, an invertible matrix.

(c) Similarly, Id[B, A] = A−1 B. So, Id[B, A] · Id[A, B] = (A−1 B)(B −1 A) = In .
1. Let T (x, y)T = (x + y, x − y)T and A = (e1 , e1 + e2 ) be the ordered

Example 5.5.6.
" # " #
1 1 1 1
basis of R . Then, for S =
2 , T (x) = Sx. Further, if A = is the basis
1 −1 0 1
matrix of A then using Remark 5.5.5.3a, we obtain
" #" #" # " #" # " #
−1 1 −1 1 1 1 1 1 −1 1 2 0 2
T [A, A] = A SA = = = .
0 1 1 −1 0 1 0 1 1 0 1 0
2. [Finding T from T [A, B]] Let V and W be vector spaces over F with ordered bases A and
B, respectively. Suppose we are given the matrix S = T [A, B]. Then to determine the
corresponding T ∈ L(V, W), we go back to the symbolic expression in Equation (5.4.1)
and Theorem 5.5.2. We see that
(a) T (v) = B[T (v)]B = BT [A, B][v]A = BS[v]A .
(b) In particular, if V = W = Fn and A = B then T (v) = BSB −1 v.
(c) Further, if B is the standard ordered basis then T (v) = Sv.
Exercise 5.5.7. 1. Relate Remark 5.5.5.3 with Theorem 5.4.9 as Id is the identity map.
2. Verify Remark 5.5.5 from different examples in Example 5.5.4.

3. Let T ∈ L(R2 ) represent the reflection about the line y = mx. Find [T ].
4. Let T ∈ L(R3 ) represent the reflection about/across the X-axis. Find [T ]. What about the
reflection across the XY -plane?
5. Let T ∈ L(R3 ) represent the counter-clockwise rotation around the positive Z-axis by an
angle θ, 0 ≤ θ < 2π. Findits matrix with respect to the standard ordered basis of R3 .
cos θ − sin θ 0
 
[Hint: Is  sin θ cos θ 0

 the required matrix?]
0 0 1
6. Define a function D ∈ L(R[x; n]) by D(f (x)) = f 0 (x). Find the matrix of D with respect
to the standard ordered basis of R[x; n]. Observe that Rng(D) ⊆ R[x; n − 1].
5.6 Similarity of Matrices

Let V be a vector space over F with dim(V) = n and ordered basis B. Then any T ∈ L(V)
corresponds to a matrix in Mn (F). Then in Remark 5.5.5.3 we have already seen that if A is
the standard ordered basis of Fn and B is any ordered basis of Fn with basis matrix B then
T [B, B] = B −1 T [A, A]B. Similarly, if C is any other ordered basis of Fn with basis matrix C
then T [C, C] = C −1 T [A, A]C and thus
T
AF
T [C, C] = C −1 T [A, A] C = C −1 (B T [B, B] B −1 )C = (B −1 C)−1 T [B, B] (B −1 C).

DR
This idea can be generalized to any finite dimensional vector space. To do so, we start with
the matrix of the composition of two linear transformations. This also helps us to relate matrix
multiplication with composition of two functions.
T [B, C]m×n S[C, D]p×m

(V, B, n) (W, C, m) (Z, D, p)
(ST )[B, D]p×n = S[C, D] · T [B, C]
Figure 5.3: Composition of Linear Transformations
Theorem 5.6.1 (Composition of Linear Transformations). Let V, W and Z be finite dimen-

sional vector spaces over F with ordered bases B, C and D, respectively. Also, let T ∈ L(V, W)
and S ∈ L(W, Z). Then S ◦ T = ST ∈ L(V, Z) (see Figure 5.3). Then
(ST ) [B, D] = S[C, D] · T [B, C].
Proof. Let B = (u1 , . . . , un ), C = (v1 , . . . , vm ) and D = (w1 , . . . , wp ) be the ordered bases of

V, W and Z, respectively. As (ST )(u) ∈ Z, using Theorem 5.5.2, we note that

[(ST )(u)]D = S T (u) D = S[C, D] · [T (u)]C = S[C, D] · (T [B, C] · [u]B ) .
5.6. SIMILARITY OF MATRICES 139
So, for all u ∈ V, we get (S[C, D] · T [B, C]) [u]B = [(ST )(u)]D = (ST )[B, D] [u]B . Hence
(ST ) [B, D] = S[C, D] · T [B, C].
As an immediate corollary of Theorem 5.6.1 we see that the matrix of the inverse linear
transform is the inverse of the matrix of the linear transform, whenever the inverse exists.
Theorem 5.6.2 (Inverse of a Linear Transformation). Let V is a vector space with dim(V) = n.
If T ∈ L(V) is invertible then for any ordered basis B and C of the domain and co-domain,
respectively, one has (T [C, B])−1 = T −1 [B, C]. That is, the inverse of the coordinate matrix of
T is the coordinate matrix of the inverse linear transform.
Proof. As T is invertible, T T −1 = Id. Thus, Remark 5.5.5.3 and Theorem 5.6.1 imply
In = Id[B, B] = (T T −1 )[B, B] = T [C, B] · T −1 [B, C].
Hence, by definition of inverse, T −1 [B, C] = (T [C, B])−1 and the required result follows.
Exercise 5.6.3. Find the matrix of the linear transformations given below.
1. Let B = x1 , x2 , x3 be an ordered basis of R3 . Now, define T ∈ L(R3 ) by T (x1 ) = x2 ,

T (x2 ) = x3 and T (x3 ) = x1 . Determine T [B, B]. Is T invertible?
2. Let B = 1, x, x2 , x3 be an ordered basis of R[x; 3] and define T ∈ L(R[x; 3]) by T (1) = 1,

T
T (x) = 1 + x, T (x2 ) = (1 + x)2 and T (x3 ) = (1 + x)3 . Prove that T is invertible. Also,
AF
find T [B, B] and T −1 [B, B].

DR
Let V be a finite dimensional vector space. Then, the next result answers the question “what
happens to the matrix T [B, B] if the ordered basis B changes to C?”
T [B, B]
(V, B) (V, B)
Id ◦ T
Id[B, C] Id[B, C]
T ◦ Id
(V, C) (V, C)
T [C, C]
Figure 1: Commutative Diagram for Similarity of Matrices
Figure 5.4: T [C, C] = Id[B, C] · T [B, B] · (Id[B, C])−1 - Similarity of Matrices
Theorem 5.6.4. Let B = (u1 , . . . , un ) and C = (v1 , . . . , vn ) be two ordered bases of V and Id
the identity operator. Then, for any linear operator T ∈ L(V)
T [C, C] = Id[B, C] · T [B, B] · (Id[B, C])−1 . (5.6.1)
Proof. As Id is the identity operator, the composite functions (T ◦ Id), (Id ◦ T ) from (V, B) to
(V, C) are equal (see Figure 5.4 for clarity). Hence, their matrix representations with respect to
ordered bases B and C are equal. Thus, (T ◦ Id)[B, C] = T [B, C] = (Id ◦ T )[B, C]. Thus, using
Theorem 5.6.1, we get
Id[B, C] · T [B, B] = T [B, C] = T [C, C]Id[B, C].
Hence, using Theorem 5.6.2, the required result follows.

Let V be a vector space and let T ∈ L(V). If dim(V) = n then every ordered basis B of V
gives an n × n matrix T [B, B]. So, as we change the ordered basis, the coordinate matrix of
T changes. Theorem 5.6.4 tells us that all these matrices are related by an invertible matrix.
Thus, we are led to the following definitions.
Definition 5.6.5. Let V be a vector space with ordered bases B and C. If T ∈ L(V) then,
T [C, C] = Id[B, C] · T [B, B] · Id[C, B]. The matrix Id[B, C] is called the change of basis matrix
(also, see Theorem 5.4.9) from B to C.
Definition 5.6.6. Let X, Y ∈ Mn (C). Then, X and Y are said to be similar if there exists a
non-singular matrix P such that P −1 XP = Y ⇔ X = P Y P −1 ⇔ XP = P Y .
Example 5.6.7. Let B = 1 + x, 1 + 2x + x2 , 2 + x and C = 1, 1 + x, 1 + x + x2 be ordered

bases of R[x; 2]. Then, verify that Id[B, C]−1 = Id[C, B], as
 
−1 1 −2
T
Id[C, B] = [[1]B , [1 + x]B , [1 + x + x2 ]B ] = 

 
0 0 1  and
AF
 
1 0 1
DR
 
0 −1 1
Id[B, C] = [[1 + x]C , [1 + 2x + x2 ]C , [2 + x]C ] = 
 
1 1 1 .
 
0 1 0
Exercise 5.6.8. 1. Let A ∈ Mn (R) such that tr(A) = 0. Then prove that there exists a
non-singular matrix S such that SAS −1 = B with B = [bij ] and bii = 0, for 1 ≤ i ≤ n.
2. Let V be a vector space with dim(V) = n. Let T ∈ L(V) satisfy T n−1 6= 0 but T n = 0.
Then, use Exercise 5.1.13.2 to get an ordered basis B = u, T (u), . . . , T n−1 (u) of V.

 
0 0 0 ··· 0
1 0 0 · · · 0
 
 
0 1 0 · · · 0
(a) Now, prove that T [B, B] =  .
. . . . . .. 
. . . .
. 
0 0 ··· 1 0
(b) Let A ∈ Mn (C) satisfy An−1 6= 0 but An = 0. Then, prove that A is similar to the
matrix given in Part 2a.
3. Let A be an ordered basis of a vector space V over F with dim(V) = n. Then prove that
the set of all possible matrix representations of T is given by (also see Definition 5.6.5)
{S · T [A, A] · S −1 | S ∈ Mn (F) is an invertible matrix}.

5.7. ORTHOGONAL PROJECTIONS AND APPLICATIONS 141
4. Let B1 (α, β) = {(x, y)T ∈ R2 : (x − α)2 + (y − β)2 ≤ 1}. Then, can we get a linear
transformation T ∈ L(R2 ) such that T (S) = W , where S and W are given below?
(a) S = B1 (0, 0) and W = B1 (1, 1).

(b) S = B1 (0, 0) and W = B1 (.3, 0).
(c) S = B1 (0, 0) and W = hull(±e1 , ±e2 ), where hull means the convex hull.
(d) S = B1 (0, 0) and W = {(x, y)T ∈ R2 : x2 + y 2 /4 = 1}.
(e) S = hull(±e1 , ±e2 ) and W is the interior of a pentagon.
5. Let V, W be vector spaces over F with dim(V) = n and dim(W) = m and ordered bases
B and C, respectively. Define IB,C : L(V, W) → Mm,n (F) by IB,C (T ) = T [B, C]. Show
that IB,C is an isomorphism. Thus, when bases are fixed, the number of m × n matrices
is same as the number of linear transformations.
6. Define T ∈ L(R3 ) by T ((x, y, z)T ) = (x + y + 2z, x − y − 3z, 2x + 3y + z)T . Let B be the

standard ordered basis and C = (1, 1, 1)T , (1, −1, 1)T , (1, 1, 2)T be another ordered basis

of R3 . Then find
(a) matrices T [B, B] and T [C, C].

T
(b) the matrix P such that P −1 T [B, B] P = T [C, C]. Note that P = Id[C, B].
AF
DR
5.7 Orthogonal Projections and Applications
As an application of the ideas and results related with orthogonality, we would like to go back
to the system of linear equations. So, recall that we started with the solution set of the linear
system Ax = b, for A ∈ Mm,n (C), x ∈ Cn and b ∈ Cm . We saw that if b ∈ Col(A) then the
system Ax = b is consistent and one can use the Gauss-Jordan method to get the solution set
of Ax = b. If the system is inconsistent can we talk of the ‘best possible solution’ ? How do we
define ‘Best’ ?
In most practical applications, the linear systems are inconsistent due to various reasons.
The reasons could be related with human error, or computational/rounding-off error or missing
data or there is not enough time to solve the whole linear system. So, we need to go bound
consistent linear systems. In quite a few such cases, we are interested in finding a point x ∈ Rn
such that the error vector, defined as kb − Axk has the least norm. Thus, we consider the
problem of finding x0 ∈ Rn such that
kb − Ax0 k = min{kb − Axk : x ∈ Rn }, (5.7.1)
i.e., we try to find the vector x0 ∈ Rn which is nearest to b.

To begin with, recall the following result from Page 105.
Theorem 5.7.1 (Decomposition). Let V be an ips having W as a finite dimensional subspace.

k
Suppose {f1 , . . . , fk } is an orthonormal basis of W. Then, for each b ∈ V, y =
P
hb, fi ifi is the
i=1
closest point in W from b. Thus
min{kb − Axk : x ∈ Rn } = kb − yk.
We now give a definition and then an implication of Theorem 5.7.1.
Definition 5.7.2. Let W be a finite dimensional subspace of an ips V. Then, by Theorem 5.7.1,
for each v ∈ V there exist unique vectors w ∈ W and u ∈ W⊥ with v = w + u. We thus define
the orthogonal projection of V onto W, denoted PW , by
PW : V → W by PW (v) = w.
The vector w is called the projection of v on W.
So, note that the solution x0 ∈ Rn satisfying kb − Ax0 k = min{kb − Axk : x ∈ Rn } is the
projection of b on the Col(A).
Remark 5.7.3. Let A ∈ Mm,n (R) and W = Col(A). Then, to find the orthogonal projection
PW (b), we can use either of the following ideas:
k
P
1. Determine an orthonormal basis {f1 , . . . , fk } of Col(A). Then PW (b) = hb, fi ifi . Note
T
i=1
that
AF
k k k k
!
DR
X X X X
fi (fiT b) = fi fiT b = fi fiT

x0 = PW (b) = hb, fi ifi = b = P b,
i=1 i=1 i=1 i=1
k
fi fiT is called the projection matrix of Rm onto Col(A).
P
where P =
i=1
2. By Theorem 3.6.5.2, Col(A) = Null(AT )⊥ . Hence, for b ∈ Rm there exists unique
u ∈ Col(A) and v ∈ Null(AT ) such that b = u + v. Thus, using Definition 5.7.2 and
Theorem 5.7.1, PW (b) = u.
We now give another method to obtain the vector x0 of Equation 5.7.1.
Corollary 5.7.4. Let A ∈ Mm,n (R) and b ∈ Rm . Then, x0 is a least square solution of Ax = b
if and only if x0 is a solution of the system AT Ax = AT b.
Proof. As b ∈ Rm , by Remark 5.7.3, there exists y ∈ Col(A) and v ∈ Null(AT ) such that
b = y + v and min{kb − wk | w ∈ Col(A)} = kb − yk. As y ∈ Col(A), there exists x0 ∈ Rn
such that Ax0 = y, i.e.,
min{kb − wk | w ∈ Col(A)} = kb − yk = kb − Ax0 k.
Hence (AT A)x0 = AT (Ax0 ) = AT y = AT (b − v) = AT b − 0 = AT b (as v ∈ Null(AT )).

Conversely, let x1 ∈ Rn be a solution of AT Ax = AT b, i.e., AT (Ax1 − b) = 0. To show
min{kb − Axk | x ∈ Rn } = kb − Ax1 k.

5.7. ORTHOGONAL PROJECTIONS AND APPLICATIONS 143
Note that AT (Ax1 − b) = 0 implies
0 = (x − x1 )T 0 = (x − x1 )T AT (Ax1 − b) = (Ax − Ax1 )T (Ax1 − b) = hAx1 − b, Ax − Ax1 i.
Thus, the vectors b − Ax1 and Ax1 − Ax are orthogonal and hence
kb − Axk2 = kb − Ax1 + Ax1 − Axk2 = kb − Ax1 k2 + kAx1 − Axk2 ≥ kb − Ax1 k2 .
Thus, min{kb − Axk | x ∈ Rn } = kb − Ax1 k. Hence, the required result follows.
We now give two examples to relate the above discussions.

Example 5.7.5. 1. Determine the projection of (1, 1, 1)T on Null ([1, 1, −1]).
Solution: Here A = [1, 1, −1], a basis of Null(A) equals {(1, −1, 0)T , (1, 0, 1)T }, which
is not an orthonormal set. Also, a basis of Col(AT ) equals {(1, 1, −1)T }.
1 1
(a) Method 1: Observe that { √ (1, −1, 0)T , √ (1, 1, 2)T } is a basis of Null(A). Thus,
2  6    
1 h 1 2/3 −1/3 1/3
1   1 −1 0 + 1 1 1 1 2 = −1/3 2/3 1/3
i  h i  
the projection matrix P =  −1
2   6    
0 2 1/3 1/3 1/3
   
1 2/3
   
and P 1 = 2/3.
  
T
1 4/3
AF
 
1 1 1
DR
−1 0 1  form a basis of R . Then

3
 
(b) Method 2: Then the columns of B =  
0 1 −1
   
−2 1
1    T
x=  4  is a solution of Bx = 
1. Thus, we see that (1, 1, 1) = u + v, where

3 
1 1
−2

1 T T T 4 T 2
u = (1, 1, −1) ∈ Col(A ) and v = (1, −1, 0) + (1, 0, 1) = (1, 1, 2)T ∈
3 3 3 3
T
2 2 4
Null(A). Thus, the required projection equals v = , , .
3 3 3
 
1 1
 
(c) Method 3: Since we want the projection on Null(A). Consider B =   −1 0 .

0 1
Then Null(A) =Col(B).
 Thus, we need the vector x0 , a solution of the linear sys-
1 " # " #
2 1 0
tem B T Bx = B T 
 
1. Or equivalently, we need the solution of 1 2 x = 2 . The

1
" #
2 −1 2 2 4 T

solution x0 = . Thus, the projection vector equals Bx0 = v = , , .
3 2 3 3 3
2. Find the foot of the perpendicular from the point v = (1, 2, 3, 4)T on the plane generated
by the vectors (1, 1, 0, 0)T , (1, 0, 1, 0)T and (0, 1, 1, 1)T .
(a) Method 1: Note that the three vectors lie on the plane x − y − z − 2w = 0. Then
r = (1, −1, −1, 2)T is the normal vector of the plane. Hence
4 1
v − Projr v = (1, 2, 3, 4)T − (1, −1, −1, 2)T = (3, 18, 25, 20)T
7 7
is the required projection of v.
(b) Method 2: Using the Gram-Schmidt process, we get
1 1 1
w1 = √ (1, 1, 0, 0)T , w2 = √ (1, −1, 2, 0)T , w3 = √ (−2, 2, 2, 3)T
2 6 21
as an orthonormal basis of the plane generated by the vectors (1, 1, 0, 0)T , (1, 0, 1, 0)T
and (0, 1, 1, 1)T . Thus, the projection matrix equals
   
6/7 1/7 1/7 −2/7 3
3 
6/7 −1/7 2/7 
  
 1/7 118.
X
T

P = wi wi =   and P v =
1/7 −1/7 6/7 2/7  7 25

i=1   
−2/7 2/7 2/7 3/7 20
 
1 1 0
 
1 0 1
(c) Method 3: Let A =   . Then we need x0 satisfying (AT A)x = AT b. Here
 0 1 1

0 0 1
T
AF
     
2 1 1 3 5 −2 −1
−1 1
AT A =  T T
    
1 2 1 and A b = 4. Note that (A A) = 7 −2 5 −1 and
DR
    
1 1 3 9 −1 −1 3
T T
hence the solution of the system (A A)x = A b equals
    
5 −2 −1 2 1 1 −2
T −1 T 1   1 
x = (A A) (A b) =  −2 5 −1 1 2 1 = 7  5 .
   
7
−1 −1 3 1 1 3 20
   
1 1 0   3
  −2  
1 0 1 1 118 is the nearest vector to v = (1, 2, 3, 4)T .
 
Thus, Ax =  ·  5 =
0 1 1 7   7 25

20
  
0 0 1 20
Exercise 5.7.6. 1. Let W = {(x, y, z, w) ∈ R4 : x = y, z = w} be a subspace of R4 .
Determine the matrix of the orthogonal projection.
2. Let PW1 and PW2 be the orthogonal projections of R2 onto W1 = {(x, 0) : x ∈ R} and
W2 = {(x, x) : x ∈ R}, respectively. Note that PW1 ◦ PW2 is a projection onto W1 . But,
it is not an orthogonal projection. Hence or otherwise, conclude that the composition of
two orthogonal projections need not be an orthogonal projection?
" #
1 1
3. Let A = . Then, A is idempotent but not symmetric. Now, define P : R2 → R2 by
0 0
P (v) = Av, for all v ∈ R2 . Then,
5.8. ORTHOGONAL OPERATOR AND RIGID MOTION∗ 145
(a) P is idempotent.
(b) Null(P ) ∩ Rng(P ) = Null(A) ∩ Col(A) = {0}.
(c) R2 = Null(P ) + Rng(P ). But, (Rng(P ))⊥ = (Col(A))⊥ 6= Null(A).
(d) Since (Col(A))⊥ 6= Null(A), the map P is not an orthogonal projector. In this
case, P is called a projection of R2 onto Rng(P ) along Null(P ).
4. Find all 2 × 2 real matrices A such that A2 = A. Hence, or otherwise, determine all
projection operators of R2 .
5. Let W be an (n − 1)-dimensional subspace of Rn with ordered basis BW = [f1 , . . . , fn−1 ].
Suppose B = [f1 , . . . , fn−1 , fn ] is an orthogonal ordered basis of Rn obtained by extending
n−1
BW . Now, define a function Q : Rn → Rn by Q(v) = hv, fn ifn −
P
hv, fi ifi . Then,
i=1
(a) Q fixes every vector in W⊥ .

(b) Q sends every vector w ∈ W to −w.
(c) Q ◦ Q = In .
The function Q is called the reflection operator with respect to W⊥ .

6. Let {f1 , . . . , fk } be an orthonormal basis of a subspace W of Rn . If {f1 , . . . , fn } is an
k n
T
extended orthonormal basis of Rn , define PW = fi fiT and PW⊥ = fi fiT . Then

P P
AF
i=1 i=k+1
prove that
DR
(a) In − PW = PW⊥ .
(b) (PW )T = PW and (PW⊥ )T = PW⊥ . That is, PW and PW⊥ are symmetric.
(c) (PW )2 = PW and (PW⊥ )2 = PW⊥ . That is, PW and PW⊥ are idempotent.
(d) PW ◦ PW⊥ = PW⊥ ◦ PW = 0.
5.8 Orthogonal Operator and Rigid Motion∗

We now give the definition and a few properties of an orthogonal operator in Rn .
Definition 5.8.1. A linear operator T : Rn → Rn is said to be an orthogonal operator if

kT (x)k = kxk, for all x ∈ Rn .
Example 5.8.2. Prove that the following maps T are orthogonal operators.
1. Fix a unit vector a ∈ Rn and define T : Rn → Rn by T (x) = 2hx, aia − x, for all x ∈ Rn .

Solution: Note that Proja (x) = hx, aia. So, hx, aia, x − hx, aia = 0 and
kxk2 = kx − hx, aia + hx, aiak2 = kx − hx, aiak2 + khx, aiak2 .
Thus, kx − hx, aiak2 = kxk2 − khx, aiak2 and hence
kT (x)k2 = k(hx, aia) + (hx, aia − x)k2 = khx, aiak2 + kx − hx, aiak2 = kxk2 .
" #" #
cos θ − sin θ x
2. Fix θ, 0 ≤ θ < 2π and define T : R2 → R2
by T (x) = , for all x ∈ R2 .
sin θ cos θ y
" # " #
x cos θ − y sin θ p x
Solution: Note that kT (x)k = k k = x2 + y 2 = k k.
x sin θ + y cos θ y
We now show that an operator is orthogonal if and only if it preserves the angle.
Theorem 5.8.3. Let T ∈ L(Rn ). Then, the following statements are equivalent.
1. T is an orthogonal operator.
2. hT (x), T (y)i = hx, yi, for all x, y ∈ Rn , i.e., T preserves inner product.
Proof. 1 ⇒ 2 Let T be an orthogonal operator. Then, kT (x + y)k2 = kx + yk2 . So,

kT (x)k2 + kT (y)k2 + 2hT (x), T (y)i = kT (x) + T (y)k2 = kT (x + y)k2 = kxk2 + kyk2 + 2hx, yi.
Thus, using definition again hT (x), T (y)i = hx, yi.
2 ⇒ 1 If hT (x), T (y)i = hx, yi, for all x, y ∈ Rn then T is an orthogonal operator as
kT (x)k2 = hT (x), T (x)i = hx, xi = kxk2 .
As an immediate corollary, we obtain the following result.
Corollary 5.8.4. Let T ∈ L(Rn ). Then, T is an orthogonal operator if and only if “for every
orthonormal basis {u1 , . . . , un } of Rn , {T (u1 ), . . . , T (un )} is an orthonormal basis of Rn ”.
T
Thus, if B is an orthonormal ordered basis of Rn then T [B, B] is an orthogonal matrix.

AF
DR
Definition 5.8.5. A map T : Rn → Rn is said to be an isometry or a rigid motion if

kT (x) − T (y)k = kx − yk, for all x, y ∈ Rn . That is, an isometry is distance preserving.
Observe that if T and S are two rigid motions then ST is also a rigid motion. Furthermore,
it is clear from the definition that every rigid motion is invertible.
Example 5.8.6. The maps given below are rigid motions/isometry.
1. Let k · k be a norm in Rn . If a ∈ Rn then the translation map Ta : Rn → Rn defined by

Ta (x) = x + a for all x ∈ Rn , is an isometry/rigid motion as
kTa (x) − Ta (y)k = k (x + a) − (y + a) k = kx − yk.
2. Theorem 5.8.3 implies that every orthogonal operator is an isometry.
We now prove that every rigid motion that fixes origin is an orthogonal operator.
Theorem 5.8.7. The following statements are equivalent for any map T : Rn → Rn .
1. T is a rigid motion that fixes origin.
2. T is linear and hT (x), T (y)i = hx, yi, for all x, y ∈ Rn (preserves inner product).
3. T is an orthogonal operator.
5.8. ORTHOGONAL OPERATOR AND RIGID MOTION∗ 147
Proof. We have already seen the equivalence of Part 2 and Part 3 in Theorem 5.8.3. Let us now
prove the equivalence of Part 1 and Part 2/Part 3.
If T is an orthogonal operator then T (0) = 0 and kT (x) − T (y)k = kT (x − y)k = kx − yk.
This proves Part 3 implies Part 1.
We now prove Part 1 implies Part 2. So, let T be a rigid motion that fixes 0. Thus,
T (0) = 0 and kT (x) − T (y)k = kx − yk, for all x, y ∈ Rn . Hence, in particular for y = 0, we
have kT (x)k = kxk, for all x ∈ Rn . So,
kT (x)k2 + kT (y)k2 − 2hT (x), T (y)i = kT (x) − T (y)k2 = kx − yk2 = kxk2 + kyk2 − 2hx, yi.
Thus, using kT (x)k = kxk, for all x ∈ Rn , we get hT (x), T (y)i = hx, yi, for all x, y ∈ Rn . Now,
to prove T is linear, we use hT (x), T (y)i = hx, yi, for all x, y ∈ Rn , in 3-rd and 4-th line below
to get
kT (x + y) − (T (x) + T (y)) k2 = hT (x + y) − (T (x) + T (y)) , T (x + y) − (T (x) + T (y))i

= hT (x + y), T (x + y)i − 2 hT (x + y), T (x)i
−2 hT (x + y), T (y)i + hT (x) + T (y), T (x) + T (y)i
= hx + y, x + yi − 2hx + y, xi − 2hx + y, yi
+hT (x), T (x)i + 2hT (x), T (y)i + hT (y), T (y)i
= −hx + y, x + yi + hx, xi + 2hx, yi + hy, yi = 0.
T
Thus, T (x+y)−(T (x) + T (y)) = 0 and hence T (x+y) = T (x)+T (y). A similar calculation
AF
gives T (αx) = αT (x) and hence T is linear.

DR
Exercise 5.8.8. 1. Let A, B ∈ Mn (C). Then, A and B are said to be

(a) Orthogonally Congruent if B = S T AS, for some orthogonal matrix S.
(b) Unitarily Congruent if B = S ∗ AS, for some unitary matrix S.
Prove that Orthogonal and Unitary congruences are equivalence relations on Mn (R) and
Mn (C), respectively.
2. Let x ∈ C2 . Identify it with the complex number x = x1 + ix2 . If we rotate x by a
counterclockwise rotation θ, 0 ≤ θ < 2π then, we have
xeiθ = (x1 + ix2 ) (cos θ + i sin θ) = x1 cos θ − x2 sin θ + i[x1 sin θ + x2 cos θ].
Thus, the corresponding vector in R2 is

" # " #" #
x1 cos θ − x2 sin θ cos θ − sin θ x1
= .
x1 sin θ + x2 cos θ sin θ cos θ x2
" #
cos θ − sin θ
Is the matrix, , the matrix of the corresponding rotation? Justify.
sin θ cos θ
" #
cos θ sin θ
3. Let A ∈ M2 (R) and T (θ) = , for θ ∈ R. Then, A is an orthogonal matrix
− sin θ cos θ
" #
0 1
if and only if A = T (θ) or A = T (θ), for some θ ∈ R.
1 0
4. Let A ∈ Mn (R). Then, the following statements are equivalent.
(a) A is an orthogonal matrix.

(b) A−1 = AT .
(c) AT is orthogonal.
(d) the columns of A form an orthonormal basis of the real vector space Rn .
(e) the rows of A form an orthonormal basis of the real vector space Rn .
(f ) for any two vectors x, y ∈ Rn , hAx, Ayi = hx, yi Orthogonal matrices preserve
angle.
(g) for any vector x ∈ Rn , kAxk = kxk Orthogonal matrices preserve length.
5. Let U ∈ Mn (C). Then, prove that the following statements are equivalent.
(a) U is a unitary matrix.

(b) U −1 = U ∗ .
(c) U ∗ is unitary.
(d) the columns of U form an orthonormal basis of the complex vector space Cn .
(e) the rows of U form an orthonormal basis of the complex vector space Cn .
(f ) for any two vectors x, y ∈ Cn , hU x, U yi = hx, yi Unitary matrices preserve
T
AF
angle.
(g) for any vector x ∈ Cn , kU xk = kxk Unitary matrices preserve length.
DR
|aij |2 = |bij |2 .
P P
6. If A = [aij ] and B = [bij ] are unitarily equivalent then prove that
ij ij
7. Let U be a unitary matrix and for every x ∈ Cn , define
kxk1 = max{|xi | : xT = [x1 , . . . , xn ]}.
Then, is it necessary that kU xk1 = kxk1 ?
5.9 Dual Space*

Definition 5.9.1. Let V be a vector space over F. Then a map T ∈ L(V, F) is called a linear
functional on V.
Example 5.9.2. 1. Let a ∈ Cn be fixed. Then, T (x) = a∗ x is a linear function from Cn to
C.
2. Define T (A) = tr(A), for all A ∈ Mn (R). Then, T is a linear functional from Mn (R) to R.
Rb
3. Define T (f ) = f (t)dt, for all f ∈ C([a, b], R). Then, T is a linear functional from
a
L(C([a, b], R) to R.
Rb
4. Define T (f ) = t2 f (t)dt, , for all f ∈ C([a, b], R). Then, T is a linear functional from
a
L(C([a, b], R) to R.
5.9. DUAL SPACE* 149
5. Define T : C3 → C by T ((x, y, z)T ) = x. Is it a linear functional?

6. Let B be a basis of a vector space V over F. For a fixed element u ∈ B, define
(
1 if x = u
T (x) =
0 if x ∈ B \ u.
Now, extend T linearly to all of V. Does, T give rise to a linear functional?
Definition 5.9.3. Let V be a vector space over F. Then L(V, F) is called the dual space of
V and is denoted by V∗ . The double dual space of V, denoted V∗∗ , is the dual space of V∗ .
We first give an immediate corollary of Theorem 5.3.12.
Corollary 5.9.4. Let V and W be vector spaces over F with dim V = n and dim W = m.
1. Then L(V, W) ∼
= Fmn . Moreover, {fij |1 ≤ i ≤ n, 1 ≤ j ≤ m} is a basis of L(V, W).
2. In particular, if W = F then L(V, F) = V∗ ∼ = Fn . Moreover, if({v1 , . . . , vn } is a basis of
1, if k = i
V then the set {fi |1 ≤ i ≤ n} is a basis of V∗ , where fi (vk ) = The basis
0, k 6= i.
{fi |1 ≤ i ≤ n} is called the dual basis of Fn .
Exercise 5.9.5. Let V be a vector space. Suppose there exists v ∈ V such that f (v) = 0, for
all f ∈ V∗ . Then prove that v = 0.
So, we see that V∗ can be understood through a basis of V. Thus, one can understand V∗∗
T
again via a basis of V∗ . But, the question arises “can we understand it directly via the vector
AF
space V itself?” We answer this in affirmative by giving a canonical isomorphism from V to V∗∗ .
DR
To do so, for each v ∈ V, we define a map Lv : V∗ → F by Lv (f ) = f (v), for each f ∈ V∗ . Then

Lv is a linear functional as
Lv (αf + g) = (αf + g) (v) = αf (v) + g(v) = αLv (f ) + Lv (g).
So, for each v ∈ V, we have obtained a linear functional Lv ∈ V∗∗ . Note that, if v 6= w then,
Lv 6= Lw . Indeed, if Lv = Lw then, Lv (f ) = Lw (f ), for all f ∈ V∗ . Thus, f (v) = f (w), for all
f ∈ V∗ . That is, f (v − w) = 0, for each f ∈ V∗ . Hence, using Exercise 5.9.5, we get v − w = 0,
or equivalently, v = w.
We use the above argument to give the required canonical isomorphism.
Theorem 5.9.6. Let V be a vector space over F. If dim(V) = n then the canonical map
T : V → V∗∗ defined by T (v) = Lv is an isomorphism.
Proof. Note that for each f ∈ V∗ ,
Lαv+u (f ) = f (αv + u) = αf (v) + f (u) = αLv (f ) + Lu (f ) = (αLv + Lu ) (f ).
Thus, Lαv+u = αLv +Lu . Hence, T (αv+u) = αT (v)+T (u). Thus, T is a linear transformation.
For verifying T is one-one, assume that T (v) = T (u), for some u, v ∈ V. Then, Lv = Lu . Now,
use the argument just before this theorem to get v = u. Therefore, T is one-one.
Thus, T gives an inclusion (one-one) map from V to V∗∗ . Further, applying Corollary 5.9.4.2
to V∗ , gives dim(V∗∗ ) = dim(V∗ ) = n. Hence, the required result follows.
We now give a few immediate consequences of Theorem 5.9.6.
Corollary 5.9.7. Let V be a vector space of dimension n with basis B = {v1 , . . . , vn }.

1. Then, a basis of V∗∗ , the double dual of V, equals D = {Lv1 , . . . , Lvn }. Thus, for each
T ∈ V∗∗ there exists x ∈ V such that T (f ) = f (x), for all f ∈ V∗ . Or equivalently, there
exists x ∈ V such that T = Tx .
2. If C = {f1 , . . . , fn } is the dual basis of V∗ defined using the basis B (see Corollary 5.9.4.2)
then D is indeed the dual basis of V∗∗ obtained using the basis C of V∗ . Thus, each basis
of V∗ is the dual basis of some basis of V.
Proof. Part 1 is direct as T : V → V∗∗ was a canonical inclusion map. For Part 2, we need to
show that
( (
1, if j = i 1, if j = i
Lvi (fj ) = or equivalently fj (vi ) =
0, if j 6= i 0, if j 6= i
which indeed holds true using Corollary 5.9.4.2.
Let V be a finite dimensional vector space. Then Corollary 5.9.7 implies that the spaces V
and V∗ are naturally dual to each other.
We are now ready to prove the main result of this subsection. To start with, let V and W
be vector spaces over F. Then, for each T ∈ L(V, W), we want to define a map Tb : W∗ → V∗ .
So, if g ∈ W∗ then, Tb(g) a linear functional
from V to F. So, we need to be evaluate T (g) at
b
an element of V. Thus, we define Tb(g) (v) = g (T (v)), for all v ∈ V. Now, we note that
Tb ∈ L(W∗ , V∗ ), as for every g, h ∈ W∗ ,
T
AF

Tb(αg + h) (v) = (αg + h) (T (v)) = αg (T (v)) + h (T (v)) = αTb(g) + Tb(h) (v),
DR
for all v ∈ V implies that Tb(αg + h) = αTb(g) + Tb(h).
Theorem 5.9.8. Let V and W be vector spaces over F with ordered bases A = (v1 , . . . , vn )
and B = (w1 , . . . , wm ), respectively. Also, let A∗ = (f1 , . . . , fn ) and B ∗ = (g1 , . . . , gm ) be the
corresponding ordered bases of the dual spaces V∗ and W∗ , respectively. Then,
Tb[B ∗ , A∗ ] = (T [A, B])T ,
the transpose of the coordinate matrix T .

hh i h i i
Proof. Note that we need to compute Tb[B ∗ , A∗ ] = Tb(g1 ) ∗
, . . . , T
b(gm ) and prove that
A A∗
it equals the transpose of the matrix T [A, B]. So, let
 
a11 a12 ··· a1n
 
 a21 a22 · · · a2n 
T [A, B] = [[T (v1 )]B , . . . , [T (vn )]B ] =  . .
 
 .. .. .. .. 
 . . . 

am1 am2 · · · amn
Thus, to prove the required result, we need to show that
 
aj1
n
 
h i  aj2  X
Tb(gj ) ∗ = [f1 , . . . , fn ]  .  = ajk fk , for 1 ≤ j ≤ m. (5.9.1)
 
A  .. 
  k=1
ajn
5.10. SUMMARY 151
n n

P P
Now, recall that the functionals fi ’s and gj ’s satisfy αk fk (vt ) = αk (fk (vt )) = αt ,
k=1 k=1
for 1 ≤ t ≤ n and [gj (w1 ), . . . , gj (wm )] = eTj , a row vector with 1 at the j-th place and 0,
elsewhere. So, let B = [w1 , . . . , wm ] and evaluate Tb(gj ) at vt ’s, the elements of A.

Tb(gj ) (vt ) = gj (T (vt )) = gj (B [T (vt )]B ) = [gj (w1 ), . . . , gj (wm )] [T (vt )]B
 
a1t
n
  !
a 
T  2t 
X
= ej  .  = ajt = ajk fk (vt ).
 .. 
  k=1
amt
n
P
Thus, the linear functional Tb(gj ) and ajk fk are equal at vt , for 1 ≤ t ≤ n, the basis vectors
k=1
n
of V. Hence Tb(gj ) =
P
ajk fk which gives Equation (5.9.1).
k=1
Remark 5.9.9. The proof of Theorem 5.9.8 also shows the following.
1. For each T ∈ L(V, W) there exists a unique map Tb ∈ L(W∗ , V∗ ) such that

Tb(g) (v) = g (T (v)) , for each g ∈ W∗ .
2. The coordinate matrices T [A, B] and Tb[B ∗ , A∗ ] are transpose of each other, where the
T
ordered bases A∗ of V∗ and B ∗ of W∗ correspond, respectively, to the ordered bases A of

AF
V and B of W.
DR
3. Thus, the results on matrices and its transpose can be re-written in the language of a
vector space and its dual space.
5.10 Summary
T
AF
DR
Chapter 6
Eigenvalues, Eigenvectors and

Diagonalizability
6.1 Introduction and Definitions
Note that we have been trying to solve the linear system Ax = b. But, in most cases, we are
not able to solve it because of certain restrictions. Hence in the last chapter, we looked at the
nearest solution or obtained the projection of b on the column space of A.
T
AF
These problems arise from the fact that either our data size is too large or there are missing
informations (data is incomplete or the data has ambiguities or the data is inaccurate) or the
DR
data is coming too fast in the sense that our computational power doesn’t match the speed at
which data is received or it could be any other reason. So, to take care of such issues, we either
work with a submatrix of A or with the matrix AT A. We also try to concentrate on only a few
important aspects depending on our past experience.
Thus, we need to find certain set of critical vectors/directions associated with the given
linear system. Hence, in this chapter, all our matrices will be square matrices. They will have
real numbers as entries for convenience. But, we need to work over complex numbers. Hence,
we will be working with Mn (C) and x = (x1 , . . . , xn )T ∈ Cn , for some n ∈ N. Further, Cn will
be considered only as a complex vector space. We start with an example for motivation.
Example 6.1.1. Let A be a real symmetric matrix. Consider the following problem:
Maximize (Minimize) xT Ax such that x ∈ Rn and xT x = 1.
To solve this, consider the Lagrangian
n X
n n
!
X X
T T
L(x, λ) = x Ax − λ(x x − 1) = aij xi xj − λ x2i −1 .
i=1 j=1 i=1
153
154 CHAPTER 6. EIGENVALUES, EIGENVECTORS AND DIAGONALIZABILITY
Partially differentiating L(x, λ) with respect to xi for 1 ≤ i ≤ n, we get
∂L
= 2a11 x1 + 2a12 x2 + · · · + 2a1n xn − 2λx1 ,
∂x1
.. ..
.=.
∂L
= 2an1 x1 + 2an2 x2 + · · · + 2ann xn − 2λxn .
∂xn
Therefore, to get the points of extremum, we solve for
T
T ∂L ∂L ∂L ∂L
0 = , ,..., = = 2(Ax − λx).
∂x1 ∂x2 ∂xn ∂x
Thus, to solve the extremal problem, we need λ ∈ R, x ∈ Rn such that x 6= 0 and Ax = λx.
Note that we could have started with a Hermitian matrix and arrived at a similar situation.
So, in previous chapters, we had looked at Ax = b, where A and b were known. Here, we need
to solve Ax = λx with x 6= 0. Note that 0 is already a solution and is not of interest to us.
Further, we will see that we are interested in only those solutions of Ax = λx which are linearly
independent. To proceed further, let us take a few examples, where we will try to look at what
does the system Ax = b imply?
" # " # " #
1 2 9 −2 x
Example 6.1.2. 1. Let A = ,B= and x = .
2 1 −2 6 y
T
AF
" # " # " #

1 1 1
(a) Then A magnifies the nonzero vector three times as A =3 and behaves
DR
1 1 1
" # " # " # " #
1 1 1 1
by changing the direction of as A = −1 . Further, the vectors
−1 −1 −1 1
" #
1
and are orthogonal.
−1
" # " # " # " # " # " #
1 −2 1 1 2 2
(b) B magnifies both the vectors and as B =5 and B = 10 .
2 1 2 2 −1 −1
" # " #
1 2
Here again, the vectors and are orthogonal.
2 −1
(x + y)2 (x − y)2
(c) xT Ax = 3 − . Here, the displacements occur along perpendicular
2 2 " # " #
1 1
lines x + y = 0 and x − y = 0, where x + y = (x, y) and x − y = (x, y) .
1 −1
(x + 2y)2 (2x − y)2
Whereas xT Bx = 5 + 10 . Here also the maximum/minimum
5 5
displacements occur
" # along the orthogonal
" # lines x + 2y = 0 and 2x − y = 0, where
1 2
x + 2y = (x, y) and 2x − y = (x, y) .
2 −1
(d) the curve xT Ax = 10 represents a hyperbola, where as the curve xT Bx = 10 rep-
resents an ellipse (see the left two curves in Figure 6.1 drawn using the package
“Sagemath”).
6.1. INTRODUCTION AND DEFINITIONS 155
Figure 6.1: A Hyperbola and two Ellipses (first one has orthogonal axes)
.
In the above two examples we looked at symmetric matrices. What if our matrix is not
symmetric?
" #
7 −2
2. Let C = , a non-symmetric matrix. Then, does there exist a non-zero x ∈ C2
2 2
which gets magnified by C?
We need x 6= 0 and α ∈ C such that Cx = αx ⇔ [αI2 −C]x = 0. As x 6= 0, [αI2 −C]x = 0
has a solution if and only if det[αI − A] = 0. But,
T
AF
" #!
α−7 2
det[αI − A] = det = α2 − 9α + 18.
DR
−2 α − 2
" # " #
2 1
So α = 6, 3. For α = 6, verify that x = 6= 0 satisfies Cx = 6x. Similarly, x =
1 2
satisfies Cx = 3x. In this example,
" # " #
2 1
(a) we still have magnifications in the directions and .
1 2
(b) the maximum/minimum displacements do not occur along the lines 2x + y = 0 and
x + 2y = 0 (see the third curve in Figure 6.1). Note that
" #
7 0
{x ∈ R2 : xT Ax = 10} = {x ∈ R2 : xT x = 10},
0 2
" #
7 0
where is a symmetrization of A.
0 2
(c) the lines 2x + y = 0 and x + 2y = 0 are not orthogonal.
We observe the following about the matrices A, B and C that appear above:
1. det(A) = −3 = 3 × −1, det(B) = 50 = 5 × 10 and det(C) = 18 = 6 × 3.
2. tr(A) = 2 = 3 − 1, tr(B) = 15 = 5 + 10 and det(C) = 9 = 6 + 3.
(" # " #) (" # " #) (" # " #)
1 1 1 2 2 1
3. The sets , , , and , are linearly independent.
1 −1 2 −1 1 2
" # " #
1 1
4. If v1 = and v2 = and S = [v1 , v2 ] then
1 −1
" # " #
3 0 3 0
(a) AS = [Av1 , Av2 ] = [3v1 , −v2 ] = S ⇔ S −1 AS = = diag(3, −1).
0 −1 0 −1
1 1
(b) Let u1 = √ v1 and u2 = √ v2 . Then, u1 and u2 are orthonormal unit vectors,
2 2
i.e., if U = [u1 , u2 ] then I = U U ∗ = u1 u∗1 + u2 u∗2 and A = 3u1 u∗1 − u2 u∗2 .
" # " #
1 2
5. If v1 = and v2 = and S = [v1 , v2 ] then
2 −1
" # " #
5 0 5 0
(a) AS = [Av1 , Av2 ] = [5v1 , 10v2 ] = S ⇔ S −1 AS = = diag(3, −1).
0 10 0 10
1 1
(b) Let u1 = √ v1 and u2 = √ v2 . Then, u1 and u2 are orthonormal unit vectors,
5 5
i.e., if U = [u1 , u2 ] then I = U U ∗ = u1 u∗1 + u2 u∗2 and A = 5u1 u∗1 + 10u2 u∗2 .
" # " # " #
2 1 6 0
6. If v1 = and v2 = and S = [v1 , v2 ] then S −1 CS = = diag(6, 3).
1 2 0 3
Thus, we see that given A ∈ Mn (C), the number λ ∈ C and x ∈ Cn , x 6= 0 satisfying Ax = λx

have certain nice properties. For example, all of them are similar to diagonal matrices. That
is, for each matrix discussed above, there exists a basis of C2 with respect to which the matrix
T
representation is a diagonal matrix. To understand the ideas better, we start with the following
AF
definitions.
DR
Definition 6.1.3. Let A ∈ Mn (C). Then the equation
Ax = λx ⇔ (A − λIn )x = 0 (6.1.1)
is called the eigen-condition.

1. An α ∈ C is called a characteristic value/root or eigenvalue or latent root of A if
there exists x 6= 0 satisfying Ax = αx.
2. A x 6= 0 satisfying Equation (6.1.1) is called a characteristic vector or eigenvector or
invariant/latent vector of A corresponding to λ.
3. The tuple (α, x) with x 6= 0 and Ax = αx is called an eigen-pair or characteristic-pair.
4. For an eigenvalue α ∈ C, Null(A − αI) = {x ∈ Rn |Ax = αx} is called the eigen-space
or characteristic vector space of A corresponding to α.
Theorem 6.1.4. Let A ∈ Mn (C) and α ∈ C. Then α is an eigenvalue of A if and only if

det(A − αIn ) = 0.
Proof. Let B = A − αIn . Then, by definition, α is an eigenvalue of A if any only if the system
Bx = 0 has a non-trivial solution. By Theorem 2.6.3 this holds if and only if det(B) = 0.
Definition 6.1.5. Let A ∈ Mn (C). Then det(A − λI) is a polynomial of degree n in λ and is
called the characteristic polynomial of A, denoted PA (λ), or in short P (λ).
1. The equation PA (λ) = 0 is called the characteristic equation of A.
2. The multi-set (collection with multiplicities) {α ∈ C : PA (α) = 0} is called the spectrum

of A, denoted σ(A). Hence, σ(A) contains all the eigenvalues of A containing multiplicities.
3. The Spectral Radius, denoted ρ(A), of A ∈ Mn (C) equals max{|α| : α ∈ σ(A)}.
We thus observe the following.
Remark 6.1.6. Let A ∈ Mn (C). Then A is singular if and only if 0 ∈ σ(A). Further, the
following statements hold.
1. If α ∈ σ(A) then
(a) {0} $ Null(A − αI). Therefore, if Rank(A − αI) = r then r < n. Hence, by
Theorem 2.6.3, the system (A − αI)x = 0 has n − r linearly independent solutions.
(b) v ∈ Null(A−αI) if and only if cv ∈ Null(A−αI), for c 6= 0. Thus, an eigenvector
v of A is in some sense a line ` = Span({v}) that passes through 0 and v and has
the property that the image of ` is either ` itself or 0.
r
ci xi ∈ Null(A − αI), for all ci ∈ C. Hence, if
P
(c) If x1 , . . . , xr ∈ Null(A − αI) then
i=1
S is a collection of eigenvectors then, we necessarily want the set S to be linearly
independent.
T
2. α ∈ σ(A) if and only if α is a root of PA (x) ∈ C[x]. As deg(PA (x)) = n, A has exactly n
AF
eigenvalues in C, including multiplicities.

DR
3. Let (α, x) be an eigen-pair of A ∈ Mn (R). If α ∈ R then x ∈ Rn .
4. Let (α, x) be an eigen-pair of A. Then A2 x = A(Ax) = A(αx) = α(Ax) = α2 x. Thus,

the polynomial f (A) = b0 I + b1 A + · · · + bk Ak (in A) has (f (α), x) as an eigen-pair.
Almost all books in mathematics differentiate between characteristic value and eigenvalue
as the ideas change when one moves from complex numbers to any other scalar field. We give
the following example for clarity.
Remark 6.1.7. Let A ∈ M2 (F). Then, A induces a map T ∈ L(F2 ) defined by T (x) = Ax, for
all x ∈ F2 . We use this idea to understand the difference.
" #
0 1
1. Let A = . Then, pA (λ) = λ2 + 1.
−1 0
(a) A has no characteristic value in R as λ2 + 1 = 0 has no root in R.

(b) A has ±i as the roots of P (λ) = 0 in C. Hence, verify that A has (i, (1, i)T ) and
(−i, (i, 1)T ) as eigen-pairs or characteristic-pairs.
" #
1 2 √
2. Let A = . Then 2 ± 3 are the roots of the characteristic equation.
1 3
(a) Hence A has no characteristic value over Q.
(b) But A has characteristic values or eigenvalues over R.
Let us look at some more examples. Also, as stated earlier, we look at roots of the charac-
teristic equation over C.
Example 6.1.8. 1. Let A = diag(d1 , . . . , dn ) with di ∈ C, 1 ≤ i ≤ n. Then, p(λ) =

n
Q
(λ − di ) and thus verify that (d1 , e1 ), . . . , (dn , en ) are the eigen-pairs.
i=1
" #
1 0
2. Let A = . Then, 1 is a repeated eigenvalue of A. In this case, (A − I2 )x = 0 has a
0 1
solution for every x ∈ C2 . Hence, any two linearly independent vectors xT , yT ∈ C2
gives (1, x) and (1, y) as the two eigen-pairs for A. In general, if S = {x1 , . . . , xn } is a
basis of Cn then (1, x1 ), . . . , (1, xn ) are eigen-pairs of In , the identity matrix.
" #
1 1
3. Let A = . Then, p(λ) = (1 − λ)2 . Hence, σ(A) = {1, 1}. But the complete solution
0 1
of the system (A − I2 )x = 0 equals x = ce1 , for c ∈ C. Hence using Remark 6.1.6.2, e1 is
an eigenvector. Therefore, 1 is a repeated eigenvalue whereas there is only one
eigenvector.
n
Q
4. Let A = (aij ) be an n × n triangular matrix. Then, p(λ) = (λ − aii ) and thus verify
i=1
that σ(A) = {a11 , a22 , . . . , ann }. What can you say about the eigenvectors if the diagonal
entries of A are all distinct?
T
" # " #! " #!

AF
1 −1 i 1
5. Let A = . Then, 1 + i, and 1 − i, are the eigen-pairs of A.
1 1 1 i
DR
 
0 1 0
 
6. Let A = 0 0 1 . Then, σ(A) = {0, 0, 0} with e1 as the only eigenvector.

0 0 0
   
0 1 0 0 0 x1
   
0 0 1 0 0 x2 
   
7. Let A = 0 0 0 0 0. Then, σ(A) = {0, 0, 0, 0, 0}. Note that Ax3  = 0 implies
   
   
0 0 0 0 1 x 
   4
0 0 0 0 0 x5
x2 = 0 = x3 = x5 . Thus, e1 and e4 are the only eigenvectors. Note that the diagonal
blocks of A are nilpotent matrices.
Exercise 6.1.9. 1. Prove that the matrices A and AT have the same set of eigenvalues.
Construct a 2 × 2 matrix A such that the eigenvectors of A and AT are different.
2. Prove that λ ∈ C is an eigenvalue of A if and only if λ ∈ C is an eigenvalue of A∗ .
3. Let A ∈ Mn (R) be an idempotent matrix. Determine its eigenvalues and eigenvectors.
4. Let A be a nilpotent matrix. Then, prove that its eigenvalues are all 0.
5. Let J = 11T ∈ Mn (C). Then, J is a matrix with each entry 1. Show that
(a) (n, 1) is an eigenpair for J.

(b) 0 ∈ σ(J) with multiplicity n − 1. Find a set of n − 1 linearly independent eigenvectors
for 0 ∈ σ(J).
6. Let A = [aij ] ∈ Mn (R), where aij = a, if i = j and b, otherwise. Then, verify that
A = (a − b)I + bJ. Hence, or otherwise determine the eigenvalues and eigenvectors of J.
7. Let A ∈ Mn (R) be a matrix of rank 1. Determine its eigen-pairs.

" # " #
cos θ − sin θ cos θ sin θ
8. For a fixed θ ∈ R, find eigen-pairs of A = and R = .
sin θ cos θ sin θ − cos θ
9. Let A ∈ Mn (C) satisfy kAxk ≤ kxk for all x ∈ Cn . Then prove that every eigenvalue of
A lies between −1 and 1.
n
10. Let A = [aij ] ∈ Mn (C) with
P
aij = a, for all 1 ≤ i ≤ n. Then, prove that a is an
j=1
eigenvalue of A with corresponding eigenvector 1 = [1, 1, . . . , 1]T .
" #
B 0
11. Let B ∈ Mn (C) and C ∈ Mm (C). Let Z = . Then
0 C
" #!
x
(a) (α, x) is an eigen-pair for B implies α, is an eigen-pair for Z.
0
T
AF
" #!
0
(b) (β, y) is an eigen-pair for C implies β, is an eigen-pair for Z.
DR
Definition 6.1.10. Let A ∈ L(Cn ). Then, a vector y ∈ Cn \ {0} satisfying y∗ A = λy∗ is called
a left eigenvector of A for λ.
" # " # " # " # " #
7 −2 2 1 2 1
Example 6.1.11. Let A = , x= , y = , u= and v = . Then
2 2 1 2 −1 −2
verify that (6, x) and (3, y) are (right) eigen-pairs of A and (6, u), (3, v) are left eigen-pairs of
A. Note that xT v = 0 and yT u = 0. This is true in general and is proved next.
Theorem 6.1.12. [Principle of bi-orthogonality] Let (λ, x) be a (right) eigen-pair and (µ, y)
be a left eigen-pair of A. If λ 6= µ then y is orthogonal to x.
Proof. Verify that µy∗ x = (y∗ A)x = y∗ (Ax) = y∗ (λx) = λy∗ x. Thus y∗ x = 0.
Exercise 6.1.13. 1. Let Ax = λx and x∗ A = µx∗ . Then µ = λ.
2. Let S be a non-singular matrix such that its columns are left eigenvectors of A. Then,
prove that the columns of (S ∗ )−1 are right eigenvectors of A.
Definition 6.1.14. Let T ∈ L(Cn ). Then α ∈ C is called an eigenvalue of T if there exists

v ∈ Cn with v 6= 0 such that T (v) = αv.
Proposition 6.1.15. Let T ∈ L(Cn ) and let B be an ordered basis in Cn . Then (α, v) is an
eigen-pair of T if and only if (α, [v]B ) is an eigen-pair of A = T [B, B].
Proof. By definition, T (v) = αv if and only if [T v]B = [αv]B . Or equivalently, α ∈ σ(T ) if and
only if A[v]B = α[v]B . Thus, the required result follows.
Thus, the spectrum of a linear operator is independent of the choice of basis.
Remark 6.1.16. We give two examples to show that a linear operator on an infinite
dimensional vector space need not have an eigenvalue.
1. Let V be the space of all real sequences (see Example 3.1.4.7) and define T ∈ L(V) by
T (a0 , a1 , . . .) = (0, a1 , a2 , . . .).
Let if possible (α, x) be an eigen-pair of T with x = (x1 , x2 , . . .). Then
T (x) = αx ⇔ (0, x1 , x2 , . . .) = α(x1 , x2 , . . .) = (αx1 , αx2 , . . .).
So, if α 6= 0 then x1 = 0. This in turn implies x = 0, a contradiction. If α = 0 then

(0, x1 , x2 , . . .) = (0, 0, . . .) ⇒ x = 0, a contradiction. Hence, T doesn’t have an eigenvalue.
2. Recall the map T ∈ L(C[x]) defined by T (f (x)) = xf (x), for all f (x) ∈ C[x].
T has an eigen-pair (α, f (x)) ⇔ xf (x) = αf (x) ⇔ (x − α)f (x) = 0. As x is an indeter-
minate, f (x) is the zero polynomial. Hence, T cannot have an eigenvector.
We now prove the observations that det(A) is the product of eigenvalues and tr(A) is the
sum of eigenvalues.
T
AF
Theorem 6.1.17. Let λ1 , . . . , λn , not necessarily distinct, be the A = [aij ] ∈ Mn (C). Then,
n
Q Pn n
P
det(A) = λi and tr(A) = aii = λi .
DR
i=1 i=1 i=1
Proof. Since λ1 , . . . , λn are the eigenvalues of A, by definition,

n
Y
det(A − xIn ) = (−1)n (x − λi ) (6.1.2)
i=1
is an identity in x as polynomials. Therefore, by substituting x = 0 in Equation (6.1.2), we get

det(A) = (−1)n (−1)n ni=1 λi = ni=1 λi . Also,
Q Q
 
a11 − x a12 ··· a1n
 
 a21 a22 − x · · · a2n 
det(A − xIn ) =  . (6.1.3)
 
 .. .. .. .. 
 . . . 

an1 an2 · · · ann − x
= a0 − xa1 + · · · + (−1)n−1 xn−1 an−1 + (−1)n xn (6.1.4)
for some a0 , a1 , . . . , an−1 ∈ C. Then, an−1 , the coefficient of (−1)n−1 xn−1 , comes from the term
(a11 − x)(a22 − x) · · · (ann − x).

n
P
So, an−1 = aii = tr(A), the trace of A. Also, from Equation (6.1.2) and (6.1.4), we have
i=1
n
Y
n−1 n−1 n n n
a0 − xa1 + · · · + (−1) x an−1 + (−1) x = (−1) (x − λi ).
i=1
6.2. SPECTRUM OF A MATRIX 161
Therefore, comparing the coefficient of (−1)n−1 xn−1 , we have

n n
( )
X X
tr(A) = an−1 = (−1) (−1) λi = λi .
i=1 i=1
Hence, we get the required result.
Exercise 6.1.18. 1. Let A ∈ Mn (C). Then, A is invertible if and only if 0 is not an

eigenvalue of A.
2. Let A ∈ Mn (R). Then, prove that
(a) if α ∈ σ(A) then αk ∈ σ(Ak ), for all k ∈ N.

(b) if A is invertible and α ∈ σ(A) then αk ∈ σ(Ak ), for all k ∈ Z.
3. Let A be a 3 × 3 orthogonal matrix (AAT = I). If det(A) = 1, then prove that there exists
v ∈ R3 \ {0} such that Av = v.
4. Let A ∈ M2n+1 (R) with AT = −A. Then, prove that 0 is an eigenvalue of A.
6.2 Spectrum of a Matrix

Definition 6.2.1. Let A ∈ Mn (C). Then, for α ∈ σ(A)
T
AF
1. the algebraic multiplicity of α, denoted Alg.Mulα (A), is the multiplicity of α as a

DR
root of the characteristic polynomial or the number of times α ∈ σ(A).

2. the geometric multiplicity of α, denoted Geo.Mulα (A), equals dim(Null(A − αI)).
 
1 1 0
 
Example 6.2.2. 1. Let A =  0 1 1. Then σ(A) = {1, 1, 1, }. Hence, the algebraic

0 0 1
multiplicity of 1 is 3, i.e., Alg.Mul1 (A) = 3. Verify that Geo.Mul1 (A) = 1.
 
3 1 1
 
0 3 1
 0 0 
0 0 3
 

 
 2 1 0 
2. Let A =  . Then A is an upper triangular matrix and thus
 

 0 0 2 1 0 

0 0 2
 
 
 
 3 1 
 0 0 
0 3
σ(A) = {3, 3, 3, 3, 3, 2, 2, 2}, Alg.Mul3 (A) = 5 and Alg.Mul2 (A) = 3. Verify that
Rank(A − 3I) = 6, Rank(A − 2I) = 7 ⇒ Geo.Mul3 (A) = 2 and Geo.Mul2 (A) = 1.
We now show that for any eigenvalue α, the algebraic and geometric multiplicities do not
change under similarity transformation, or equivalently, under change of basis.
Theorem 6.2.3. Let A and B be two similar matrices. Then,

1. α ∈ σ(A) if and only if α ∈ σ(B).
2. for each α ∈ σ(A), Alg.Mulα (A) = Alg.Mulα (B) and Geo.Mulα (A) = Geo.Mulα (B).
Proof. Since A and B are similar, there exists an invertible matrix S such that A = SBS −1 .
So, α ∈ σ(A) if and only if α ∈ σ(B) as
det(A − xI) = det(SBS −1 − xI) = det S(B − xI)S −1

= det(S) det(B − xI) det(A−1 ) = det(B − xI). (6.2.5)
Note that Equation (6.2.5) also implies that Alg.Mulα (A) = Alg.Mulα (B). We will now
show that Geo.Mulα (A) = Geo.Mulα (B).
So, let Q1 = {v1 , . . . , vk } be a basis of Null(A − αI). Then, B = SAS −1 implies that
Q2 = {Sv1 , . . . , Svk } ⊆ Null(B − αI). Since Q1 is linearly independent and S is invertible, we
get Q2 is linearly independent. So, Geo.Mulα (A) ≤ Geo.Mulα (B). Now, we can start with
eigenvectors of B and use similar arguments to get Geo.Mulα (B) ≤ Geo.Mulα (A). Hence
the required result follows.
Remark 6.2.4. 1. Let A = S −1 BS. Then, from the proof of Theorem 6.2.3, we see that x
is an eigenvector of A for λ if and only if Sx is an eigenvector of B for λ.
2. Let A and B be two similar

" " σ(A)
matrices then
# # = σ(B). But, the converse is not true.
0 0 0 1
T
For example, take A = and B = .

AF
0 0 0 0
3. Let A ∈ Mn (C). Then, for any invertible matrix B, the matrices AB and BA =
DR
B(AB)B −1 are similar. Hence, in this case the matrices AB and BA have
(a) the same set of eigenvalues.
(b) Alg.Mulα (AB) = Alg.Mulα (BA), for each α ∈ σ(A).
(c) Geo.Mulα (AB) = Geo.Mulα (BA), for each α ∈ σ(A).
We will now give a relation between the geometric multiplicity and the algebraic multiplicity.
Theorem 6.2.5. Let A ∈ Mn (C). Then, for α ∈ σ(A), Geo.Mulα (A) ≤ Alg.Mulα (A).
Proof. Let Geo.Mulα (A) = k. So, suppose that {v1 , . . . , vk } is an orthonormal basis of
Null(A − αI). Extend it to get {v1 , . . . , vk , vk+1 , . . . , vn } as an orthonormal basis of Cn . Put
P = [v1 , . . . , vk , vk+1 , . . . , vn ]. Then P ∗ = P −1 and
P ∗ AP = P ∗ [Av1 , . . . , Avk , Avk+1 , . . . , Avn ]

   
v1∗ α ··· 0 ∗ ··· ∗
 .. 
..

. 0 . 0 ∗ · · · ∗
   

   
 v∗  0 · · · α ∗ · · · ∗
=  ∗ k  [αv1 , . . . , αvk , ∗, . . . , ∗] =  .
   
vk+1 
 
0
 · · · 0 ∗ · · · ∗

 .. .
 ..


 .  


vn∗ 0 ··· 0 ∗ ··· ∗
6.2. SPECTRUM OF A MATRIX 163
Now, if we denote the lower diagonal submatrix as B then P ∗ = P −1 implies
PA (x) = det(A − xI) = det(P ∗ AP − xI) = (α − x)k det(B − xI). (6.2.6)
So, Alg.Mulα (A) = Alg.Mulα (P ∗ AP ) ≥ k = Geo.Mulα (A).

As a corollary to the above result, one obtains the following observations.
Remark 6.2.6. Let A ∈ Mn (C).

1. Then, for each α ∈ σ(A), dim(Null(A − αI)) ≥ 1. So, we have at least one eigenvector.
2. If Alg.Mulα (A) = r then dim(Null(A − αI)) ≤ r. Thus, A may not have r linearly
independent eigenvectors.
 
1 2 3
1
 
Exercise 6.2.7. 1. Let A = 3 2 1. Then 6, x1 = 3 1 is an eigen-pair of A. Let
  √
2 3 1
h i
(x1 , e1 , e2 ) be an ordered basis of C3 . Put X = x1 e1 e2 . Compute X −1 AX. Can
you now find the remaining eigenvalues of A?
2. Let A ∈ Mm×n (R) and B ∈ Mn×m (R).

(a) If α ∈ σ(AB) and α 6= 0 then
i. α ∈ σ(BA).
T
ii. Alg.Mulα (AB) = Alg.Mulα (BA).

AF
iii. Geo.Mulα (AB) = Geo.Mulα (BA).

DR
(b) If 0 ∈ σ(AB) and n = m then Alg.Mul0 (AB) = Alg.Mul0 (BA) as there are n
eigenvalues, counted with multiplicity.
(c) Give an example to show that Geo.Mul0 (AB) need not equal Geo.Mul0 (BA) even
when n = m.
3. Let A ∈ Mn (R) be an invertible matrix and let x, y ∈ Rn with x 6= 0 and yT A−1 x 6= 0.

Define B = xyT A−1 . Then, prove that
(a) λ0 = yT A−1 x is an eigenvalue of B of multiplicity 1.

(b) 0 is an eigenvalue of B of multiplicity n − 1 [Hint: Use Exercise 6.2.7.2a].
(c) 1 + αλ0 is an eigenvalue of I + αB of multiplicity 1, for any α ∈ R, α 6= 0.
(d) 1 is an eigenvalue of I + αB of multiplicity n − 1, for any α ∈ R, α 6= 0.
(e) det(A + αxyT ) equals (1 + αλ0 ) det(A), for any α ∈ R, α 6= 0. This result is known
as the Shermon-Morrison formula for determinant.
4. Let A, B ∈ M2 (R) such that det(A) = det(B) and tr(A) = tr(B).
(a) Do A and B have the same set of eigenvalues?

(b) Give examples to show that the matrices A and B need not be similar.
5. Let A, B ∈ Mn (R). Also, let (λ1 , u) and (λ2 , v) are eigen-pairs of A and B, respectively.
(a) If u = αv for some α ∈ R then (λ1 + λ2 , u) is an eigen-pair for A + B.

(b) Give an example to show that if u and v are linearly independent then λ1 + λ2 need
not be an eigenvalue of A + B.
6.3 Basic Results on Diagonalization

Let A ∈ Mn (C) and let T ∈ L(Cn ) be defined by T (x) = Ax, for all x ∈ Cn . In this section, we
first find conditions under which one can obtain a basis B of Cn such that T [B, B] is a diagonal
matrix. To start with, we have the following definition.
Definition 6.3.1. A matrix A ∈ Mn (C) is said to be diagonalizable if A is similar to a

diagonal matrix. Or equivalently, P −1 AP = D ⇔ AP = P D, for some diagonal matrix D and
invertible matrix P . Or equivalently, there exists an ordered basis B of Cn such that A[B, B] is
a diagonal matrix.
Example 6.3.2. 1. Let A ∈ Mn (C) be a diagonalizable matrix. Then, by definition, A is
similar to D = diag(d1 , . . . , dn ). Thus, by Remark 6.2.4, σ(A) = σ(D) = {d1 , . . . , dn }.
" #
0 1
2. Let A = . Then, A cannot be diagonalized.
0 0
Solution: A is diagonalizable implies A is similar to a diagonal matrix D with diagonal
T
entries {d1 , d2 } = {0, 0}. Hence D = 0 ⇒ A = SDS −1 = 0, a contradiction.

AF
 
2 1 1
DR
 
3. Let A = 0 2 1. Then, A cannot be diagonalized.

0 0 2
Solution: A is diagonalizable implies A is similar to a diagonal matrix D with diagonal
entries {d1 , d2 , d3 } = {2, 2, 2}. Hence, D = 2I3 ⇒ A = SDS −1 = 2I3 , a contradiction.
" # " #! " #!
0 1 i −i
4. Let A = . Then, i, and −i, are two eigen-pairs of A. Define
−1 0 1 1
" # " #
i −i −i 0
U = √12 . Then, U ∗ U = I2 = U U ∗ and U ∗ AU = .
1 1 0 i
Theorem 6.3.3. Let A ∈ Mn (R).

1. Let S be an invertible matrix such that S −1 AS = diag(d1 , . . . , dn ). Then, for 1 ≤ i ≤ n,
the i-th column of S is an eigenvector of A corresponding to di .
2. Then, A is diagonalizable if and only if A has n linearly independent eigenvectors.
Proof. Let S = [u1 , . . . , un ]. Then S −1 AS = D implies AS = SD. Thus
[Au1 , . . . , Aun ] = A [u1 , . . . , un ] = AS = SD = S diag(d1 , . . . , dn ) = [d1 u1 , . . . , dn un ] .
Or equivalently, Aui = di ui , for 1 ≤ i ≤ n. As S is invertible, {u1 , . . . , un } are linearly

independent. Hence, (di , ui ), for 1 ≤ i ≤ n, are eigen-pairs of A. This proves Part 1 and “only
if” part of Part 2.
6.3. BASIC RESULTS ON DIAGONALIZATION 165
Conversely, let {u1 , . . . , un } be n linearly independent eigenvectors of A corresponding to

eigenvalues α1 , . . . , αn . Then, by Corollary 3.3.16, S = [u1 , . . . , un ] is non-singular and
 
α1 0 · · · 0
 
 0 α2 · · · 0 
AS = [Au1 , . . . , Aun ] = [α1 u1 , . . . , λn un ] = [u1 , . . . , un ]  .  = SD,
 
 .. .
.. . .
. . .. 
 
0 0 · · · αn
where D = diag(α1 , . . . , αn ). Therefore, S −1 AS = D. This implies A is diagonalizable.

The next result implies that the eigenvectors corresponding to distinct eigenvalues are lin-
early independent. A proof is given for clarity. A separate proof appears later in Corollary 6.3.7.
Theorem 6.3.4. Let (α1 , v1 ), . . . , (αk , vk ) be k eigen-pairs of A ∈ Mn (C) with αi ’s distinct.

Then, {v1 , . . . , vk } is linearly independent.
Proof. Let {v1 , . . . , vk } be linearly dependent. Then, there exists a smallest ` ∈ {1, . . . , k − 1}
and c 6= 0 such that v`+1 = c1 v1 + · · · + c` v` . So,
α`+1 v`+1 = α`+1 c1 v1 + · · · + α`+1 c` v` . (6.3.1)
and
α`+1 v`+1 = Av`+1 = A (c1 v1 + · · · + c` v` ) = α1 c1 v1 + · · · + α` c` v` . (6.3.2)

T
Now, subtracting Equation (6.3.2) from Equation (6.3.1) gives

AF
0 = (α`+1 − α1 ) c1 v1 + · · · + (α`+1 − α` ) c` v` .
DR
So, v` ∈ LS(v1 , . . . , v`−1 ), a contradiction to the choice of `. Thus, the required result follows.
An immediate corollary of Theorem 6.3.3 and Theorem 6.3.4 is stated next without proof.
Corollary 6.3.5. Let A ∈ Mn (C) have n distinct

 eigenvalues. Then, A is diagonalizable.
1 2 3 4
 
0 2 1 2
Remark 6.3.6. 1. Let A =  . Then σ(A) = {1, 2, 3, 7}, which are distinct.
0 0 3 5
 
0 0 0 7
Hence, A is diagonalizable.
2. The converse of Theorem 6.3.4 is not true as In has n linearly independent eigenvectors
corresponding to the eigenvalue 1, repeated n times.
Corollary 6.3.7. Let α1 , . . . , αk be k distinct eigenvalues A ∈ Mn (C). Also, for 1 ≤ i ≤ k, let

Pk
dim(Null(A − αi In )) = ni . Then, A has ni linearly independent eigenvectors.
i=1
Proof. For 1 ≤ i ≤ k, let Si = {ui1 , . . . , uini } be a basis of Null(A − αi In ). Then, we need to
k
k
S Q
prove that Si is linearly independent. To do so, denote pj (A) = (A − αi In ) / (A − αj In ),
i=1 i=1
for 1 ≤ j ≤ k. Then, note that pj (A) is a polynomial in A of degree k − 1 and

 0, if u ∈ Null(A − αi In ), for some i 6= j
pj (A)u = Q
(αj − αi )u if u ∈ Null(A − αj In ) (6.3.3)

i6=j
k
S
So, to prove that Si is linearly independent, consider the linear system
i=1
c11 u11 + · · · + c1n1 u1n1 + · · · + ck1 uk1 + · · · + cknk uknk = 0
in the variables cij ’s. Now, applying the matrix pj (A) and using Equation (6.3.3), we get
Y
(αj − αi ) cj1 uj1 + · · · + cjnj ujnj = 0.
i6=j
Q
But (αj − αi ) 6= 0 as αi ’s are distinct. Hence, cj1 uj1 + · · · + cjnj ujnj = 0. As Sj is a basis
i6=j
of Null(A − αj In ), we get cjt = 0, for 1 ≤ t ≤ nj . Thus, the required result follows.
Corollary 6.3.8. Let A ∈ Mn (C) with distinct eigenvalues α1 , . . . , αk . Then, A is diagonaliz-

able if and only if Geo.Mulαi (A) = Alg.Mulαi (A), for each 1 ≤ i ≤ k.
k
P
Proof. Let Alg.Mulαi (A) = mi . Then mi = n. Let Geo.Mulαi (A) = ni , for 1 ≤ i ≤ k.
i=1
k
P
Then, by Corollary 6.3.7, A has ni linearly independent eigenvectors. Also, by Theorem 6.2.5,
i=1
ni ≤ mi , for 1 ≤ i ≤ mi .
Now, let A be diagonalizable. Then, by Theorem 6.3.3, A has n linearly independent
k
P k
P
eigenvectors. As ni ≤ mi , we get n = ni ≤ mi = n. Thus ni = mi , 1 ≤ i ≤ k.
T
i=1 i=1
k k
AF
P P
Now, assume that mi = ni , for 1 ≤ i ≤ k. Then A has ni = mi = n linearly
i=1 i=1
DR
independent eigenvectors. Hence by Theorem 6.3.3, A is diagonalizable.

Definition 6.3.9. 1. A matrix A ∈ Mn (C) is called defective if for some α ∈ σ(A),
Geo.Mulα (A) < Alg.Mulα (A).
2. A matrix A ∈ Mn (C) is called non-derogatory if Geo.Mulα (A) = 1, for each α ∈ σ(A).
As a direct consequence of the above discussions, we obtain the following result.
Corollary 6.3.10. Let A ∈ Mn (C). Then,

1. A is non-defective if and only if A is diagonalizable.
2. A has distinct eigenvalues if and only if A is non-derogatory and non-defective.
       
2 1 1 1 1
       
Example 6.3.11. Let A =  1 2 1 . Then, 1,  0  and 2,  1 
       
 are the only

0 −1 1 −1 −1
eigen-pairs. Hence, by Theorem 6.3.3, A is not diagonalizable.
Exercise 6.3.12. 1. A strictly upper triangular matrix is not diagonalizable.
2. A be diagonalizable if and only if A + αI is diagonalizable for every α ∈ C.
3. Let A be an n×n matrix with λ ∈ σ(A) with alg.mulλ (A) = m. If Rank[A−λI] 6= n−m
then prove that A is not diagonalizable.
6.4. SCHUR’S UNITARY TRIANGULARIZATION AND DIAGONALIZABILITY 167
4. Let A and B be two similar matrices such that A is diagonalizable. Prove that B is
diagonalizable.
5. If σ(A) = σ(B) and both A and B are diagonalizable then prove that A is similar to B.
Thus, they are two basis representation of the same linear transformation.
" #
A 0
6. Let A ∈ Mn (R) and B ∈ Mm (R). Suppose C = . Then, prove that C is diagonal-
0 B
izable if and only if both A and B are diagonalizable.
7. Let J = 11T be an n × n. Define A = (a − b)I + bJ. Is A diagonalizable?

 
2 1 1
 
8. Is the matrix A = 
1 2 1 diagonalizable?
1 1 2
9. Let T : R5 −→ R5 be a linear operator with Rank(T − I) = 3 and
Null(T ) = {(x1 , x2 , x3 , x4 , x5 ) ∈ R5 | x1 + x4 + x5 = 0, x2 + x3 = 0}.
(a) Determine the eigenvalues of T ?

(b) For each distinct eigenvalue α of T , determine Geo.Mulα (T ).
T
(c) Is T diagonalizable? Justify your answer.

AF
10. Let A ∈ Mn (R) with A 6= 0 but A2 = 0. Prove that A cannot be diagonalized.

DR
     
1 0 −1 1 −3 3 1 3 3
     
0 0 1 , 0 −5 6 and 0 −5 6 diagonalizable?
11. Are the matrices      
0 2 0 0 −3 4 0 −3 4
12. Let A ∈ Mn (C) be a matrix of rank 1. Then

(a) A has at most one non-zero eigenvalue of algebraic multiplicity 1.
(b) find this eigenvalue and its geometric multiplicity.
(c) when is A diagonalizable?
13. Let u, v ∈ Cn such that {u, v} is a linearly independent set. Define A = uvT + vuT .
(a) Then prove that A is a symmetric matrix.
(b) Then prove that dim(Ker(A)) = n − 2.
(c) Then 0 ∈ σ(A) and has multiplicity n − 2.
(d) Determine the other eigenvalues of A.
6.4 Schur’s Unitary Triangularization and Diagonalizability

We now prove one of the most important results in diagonalization, called the Schur’s Lemma
or the Schur’s unitary triangularization.
Lemma 6.4.1. [Schur’s unitary triangularization (SUT)] Let A ∈ Mn (C). Then, there exists
a unitary matrix U such that A is similar to an upper triangular matrix. Further, if A ∈ Mn (R)
and σ(A) have real entries then U is a real orthogonal matrix.
Proof. We prove the result by induction on n. The result is clearly true for n = 1. So, let n > 1
and assume the result to be true for k < n and prove it for n.
Let (λ1 , x1 ) be an eigen-pair of A with kx1 k = 1. Now, extend it to form an orthonormal
basis {x1 , x2 , . . . , xn } of Cn and define X = [x1 , x2 , . . . , xn ]. Then, X is a unitary matrix and
 
x∗
 1∗  " #
 x2  λ ∗
1
X ∗ AX = X ∗ [Ax1 , Ax2 , . . . , Axn ] =  .  [λ1 x1 , Ax2 , . . . , Axn ] = , (6.4.4)
 
 ..  0 B
 
x∗n
where B ∈ Mn−1 (C). Now, by induction hypothesis there exists a unitary # U ∈ Mn−1 (C)
" matrix
such that U ∗ BU = T is an upper triangular matrix. Define U b = X 1 0 . As product of
0 U
unitary matrices is unitary, the matrix U
b is unitary and
" # " # " #" #" #
∗ 1 0 1 0 1 0 λ1 ∗ 1 0
U
b AUb = X ∗ AX =
0 U∗ 0 U 0 U∗ 0 B 0 U
T
" #" # " # " #

λ1 ∗ 1 0 λ1 ∗ λ1 ∗
AF
= = = .
0 U ∗B 0 U 0 U ∗ BU 0 T
DR
" #
λ1 ∗
Since T is upper triangular, is upper triangular.
0 T
Further, if A ∈ Mn (R) and σ(A) has real entries then x1 ∈ Rn with Ax1 = λ1 x1 . Now, one
uses induction once again to get the required result.
Remark 6.4.2. Let A ∈ Mn (C). Then, by Schur’s Lemma there exists a unitary matrix U
such that U ∗ AU = T = [tij ], a triangular matrix. Thus,
{α1 , . . . , αn } = σ(A) = σ(U ∗ AU ) = {t11 , . . . , tnn }. (6.4.5)
Furthermore, we can get the αi ’s in the diagonal of T in any prescribed order.
Definition 6.4.3. Let A, B ∈ Mn (C). Then, A and B are said to be unitarily equiva-
lent/similar if there exists a unitary matrix U such that A = U ∗ BU .
Remark 6.4.4. We know that if two matrices are unitarily equivalent then they are necessarily
similar as U ∗ = U −1 , for every unitary matrix U . But, similarity doesn’t imply unitary equiv-
alence (see Exercise 6.4.6.5). In numerical calculations, unitary transformations are preferred
as compared to similarity transformations due to the following main reasons:
1. A is unitary implies kAxk = kxk. This need not be true under a similarity.
2. As U −1 = U ∗ , for a unitary matrix, unitary equivalence is computationally simpler.

3. Also, computation of “conjugate transpose” doesn’t create round-off error in calculation.

" # " #
3 2 1 1
Example 6.4.5. Consider the two matrices A = and B = . Then, we show
−1 0 0 2
that they are similar but not unitarily similar.
Solution: Note that σ(A) = σ(B) = {1, 2}. As the eigenvalues are distinct, by Theo-
rem 6.3.5, the matrices A and B are diagonalizable and hence " #there exists invertible matrices
1 0
S and T such that A = SΛS −1 , B = T ΛT −1 , where Λ = . Thus A = ST −1 B(ST −1 )−1 .
0 2
|aij |2 6= |bij |2 . Hence by Exercise 5.8.8.6, they cannot be unitarily similar.
P P
But
Exercise 6.4.6.
|tij |2 = tr(A∗ A) − |λi |2 .
P P
1. If A is unitarily similar to a triangular matrix T = [tij ] then
i<j
2. Consider
 the following 6 matrices.
√   √   √ 
2 −1 3 2 2 1 3 2 2 0 3 2
 √   √   √ 
M1 = 
0 1 , M2 = 0 1
2  − 2, M3 = 1
 1 2,
0 0 3 0 0 3 0 0 1
 √     
2 0 3 2 1 1 4 2 1 4
 √     
M4 = −1 1 − 2, M5 = 0 2
   2 and M6 = 0
 
.
1 2
0 0 1 0 0 3 0 0 1
T
Now, use the exercises given below to conclude that the upper triangular matrix obtained
AF
in the “Schur’s Lemma” need not be unique.

DR
(a) Prove that M1 , M2 and M5 are unitarily equivalent.

(b) Prove that M3 , M4 and M6 are unitarily equivalent.
(c) Do the above results contradict Exercise 5.8.8.5c? Give reasons for your answer.
   √ 
1 1 1 2 −1 2
   
3. Prove that A = 0 2 1 and B = 0 1
  
 are unitarily equivalent.
0 
0 0 3 0 0 3
4. Let A ∈ Mn (C). Then, Prove that if x∗ Ax = 0, for all x ∈ Cn , then A = 0. Do these
results hold for arbitrary matrices?
" # " #
4 4 10 9
5. Show that the matrices A = and B = are similar. Is it possible to find
0 4 −4 −2
a unitary matrix U such that A = U ∗ BU ?
We now use Lemma 6.4.1 to give another proof of Theorem 6.1.17.

n n
Corollary 6.4.7. Let A ∈ Mn (C). If α1 , . . . , αn ∈ σ(A) then det(A) =
Q P
αi and tr(A) = αi .
i=1 i=1
Proof. By Schur’s Lemma there exists a unitary matrix U such that U ∗ AU = T = [tij ], a
n
Q n
Q
triangular matrix. By Remark 6.4.2, σ(A) = σ(T ). Hence, det(A) = det(T ) = tii = αi
i=1 i=1
n n
and tr(A) = tr(A(UU∗ )) = tr(U∗ (AU)) = tr(T) =
P P
tii = αi .
i=1 i=1
6.4.1 Diagonalizability of some Special Matrices
We now use Schur’s unitary triangularization Lemma to state the main theorem of this sub-
section. Also, recall that A is said to be a normal matrix if AA∗ = A∗ A. Further, Hermitian,
skew-Hermitian and scalar multiples of Unitary matrices are examples of normal matrices.
Theorem 6.4.8. [Spectral Theorem for Normal Matrices] Let A ∈ Mn (C). If A is a normal
matrix then there exists a unitary matrix U such that U ∗ AU = diag(α1 , . . . , αn ).
Proof. By Schur’s Lemma there exists a unitary matrix U such that U ∗ AU = T = [tij ], a
triangular matrix. Since A is a normal
T ∗ T = (U ∗ AU )∗ (U ∗ AU ) = U ∗ A∗ AU = U ∗ AA∗ U = (U ∗ AU )(U ∗ AU )∗ = T T ∗ .
Thus, we see that T is an upper triangular matrix with T ∗ T = T T ∗ . Thus, by Exercise 1.3.13.8,
T is a diagonal matrix and this completes the proof.
We re-write Theorem 6.4.8 in another form to indicate that A can be decomposed into linear
combination of orthogonal projectors onto eigen-spaces. Thus, it is independent of the choice
of eigenvectors. This remark is also valid for Hermitian, skew-Hermitian and Unitary matrices.
Remark 6.4.9. Let A ∈ Mn (C) be a normal matrix with eigen-pairs α1 , . . . , αn .

1. Then, there exists a unitary matrix U = [u1 , . . . , un ] such that
T
AF
(a) ui is an eigenvector of A for αi , 1 ≤ i ≤ n.

(b) In = U ∗ U = U U ∗ = u1 u∗1 + · · · + un u∗n .
DR
(c) the columns of U form a set of orthonormal eigenvectors for A (use Theorem 6.3.3).
(d) A = A · In = A (u1 u∗1 + · · · + un u∗n ) = α1 u1 u∗1 + · · · + αn un u∗n .
2. Let the distinct eigenvalues of A be α1 , . . . , αk with respective eigen-spaces W1 , . . . , Wk .

(a) Then each eigenvector belongs to some Wi . So, Wi ’s are orthogonal to each other.
(b) Hence Cn = W1 ⊕ · · · ⊕ Wk .
(c) If Pαi is the orthogonal projector onto Wi , 1 ≤ i ≤ k, then A = α1 P1 + · · · + αk Pk .
Thus, A depends only on the eigen-spaces and not on the computed eigenvectors.
Theorem 6.4.8 also implies that if A ∈ Mn (C) is a normal matrix then after a rotation or
reflection of axes (unitary transformation), the matrix A basically looks like a diagonal matrix.
As a special case, we now give the spectral theorem for Hermitian matrices.
Theorem 6.4.10. [Spectral Theorem for Hermitian Matrices] Let A ∈ Mn (C) be a Hermitian
matrix. Then Remark 6.4.9 holds. Further, all the eigenvalues of A are real.
Proof. The first part is immediate from Theorem 6.4.8 as Hermitian matrices are also normal
matrices. Let (α, x) be an eigen-pair. To show, α is a real number.
As A∗ = A and Ax = αx, we have x∗ A = x∗ A∗ = (Ax)∗ = (αx)∗ = αx∗ . Hence,
αx∗ x = x∗ (αx) = x∗ (Ax) = (x∗ A)x = (αx∗ )x = αx∗ x.

As x is an eigenvector, x 6= 0. Hence, kxk2 = x∗ x 6= 0. Thus α = α, i.e., α ∈ R.

As an immediate corollary of Theorem 6.4.10 and the second part of Lemma 6.4.1, we give
the following result without proof.
Corollary 6.4.11. Let A ∈ Mn (R) be symmetric. Then there exists an orthogonal matrix
P and real numbers α1 , . . . , αn such that A = P diag(α1 , . . . , αn )P T . Or equivalently, A is
diagonalizable using orthogonal matrix.
Exercise 6.4.12. 1. Let A be a normal matrix. If all the eigenvalues of A are 0 then prove
that A = 0. What happens if all the eigenvalues of A are 1?
2. Let A be a skew-symmetric matrix. Then A is unitarily diagonalizable and the eigenvalues

of A are either zero or purely imaginary.
3. Characterize all normal matrices in M2 (R).
4. Let σ(A) = {λ1 , . . . , λn }. Then, prove that the following statements are equivalent.
(a) A is normal.
(b) A is unitarily diagonalizable.
|aij |2 = |λi |2 .
P P
(c)
i,j i
(d) A has n orthonormal eigenvectors.
T
5. Let A be a normal matrix with (λ, x) as an eigen-pair. Then,

AF
(a) (A∗ )k x for k ∈ Z+ is also an eigenvector corresponding to λ.

DR
(b) (λ, x) is an eigen-pair for A∗ . [Hint: Verify kA∗ x − λxk2 = kAx − λxk2 .]
6. Let A be an n × n unitary matrix. Then,
(a) |λ| = 1 for any eigenvalue λ of A.

(b) the eigenvectors x, y corresponding to distinct eigenvalues are orthogonal.
7. Let A be a 2 × 2 orthogonal matrix. Then, prove the following:

" #
cos θ − sin θ
(a) if det(A) = 1 then A = , for some θ, 0 ≤ θ < 2π. That is, A
sin θ cos θ
counterclockwise rotates every point in R2 by an angle θ.
" #
cos θ sin θ
(b) if det A = −1 then A = , for some θ, 0 ≤ θ < 2π. That is, A
sin θ − cos θ
reflects every point in R2 about a line passing through origin. Determine "this line.
#
1 0
Or equivalently, there exists a non-singular matrix P such that P −1 AP = .
0 −1
8. Let A be a 3 × 3 orthogonal matrix. Then, prove the following:
(a) if det(A) = 1 then A is a rotation about a fixed axis, in the sense that A has an
eigen-pair (1, x) such that the restriction of A to the plane x⊥ is a two dimensional
rotation in x⊥ .
(b) if det A = −1 then A corresponds to a reflection across a plane P , followed by a

rotation about the line through origin that is orthogonal to P .
9. Let A be a normal matrix. Then, prove that Rank(A) equals the number of nonzero
eigenvalues of A.
10. [Equivalent characterizations of Hermitian matrices] Let A ∈ Mn (C). Then, the fol-
lowing statements are equivalent.
(a) The matrix A is Hermitian.
(b) The number x∗ Ax is real for each x ∈ Cn .
(c) The matrix A is normal and has real eigenvalues.
(d) The matrix S ∗ AS is Hermitian for each S ∈ Mn (C).
6.4.2 Cayley Hamilton Theorem
Let A ∈ Mn (C). Then, in Theorem 6.1.17, we saw that
PA (x) = det(xI − A) = xn − an−1 xn−1 + an−2 xn−2 + · · · + (−1)n−1 a1 x + (−1)n a0 (6.4.6)
for certain ai ∈ C, 0 ≤ i ≤ n − 1. Also, if α is an eigenvalue of A then PA (α) = 0. So,

xn − an−1 xn−1 + an−2 xn−2 + · · · + (−1)n−1 a1 x + (−1)n a0 = 0 is satisfied by n complex numbers
which are eigenvalues of A. It turns out that the expression
T
AF
An − an−1 An−1 + an−2 An−2 + · · · + (−1)n−1 a1 A + (−1)n a0 I = 0

DR
holds true as a matrix identity. This is a celebrated theorem called the Cayley Hamilton
theorem. We give a proof using Schur’s unitary triangularization. To do so, we look at
multiplication of certain upper triangular matrices.
Lemma 6.4.13. Let A1 , . . . , An ∈ Mn (C) be upper triangular matrices such that the (i, i)-th
entry of Ai equals 0, for 1 ≤ i ≤ n. Then A1 A2 · · · An = 0.
Proof. We use induction to prove that the first k columns of A1 A2 · · · Ak is 0, for 1 ≤ k ≤ n.

The result is clearly true for k = 1 as the first column of A1 is 0. For clarity, we show that the
first two columns of A1 A2 is 0. Let B = A1 A2 . Then, using A1 [:, 1] = 0 and (A2 )ji = 0, for
i = 1, 2, j ≥ 2, we get
B[:, i] = A1 [:, 1](A2 )1i + A1 [:, 2](A2 )2i + · · · + A1 [:, n](A2 )ni = 0 + · · · + 0 = 0.
So, assume that the first n − 1 columns of C = A1 · · · An−1 is 0. To show B = CAn = 0. As

n − 1 columns of C are zero, C[:, 1](An )1i + C[:, 2](An )2i + · · · + C[:, n − 1](An )(n−1)i = 0, for
1 ≤ i ≤ n − 1. Also C[:, n](An )ni = 0 as the last row of An = 0T . Thus
B[:, i] = C[:, 1](An )1i + C[:, 2](An )2i + · · · + C[:, n](An )ni = 0 + · · · + 0 = 0.
Hence, by the induction hypothesis the required result follows.

We now prove the Cayley Hamilton Theorem using Schur’s unitary triangularization.
Theorem 6.4.14. [Cayley Hamilton Theorem] Let A ∈ Mn (C). Then A satisfies its charac-
teristic equation, i.e., if PA (x) = det(xIn − A) = xn − an−1 xn−1 + · · · + (−1)n−1 a1 x + (−1)n a0
then
An − an−1 An−1 + · · · + (−1)n−1 a1 A + (−1)n a0 I = 0
holds true as a matrix identity.

n
Q
Proof. Let σ(A) = {α1 , . . . , αn } then PA (x) = (x − αi ). And, by Schur’s unitary triangular-
i=1
ization there exists a unitary matrix U such that U ∗ AU = T , an upper triangular matrix with
tii = αi , for 1 ≤ i ≤ n. Now, observe that if Ai = T − αi I then the Ai ’s satisfy the conditions
of Lemma 6.4.13. Hence
(T − α1 I) · · · (T − αn I) = 0.
Therefore,
n
Y n
Y h i
PA (A) = (A − αi I) = (U T U ∗ − αi U IU ∗ ) = U (T − α1 I) · · · (T − αn I) U ∗ = U 0U ∗ = 0.
i=1 i=1

We now give some examples and then implications of the Cayley Hamilton Theorem.
" #
1 2
Remark 6.4.15. 1. Let A = . Then, PA (x) = x2 + 2x − 5. Hence, verify that
1 −3
T
AF
" # " # " #

2 3 −4 1 2 1 0
A + 2A − 5I2 = +2 −5 = 0.
−2 11 1 −3 0 1
DR
" #
1 1 3 2
Further, A2 = −2A + 5I implies A−1 = (A + 2I2 ) = and
5 5 1 −1
A3 = A(A2 ) = A(−2A+5I) = −2A2 +5I = −2(−2A+5I)+5I = 4A−10I +5I = 4A−5I.
Now, use induction to show Am ∈ LS(I, A), for all m ≥ 1.

" #
3 1
2. Let A = . Then, PA (t) = t2 − 3t − 2. So, PA (A) = 0 ⇒ A2 = 3A + 2I. Thus,
2 0
A−1 = A−3I m
2 . Further, induction implies A ∈ LS(I, A), for all m ≥ 1.
" #
0 1
3. Let A = . Then, PA (x) = x2 . So, even though A 6= 0, A2 = 0.
0 0
 
0 0 1
3 3 2
 
0 0 0, PA (x) = x . So, Cayley Hamilton theorem ⇒ A = 0. Here A = 0.
4. For A =  
0 0 0
 
1 0 0
3
 
5. For A = 0 1 , PA (t) = (t − 1) . So PA (A) = 0. But, observe that q(A) = 0, where
1

0 0 1
q(t) = (t − 1)2 .
6. Let A ∈ Mn (C) with PA (x) = xn − an−1 xn−1 + · · · + (−1)n−1 a1 x + (−1)n a0 .
(a) Then, for any ` ∈ N, the division algorithm gives α0 , α1 , . . . , αn−1 ∈ C and a poly-
nomial f (x) with coefficients from C such that
x` = f (x)PA (x) + α0 + xα1 + · · · + xn−1 αn−1 .
Hence, by the Cayley Hamilton theorem, A` = α0 I + α1 A + · · · + αn−1 An−1 .

i. Thus, to compute any power of A, one needs to apply the division algorithm to
get αi ’s and know Ai , for 1 ≤ i ≤ n − 1. This is quite helpful in numerical
computation as computing powers takes much more time than division.
ii. Note that LS I, A, A2 , . . . is a subspace of Mn (C). Also, dim (Mn (C)) = n2 .

But, the above argument implies that dim LS I, A, A2 , . . . ≤ n.

iii. In the language of graph theory, it says the following: “Let G be a graph on n
vertices and A its adjacency matrix. Suppose there is no path of length n − 1 or
less from a vertex v to a vertex u in G. Then, G doesn’t have a path from v to u
of any length. That is, the graph G is disconnected and v and u are in different
components of G.”
(b) Suppose A is non-singular. Then, by definition a0 = det(A) 6= 0. Hence,
1
A−1 = a1 I − a2 A + · · · + (−1)n−2 an−1 An−2 + (−1)n−1 An−1 .

a0
T
This matrix identity can be used to calculate the inverse.

AF
(c) The above also implies that if A is invertible then A−1 ∈ LS I, A, A2 , . . . . That is,

DR
A−1 is a linear combination of the vectors I, A, . . . , An−1 .
Exercise 6.4.16. Miscellaneous Exercises:

1. 
Use the Cayley-Hamilton
  theorem
 to compute 
the inverse of following matrices:
2 3 4 −1 −1 1 1 −2 −1
     
5
 6 7,  1 −1 1 and −2 1 −1
   
.
1 1 2 0 1 1 0 −1 2
2. Let A, B ∈ M2 (C) such that A = AB − BA. Then, prove that A2 = 0.
3. Let A, B ∈ Mn (C) be upper triangular matrices with the top leading principal submatrix of
A of size k being 0. If B[k + 1, k + 1] = 0 then prove that the leading principal submatrix
of size k + 1 of AB is 0.
" # " #! " #!
0 B x x
4. Let B ∈ Mm,n (C) and A = . Then λ, is an eigen-pair ⇔ −λ,
BT 0 y −y
is an eigen-pair.
" #
B C
5. Let B, C ∈ Mn (R). Define A = . Then, prove the following:
−C B
" #
x
(a) if s is a real eigenvalue of A with corresponding eigenvector then s is also an
y
" #
−y
eigenvalue corresponding to the eigenvector .
x
6.5. QUADRATIC FORMS 175
" #
x + iy
(b) if s + it is a complex eigenvalue of A with corresponding eigenvector then
−y + ix
" #
x − iy
s − it is also an eigenvalue of A with corresponding eigenvector .
−y − ix
(c) (s + it, x + iy) is an eigen-pair of B+iC if and only if (s − it, x − iy) is an eigen-pair
of B − iC.
" #!
x + iy
(d) s + it, is an eigen-pair of A if and only if (s + it, x + iy) is an eigen-
−y + ix
pair of B + iC.
(e) det(A) = | det(B + iC)|2 .
The next section deals with quadratic forms which helps us in better understanding of conic
sections in analytic geometry.
6.5 Quadratic Forms

Definition 6.5.1. Let A ∈ Mn (C). Then A is said to be
1. positive semi-definite (psd) if x∗ Ax ∈ R and x∗ Ax ≥ 0, for all x ∈ Cn .
2. positive definite (pd) if x∗ Ax ∈ R and x∗ Ax > 0, for all x ∈ Cn \ {0}.
T
3. negative semi-definite (nsd) if x∗ Ax ∈ R and x∗ Ax ≤ 0, for all x ∈ Cn .

AF
4. negative definite (nd) if x∗ Ax ∈ R and x∗ Ax < 0, for all x ∈ Cn \ {0}.

DR
5. indefinite if x∗ Ax ∈ R and there exist x, y ∈ Cn such that x∗ Ax < 0 < y∗ Ay.
Lemma 6.5.2. Let A ∈ Mn (C). Then A is Hermitian if and only if at least one of the following
statements hold:
1. S ∗ AS is Hermitian for all S ∈ Mn .
2. A is normal and has real eigenvalues.
3. x∗ Ax ∈ R for all x ∈ Cn .
Proof. Let S ∈ Mn , (S ∗ AS)∗ = S ∗ A∗ S = S ∗ AS. Thus S ∗ AS is Hermitian.

Suppose A = A∗ . Then, A is clearly normal as AA∗ = A2 = A∗ A. Further, if (λ, x) is an
eigenpair then λx∗ x = x∗ Ax ∈ R implies λ ∈ R.
For the last part, note that x∗ Ax ∈ C. Thus x∗ Ax = (x∗ Ax)∗ = x∗ A∗ x = x∗ Ax, we get
Im(x∗ Ax) = 0. Thus, x∗ Ax ∈ R.
If S ∗ AS is Hermitian for all S ∈ Mn then taking S = In gives A is Hermitian.
If A is normal then A = U ∗ diag(λ1 , . . . , λn )U for some unitary matrix U . Since λi ∈ R,
A∗ = (U ∗ diag(λ1 , . . . , λn )U )∗ = U ∗ diag(λ1 , . . . , λn )U = U ∗ diag(λ1 , . . . , λn )U = A. So, A is
Hermitian.
If x∗ Ax ∈ R for all x ∈ Cn then aii = e∗i Aei ∈ R. Also, aii +ajj +aij +aji = (ei +ej )∗ A(ei +
ej ) ∈ R. So, Im(aij ) = −Im(aji ). Similarly, aii + ajj + iaij − iaji = (ei + iej )∗ A(ei + iej ) ∈ R
implies that Re(aij ) = Re(aji ). Thus, A = A∗ .
Remark 6.5.3. Let A ∈ Mn (R). Then the condition x∗ Ax ∈ R, for all x ∈ Cn , in Defini-
tion 6.5.8 implies AT = A, i.e., A is a symmetric matrix. But, when we study matrices over
R, we seldom consider vectors from Cn . So, in such cases, we assume A is symmetric.
" # " #
2 1 3 1+i
Example 6.5.4. 1. Let A = or A = . Then, A is positive definite.
1 2 1−i 4
" # "√ #
1 1 2 1+i
2. Let A = or A = √ . Then, A is positive semi-definite but not positive
1 1 1−i 2
definite.
" # " #
−2 1 −2 1 − i
3. Let A = or A = . Then, A is negative definite.
1 −2 1 + i −2
" # " #
−1 1 −2 1 − i
4. Let A = or A = . Then, A is negative semi-definite.
1 −1 1 + i −1
" # " #
0 1 1 1+i
5. Let A = or A = . Then, A is indefinite.
1 −1 1−i 1
Theorem 6.5.5. Let A ∈ Mn (C). Then, the following statements are equivalent.
1. A is positive semi-definite.
2. A∗ = A and each eigenvalue of A is non-negative.
T
3. A = B ∗ B for some B ∈ Mn (C).

AF
Proof. 1 ⇒ 2: Let A be positive semi-definite. Then, by Lemma 6.5.2, A is Hermitian. If

DR
(α, v) is an eigen-pair of A then αkvk2 = α(v∗ v) = v∗ (αv) = v∗ Av ≥ 0. So, α ≥ 0.

2 ⇒ 3: Let σ(A) = {α1 , . . . , αn }. Then, by spectral theorem, there exists a unitary
matrix U such that U ∗ AU = D with D = diag(α1 , . . . , αn ). As αi ≥ 0, for 1 ≤ i ≤ n, define
1 √ √ 1 1
D 2 = diag( α1 , . . . , αn ). Then, A = U D 2 [D 2 U ∗ ] = B ∗ B.
3 ⇒ 1: Let A = B ∗ B. Then, for x ∈ Cn , x∗ Ax = x∗ B ∗ Bx = kBxk2 ≥ 0. Thus, the
A similar argument gives the next result and hence the proof is omitted.
1. A is positive definite.
2. A∗ = A and each eigenvalue of A is positive.
3. A = B ∗ B for a non-singular matrix B ∈ Mn (C).
Remark 6.5.7. Let A ∈ Mn (C) be Hermitian with eigenvalues λ1 ≥ λ2 ≥ · · · ≥ λn . Then, there

exists a unitary matrix U = [u1 , u2 , . . . , un ] and a diagonal matrix D = diag(λ1 , λ2 , . . . , λn )
such that A = U DU ∗ . Now, for 1 ≤ i ≤ n, define αi = max{λi , 0} and βi = min{λi , 0}. Then
1. for D1 = diag(α1 , α2 , . . . , αn ), the matrix A1 = U D1 U ∗ is positive semi-definite.
2. for D2 = diag(β1 , β2 , . . . , βn ), the matrix A2 = U D2 U ∗ is positive semi-definite.
3. A = A1 − A2 . The matrix A1 is generally called the positive semi-definite part of A.
Definition 6.5.8. Let A = [aij ] ∈ Mn (C) be a Hermitian matrix and let x, y ∈ Cn . Then, a
sesquilinear form in x, y ∈ Cn is defined as H(x, y) = y∗ Ax. In particular, H(x, x), denoted
H(x), is called a Hermitian form. In case A ∈ Mn (R), H(x) is called a quadratic form.
Remark 6.5.9. Observe that

1. if A = In then the bilinear/sesquilinear form reduces to the standard inner product.
2. H(x, y) is ‘linear’ in the first component and ‘conjugate linear’ in the second component.
3. the quadratic form H(x) is a real number. Hence, for α ∈ R, the equation H(x) = α,
represents a conic in Rn .
Example 6.5.10. 1. Let A ∈ Mn (R). Then, f (x, y) = yT Ax, for x, y ∈ Rn , is a bilinear
form on Rn .
" # " #
1 2−i ∗ x
2. Let A = . Then, A = A and for x = ∈ C2 , verify that
2+i 2 y
H(x) = x∗ Ax = |x|2 + 2|y|2 + 2Re ((2 − i)xy)
where ‘Re’ denotes the real part of a complex number, is a sesquilinear form.
6.5.1 Sylvester’s law of inertia

T
The main idea of this section is to express H(x) as sum or difference of squares. Since H(x) is
AF
a quadratic in x, replacing x by cx, for c ∈ C, just gives a multiplication factor by |c|2 . Hence,
DR
one needs to study only the normalized vectors. Let us consider Example 6.1.2 again. There
we see that
(x + y)2 (x − y)2
xT Ax = 3 − = (x + 2y)2 − 3y 2 , and (6.5.1)
2 2
(x + 2y)2 (2x − y)2 2y 50y 2
xT Bx = 5 + 10 = (3x − )2 + . (6.5.2)
5 5 3 9
Note that both the expressions in Equation (6.5.1) is the difference of two non-negative terms.
Whereas, both the expressions in Equation (6.5.2) consists of sum of two non-negative terms.
Is the number of non-negative terms, appearing in the above expressions, just a coincidence?
For a better understanding, we define inertia of a Hermitian matrix.
Definition 6.5.11. Let A ∈ Mn (C) be a Hermitian matrix. The inertia of A, denoted i(A),
is the triplet (i+ (A), i− (A), i0 (A)), where i+ (A) is the number of positive eigenvalues of A,
i− (A) is the number of negative eigenvalues of A and i0 (A) is the nullity of A. The difference
i+ (A) − i− (A) is called the signature of A.
Exercise 6.5.12. Let A ∈ Mn (C) be a Hermitian matrix. If the signature and the rank of A
is known then prove that one can find out the inertia of A.
To proceed with the earlier discussion, let A ∈ Mn (C) be Hermitian with eigenvalues
α1 , . . . , αn . Then, by Theorem 6.4.10, U ∗ AU = D = diag(α1 , . . . , αn ), for some unitary matrix
 
z1
.
U . Let x = U z. Then, kxk = 1 implies kzk = 1. Thus, if z =  .
 .  then
zn
n p r
√ p 2
| αi zi |2 −
X X X
H(x) = z∗ U ∗ AU z = z∗ Dz = αi |zi |2 = |αi | zi , (6.5.3)

i=1 i=1 i=p+1
where α1 , . . . , αp > 0, αp+1 , . . . , αr < 0 and αr+1 , . . . , αn = 0, where p = i+ (A) and r − p =

i− (A). Thus, we see that the possible values of H(x) seem to depend only on the positive and
negative eigenvalues of A. Since U is an invertible matrix, the components zi ’s of z = U −1 x =
U ∗ x are commonly known as the linearly independent linear forms. Note that each zi is a
linear expression in the components of x.
As a next result, we show that in any expression of H(x) as a sum or difference of n absolute
squares of linearly independent linear forms, the number p (respectively, r − p) gives the number
of positive (respectively, negative) eigenvalues of A. This is popularly known as the ‘Sylvester’s
law of inertia’.
Lemma 6.5.13. [Sylvester’s Law of Inertia] Let A ∈ Mn (C) be a Hermitian matrix and let
x ∈ Cn . Then, every Hermitian form H(x) = x∗ Ax, in n variables can be written as
H(x) = |y1 |2 + · · · + |yp |2 − |yp+1 |2 − · · · − |yr |2

T
where y1 , . . . , yr are linearly independent linear forms in the components of x and the integers
AF
p and r satisfying 0 ≤ p ≤ r ≤ n, depend only on A.

DR
Proof. Equation (6.5.3) implies that H(x) has the required form. We only need to show that
p and r are uniquely determined by A. Hence, let us assume on the contrary that there exist
p, q, r, s ∈ N with p > q such that
H(x) = |y1 |2 + · · · + |yp |2 − |yp+1 |2 − · · · − |yr |2 (6.5.4)

= |z1 |2 + · · · + |zq |2 − |zq+1 |2 − · · · − |zs |2 , (6.5.5)
   
" # " # y1 z1
Y1 Z1 . .
where y = = M x, z = = N x with Y1 =  . .
Y2 Z2  .  and Z1 =  .  for some invertible
yp zq
matrices M and N . Now" the invertibility
# of M and N implies z = By,
" # some
for " invertible
# " matrix
#
B1 B2 Z1 B 1 B 2 Y1
B. Decompose B = , where B1 is a q × p matrix. Then = . As
B3 B4 Z2 B 3 B 4 Y2
 
y˜1
.
p > q, the homogeneous linear system B1 Y1 = 0 has a nontrivial solution, say Y f1 =  .  and
.
y˜p
" #
Y
f1
consider ye = . Then for this choice of y e , Z1 = 0 and thus, using Equations (6.5.4) and
0
(6.5.5), we have
H(ỹ) = |y˜1 |2 + |y˜2 |2 + · · · + |y˜p |2 − 0 = 0 − (|zq+1 |2 + · · · + |zs |2 ).

Now, this can hold only if Y

f1 = 0, a contradiction to Y
f1 being a non-trivial solution. Hence
p = q. Similarly, the case r > s can be resolved. This completes the proof of the lemma.
Remark 6.5.14. Since A is Hermitian, Rank(A) equals the number of nonzero eigenvalues.
Hence, Rank(A) = r. The number r is called the rank and the number r − 2p is called the
inertial degree of the Hermitian form H(x).
Do we need ∗-congruence at this stage?

We now look at another form of the Sylvester’s law of inertia. We start with the following
definition.
Definition 6.5.15. Let A, B ∈ Mn (C). Then, A is said to be ∗-congruent (read star-

congruent) to B if there exists an invertible matrix S such that A = S ∗ BS.
Theorem 6.5.16. [Second Version: Sylvester’s Law of Inertia] Let A, B ∈ Mn (C) be

Hermitian. Then, A is ∗-congruent to B if and only if i(A) = i(B).
Proof. By spectral theorem U ∗ AU = ΛA and V ∗ BV = ΛB , for some unitary matrices U, V

and diagonal matrices ΛA , ΛB of the form diag(+, · · · , +, −, · · · , −, 0, · · · , 0). Thus, there exist
invertible matrices S, T such that S ∗ AS = DA and T ∗ BT = DB , where DA , DB are diagonal
matrices of the form diag(1, · · · , 1, −1, · · · , −1, 0, · · · , 0).
If i(A) = i(B), then it follows that DA = DB , i.e., S ∗ AS = T ∗ BT and hence A =
T
AF
(T S −1 )∗ B(T S −1 ).
Conversely, suppose that A = P ∗ BP , for some invertible matrix P , and i(B) = (k, l, m).
DR
As T ∗ BT = DB , we have, A = P ∗ (T ∗ )−1 DB T −1 P = (T −1 P )∗ DB (T −1 P ). Now, let X =

(T −1 P )−1 . Then, A = (X −1 )∗ DB X −1 and we have the following observations.
1. As rank and nullity do not change under similarity transformation, i0 (A) = i0 (DB ) = m
as i(B) = (k, l, m).
2. Using i(B) = (k, l, m), we also have
X[:, k + 1]∗ AX[:, k + 1] = X[:, k + 1]∗ (X −1 )∗ DB (X −1 ) X[:, k + 1] = e∗k+1 DB ek+1 = −1.

Similarly, X[:, k + 2]∗ AX[:, k + 2] = · · · = X[:, k + l]∗ AX[:, k + l] = −1. As the vectors
X[:, k + 1], . . . , X[:, k + l] are linearly independent, using 9.7.10, we see that A has at least
l negative eigenvalues.
3. Similarly, X[:, 1]∗ AX[:, 1] = · · · = X[:, k]∗ AX[:, k] = 1. As X[:, 1], . . . , X[:, k] are linearly
independent, using 9.7.10 again, we see that A has at least k positive eigenvalues.
Thus, it now follows that i(A) = (k, l, m).
6.5.2 Applications in Eculidean Plane
We now obtain conditions on the eigenvalues of A, corresponding to the associated quadratic

form, to characterize conic sections in R2 , with respect to the standard inner product.
Definition 6.5.17. Let f (x, y) = ax2 + 2hxy + by 2 + 2f x + 2gy + c be a general quadratic in

x and y, with coefficients from R. Then,
" #" #
h i a h x
T
H(x) = x Ax = x, y = ax2 + 2hxy + by 2
h b y
is called the associated quadratic form of the conic f (x, y) = 0.
Proposition 6.5.18. Consider the quadratic f (x, y) = ax2 + 2hxy + by 2 + 2gx + 2f y + c, for
a, b, c, g, f, h ∈ R. If (a, b, h) 6= (0, 0, 0) then f (x, y) = 0 represents
1. a parabola or a pair of parallel lines if ab − h2 = 0,
2. a hyperbola or a pair of perpendicular lines if ab − h2 < 0,
3. an ellipse or a circle or a point (point of intersection of a pair of perpendicular lines) if

ab − h2 > 0.
" #
a h
Proof. Consider the associated quadratic ax2 + 2hxy + by 2 with A = as the associated
h b
symmetric matrix. Then, by Corollary 6.4.11, A = U diag(α1 , α2 )U T , where U = [u1 , u2 ] is an
orthogonal matrix, with (α1 , u1 ) and (α2 , u2 ) as eigen-pairs of A. As (a, b, h) 6= (0, 0, 0) at least
one of α1 , α2 6= 0. Also,
T
" # " # " #" #

h i α1 0 T x
h i α
1 0 u
AF
xT Ax = x, y U U = u v = α1 u2 + α2 v 2 ,
0 α2 y 0 α2 v
DR
" #
u
where = U T x. The lines u = 0, v = 0 are the two linearly independent linear forms, which
v
correspond to two perpendicular lines passing through the origin in the (x, y)-plane. In terms
of u, v, f (x, y) reduces to f (u, v) = α1 u2 + α2 v 2 + d1 u + d2 v + c, for some choice of d1 , d2 ∈ R.
We now look at different cases:
1. if α1 = 0 and α2 6= 0 then ab − h2 = det(A) = α1 α2 = 0. In this case,
d2 2

f (u, v) = 0 ⇔ α2 v + = c1 − d1 u,
2α2
for some c1 ∈ R.
d2
(a) If d1 = 0, the quadratic corresponds to either the same line v + = 0, two parallel
2α2
lines or two imaginary lines, depending on whether c1 = 0, c1 α2 > 0 and c1 α2 < 0,
respectively.
(b) If d1 6= 0, the quadratic corresponds to a parabola of the form V 2 = 4aU , for some
translate U = u + α and V = v + β.
2. If α1 α2 < 0 then ab − h2 = det(A) = λ1 λ2 < 0. If α2 = −β2 < 0, for β2 >

0 then the quadratic reduces to α1 (u + d1 − β2 (v + d2 )2 = d3 , or equivalently, to
)2
√ √ √ √
α1 (u + d1 ) + β2 (v + d2 ) · α1 (u + d1 ) − β2 (v + d2 ) = d3 , for some d1 , d2 , d3 ∈ R.

Thus, the quadratic corresponds to

(a) a pair of perpendicular lines u + d1 = 0 and v + d2 = 0 whenever d3 = 0.

(b) a hyperbola with orthogonal principal axes u + d1 = 0 and v + d2 = 0 whenever
d3 6= 0. In particular, if d3 > 0 then the corresponding equation equals
α1 (u + d1 )2 α2 (v + d2 )2
− = 1.
d3 d3
3. If α1 α2 > 0 then ab − h2 = det(A) = α1 α2 > 0. Here, the quadratic reduces to α1 (u +

d1 )2 + α2 (v + d2 )2 = d3 , for some d1 , d2 , d3 ∈ R. Thus, the quadratic corresponds to
(a) a point which is the point of intersection of the pair of orthogonal lines u + d1 = 0
and v + d2 = 0 if d3 = 0.
(b) an empty set if α1 d3 < 0.
(c) an ellipse or circle with u + d1 = 0 and v + d2 = 0 as the orthogonal principal axes
if α1 d3 > 0 with the corresponding equation
α1 (u + d1 )2 α2 (v + d2 )2
+ = 1.
d3 d3
Thus, we have considered all the possible cases and the required result follows.
" # " #
u T x
Remark 6.5.19. Observe that the linearly independent forms =U are functions of
v y
T
the eigenvectors u1 and u2 . Further, the linearly independent forms together with the shifting
AF
of the origin give us the principal axes of the corresponding conic.

DR
Example 6.5.20. 1. Let"H(x)# = x2 + y 2 + 2xy be the associated

" √ quadratic
#! form
" for√a class
#!
1 1 1/ 2 1/ 2
of curves. Then, A = and the eigen-pairs are 2, √ and 0, √ .
1 1 1/ 2 −1/ 2
In particular, for
(a) f (x, y) = x2 + 2xy + y 2 − 8x − 8y + 16, we have f (x, y) = 0 ⇔ (x + y − 4)2 = 0, just
one line.
(b) f (x, y) = x2 + 2xy + y 2 − 8x − 8y, we have f (x, y) = 0 ⇔ (x + y − 8) · (x + y) = 0, a
pair of parallel lines.
(c) f (x, y) = x2 + 2xy + y 2 − 6x − 10y − 3, we have
x+y 2 x−y 2 √ √ x−y

x+y
f (x, y) = 0 ⇔ 2 √ +0 √ =8 2 √ −2 2 √ +3
2 2 2 2
x+y−4 2 √ x − y − 19/2

⇔ √ =− 2 √ ,
2 2
a parabola with principal axes x + y = 4, 2x − 2y = 19 and directrix x − y = 10.
2 2
2. Let H(x)
" # − 5y + 20xy be the associated
= 10x " quadratic
√ #!
form for "
a class of#!
√
curves. Then
10 10 2/ 5 1/ 5
A= and the eigen-pairs are 15, √ and −10, √ . So, for
10 −5 1/ 5 −2/ 5
Figure 6.2: Conic x2 + 2xy + y 2 − 6x − 10y = 3
(a) f (x, y) = 10x2 − 5y 2 + 20xy + 16x − 2y + 1, we have f (x, y) = 0 ⇔ 3(2x + y + 1)2 −

2(x − 2y − 1)2 = 0, a pair of perpendicular lines.
(b) f (x, y) = 10x2 − 5y 2 + 20xy + 16x − 2y + 19, we have
x − 2y − 1 2 2x + y + 1 2

f (x, y) = 0 ⇔ − √ = 1,
3 6
a hyperbola.
(c) f (x, y) = 10x2 − 5y 2 + 20xy + 16x − 2y − 17, we have
2x + y + 1 2 x − 2y − 1 2
T

f (x, y) = 0 ⇔ √ − = 1,
AF
6 3
DR
a hyperbola.
Figure 6.3: Conic 10x2 − 5y 2 + 20xy + 16x − 2y = c, c = −1, c = −19 and c = 17
3. Let H(x) = 2 2
" # 6x + 9y + 4xy be the associated
" √quadratic
#! form"for a class
√ #!
of curves. Then,
6 2 1/ 5 2/ 5
A= , and the eigen-pairs are 10, √ and 5, √ . So, for
2 9 2/ 5 −1/ 5
(a) f (x, y) = 6x2 + 9y 2 + 4xy + 10y − 53, we have
x + 2y + 1 2 2x − y − 1 2

f (x, y) = 0 ⇔ + √ = 1,
5 5 2
an ellipse.
Figure 6.4: Conic 6x2 + 9y 2 + 4xy + 10y = 53

.
Exercise 6.5.21. Sketch the graph of the following surfaces:
1. x2 + 2xy + y 2 + 6x + 10y = 3.
2. 2x2 + 6xy + 3y 2 − 12x − 6y = 5.
3. 4x2 − 4xy + 2y 2 + 12x − 8y = 10.
4. 2x2 − 6xy + 5y 2 − 10x + 4y = 7.

T
AF
6.5.3 Applications in Eculidean Space

DR
As a last application,
 we consider
 a
quadratic
 in 3 variables,
 namely x1 , x2 and x3 . To do so,
a h g x l y
   1    1
let A = h b f , x = x2 , b = m and y = y2 
      
 with
g f c x3 n y3
f (x1 , x2 , x3 ) = xT Ax + 2bT x + q
= ax21 + bx22 + cx23 + 2hx1 x2 + 2gx1 x3 + 2f x2 x3
+2lx1 + 2mx2 + 2nx3 + q (6.5.6)
Then, we observe the following:
1. As A is symmetric, P T AP = diag(α1 , α2 , α3 ), where P = [u1 , u2 , u3 ] is an orthogonal

matrix and (αi , ui ), for i = 1, 2, 3 are eigen-pairs of A.
2. Let y = P T x. Then, f (x1 , x2 , x3 ) reduces to
g(y1 , y2 , y3 ) = α1 y12 + α2 y22 + α3 y32 + 2l1 y1 + 2l2 y2 + 2l3 y3 + q. (6.5.7)
3. Depending on the values of αi ’s, rewrite g(y1 , y2 , y3 ) to determine the center and the
planes of symmetry of f (x1 , x2 , x3 ) = 0.
Example 6.5.22. Determine the following quadrics f (x, y, z) = 0, where

1. f (x, y, z) = 2x2 + 2y 2 + 2z 2 + 2xy + 2xz + 2yz + 4x + 2y + 4z + 2.
2. f (x, y, z) = 3x2 − y 2 + z 2 + 10.
3. f (x, y, z) = 3x2 − y 2 + z 2 − 10.
4. f (x, y, z) = 3x2 − y 2 + z − 10.

     
2 1 1 2 √1 √1 √1
 13 −1
2 6
√1 
   
Solution: (1) Here A =  1 2 1, b = 1 and q = 2. So, verify P =  3
   √ √
2 6
and
1 1 2 2 √1 −2
3
0 √
6
P T AP = diag(4, 1, 1). Hence, f (x, y, z) = 0 reduces to
x+y+z 2 x−y 2 x + y − 2z 2

4 √ + √ + √ = −(4x + 2y + 4z + 2).
3 2 6
4(x + y + z) + 5 2 x−y+1 2 x + y − 2z − 1 2

√ √ √ 9
Or equivalently to 4 + + = 12 . So, the
4 3 2 6
principal axes of the quadric (an ellipsoid) are 4(x + y + z) = −5, x − y = 1 and x + y − 2z = 1.
y2 3x2 z2
Part 2 Here f (x, y, z) = 0 reduces to 10 − 10 − 10 = 1 which is the equation of a
hyperboloid consisting of two sheets with center 0 and the axes x, y and z as the principal axes.
3x2 y2 z2
Part 3 Here f (x, y, z) = 0 reduces to 10 − 10 + 10 = 1 which is the equation of a
hyperboloid consisting of one sheet with center 0 and the axes x, y and z as the principal axes.
T
Part 4 Here f (x, y, z) = 0 reduces to z = y 2 −3x2 +10 which is the equation of a hyperbolic
AF
paraboloid.
DR
Figure 6.5: Ellipsoid, hyperboloid of two sheets and one sheet, hyperbolic paraboloid
.
6.6 Singular Value Decomposition

In Theorem 6.4.10, we saw that if A ∈ Mn (C) is a Hermitian matrix then we can find a unitary
matrix U such that A = U DU ∗ , where D is a diagonal matrix. That is, after a rotation
or reflection of axes, the matrix A basically looks like a diagonal matrix. We also saw it’s
applications in Section 6.5. In this section, the idea is to have a similar understanding for any
6.6. SINGULAR VALUE DECOMPOSITION 185
matrix A. We will do it over complex numbers and hence, the ideas from Theorem 6.4.10 will
be used. We start with the following result.
Lemma 6.6.1. Let A ∈ Mm,n (C) with m ≤ n and RankA = k ≤ m. Then A = U DV ∗ , where
1. U is a unitary matrix and is obtained from the spectral decomposition of AA∗ = U ΛU ∗

with λ11 ≥ · · · ≥ λmm ≥ 0 are the eigenvalues of AA∗ ,
2. D = Λ1/2 , and
3. V ∗ is formed by taking the first k rows of U ∗ A and adding m − k new rows so that V ∗ has
orthonormal rows.
If A is real, then U and V may be chosen to have real entries.
Proof. Note that AA∗ is an m × m Hermitian matrix. Thus, for any x ∈ Cm ,
x∗ (AA∗ )x = (x∗ A)(A∗ x) = (A∗ x)∗ (A∗ x) = kA∗ xk2 ≥ 0.
Hence, the matrix AA∗ is a positive semi-definite matrix. Therefore, all it’s eigenvalues are
non-negative. So, by the spectral theorem, Theorem 6.4.10, AA∗ = U ΛU ∗ , where λii ≥ 0 are
in decreasing order. As RankA = k, λii > 0, for 1 ≤ i ≤ k and λii = 0, for k + 1 ≤ i ≤ m. Now,
T
let Σ = [σij ] be the diagonal matrix with

AF
( √
DR
1/ λii , if i ≤ k
σii =
1, otherwise.
Then, we see that the matrix X = ΣU ∗ A is an m × n matrix with

" #
∗ ∗ ∗ ∗ Ik 0∗
XX = (ΣU A)(A U Σ) = ΣU (U ΛU )U Σ = . (6.6.8)
0 0
" #
Ik 0
As XX ∗ = , the first k-rows of X form an orthonormal set. Note that the first k rows
0 0
of the matrix X are given by
1 1
X[1, :] = √ (U ∗ A)[1, :], . . . , X[k, :] = √ (U ∗ A)[k, :].
λ11 λkk
Or equivalently,
(U ∗ A)[1, :] = X[1, :] λ11 , . . . , (U ∗ A)[k, :] = X[k, :] λkk .

p p
(6.6.9)
Now, take these k rows of X and add m − k many rows to form V ∗ , so that the rows of V ∗ are
orthonormal, i.e., V ∗ V = Im . Also, using (6.6.8), we see that (ΣU ∗ AA∗ U Σ)k+1,k+1 = 0. Thus,
(ΣU ∗ A)[k + 1, :](A∗ U Σ)[:, k + 1] = 0 ⇒ (ΣU ∗ A)[k + 1, :] = 0.

This in turn implies that (U ∗ A)[k + 1, :] = 0. Similarly, (U ∗ A)[j, :] = 0, for k + 2 ≤ j ≤ m.

Thus, using (6.6.9) and the definition of the matrix V ∗ , we get
 √   
X[1, :] λ11 X[1, :]
 ..   .. 
. .
   
   
X[k, :]√λ 
   
kk 
 X[k, :] 
U ∗A =   = DV ∗
p p
= diag( λ , . . . , λ , 0, . . . 0)
  
11 kk
T V ∗ [k + 1, :]
 

 0 
  
 ..   .. 

 . 


 . 

0T V ∗ [m, :]
√ √
where D = diag( λ11 , . . . , λkk , 0, . . . 0). Thus, we have A = U DV ∗ .
We already know that in spectral theorem, that if A is real symmetric, we could choose U
to be a real orthogonal matrix and that makes the first k rows of V to have real entries. We
can always choose the next m − k vectors to also have real entries.
It is important to note that
A∗ A = (U DV ∗ )∗ (U DV ∗ ) = (V DU ∗ )(U DV ∗ ) = V D2 V ∗ ,
where D2 = diag(λ11 , . . . , λkk , 0, . . . 0) are the eigenvalues of A∗ A and the columns of V are the
corresponding eigenvectors.
T
AF
Corollary 6.6.2. [Polar decomposition] Let A ∈ Mm,n (C) with m ≤ n. Then A = P W , for
some positive semi-definite matrix P with RankP = RankA and a matrix W having orthonormal
DR
rows. In fact, P = (AA∗ )1/2 .
Proof. By Lemma 6.6.1, A = U DV ∗ , where U is a unitary matrix which is obtained from

the spectral decomposition of AA∗ = U ΛU ∗ , D = Λ1/2 , and V ∗ has orthonormal rows. Then
A = (U DU ∗ )(U V ∗ ). Notice that the matrix U V ∗ also has orthonormal rows. Note that
U DU ∗ = U Λ1/2 U ∗ = (AA∗ )1/2 . So, if we put P = U DU ∗ and W = U V ∗ then, we see that
A = P W with P = (AA∗ )1/2 is positive semi-definite matrix with RankP = RankA and a matrix
W having orthonormal rows.
Corollary 6.6.3. [Singular value decomposition] Let A ∈ Mm,n (C) with m ≤ n and RankA =
k ≤ m. Then A = U DV ∗ , where
1. U is an m × m unitary matrix and is obtained from the spectral decomposition of AA∗ =

U ΛU ∗ with λ11 ≥ · · · ≥ λmm ≥ 0 are the eigenvalues of AA∗ ,
h i
2. D = Λ1/2 0m,n−m , and
3. V ∗ is formed by taking the first k rows of U ∗ A and adding n − k new rows so that V is
an n × n unitary matrix.
If A is real, then U and V may be chosen to have real entries.
Proof. Follows from Lemma 6.6.1.

6.6. SINGULAR VALUE DECOMPOSITION 187
√ √
Definition 6.6.4. Let A ∈ Mm,n . In view of Corollary 6.6.3, the values λ11 , . . . , λrr , where
r = min{m, n}, are called the singular values of A. (Sometimes only the nonzero λii ’s are
understood to be the singular values of A).
Let A ∈ Mm,n (C). Then, by the singular value decomposition of A we mean writing
A = U ΣV ∗ , where U ∈ Mm (C), V ∈ Mn (C) are unitary matrices and Σ ∈ Mm,n (R) with Σii as
the singular values of A, for 1 ≤ i ≤ RankA and the remaining entries of Σ being 0.
In Corollary 6.6.3, we saw that the matrix U is obtained as the unitary matrix in the spectral
decomposition of AA∗ , the Σii ’s are the square-root of the eigenvalues of AA∗ , and V ∗ is formed
by taking the first r = RankA rows of U ∗ A and adding n − k new rows so that V ∗ is a unitary
matrix.
Now, let us go backi to matrixh multiplication and ∗
h i try to understand A = U ΣV . So, let
U = u1 u2 · · · um and V = v1 v2 · · · vm . Then,
√ 
λ11 0 0 · · · 0 0  ∗
√  v1
0 λ22 0 · · · 0 0 ∗ 


h i √  v2 
A = U ΣV ∗ = u1 u2 · · · um  0 0 λ33 · · · 0 0
  
 . 
 .. .. .. . . .. ..  .. 

 . . . . . .  
 ∗
vn
0 0 0 ··· 0 0
λ11 u1 v1∗ + λ22 u2 v2∗ + · · · + λmm um vm
∗
p p p
= . (6.6.10)
T
AF
Now, recall that if A is an m × n matrix then we can associate a linear transformation

T : Cn → Cm such that T (x) = Ax, for all x ∈ Cn . Thus, the singular value decomposition
DR
states that there exist unitary matrices (rotation or reflection matrices) in both the domain
(Cn , corresponds to V ∗ ) and co-domain (Cm , corresponds to U ) such that the matrix of A with
respect to these ordered bases is a diagonal matrix and the diagonal entries consist of just the
singular values, including zeros.
√ √ √
We also note that if r = RankA then A = λ11 u1 v1∗ + λ22 u2 v2∗ + · · · + λrr ur vr∗ . Thus,
A = U1 Σ1 V1∗ , where U1 is a submatrix of U consisting of the first r orthonormal columns, Σ1
is a diagonal matrix of non-zero singular values and V1∗ is a submatrix of V ∗ consisting of the
first r orthonormal rows. More specifically,
√  
λ11 0 v∗··· 0
 √  1∗ 
h i 0 λ22 · · · 0 v2 
 
A = u1 u2 · · · ur   . .

.. .. .
 0 0 .  . 
. 
√

0 0 ··· λrr vr∗
" # " #
2 1 1 6 3
Example 6.6.5. Let A = . Then, AAT = . Thus, AAT = U DU T , where
1 2 −1 3 6
" # " # " #
1 1 1 9 0 3 0 0
U=√ and D = . Hence, Σ = √ . Here,
2 1 −1 0 3 0 3 0
#" 1
√1
" # " #
T 1 3 3 0 3 0 √
2 2
0
U A= √ = √ .
2 1 −1 2 0 3 √16 − √16 √26
   
√1 √1 0 5 4 1
 12 2
Thus, V T = − √16 √2  and it’s rows are the eigenvectors of AT A = 4
  
5 −1.
√
 6 6 
√1 −1
√ −1
√ 1 −1 2
3 3 3
In actual computations, the values of m and n could be very large. Also, the largest and
the smallest eigenvalues or the rows and columns of A that are of interest to us may be very
small. So, in such cases, we compute the singular value decomposition to relate the above ideas
or to find clusters which have maximum influence on the problem being looked. For example,
in the above computation, the singular value 3 is the larger of the two singular values. So, if
we are looking at the largest deviation or movement etc. then we need to concentrate on the
singular value 3. Then, using equation (6.6.10), note that 3 is associated with the first column
√
of U and the first row of V T . Similarly, 3 is associated with the second column of U and the
second row of V T .
Note that in any computation, we need to decompose our problem into sub-problems. If
the decomposition into sub-problems is possible through orthogonal decomposition then in some
sense the sub-problems can be handled separately. That’s how the singular value decomposition
helps us in applications. This is the reason, that with slight change, SVD is also called “factor
analysis” or “principal component analysis” and so on.
Exercise 6.6.6. 1. Let A ∈ Mm,n (C) with m ≥ n. Then A = W Q, for some positive
semi-definite matrix Q and a matrix W of orthonormal columns.
T
AF
2. Let A ∈ Mn,1 (C). Illustrate the polar decomposition and the singular value decompositions
for A = ei and for A = e1 + 2e2 + · · · + nen .
DR
3. Let A ∈ Mm,n (C) with RankA = r. If d1 , . . . , dr are the non-zero singular values of A
then, there exist Σ ∈ Mm,n (R),"and unitary
# matrices U ∈ Mm (C) and V ∈ Mn (C) such
Σ1 0
that A = U ΣV ∗ , where Σ = with Σ1 = diag(d1 , . . . , dr ). Then, prove that
0 0
" #
−1
Σ1 0
G = V DU ∗ , for D = ∈ Mn,m (C) is the pseudo-inverse of A.
0 0
Chapter 7
Jordan canonical form
7.1 Jordan canonical form theorem

We start this chapter with the following theorem which generalizes the Schur Upper triangu-
larization theorem.
Theorem 7.1.1. [Generalized Schur’s theorem] Let A ∈ Mn (C). Suppose λ1 , . . . , λk are

the distinct eigenvalues of A with multiplicities m1 , . . . , mk , respectively. Then, there exists a
non-singular matrix W such that
T
k
AF
M
W −1 AW = Ti , where, Ti ∈ Mmi (C), for 1 ≤ i ≤ k
DR
i=1
and Ti ’s are upper triangular matrices with constant diagonal λi . If A has real entries with real
eigenvalues then W can be chosen to have real entries.
Proof. By Schur Upper Triangularization (see Lemma 6.4.1), there exists a unitary matrix U
such that U ∗ AU = T , an upper triangular matrix with diag(T ) = (λ1 , . . . , λ1 , . . . , λk , . . . , λk ).
Now, for any upper triangular matrix B, a real number α and i < j, consider the matrix
F (B, i, j, α) = Eij (−α)BEij (α), where the matrix Eij (α) is defined in Definition 2.2.5. Then,
for 1 ≤ k, ` ≤ n,


 Bij − αBjj + αBii , whenever k = i, ` = j


 B − αB , whenever ` 6= j
i` j`
(F (B, i, j, α))k` = (7.1.1)


 Bkj + αBki , whenever k 6= i

Bk` , otherwise.

Now, using Equation (7.1.1), the diagonal entries of F (T, i, j, α) and T are equal and
(
Tij , whenever Tjj = Tii
(F (T, i, j, α))ij = Tij
0, whenever Tjj 6= Tii and α = Tjj −Tii .
Thus, if we denote the matrix F (T, i, j, α) by T1 then (F (T1 , i − 1, j, α))i−1,j = 0, for some
choice of α, whenever (T1 )i−1,i−1 6= Tjj . Moreover, this operation also preserves the 0 created by
189
190 CHAPTER 7. JORDAN CANONICAL FORM
F (T, i, j, α) at (i, j)-th place. Similarly, F (T1 , i, j + 1, α) preserves the 0 created by F (T, i, j, α)
at (i, j)-th place. So, we can successively apply the following sequence of operations to get
T → F (T, m1 , m1 +1, α) = T1 → F (T1 , m1 −1, m1 +1, β) → · · · → F (Tm1 −1 , 1, m1 +1, γ) = Tm1 ,
where α, β, . . . , γ are appropriately chosen and Tm1 [:, m1 + 1] = λ2 em1 +1 . Thus, observe that
the above operation can be applied for different choices of i and j with i < j to get the required
result.
Practice 7.1.2. Apply Theorem 7.1.1 to the matrix given below for better understanding.
 
1 2 3 4 5 6 7 8 9
 
0 1 2 3 4 5 6 7 8
 
0 0 1 2 3 4 5 6 7
 
 
0 0 0 2 3 4 5 6 7
 
0 0 0 0 2 3 4 5 6.
 
 
0 0 0 0 0 2 3 4 5
 
 
0 0 0 0 0 0 3 4 5
 
0 0 0 0 0 0 0 3 4
 
0 0 0 0 0 0 0 0 3
Definition 7.1.3. 1. Let λ ∈ C and k be a positive integer. Then, by the Jordan block
T
Jk (λ) ∈ Mk (C), we understand the matrix

AF
 
λ 1
DR
.. ..
 
. .
 
.
 


 λ 1 
λ
2. A Jordan matrix is a direct sum of Jordan blocks. That is, if A is a Jordan matrix
having r blocks then there exist positive integers ki ’s and complex numbers λi ’s (not
necessarily distinct), for 1 ≤ i ≤ r such that
A = Jk1 (λ1 ) ⊕ · · · ⊕ Jkr (λr ).
We now give some examples of Jordan matrices with diagonal entries 0.

h i
Example 7.1.4. 1. J1 (0) = 0 is the only Jordan matrix of size 1.
" # " #
0 0 0 1
2. J1 (0) ⊕ J1 (0) = and J2 (0) = are Jordan matrices of size 2.
0 0 0 0
3. Even though , J1 (0) ⊕ J2 (0) and J2 (0) ⊕ J1 (0) are two Jordan matrices of size 3, we do
not differentiate between them as they are similar (use permutations).
     
0 0 0 0 1 0 0 1 0
     
4. J1 (0) ⊕ J1 (0) ⊕ J1 (0) = 
0 0 0, J2 (0) ⊕ J1 (0) = 0 0 0 and J3 (0) = 0 0 1
    
0 0 0 0 0 0 0 0 0
are Jordan matrices of size 3.
7.1. JORDAN CANONICAL FORM THEOREM 191
5. Observe that the number of Jordan matrices of size 4 with 0 on the diagonal are 5.
We now give some properties of the Jordan blocks. The proofs are immediate and hence left
for the reader. They will be used in the proof of subsequent results.
Remark 7.1.5. [Jordan blocks] Fix a positive integer k. Then,

1. Jk (λ) is an upper triangular matrix with λ as an eigenvalue.
2. Jk (λ) = λIk + Jk (0).
3. Alg.Mulλ (Jk (λ)) = k.
4. The matrix Jk (0) satisfies the following properties.
(a) Rank((Jk (0)i ) = k − i, for 1 ≤ i ≤ k.
" #
0 0
(b) JkT (0)Jk (0) = .
0 Ik−1
(c) Jk (0)p = 0 whenever p ≥ k.
(d) Jk (0)ei = ei−1 for i = 2, . . . , k.
" #
x1
(e) I − JkT (0)Jk (0) x = = hx, e1 ie1 .
0
5. Thus, using Remark 7.1.5.4d Geo.Mulλ (Jk (λ)) = 1.

Exercise 7.1.6. 1. Fix a positive integer k and a complex number λ. Then, prove that
T
(a) Rank(Jk (λ) − λIk ) = k − 1.

AF
(b) Rank(Jk (λ) − αIk ) = k, whenever α 6= λ. Or equivalently, for all α 6= λ the matrix
DR
Jk (λ) − αIk is invertible.

(c) for 1 ≤ i ≤ k, Rank((Jk (λ) − λIk )i ) = k − i.
(d) for α 6= λ, Rank((Jk (λ) − αIk )i ) = k, for all i.
2. Let J be a Jordan matrix that contains ` Jordan blocks for λ. Then, prove that
(a) Rank(J − λI) = n − `.
(b) J has ` linearly independent eigenvectors for λ.
(c) Rank(J − λI) ≥ Rank((J − λI)2 ) ≥ Rank((J − λI)3 ) ≥ · · · .
3. Let A ∈ Mn (C). Then, prove that AJn (λ) = Jn (λ)A if and only if AJn (0) = Jn (0)A.
Definition 7.1.7. Let J be a Jordan matrix containing Jt (λ), for some positive integer t
and some complex number λ. Then, the smallest value of k for which Rank((J − λI)k ) stops
decreasing is the order of the largest Jordan block Jk (λ) in J. This number k is called the
index of the eigenvalue λ.
Lemma 7.1.8. Let A ∈ Mn (C) be strictly upper triangular. Then, A is similar to a direct sum
of Jordan blocks. Or equivalently, there exists integers n1 ≥ . . . ≥ nm ≥ 1 and a non-singular
matrix S such that
A = S −1 Jn1 (0) ⊕ · · · ⊕ Jnm (0) S.
If A ∈ Mn (R) then S can be chosen to have real entries.

Proof. We will prove the result by induction on n. For n = 1, the statement is trivial. So, let
the result be # matrices of size ≤ n − 1 and let A ∈ Mn (C) be strictly upper triangular.
" true for
0 a T
Then, A = . By induction hypothesis there exists an invertible matrix S1 such that
0 A1
m
A1 = S1−1 Jn1 (0) ⊕ · · · ⊕ Jnm (0) S1 with
X
ni = n − 1.
i=1
Thus,
 
" # " # " #" #" # " # 0 aT1 aT2
1 0 1 0 1 0 0 aT 1 0 0 a T S1  
A = = =
0 Jn1 (0) 0 ,

0 S1−1 0 S1 0 S1−1 0 A1 0 S1 0 S −1 A1 S1
0 0 J
h i
where S1−1 Jn1 (0) ⊕ · · · ⊕ Jnm (0) S1 = Jn1 (0) ⊕ J and aT S1 = aT1 aT2 . Now, writing Jn1 to
mean Jn1 (0) and using Remark 7.1.5.4e, we have
     
1 −aT1 JnT1 0 0 aT1 aT2 1 aT1 JnT1 0 0 ha1 , e1 ieT1 aT2
     
 = 0 .
0 In1 0
0 Jn1 0 0 In1 0 Jn1 0
  

0 0 I 0 0 J 0 0 I 0 0 J
So, we now need to consider  two cases depending on whether ha1 , e

1 i = 0 or ha1 ,e1 i =
6 0. In the
0 0 a2 T J 0 0
T
   n1 
first case, A is similar to 0 Jn1 0 . This in turn is similar to  0 0 aT2 
AF
 by permuting
  
0 0 J 0 0 J
DR
the first row and column. At this stage, one can apply induction and if necessary do a block
permutation, in order to keep the block sizes in decreasing order.
So, let us now assume that ha1 , e1 i =
6 0. Then, writing α = ha1 , e1 i, we have
     
1 T aT T T
0 0 0 αe 1 2 α 0 0 0 e1 a 2
" #
α      Jn1 +1 e a
1 2
T
 0 I 0  = 0 Jn1 0  ≡ .
 0 I 0 0 Jn 0    
1

1
 0 J
0 0 αI 0 0 J 0 0 αI 0 0 J
Now, using Remark 7.1.5.4c, verify that

" #" #" # " #
I ei+1 aT2 J i−1 Jn1 +1 ei aT2 J i−1 I −ei+1 aT2 J i−1 Jn1 +1 ei+1 aT2 J i
= , for i ≥ 1.
0 I 0 J 0 I 0 J
Hence, for p = n − n1 − 1, we have

" # " #" #" # " # " #
I ep+1 aT2 J p−1 I e2 aT Jn1 +1 e1 aT I −e2 aT I −ep+1 aT J p−1 Jn1 +1 0
··· ··· = .
0 I 0 I 0 J 0 I 0 I 0 J
If necessary, we need to do a block permutation, in order to keep the block sizes in decreasing
order. Hence, the required result follows.
   
0 1 1 0 1 2
   
Practice 7.1.9. Convert  0 0 1  to J3 (0) and 0 0 0 to J2 (0) ⊕ J1 (0).
   
0 0 0 0 0 0
Corollary 7.1.10. A ∈ Mn (C). Then, A is similar to J, a Jordan matrix.
Proof. Let λ1 , . . . , λk be the distinct eigenvalues of A with algebraic multiplicities m1 , . . . , mk .

k
By Theorem 7.1.1, there exists a non-singular matrix S such that S −1 AS =
L
Ti , where Ti
i=1
is upper triangular with diagonal (λi , . . . , λi ). Thus Ti − λi Imi is a strictly upper triangular
matrix. Thus, by Theorem 7.1.8, there exist a non-singular matrix Si such that

Si−1 Ti − λi Imi Si = J(0),
a Jordan matrix with 0 on the diagonal and the size of the Jordan blocks decreases as we move
down the diagonal. So, Si−1 Ti Si = J(λi ) is a Jordan matrix with λi on the diagonal and the
size of the Jordan blocks
k decreases
as we move down the diagonal.
Si . Then, verify that W −1 AW is a Jordan matrix.
L
Now, take W = S
i=1
Let A ∈ Mn (C). Suppose λ ∈ σ(A) and J is a Jordan matrix that is similar to A. Then, for
each fixed i, 1 ≤ i ≤ n, by ì (λ), we denote the number of Jordan blocks Jk (λ) in J for which
k ≥ i. Then, the next result uses Exercise 7.1.6 to determine the number ì (λ).
Remark 7.1.11. Let A ∈ Mn (C). Suppose λ ∈ σ(A) and J is a Jordan matrix that is similar
to A. Then, for 1 ≤ k ≤ n,
`k (λ) = Rank(A − λI)k−1 − Rank(A − λI)k .

T
Proof. In view of Exercise 7.1.6, we need to consider only the Jordan blocks Jk (λ), for different
AF
n
L
values of k. Hence, without loss of generality, let us assume that J = ai Ji (λ), where ai ’s are
DR
i=1
non-negative integers and J contains exactly ai copies of the Jordan block Ji (λ), for 1 ≤ i ≤ n.
Then, by definition and Exercise 7.1.6, we observe the following:
P
1. n = iai .
i≥1
P
2. Rank(J − λI) = (i − 1)ai .
i≥2
3. Rank((J − λI)2 ) =
P
(i − 2)ai .
i≥3
4. In general, for 1 ≤ k ≤ n, Rank((J − λI)k ) =

P
(i − k)ai .
i≥k+1
Thus, writing ì in place of ì (λ), we get

X X X
`1 = ai = iai − (i − 1)ai = n − Rank(J − λI),
i≥1 i≥1 i≥2
X X X
`2 = ai = (i − 1)ai − (i − 2)ai = Rank(J − λI) − Rank((J − λI)2 ),
i≥2 i≥2 i≥3
..
.
X X X
`k = ai = (i − (k − 1))ai − (i − k)ai = Rank((J − λI)k−1 ) − Rank((J − λI)k ).
i≥k i≥k i≥k+1
Now, the required result follows as rank is invariant under similarity operation and the matrices
J and A are similar.
Lemma 7.1.12. [Similar Jordan matrices] Let J and J 0 be two similar Jordan matrices of
size n. Then, J is a block permutation of J 0 .
Proof. For 1 ≤ i ≤ n, let ì and `0i be, respectively, the number of Jordan blocks of J and J 0
of size at least i corresponding to λ. Since J and J 0 are similar, the matrices (J − λI)i and
(J 0 − λI)i are similar for all i, 1 ≤ i ≤ n. Therefore, their ranks are equal for all i ≥ 1 and
hence, ì = `0i for all i ≥ 1. Thus the required result follows.
We now state the main result of this section which directly follows from Lemma 6.4.1,
Theorem 7.1.1 and Corollary 7.1.10 and hence the proof is omitted.
Theorem 7.1.13. [Jordan canonical form theorem] Let A ∈ Mn (C). Then, A is similar to
a Jordan matrix J, which is unique up to permutation of Jordan blocks. If A ∈ Mn (R) and has
real eigenvalues then the similarity transformation matrix S may be chosen to have real entries.
This matrix J is called the the Jordan canonical form of A, denoted Jordan CF(A).
We now start with a few examples and observations.
Example 7.1.14. Let us use the idea from Lemma 7.1.11 to find the Jordan Canonical Form
of the following matrices.
 
0 0 1 0
 
0 0 0 1
1. Let A = J4 (0)2 =  .
T
0 0 0 0
AF
 
0 0 0 0
DR
Solution: Note that `1 = 4 − Rank(A − 0I) = 2. So, there are two Jordan blocks.
Also, `2 = Rank(A − 0I) − Rank((A − 0I)2 ) = 2. So, there are at least 2 Jordan blocks of
size 2. As there are exactly two Jordan blocks, both the blocks must have size 2. Hence,
Jordan CF(A) = J2 (0) ⊕ J2 (0).
 
1 1 0 1
 
0 1 1 1
2. Let A1 = 
 .
0 0 1 1 

0 0 0 1
Solution: Let B = A1 − I. Then, `1 = 4 − Rank(B) = 1. So, B has exactly one Jordan
block and hence A1 is similar to J4 (1).
 
1 1 0 1
 
0 1 1 1
3. A2 = 0
.
 0 1 0 
0 0 0 1
Solution: Let C = A2 − I. Then, `1 = 4 − Rank(C) = 2. So, C has exactly two Jordan
blocks. Also, `2 = Rank(C) − Rank(C 2 ) = 1 and `3 = Rank(C 2 ) − Rank(C 3 ) = 1. So, there
is at least 1 Jordan blocks of size 3.
Thus, we see that there are two Jordan blocks and one of them is of size 3. Also, the size
of the matrix is 4. Thus, A2 is similar to J3 (1) ⊕ J1 (1).
4. Let A = J4 (1)2 ⊕ A1 ⊕ A2 , where A1 and A2 are given in the previous exercises.

Solution: One can directly get the answer from the previous exercises as the matrix A is
already in the block diagonal form. But, we compute it again for better understanding.
Let B = A − I. Then, `1 = 16 − Rank(B) = 5, `2 = Rank(B) − Rank(B 2 ) = 11 − 7 = 4,
`3 = Rank(B 2 ) − Rank(B 3 ) = 7 − 3 = 4 and `4 = Rank(B 3 ) − Rank(B 4 ) = 3 − 0 = 3.
Hence, J4 (1) appears thrice (as `4 = 3 and `5 = 0), J3 (1) also appears once (as `3 −`4 = 1),
J2 (1) does not appear as (as `2 − `3 = 0) and J1 (1) appears once (as `1 − `2 = 1). Thus,
Remark 7.1.15. [Observations about Jordan CF(A)]

1. What are the steps to find Jordan CFA?
Ans: Let λ1 , . . . , λk be the distinct eigenvalues of A. Now, apply the Schur Upper Tri-
angularization Lemma (see Lemma 6.4.1) to get an upper triangular matrix, say T such
that the diagonal entries of T are λ1 , . . . , λ1 , λ2 , . . . , λ2 , . . . , λk , . . . , λk . Now, apply sim-
L k
ilarity transformations (see Theorem 7.1.1) to get T = Ti , where each diagonal entry
i=1
of Ti is λi . Then, for each i, 1 ≤ i ≤ k, use Theorem 7.1.8 to get an invertible matrix
Si such that Si−1 (Ti − λi I)Si = Jei , a Jordan matrix. Thus, we obtain a Jordan matrix
k
Ji = Jei + λi I = Si−1 Ti Si , where each diagonal entry of Ji is λi . Hence, S =
L
Si converts
T
i=1
k
AF
L
T = Ti into the required Jordan matrix.
i=1
DR
n
2. Let A ∈ Mn (C) be a diagonalizable matrix. Then, by definition, A is similar to
L
λi ,
i=1
n
L
where λi ∈ σ(A), for 1 ≤ i ≤ n. Thus, Jordan CF(A) = λi , up to a permutation of
i=1
λi ’s.
3. In general, the computation
" # of Jordan CF(A) is not numerically stable. To understand
0
this, let A = . Then, A is diagonalizable as A has distinct eigenvalues. So,
1 0
" #
0
Jordan CF(A ) = .
0 0
" # " #
0 0 0 1
Whereas, for A = , we know that Jordan CF(A) = 6= lim Jordan CF(A ).
1 0 0 0 →0
Thus, a small change in the entries of A may change Jordan CF(A) significantly.
4. Let A ∈ Mn (C) and > 0 be given. Then, there exists an invertible matrix S such
k
that S −1 AS =
L
Jni (λi , ), where Jni (λi , ) is obtained from Jni (λi ) by replacing each
i=1
off diagonal entry 1 by an . To get this, define Di() = diag(1, , 2 , . . . , ni −1 ), for
k
(Di())−1 Jni (λi )Di() .
L
1 ≤ i ≤ k. Now compute
i=1
5. Let Jordan CF(A) contain ` Jordan blocks for λ. Then, A has ` linearly independent
eigenvectors for λ.
For if, A has at least ` + 1 linearly independent eigenvectors for λ, then dim(Null(A −
λI)) > `. So, Rank(A − λI) < n − `. But, the number of Jordan blocks for λ in A is `.
Thus, we must have Rank(J − λI) = n − `, a contradiction.
6. Let λ ∈ σ(A). Then, by Remark 7.1.5.5, Geo.Mulλ (A) = the number of Jordan blocks
Jk (λ) in Jordan CF(A).
7. Let λ ∈ σ(A). Then, by Remark 7.1.5.3, Alg.Mulλ (A) = the sum of the sizes of all
Jordan blocks Jk (λ) in Jordan CF(A).
8. Let λ ∈ σ(A). Then, Jordan CF(A) does not  get determined
 by Alg.Mulλ (A) and
" # 0 1 0 " # " # " #
h i 0 1   0 1 0 1 0 1
Geo.Mulλ (A). For example, 0 ⊕ ⊕0 0 1 and 0 0 ⊕ 0 0 ⊕ 0 0

0 0
0 0 0
are different Jordan CFs but they have the same algebraic and geometric multiplicities.
9. Let A ∈ Mn (C). Suppose that, for each λ ∈ σ(A), the values of Rank(A − λI)k , for
k = 1, . . . , n are known. Then, using Remark 7.1.11, Jordan
h i CF(A) can be computed.
But, note here that finding rank is numerically unstable as has rank 1 but it converges
h i
to 0 which has a different rank.
Theorem 7.1.16. [A is similar to AT ] Let A ∈ Mn (C). Then, A is similar to AT .

 
1
T
. . . Then, observe that K −1 = K and KJn (a)K = Jn (a)T , as the

AF
 
Proof. Let Kn =   . 
1
DR
(i, j)-th entry of A goes to (n − i + 1, n − j + 1)-th position in KAK. Hence,

hM i hM i hM i hM iT
Kni Jni (λi ) Kni = Jni (λi ) .
Thus, J is similar to J T . But, A is similar to J and hence AT is similar to J T and finally we

get A is similar to AT . Therefore, the
 required result follows.
    
−2 0 −1 2 2 

 1 0 
      
 −2 2 −1 1 1  

 0 1 

      
Exercise 7.1.17. 1. Let M =  1 0 2 0 −1 , Null(M − 2I) = LS 0 , 0
     
      
 −3 0 −1 4 1  

 1 0
  
    

−4 0 −1 2
 
4  1 0 
and σ(M ) = {2, 2, 2, 2, 2}. Then determine the Jordan canonical form of M .
2. Fix k ∈ N and let λ 6= 0.
 
1/λ −1/λ2 1/λ3 · · · (−1)k+1 /λk
.. ..
 
 0

1/λ . k k−1
. (−1) /λ 

..
 
(a) Then prove that Jk (λ)−1 = 0
 .. , for all k ≥ 1.

 0 . . 
 .. 
 0
 0 . 1/λ −1/λ2  
0 0 0 1/λ

1
(b) Show that Jk (λ)−1 − Ik is a nilpotent matrix of index k.
λ
7.2. MINIMAL POLYNOMIAL 197

1
(c) Use the previous part to conclude Jordan CF Jk (λ)−1 − Ik = Jk (0).
λ

1
(d) Therefore, prove that Jordan CF Jk (λ)−1 = Jk

.
λ
Lk
(e) Further, let Jordan CF(A) = Jni (λ) for some integers n1 ≥ · · · ≥ nk ≥ 1. Then
i=1
k
CF(A−1 )
L
Jordan = Jni (1/λ).
i=1
k
L
(f ) Let Jordan CF(A) = Jni (λi ) for some integers n1 ≥ · · · ≥ nk ≥ 1 and λi 6= 0
i=1
k
for 1 ≤ i ≤ k. Then prove that Jordan CF(A−1 ) =
L
Jni (1/λi ).
i=1
7.2 Minimal polynomial

Recall that a polynomial p(x) = a0 + a1 x + · · · + an xn with an = 1 is called a monic polynomial.
We now have the following definition.
Definition 7.2.1. Let P (t) = tn + an−1 tn−1 + · · · + a0 be a monic polynomial in t of degree
0 0 0 · · · 0 −a0
 
1 0 0
 ··· 0 −a1  
0 1 0 · · · 0 −a
 
 2 
n. Then, the n × n matrix A =  .. . . .. .. , denoted A(n : a0 , . . . , an−1 ) or
T

0 0 . . . . 
AF
 
0 0 0
 · · · 0 −an−2  
DR
0 0 0 1 −an−1
Companion(P ), is called the companion matrix of P (t).
Remark 7.2.2. Let A ∈ Mn (C) and let f (x) = xn +an−1 xn−1 +· · ·+a1 x+a0 be its characteristic
polynomial. Then by the Cayley Hamilton Theorem, An + an−1 An−1 + · · · + a1 A + a0 I = 0.
Hence An = −(an−1 An−1 + · · · + a1 A + a0 I).
Suppose, there exists a vector u ∈ Cn such that B = u, Au, A2 u, . . . , An−1 u is an ordered

basis of Cn . Then, An u = −(an−1 An−1 u + · · · + a1 Au + a0 u) and hence the matrix of A in the

basis B equals
h i h i
A[B, B] = [Au]B [A(Au)]B · · · [A(An−1 u)]B = [Au]B [A2 u]B · · · [An u]B
 
0 0 0 ··· 0 −a0
 
1 0 0
 ··· 0 −a1 

0 1 0 ··· 0 −a2 
 
=  .. .. .. ,
.. 
0 0 . . . . 

 
0 0 0
 ··· 0 −an−2  
0 0 0 1 −an−1
the companion matrix of A.
Definition 7.2.3. Let A ∈ Mn (C). Then, the polynomial P (t) is said to annihilate (destroy)
A if P (A) = 0.
Let P (x) be the characteristic polynomial of A. Then, by the Cayley-Hamilton Theorem,

P (A) = 0. So, if f (x) = P (x)g(x), for any multiple of g(x), then f (A) = P (A)g(A) = 0g(A) =
0. Thus, there are infinitely many polynomials which annihilate A. In this section, we will
concentrate on a monic polynomial of least positive degree that annihilates A.
Definition 7.2.4. Let A ∈ Mn (C). Then, the minimal polynomial of A, denoted mA (x), is
a monic polynomial of least positive degree satisfying mA (A) = 0.
Theorem 7.2.5. Let A be the companion matrix of the monic polynomial P (t) = tn +an−1 tn−1 +
· · · + a0 . Then, P (t) is both the characteristic and the minimal polynomial of A.
Proof. Expanding det(tIn − Companion(P )) along the first row, we have
det(tIn − A(n : a0 , . . . , an−1 )) = t det(tIn−1 − A(n − 1 : a1 , . . . , an−1 )) + (−1)n+1 a0 (−1)n−1

= t2 det(tIn−2 − A(n − 2 : a2 , . . . , an−1 )) + a0 + a1 t
..
.
= P (t).
Thus, P (t) is the characteristic polynomial of A and hence P (A) = 0.

We will now show that P (t) is the minimal polynomial of A. To do so, we first observe that
Ae1 = e2 , . . . , Aen−1 = en . Thus,
T
AF
Ak e1 = ek+1 , for 1 ≤ k ≤ n − 1. (7.2.1)

DR
Now, Suppose we have a monic polynomial Q(t) = tm + bm−1 tm−1 + · · · + b0 , with m < n,
such that Q(A) = 0. Then, using Equation (7.2.1), we get
0 = Q(A)e1 = Am e1 + bm−1 Am−1 e1 + · · · + b0 Ie1 = em+1 + bm−1 em + · · · + b0 e1 ,
a contradiction to the linear independence of {e1 , . . . , em+1 } ⊆ {e1 , . . . , en }.

The next result gives us the existence of such a polynomial for every matrix A. To do so,
recall that the well-ordering principle implies that if S is a subset of natural numbers then it
contains a least element.
Lemma 7.2.6. [Existence of the minimal polynomial] Let A ∈ Mn (C). Then, there exists a
unique monic polynomial m(x) of minimum (positive) degree such that m(A) = 0. Further, if
f (x) is any polynomial with f (A) = 0 then m(x) divides f (x).
Proof. Let P (x) be the characteristic polynomial of A. Then, deg(P (x)) = n and by the
Cayley-Hamilton Theorem, P (A) = 0. So, consider the set
S = {deg(f (x)) : f (x) is a nonzero polynomial, f (A) = 0}.
Then, S is a non-empty subset of N as n ∈ S. Thus, by well-ordering principle there ex-

ists a smallest positive integer, say M , and a corresponding polynomial, say m(x), such that
deg(m(x)) = M , m(A) = 0.
7.2. MINIMAL POLYNOMIAL 199
Also, without loss of generality, we can assume that m(x) is monic and unique (non-
uniqueness will lead to a polynomial of smaller degree in S).
Now, suppose there is a polynomial f (x) such that f (A) = 0. Then, by division algorithm,
there exist polynomials q(x) and r(x) such that f (x) = m(x)q(x) + r(x), where either r(x) is
identically the zero polynomial of deg(r(x)) < M = deg(m(x)). As
0 = f (A) = m(A)q(A) + r(A) = 0q(A) + r(A) = r(A),
we get r(A) = 0. But, m(x) was the least degree polynomial with m(A) = 0 and hence r(x) is
the zero polynomial. That is, m(x) divides f (x).
As an immediate corollary, we have the following result.
Corollary 7.2.7. [Minimal polynomial divides the characteristic polynomial] Let mA (x)
and PA (x) be, respectively, the minimal and the characteristic polynomials of A ∈ Mn (C).
1. Then, mA (x) divides PA (x).
2. Further, if λ is an eigenvalue of A then mA (λ) = 0.
Proof. The first part following directly from Lemma 7.2.6. For the second part, let (λ, x) be an
eigen-pair. Then, f (A)x = f (λ)x, for any polynomial of f , implies that
mA (λ)x = mA (A)x = 0x = 0.
T
AF
But, x 6= 0 and hence mA (λ) = 0. Thus, the required result follows.

DR
we also have the following result.
Lemma 7.2.8. Let A and B be two similar matrices. Then, they have the same minimal
polynomial.
Proof. Since A and B are similar, there exists an invertible matrix S such that A = S −1 BS.
Hence, f (A) = F (S −1 BS) = S −1 f (B)S, for any polynomial f . Hence, mA (A) = 0 if and only
if mA (B) = 0 and thus the required result follows.
Theorem 7.2.9. Let A ∈ Mn (C) and let λ1 , . . . , λk be the distinct eigenvalues of A. If ni is

the size of the largest Jordan block for λi in J = Jordan CFA then
k
Y
mA (x) = (x − λi )ni .
i=1
k
(x−λi )αi , for some αi ’s with 1 ≤ αi ≤ Alg.Mulλi (A).
Q
Proof. Using 7.2.7, we see that mA (x) =
i=1
k
(J − λi I)αi = 0. But, observe that
Q
As mA (A) = 0, using Lemma 7.2.8 we have mA (J) =
i=1
for the Jordan block Jni (λi ), one has
1. (Jni (λi ) − λi I)αi = 0 if and only if αi ≥ ni , and
2. (Jnm (λm ) − λi I)αi is invertible, for all m 6= i.

k k k k
(J − λi I)ni = 0 and (x − λi )ni divides (x − λi )αi = mA (x) and (x − λi )ni
Q Q Q Q
Thus
i=1 i=1 i=1 i=1
is a monic polynomial, the result follows.
As an immediate consequence, we also have the following result which corresponds to the
converse of the above theorem.
Theorem 7.2.10. Let A ∈ Mn (C) and let λ1 , . . . , λk be the distinct eigenvalues of A. If the
k
(x − λi )ni then ni is the size of the largest Jordan block for
Q
minimal polynomial of A equals
i=1
λi in J = Jordan CFA.
Proof. It directly follows from Theorem 7.2.9.

We now give equivalent conditions for a square matrix to be diagonalizable.
1. A is diagonalizable.
2. Every zero of mA (x) has multiplicity 1.
d
3. Whenever mA (α) = 0, for some α, then mA (x)x=α 6= 0.
dx
Proof. Part 1 ⇒ Part 2. If A is diagonalizable, then each Jordan block in J = Jordan CFA
k
Q
has size 1. Hence, by Theorem 7.2.9, mA (x) = (x−λi ), where λi ’s are the distinct eigenvalues
i=1
of A.
k
T
Q
Part 2 ⇒ Part 3. Let mA (x) = (x − λi ), where λi ’s are the distinct eigenvalues of A.
AF
i=1
Then, mA (x) = 0 if and only if x = some i, 1 ≤ i ≤ k. In that case, it is easy to verify
λi , for
d
DR
that mA (x) 6= 0, for each λi .

dx
d
Part 3 ⇒ Part 1. Suppose that for each α satisfying mA (α) = 0, one has mA (α) 6= 0.
dx
Then, it follows that each zero of mA (x) has multiplicity 1. Also, using Corollary 7.2.7, each
zero of mA (x) is an eigenvalue of A and hence by Theorem 7.2.9, the size of each Jordan block
is 1. Thus, A is diagonalizable.
We now have the following remarks and observations.
Remark 7.2.12. 1. Let f (x) be a monic polynomial and A = Companion(f ) be the com-
panion matrix of f . Then, by Theorem 7.2.5) f (A) = 0 and no monic polynomial of
smaller degree annihilates A. Thus PA (x) = mA (x) = f (x), where PA (x) is the charac-
teristic polynomial and mA (x), the minimal polynomial of A.
2. Let A ∈ Mn (C). Then, A is similar to Companion(f ), for some monic polynomial f if
and only if mA (x) = f (x).
Proof. Let B = Companion (f ). Then, using Lemma 7.2.8, we see that mA (x) = mB (x).
But, by Remark 7.2.12.1, we get mB (x) = f (x) and hence the required result follows.
Conversely, assume that mA (x) = f (x). But, by Remark 7.2.12.1, mB (x) = f (x) =
PB (x), the characteristic polynomial of B. Since mA (x) = mB (x), the matrices A and
B have the same largest Jordan blocks for each eigenvalue λ. As PB = mB , we know
that for each λ, there is only one Jordan block in Jordan CFB. Thus, Jordan CFA =
Jordan CFB and hence A is similar to Companion (f ).
7.3. APPLICATIONS OF JORDAN CANONICAL FORM 201
Exercise 7.2.13. The following are some facts and questions.

1. Let A ∈ Mn (C). If PA (x) is the minimal polynomial of A then A is similar to Companion (PA )
if and only if A is nonderogatory. T/F?
2. Let A, B ∈ M3 (C) with eigenvalues 1, 2, 3. Is it necessary that A is similar to B?
3. Let A, B ∈ M3 (C) with eigenvalues 1, 1, 3. Is it necessary that A is similar to B?
4. Let A, B ∈ M4 (C) with the same minimal polynomial. Is it necessary that A is similar to
B?
5. Let A, B ∈ M3 (C) with the same minimal polynomial. Is it necessary that A is similar to
B?
6. Let A ∈ Mn (C) be idempotent and let J = Jordan CFA. Thus, J 2 = J and hence con-
clude that J must be a diagonal matrix. Hence, every idempotent matrix is diagonalizable.
7. Let A ∈ Mn (C). Suppose that mA (x)|x(x − 1)(x − 2)(x − 3). Must A be diagonalizable?
8. Let A ∈ M9 (C) be a nilpotent matrix such that A5 6= 0 but A6 = 0. Determine PA (x) and
mA (x).
9. Recall that for A, B ∈ Mn (C), the characteristic polynomial of AB and BA are the same.
That is, P"
AB (x)#= PBA (x)." However,
# they need not have the same minimal polynomial.
0 1 0 0
Take A = and B = to verify that mAB (x) 6= mBA (x).
0 0 0 1
T
AF
10. Let A ∈ Mn (C) be an invertible matrix. Then prove that if the minimal polynomial of A
equals m(x, λ1 , . . . , λk ) then the minimal polynomial of A−1 equals m(x, 1/λ1 , . . . , 1/λk ).
DR
11. Let λ an eigenvalue of A ∈ Mn (C) with two linearly independent eigenvectors. Show that
there does not exist a vector u ∈ Cn such that LS u, Au, A2 u, . . . = Cn .

We end this section with a method to compute the minimal polynomial of a given matrix.
Example 7.2.14. [Computing the minimal polynomial] Let λ1 , . . . , λk be the distinct eigen-
values of A ∈ Mn (C).
7.3 Applications of Jordan Canonical Form

In the last section, we say that the matrices if A is a square matrix then A and AT are similar.
In this section, we look at some more applications of the Jordan Canonical Form.
7.3.1 Coupled system of linear differential equations

   
x01 (t) x1 (t)
 .   . 
Consider the first order Initial Value Problem (IVP) x0 (t) =  .   . 
 .  = A .  = Ax(t),
x0n (t) xn (t)
with x(0) = 0. If A is not a diagonal matrix then the system is called coupled and is hard
to solve. Note that if A can be transformed to a nearly diagonal matrix, then the amount of
coupling among xi ’s can be reduced. So, let us look at J = Jordan CF(A) = S −1 AS. Then,
using S −1 A = JS −1 . verify that the initial problem x0 (t) = Ax(t) is equivalent to the equation
S −1 x0 (t) = S −1 Ax(t) which in turn is equivalent to y0 (t) = Jy(t), where S −1 x(t) = y(t) with
y(0) = S −1 x(0) = 0. Therefore, if y is a solution to the second equation then x(t) = Sy is a
solution to the initial problem.
When J is diagonalizable then solving the second is as easy as solving yi0 (t) = λi yi (t) for
which the required solution is given by yi (t) = yi (0)eλi t .
If J is not diagonal, then for each Jordan block, the system reduces to
y10 (t) = λy1 (t) + y2 (t), · · · , yk−1

0
(t) = λyk−1 (t) + yk (t), yk0 (t) = λyk (t).
This problem can also be solved as in this case the solution is given by yk = c0 eλt ; yk−1 =
(c0 t + c1 )eλt and so on.
7.3.2 Commuting matrices
Let P (x) be a polynomial and A ∈ Mn (C). Then, P (A)A = AP (A). What about the converse?
That is, suppose we are given that AB = BA for some B ∈ Mn (C). Does it necessarily imply
that B = P (A), for some nonzero polynomial P (x)? The answer is No as I commutes with A
for every A. We start with a set of remarks.
Theorem 7.3.1. Let A ∈ Mn (C) and B ∈ Mm (C). Then, the linear system AX − XB = 0, in
T
AF
the variable matrix X of size n × m, has a unique solution, namely X = 0 (the trivial solution),
if and only if σ(A) and σ(B) are disjoint.
DR
Proof. Let us assume that σ(A) and σ(B) are disjoint. !

Q
Since σ(A) and σ(B) are disjoint, the matrix PB (A) = [λI − A] , obtained by
λ∈σ(B)
evaluating A at the characteristic polynomial, PB (t), of B, is invertible. So, let us look at
the implication of the condition AX = XB. This condition implies that A2 X = AXB =
XBB = XB 2 and hence, P (A)X = XP (B), for any polynomial P (t). In particular, PB (A)X =
XPB (B) = X0 = 0. As PB (A) is invertible, we get X = 0.
Now, conversely assume that AX − XB = 0 has only the trivial solution X = 0. Suppose
on the contrary λ is a common eigenvalue of both A and B. So, choose nonzero vectors x ∈ Cn
and y ∈ Cm such that (λ, x) is an eigen-pair of A and (λ, y) is a left eigen-pair of B. Now,
define X0 = xyT . Then, X0 is an n × m nonzero matrix and
AX − XB = AxyT − xyT B = λxyT − λxyT = 0.
Thus, we see that if λ is a common eigenvalue of A and B then the system AX − XB = 0 has
a nonzero solution X0 , a contradiction. Hence, the required result follows.
Corollary 7.3.2. Let A ∈ Mn (C), B ∈ Mm (C) and C be an n × m matrix. Also, assume that
σ(A) and σ(B) are disjoint. Then, it can be easily verified that the system AX − XB = C, in
the variable matrix X of size n × m, has a unique solution, for any given C.
7.3. APPLICATIONS OF JORDAN CANONICAL FORM 203
Proof. Consider the linear transformation T : Mn,m (C) → Mn,m (C), defined by T (X) =
AX − XB. Then, by Theorem 7.3.1, Null(T ) = {0}. Hence, by the rank-nullity theorem, T
is a bijection and the required result follows.
Definition 7.3.3. A square matrix A is said to be of Toeplitz type if each (super/sub)-


b1 b2 b3 b4
 
a1 b1 b2 b3 
diagonal of A consists of the same element. For example, A =   is a 4 × 4
a2 a1 b1 b2 

a3 a2 a1 b1
 
b1 b2 b3 b4
 
 0 b1 b2 b3 
Toeplitz type matrix. and the matrix B =    is an upper triangular Toeplitz
 0 0 b1 b2 

0 0 0 b1
type matrix.
Exercise 7.3.4. Let Jn (0) ∈ Mn (C) be the Jordan block with 0 on the diagonal.
1. Further, if A ∈ Mn (C) such that AJn (0) = Jn (0)A then prove that A is an upper Toeplitz
type matrix.
2. Further, if A, B ∈ Mn (C) are two upper Toeplitz type matrices then prove that
(a) there exists ai ∈ C, 1 ≤ i ≤ n, such that A = a0 I + a1 Jn (0) + · · · + an Jn (0)n−1 .
T
AF
(b) P (A) is a Toeplitz matrix for any polynomial P (t).

(c) AB is a Toeplitz matrix.
DR
(d) if A is invertible then A−1 is also an upper Toeplitz type matrix.
To proceed further, recall that a matrix A ∈ Mn (C) is called non-derogatory if Geo.Mulα (A) =
1, for each α ∈ σ(A) (see Definition 6.3.9).
Theorem 7.3.5. Let A ∈ Mn (C) be a non-derogatory matrix. Then, the matrices A and B
commute if and only if B = P (A), for some polynomial P (t) of degree at most n − 1.
Proof. If B = P (A), for some polynomial P (t), then A and B commute. Conversely, suppose
that AB = BA, σ(A) = {λ1 , . . . , λk } and let J = Jordan CFA = S −1 AS be the Jordan matrix
   
Jn1 (λ1 ) B 11 · · · B 1k
.. . Now, write B = S −1 BS =  .. . . . ..
  
of A. Then, J =  .   . ., where

Jnk (λk ) B k1 · · · B kk
B is partitioned conformally with J. Note that AB = BA gives JB = BJ. Thus, verify that
Jn1 (λ1 )B 12 = [JB]12 = [BJ]12 = B 12 Jn2 (λ2 ),
and hence B 12 = 0. A similar argument gives B ij = 0, for all i 6= j. Hence, JB = BJ implies

Jni (λi )B ii = B ii Jni (λi ), for 1 ≤ i ≤ k. Or equivalently, Jni (0)Bii = Bii Jni (0), for 1 ≤ i ≤ k
(using Exercise 7.1.6.3). Now, using Exercise 7.3.4.1, we see that B ii is an upper triangular
Toeplitz type matrix.
(t − λj )nj . Then, Fi (t) is a polynomial

Q
To proceed further, for 1 ≤ i ≤ k, define Fi (t) =
j6=i
with deg(Fi (t)) = n − ni and Fi (Jnj (λj )) = 0 if j 6= i. Also, note that Fi (Jni (λi )) is a
nonsingular upper triangular Toeplitz type matrix. Hence, its inverse has the same form and
using Exercise 7.3.4.1, the matrix Fi (Jni (λi ))−1 B ii is also a Toeplitz type upper triangular
matrix. Hence,
Fi (Jni (λi ))−1 B ii = c1 I + c2 Jni (0) + · · · + cni Jni (0)ni −1 = Ri (Jni (λi )), (say).
Thus, B ii = Fi (Jni (λi ))Ri (Jni (λi )). Putting Pi (t) = Fi (t)Ri (t), for 1 ≤ i ≤ k, we see that Pi (t)
is a polynomial of degree at most n − 1 with Pi ((Jnj (λj )) = 0, for j 6= i and Pi ((Jnj (λi )) = B ii .
Taking, P = P1 + · · · + Pk , we have
   
Jn1 (λ1 ) Jn1 (λ1 )
P (J) = P1 
 ..  
 + · · · + Pk  .. 

 .   . 

Jnk (λk ) Jnk (λk )
   
B 11 0

=  .. 
+ · · · +  ...
 
 = B.
 .    
0 B kk
Hence, B = SBS −1 = SP (J)S −1 = P (SJS −1 ) = P (A) and the required result follows.
T
AF
DR
Chapter 8
Advanced Topics on
Diagonalizability and
Triangularization∗
8.1 More on the Spectrum of a Matrix
We start this subsection with a few definitions and examples. So, it will be nice to recall the
T
notations used in Section 1.5 and a few results from Appendix 9.2.
AF
Definition 8.1.1. [Principal Minor] Let A ∈ Mn (C).

DR
1. Also, let S ⊆ [n]. Then, det (A[S, S]) is called the Principal minor of A corresponding
to S.
2. By EMk (A), we denote the sum of all k × k principal minors of A.
Definition 8.1.2. [Elementary Symmetric Functions] Let k be a positive integer. Then,

the kth elementary symmetric function of the numbers r1 , . . . , rn is Sk (r1 , . . . , rn ) and is
defined as
X
Sk (r1 , . . . , rn ) = ri1 · · · rik .
i1 <···<ik
 
1 2 3 4
 
5 6 7 8
9
. Then, note that
 8 7 6

5 4 3 2
1. EM1 (A) = 1 + 6 + 7 + 2 = 16 and EM2 (A) = det A({1, 2}, {1, 2}) + det A({1, 3}, {1, 3}) +
det A({1, 4}, {1, 4}) + det A({2, 3}, {2, 3}) + det A({2, 4}, {2, 4}) + det A({3, 4}, {3, 4}) =
−80.
2. S1 (1, 2, 3, 4) = 10 and S2 (1, 2, 3, 4) = 1 · (2 + 3 + 4) + 2 · (3 + 4) + 3 · 4 = 9 + 14 + 12 = 35.
Theorem 8.1.4. Let A ∈ Mn (C) and let σ(A) = {λ1 , . . . , λn }. Then,
205
206CHAPTER 8. ADVANCED TOPICS ON DIAGONALIZABILITY AND TRIANGULARIZATION∗
n
1. the coefficient of tn−k in PA (t) =
Q
(t − λi ), the characteristic polynomial of A, is
i=1
X
(−1)k λi1 · · · λik = (−1)k Sk (λ1 , . . . , λn ). (8.1.1)
i1 <···<ik
2. EMk (A) = Sk (λ1 , . . . , λn ).
Proof. Note that by definition,

n
Y
PA (t) = (t − λi ) = tn − S1 (λ1 , . . . , λn )tn−1
i=1
+S2 (λ1 , . . . , λn )tn−2 − · · · + (−1)n Sn (λ1 , . . . , λn ) (8.1.2)
= tn − EM1 (A)tn−1 + EM2 (A)tn−2 − · · · + (−1)n EMn (A). (8.1.3)
As the second part

 is just a re-writing of
 the first, we will just prove the first part. To do so,
t − a11 · · · −a1n

let B = tI − A =  .. 
. Then, using Definition 9.2.2 in Appendix, note that
 . 
−an1 · · · t − ann
P n
Q
det B = sgnσ biσ(i) and hence each S ⊆ [n] with |S| = n − k has a contribution to the
σ i=1
coefficient of tn−k in the following way:
T
For all i ∈ S, consider all permutations σ such that σ(i) = i. Our idea is to select a ‘t’ from
AF
these biσ(i) . Since we do not want any more ‘t’, we set t = 0 for any other diagonal position. So
DR
the contribution from S to the coefficient of tn−k is det[−A(S|S)] = (−1)k det A(S|S). Hence
the coefficient of tn−k in PA (t) is
X X
(−1)k det A(S|S) = (−1)k det A[T, T ] = (−1)k Ek (A).
S⊆[n], |S|=n−k T ⊆[n], |T |=k
The proof is complete in view of Equation (8.1.2).

As a direct application, we obtain Theorem 6.1.17 which we state again.
Pn
Corollary 8.1.5. Let A ∈ Mn (C) and let σ(A) = {λ1 , . . . , λn }. Then tr(A) = 1 λi and
det A = n1 λi .
Q
Let A and B be similar matrices. Then, by Theorem 6.2.3, we know that σ(A) = σ(B).
Thus, as a direct consequence of Part 2 of Theorem 8.1.4 gives the following result.
Corollary 8.1.6. Let A and B be two similar matrices of order n. Then, EMk (A) = EMk (B)
for 1 ≤ k ≤ n.
So, the sum of principal minors of similar matrices are equal. Or in other words, the sum
of principal minors are invariant under similarity.
Corollary 8.1.7. [Derivative of Characteristic Polynomial] Let A ∈ Mn (C). Then

n
d X
PA (t) = PA0 (t) = PA(i|i) (t).
dt
i=1
8.1. MORE ON THE SPECTRUM OF A MATRIX 207
Proof. For 1 ≤ i ≤ n, let us denote A(i|i) by Ai . Then, using Equation (8.1.3), we have
n
X X X X
PAi (t) = tn−1 − EM1 (Ai )tn−2 + · · · + (−1)n−1 EMn−1 (Ai )
i=1 i i i
= ntn−1 − (n − 1)EM1 (A)tn−2 + (n − 2)EM2 (A)tn−3 − · · · + (−1)n−1 EMn−1 (A)
= PA0 (t).
Which gives us the desired result.
Corollary 8.1.8. Let A ∈ Mn (C). If Alg.Mulα (A) = 1 then Rank[A − λI] = n − 1.
Proof. As Alg.Mulα (A) = 1, PA (t) = (t − λ)q(t), where q(t) is a polynomial with q(λ) 6=
0. Thus PA0 (t) = q(t) + (t − λ)q 0 (t). Hence, PA0 (λ) = q(λ) 6= 0. Thus, by Corollary 8.1.7,
0
P
i PA(i|i) (λ) = PA (λ) 6= 0. Hence, there exists i, 1 ≤ i ≤ n such that PA(i|i) (λ) 6= 0. That is,
det[A(i|i) − λI] 6= 0 or Rank[A − λI] = n − 1.
" #
0 1
Remark 8.1.9. Converse of Corollary 8.1.8 is false. Note that for the matrix A = ,
0 0
Rank[A − 0I] = 1 = 2 − 1 = n − 1, but 0 has multiplicity 2 as a root of PA (t) = 0.
As an application of Corollary 8.1.7, we have the following result.

We now relate the multiplicity of an eigenvalue with the spectrum of a principal sub-matrix.
T
Theorem 8.1.10. [Multiplicity and Spectrum of a Principal Sub-Matrix] Let A ∈ Mn (C)

AF
and k be a positive integer. Then 1 ⇒ 2 ⇒ 3, where

DR
1. Geo.Mulλ (A) ≥ k.
2. If B is a principal sub-matrix of A of size m > n − k then λ ∈ σ(B).
3. Alg.Mulλ (A) ≥ k.
Proof. Part 1⇒ Part 2. Let {x1 , . . . , xk } be linearly independent eigenvectors for λ and
" let #
B
B ∗
be a principal sub-matrix of A of size m > n − k. Without loss, we may write A = .
∗ ∗
" #
xi1
Let us partition the xi ’s , say xi = , such that
xi2
" #" # " #
B ∗ xi1 xi1
=λ , for 1 ≤ i ≤ k.
∗ ∗ xi2 xi2
As m > n − k, the size of xi2 is less than k. Thus, the set {x12 , . . . , xk2"} is# linearly dependent
y1
(see Corollary 3.3.9). So, there is a nonzero linear combination y = of x1 , . . . , xk such
y2
that y2 = 0. Notice that y1 6= 0 and By1 = λy1 .
n
Part 2⇒ Part 3. By Corollary 8.1.7, we know that PA0 (t) =
P
PA(i|i) (t). As A(i|i) is of size
i=1
n − 1, we get PA(i|i) (λ) = 0, for all i = 1, 2, . . . , n. Thus, PA0 (λ) = 0. A similar argument now
(2) (2) d
applied to each of the A(i|i)’s, gives PA (λ) = 0, where PA (t) = PA0 (t). Proceeding on above
dt
(i)
lines, we finally get PA (λ) = 0, for i = 0, 1, . . . , k − 1. This implies that Alg.Mulλ (A) ≥ k.
Definition 8.1.11. [Moments] Fix a positive integer n and let α1 , . . . , αn be n complex

n
αik is called the k-th moment of the
P
numbers. Then, for a positive integer k, the sum
i=1
numbers α1 , . . . , αn .
Theorem 8.1.12. [Newton’s identities] Let P (t) = tn + an−1 tn−1 + · · · + a0 have zeros
n
λki . Then, for 1 ≤ k ≤ n,
P
λ1 , . . . , λn , counted with multiplicities. Put µk =
i=1
k an−k + µ1 an−k+1 + · · · + µk−1 an−1 + µk = 0. (8.1.4)
That is, the first n moments of the zeros determine the coefficients of P (t).
Proof. For simplicity of expression, let an = 1. Then, using Equation (8.1.4), we see that
k = 1 gives us an−1 = −µ1 . To compute an−2 , put k = 2 in Equation (8.1.4) to verify that
−µ2 +µ21
an−2 = 2 . This process can be continued to get all the coefficients of P (t). Now, let us
prove the n given equations.
P 1 P 0 (t)
Define f (t) = t−λi = P (t) and take |t| > max |λi |. Then, the left hand side can be
i i
re-written as
n n n h
X 1 X 1 X 1 λi i n µ
1
f (t) = = = + 2 + ··· = + 2 + ··· . (8.1.5)
t − λi λi t t t t
i=1 i=1 t 1 − i=1
t
T
Thus, using P 0 (t) = f (t)P (t), we get

AF
hn µ1 ih i
nan tn−1 + (n − 1)an−1 tn−2 + · · · + a1 = P 0 (t) = + + · · · an tn
+ · · · + a0 .
DR
t t2
Now, equating the coefficient of tn−k−1 on both sides, we get
(n − k)an−k = nan−k + µ1 an−k+1 + · · · + µk an , for 0 ≤ k ≤ n − 1
which is the required Newton’s identity.
Remark 8.1.13. Let P (t) = an tn + · · · + a1 t + a0 with an = 1. Thus, we see that we need not
find the zeros of P (t) to find the k-th moments of the zeros of P (t). It can directly be computed
recursively using the Newton’s identities.
Exercise 8.1.14. Let A, B ∈ Mn (C). Then, prove that A and B have the same eigenvalues if
and only if tr(Ak ) = tr(Bk ), for k = 1, . . . , n.
8.2 Methods for Tridiagonalization and Diagonalization

Let G = {A ∈ Mn (C) : A∗ A = I}. Then, using Exercise 5.8.8, we see that
1. for every A, B ∈ G, AB ∈ G.
2. for every A, B, C ∈ G, (AB)C = A(BC).
3. In is the identity element of G. That is, for any A ∈ G, AIn = A = In A.
4. for every A ∈ G, A−1 ∈ G.
8.2. METHODS FOR TRIDIAGONALIZATION AND DIAGONALIZATION 209
Thus, the set G forms a group with respect to multiplication. We now define this group.
Definition 8.2.1. [Unitary Group] Let G = {A ∈ Mn (C) : A∗ A = I}. Then, G forms a

multiplicative group. This group is called the unitary group.
Proposition 8.2.2. [Selection Principle of Unitary Matrices] Let {Uk : k ≥ 1} be a sequence

2
of unitary matrices. Viewing them as elements of Cn , let us assume that “for any > 0, there
exists a positive integer N such that kUk − U k < , for all k ≥ N ”. That is, the matrices Uk ’s
2
converge to U as elements in Cn . Then, U is also a unitary matrix.
n
Proof. Let A = [aij ] ∈ Mn (C) be an unitary matrix. Then |aij |2 = tr(A∗ A) = n. Thus, the
P
i,j=1
2
set of unitary matrices is a compact subset of Cn . Hence, any sequence of unitary matrices has
a convergent subsequence (Bolzano-Weierstrass Theorem), whose limit is again unitary. Thus,
For a unitary matrix U , we know that U −1 = U ∗ . Our next result gives a necessary and
sufficient condition on an invertible matrix A so that the matrix A−1 is similar to A∗ .
Theorem 8.2.3. [Generalizing a Unitary Matrix] Let A be an invertible matrix. Then A−1
is similar to A∗ if and only if there exists an invertible matrix B such that A = B −1 B ∗ .
Proof. Suppose A = B −1 B ∗ , for some invertible matrix B. Then
A∗ = B(B −1 )∗ = B(B −1 )∗ BB −1 = B(B −1 B ∗ )−1 B −1 = BA−1 B −1 .

T
AF
Conversely, let A∗ = SA−1 S −1 , for some invertible matrix S. Need to show, A = S −1 S ∗ .

We first show that there exists a nonsingular Hermitian Hθ such that A−1 = Hθ−1 A∗ Hθ , for
DR
some θ ∈ R.
Note that for any θ ∈ R, if we put Sθ = eiθ S then
Sθ A−1 Sθ−1 = A∗ and Sθ = A∗ Sθ A.
Now, define Hθ = Sθ + Sθ∗ . Then, Hθ is a Hermitian matrix and Hθ = A∗ Hθ A. Furthermore,

there are infinitely many choices of θ ∈ R such that det Hθ = 0. To see this, let us choose a
θ ∈ R such that Hθ is singular. Hence, there exists x 6= 0 such that Hθ x = 0. So,
Sθ x = −Sθ∗ x = −e−iθ S ∗ x. Or equivalently, − e2iθ x = S −1 S ∗ x.
That is, −e2iθ ∈ σ(S −1 S ∗ ). Thus, if we choose θ0 ∈ R such that −e2i(θ0 ) 6∈ σ(S −1 S ∗ ) then H( θ0 )
is nonsingular.
To get our result, we finally choose B = β(αI − A∗ )H( θ0 ) such that β 6= 0 and α = eiγ ∈
/
σ(A∗ ).
Note that with α and β chosen as above, B is invertible. Furthermore,
BA = αβH( θ0 )A − βA∗ H( θ0 )A = αβH( θ0 )A − βH( θ0 ) = βH( θ0 )(αA − I).
As we need, BA = B ∗ , we get βH( θ0 )(αA − I) = βH( θ0 )(αI − A) and thus, we need β = −βα,
which holds true if β = ei(π−γ)/2 . Thus, the required result follows.
Exercise 8.2.4. Suppose that A is similar to a unitary matrix. Then, prove that A−1 is similar
to A∗ .
8.2.1 Plane Rotations
Definition 8.2.5. [Plane Rotations] For a fixed positive integer n, consider the vector space
Rn with standard basis {e1 , . . . , en }. Also, for 1 ≤ i, j ≤ n, let Ei,j = ei eTj . Then, for θ ∈ R
and 1 ≤ i, j ≤ n, a plane rotation, denoted U (θ; i, j), is defined as
U (θ; i, j) = I − Ei,i − Ej,j + [Ei,i + Ej,j ] cos θ − Ei,j sin θ + Ej,i sin θ.
 
1
 . 

 .. 

 

 cos θ − sin θ  ← i-th row


That is, U (θ; i, j) =  .. 
, where the unmentioned
 . 

← j-th row
 

 sin θ cos θ 
 .. 

 . 

1
diagonal entries are 1 and the unmentioned off-diagonal entries are 0.
Remark 8.2.6. Note the following about the matrix U (θ; i, j), where θ ∈ R and 1 ≤ i, j ≤ n.
1. U (θ; i, j) are orthogonal.
2. Geometrically U (θ; i, j)x rotates x by the angle θ in the ij-plane.

T
AF
3. Geometrically (U (θ; i, j))T x rotates x by the angle −θ in the ij-plane.

DR
4. If y = U (θ; i, j)x then the coordinates of y are given by
(a) yi = xi cos θ − xj sin θ,

(b) yj = xi sin θ + xj cos θ, and
(c) for l 6= i, j, yl = xl .
5. Thus, for x ∈ Rn , the choice of θ for which yj = 0, where y = U (θ; i, j)x equals
(a) θ = 0, whenever xj = 0. That is, U (0; i, j) = I.

(b) θ = cot−1 − xxji , whenever xj 6= 0.
6. [Geometry] Imagine standing at 1 = (1, 1, 1)T ∈ R3 . We want to apply a plane rotation

U , so that v = U T 1 with v2 = 0. That is, the final point is on the xz-plane.
Then, we can either apply a plane rotation along the xy-plane or the yz-plane. For the
xy-plane, we need the plane z = 1 (xy plane lifted by 1). This plane contains the vector 1.
Imagine moving the tip of ~1 on this plane. Then this locus corresponds to a circle that lies
√
on the plane z = 1, has radius 2 and is centred at (0, 0, 1). That is, we draw the circle
x2 +y 2 = 1 on the xy-plane and then lifted it up by so that it lies on the plane z = 1. Thus,
note that the xz-plane cuts this circle at two points. These two points of intersections give
us the two choices for the vector v (see Figure 8.1). A similar calculation can be done for
the yz-plane.
(0, 0
, 0)
Figure 8.1: Geometry of plane rotations in R3

.
7. In general, in Rn , suppose that we want to apply plane rotation to a along the x1 x2 -plane
so that the resulting vector has 0 in the 2-nd coordinate. In that case, our circle on x1 x2 -
p h iT
plane has radius r = a21 + a22 and it gets translated by 0, 0, a3 , · · · an . So, there
h iT
are two points x on this circle with x2 = 0 and they are ±r, 0, a3 , · · · an .
T
8. Consider three mutually orthogonal unit vectors, say x, y, z. Then, x can be brought to e1
AF
by two plane rotations, namely by an appropriate U (θ1 ; 1, 3) and U (θ2 ; 1, 2). Thus,
DR
U (θ2 ; 1, 2)U (θ1 ; 1, 3)x = e1 .
In this process, the unit vectors y and z, get shifted to say,
ŷ = U (θ2 ; 1, 2)U (θ1 ; 1, 3)y and ẑ = U (θ2 ; 1, 2)U (θ1 ; 1, 3)z.
As unitary transformations preserve angles, note that ŷ(1) = ẑ(1) = 0. Now, we can apply
an appropriate plane rotation U (θ3 ; 2, 3) so that U (θ3 ; 2, 3)ŷ = e2 . Since e3 is the only
unit vector in R3 orthogonal to both e1 and e2 , it follows that U (θ3 ; 2, 3)ẑ = e3 . Thus,
h i h i
I = e1 e2 e3 = U (θ3 ; 2, 3)U (θ2 ; 1, 2)U (θ1 ; 1, 3) x y z .
Hence, any real orthogonal matrix A ∈ M3 (R) is a product of three plane rotations.
We are now ready to give another method to get the QR-decomposition of a square matrix
(see Theorem 4.6.1 that uses the Gram-Schmidt Orthonormalization Process).
Proposition 8.2.7. [QR Factorization Revisited: Square Matrix] Let A ∈ Mn (R). Then
there exists a real orthogonal matrix Q and an upper triangular matrix R such that A = QR.
Proof. We start by applying the plane rotations to A so that the positions (2, 1), (3, 1), . . . , (n, 1)
of A become zero. This means, if a21 = 0, we multiply by I. Otherwise, we use the plane rotation
U (θ; 1, 2), where θ = cot−1 (−a11 /a21 ). Then, we apply a similar technique to A so that the
(3, 1) entry of A becomes 0. Note that this plane rotation doesn’t change the (2, 1) entry of A.
We continue this process till all the entry in the first column of A, except possibly the (1, 1)
entry, is zero.
We then apply the plane rotations to make positions (3, 2), (4, 2), . . . , (n, 2) zero. Observe
that this does not disturb the zeros in the first column. Thus, continuing the above process a
finite number of times give us the required result.
Lemma 8.2.8. [QR Factorization Revisited: Rectangular Matrix] Let A ∈ Mm,n (R). Then
there exists a real orthogonal matrix Q and a matrix R ∈ Mm,n (R) in upper triangular form
such that A = QR.
Proof. If RankA < m, add some columns to A to get a matrix, say Ã such that RankÃ = m. Now
suppose that Ã has k columns. For 1 ≤ i ≤ k, let vi = Ã[:, i]. Now, apply the Gram-Schmidt
Orthonormalization Process to {v1 , . . . , vk }. For example, h suppose the
i result is a sequence of k
vectors w1 , 0, w2 , 0, 0, . . . , 0, wm , 0, . . . , 0, where Q = w1 · · · wm is real orthogonal. Then
Ã[:, 1] is a linear combination of w1 , Ã[:, 2] is also a linear combination of w1 , Ã[:, 3] is a linear
combination of w1 , w2 and so on. In general, for 1 ≤ s ≤ k, the column Ã[:, s] is a linear
m
P
combination of wi -s in the list that appear up to the s-th position. Thus, Ã[:, s] = wi ris ,
i=1
where ris = 0 for all i > s. That is, Ã = QR, where R = [rij ]. Now, remove the extra columns
of Ã and the corresponding columns in R to get the required result.
T
Note that Proposition 8.2.7 is also valid for any complex matrix. In this case the matrix Q
AF
will be unitary. This can also be seen from Theorem 4.6.1 as we need to apply the Gram-Schmidt
DR
Orthonormalization Process to vectors in Cn .

To proceed further recall that a matrix A = [aij ] ∈ Mn (C) is called a tri-diagonal matrix
if aij = 0, whenever |i − j| > 1, 1 ≤ i, j ≤ n.
Proposition 8.2.9. [Tridiagonalization of a Real Symmetric Matrix: Given’s Method]

Let A be a real symmetric. Then, there exists a real orthogonal matrix Q such that QAQT is a
tri-diagonal matrix.
Proof. If a31 6= 0, then put U1 = U (θ1 ; 2, 3), where θ1 = cot−1 (−a21 /a31 ). Notice that U1T [:
, 1] = e1 and so
(U1 AU1T )[:, 1] = (U1 A)[:, 1].
We already know that U1 A[3, 1] = 0. Hence, U1 AU1T is a real symmetric matrix with (3, 1)-
th entry 0. Now, proceed to make the (4, 1)-th entry of U1 A equal to 0. To do so, take
U2 = U (θ2 ; 2, 4). Notice that U2T (:, 1) = e1 and so
(U2 (U1 AU1T )U2T )[:, 1) = (U2 U1 AU1T )[:, 1].
But by our choice of the plane rotation U2 , we have U2 (U1 AU1T )(4, 1) = 0. Furthermore, as
U2 [3, :] = eT3 , we have
(U2 U1 AU1T )[3, 1] = U2 [3, :](U1 AU1T )[:, 1] = (U1 AU1T )[3, 1] = 0.
That is, the previous zeros are preserved.

Continuing this way, we can find a real orthogonal matrix Q such that QAQT is tri-
diagonal.
Proposition 8.2.10. [Almost Diagonalization of a Real Symmetric Matrix: Jacobi method]

Let A ∈ Mn (R) be real symmetric. Then there exists a real orthogonal matrix S, a product of
plane rotations, such that SAS T is almost a diagonal matrix.
Proof. The idea is to reduce the off-diagonal entries of A to 0 as much as possible. So, we start
with choosing i 6= j) such that i < j and |aij | is maximum. Now, put
1 aii − ajj
θ= cot−1 , U = U (θ; i, j), and B = U T AU.
2 2aij
Then, for all l, k 6= i, j, we see that
blk = U T [l, :]AU [:, k] = eTl Aek = alk

bik = U T [i, :]AU [:, k] = (cos θeTi + sin θeTj )Aek = aik cos θ + ajk sin θ
blj = U T [l, :]AU [:, j] = eTl A(− sin θei + cos θej ) = −ali sin θ + alj cos θ
bij = U T [i, :]AU [:, j] = (cos θeTi + sin θeTj )A(− sin θei + cos θej )
ajj −aii
= sin(2θ) 2 + aij cos(2θ) = 0
Thus, using the above, we see that whenever l, k 6= i, j, a2lk = b2lk and for l 6= i, j, we have
T
AF
b2il + b2lj = a2il + a2lj .

DR
As U is unitary and B = U T AU , we get |aij |2 = |bij |2 . Further, bij = 0 implies that

P P
a2ii + 2a2ij + a2jj = b2ii + 2b2ij + b2jj = b2ii + b2jj .
As the rest of the diagonal entries have not changed, we observe that the sum of the squares of
the off-diagonal entries have reduced by 2a2ij . Thus, a repeated application of the above process
makes the matrix “close to diagonal”.
8.2.2 Householder Matrices
We will now look at another class of unitary matrices, commonly called the Householder matrices
(see Exercise 1.5.5.8).
Definition 8.2.11. [Householder Matrix] Let w ∈ Cn be a unit vector. Then, the matrix
Uw = I − 2ww∗ is called a Householder matrix.
Remark 8.2.12. We observe the following about the Householder matrix Uw .

1. Uw = I − 2ww∗ is the sum of two Hermitian matrices and hence is also Hermitian.
∗ = (I − 2ww∗ )(I − 2ww∗ )T = I − 2ww∗ − 2ww∗ + 4ww∗ = I. Or equivalently,
2. Uw Uw
verify that kUw xk = kxk, for all x ∈ Cn . So Uw is unitary.
3. If x ∈ w⊥ then Uw x = x.
4. If x = cw, for some c ∈ C, then Uw x = −x.

5. Thus, if v ∈ Cn then we know that v = x + y, where x ∈ w⊥ and y = cw, for some
c ∈ C. In this case, Uw v = Uw (x + y) = x − y.
6. Geometrically, Uw v reflects the vector v along the vector w⊥ . Thus, Uw is a reflection
matrix along w⊥ (see Definition 1.4.1.7).
" #
1
Example 8.2.13. In R2 , let w = e2 . Then w⊥ is the x-axis. The vector v = = e1 + 2e2 ,
2
where e1 ∈ w⊥ and 2e2 ∈ LS(w). So
Uw (e1 + 2e2 ) = Uw v = Uw (x + y) = x − y = e1 − 2e2 .
That is, the reflection of v along the x-axis (w⊥ ).
Recall that if x, y ∈ Rn with x 6= y and kxk = kyk then, (x +"y)# ⊥ (x − y)." This # is not
1 i
true in Cn as can be seen from the following example. Take x = and y = . Then
1 −1
" # " #
1+i 1−i
h , i = (1 + i)2 6= 0. Thus, to pick the right choice for the matrix Uw , we need to
0 2
be observant of the choice of the inner product space.
T
Example 8.2.14. Let x, y ∈ Cn with x 6= y and kxk = kyk. Then, which Uw should be used
AF
to reflect y to x?
DR
1. Solution in case of Rn : Imagine the line segment joining x and y. Now, place a mirror
at the midpoint and perpendicular to the line segment. Then, the reflection of y on that
x−y
mirror is x. So, take w = kx−yk ∈ Rn . Then,
x−y
Uw y = (I − 2wwT )y = y − 2wwT y = y − 2 (x − y)T y
kx − yk2
x − y −kx − yk2
= y−2 = x.
kx − yk2 2
2. Solution in case of Cn : Suppose there is a unit vector w ∈ Cn such that (I − 2ww∗ )y =

x. Then y − x = 2ww∗ y and hence w∗ (y − x) = 2w∗ ww∗ y = 2w∗ y. Thus,
w∗ (y + x) = 0, that is, w ⊥ (y + x). (8.2.1)
Furthermore, again using w∗ (y + x) = 0, we get y − x = 2ww∗ y = −2ww∗ x. So,
2(y − x) = 2ww∗ (y − x) or y − x = ww∗ (y − x).
On the other hand, using Equation (8.2.1), we get ww∗ (y + x) = 0. So,
0 = [(y + x)∗ ww∗ ](y − x) = (y + x)∗ [ww∗ (y − x)] = (y + x)∗ (y − x).
Therefore, if such a w exists, then (y − x) ⊥ (y + x).

x−y
But, in that case, w = kx−yk will work as using above kx − yk2 = 2(y − x)∗ y and
x−y
Uw y = (I − 2ww∗ )y = y − 2ww∗ y = y − 2 (x − y)∗ y
kx − yk2
x − y −kx − yk2
= y−2 = x.
kx − yk2 2
Thus, in this case, if hx + y, x − yi =

6 0 then we will not find a w such that Uw y = x.
" # " #
1 i
For example, taking x = and y = , we have hx + y, x − yi =
6 0.
1 −1
As an application, we now prove that any real symmetric matrix can be transformed into a
tri-diagonal matrix.
" #
a vT
Proposition 8.2.15. [Householder’s Tri-Diagonalization] Let v ∈ Rn−1 and A = ∈
v B
Mn (R) be a real symmetric matrix. Then, there exists a real orthogonal matrix Q, a product of
Householder matrices, such that QT AQ is tri-diagonal.
Proof. If v = e1 then we proceed to apply our technique to the matrix B, a matrix of lower
order. So, without loss of generality, we assume that v 6= e1 .
As we want QT AQ to be tri-diagonal, we need to find a vector w ∈ Rn−1 such that Uw v =
re1 ∈ Rn−1 , where r = kvk = kUw vk. Thus, using Example 8.2.14, choose the required vector
T
w ∈ Rn−1 . Then,
AF
 
DR
" #" #" # " # a r 0 " #

1 0 a vT 1 0 a v T Uw
T   a reT
1
T
= T
=
 r ∗ ∗ = re
 ,
0 Uw v B 0 Uw Uw v Uw BUw 1 S
0 ∗ ∗
where S ∈ Mn−1 (R) is a symmetric matrix. Now, use induction on the matrix S to get the
required result.
8.2.3 Schur’s Upper Triangularization Revisited
Definition 8.2.16. Let s and t be two symbols. Then, an expression of the form
W (s, t) = sm1 tn1 . . . smk tnk where mi , ni are positive integers

k
P
is called a word in symbols s and t of degree (mi + ni ).
i=1
Remark 8.2.17. [More on Unitary Equivalence] Let s and t be two symbols and W (s, t) be
a word in symbols s and t.
1. Suppose U is a unitary matrix such that B = U ∗ AU . Then, W (A, A∗ ) = U ∗ W (B, B ∗ )U .
Thus, tr[W(A, A∗ )] = tr[W(B, B∗ )].
2. Let A and B be two matrices such that tr[W(A, A∗ )] = tr[W(B, B∗ )], for each word W .
Then, does it imply that A and B are unitarily equivalent? The answer is ‘yes’ as provided
by the following result. The proof is outside the scope of this book.
Theorem 8.2.18. [Specht-Pearcy] Let A, B ∈ Mn (C) and suppose that tr[W(A, A∗ )] =

tr[W(B, B∗ )] holds for all words of degree less than or equal to 2n2 . Then B = U ∗ AU , for
some unitary matrix U .
Exercise 8.2.19. [Triangularization via Complex Orthogonal Matrix need not be Possi-
ble] Let A ∈ Mn (C) and A = QT QT , where Q is complex orthogonal matrix and T is upper
triangular. Then, prove that
1. A has an eigenvector x such that xT x 6= 0.
" #
1 i
2. there is no orthogonal matrix Q such that QT Q is upper triangular.
i −1
Proposition 8.2.20. [Matrices with Distinct Eigenvalues are Dense in Mn (C)] Let A ∈
Mn (C). Then, for each > 0, there exists a matrix A() ∈ Mn (C) such that A() = [a()ij ] has
|aij − a()ij |2 < .
P
distinct eigenvalues and
such that U ∗ AU = T , an upper triangular matrix. Now, choose αi ’s such that tii + αi are
|αi |2 < . Now, consider the matrix A() = U (T + diag(α1 , . . . , αn )) U ∗ . Then,
P
distinct and
B = A() − A = U diag(α1 , . . . , αn )]U ∗ with
X X
|bij |2 = tr(B∗ B) = trU diag(|α1 |2 , . . . , |αn |2 )U∗ = |αi |2 < .
T
i,j i
AF

DR
Before proceeding with our next result on almost diagonalizability, we look at the following
example.
" #
1 2
Example 8.2.21. Let A = and > 0 be given. Then, determine a diagonal matrix D
0 3
such that the non-diagonal entry of D−1 AD is less than .

Solution: Choose α < and define D = diag(1, α). Then,
2
" #" #" # " #
1 0 1 2 1 0 1 2α
D−1 AD = = .
0 α1 0 3 0 α 0 3

As α < , the required result follows.
2
Proposition 8.2.22. [A matrix is Almost Diagonalizable] Let A ∈ Mn (C) and > 0 be
given. Then, there exists an invertible matrix S such that S−1 AS = T , an upper triangular
matrix with |tij | < , for all i 6= j.
such that U ∗ AU = T , an upper triangular matrix. Now, take t = 2 + max |tij | and choose α
i<j
such that 0 < α < /t. Then, if we take Dα = diag(1, α, α2 , . . . , αn−1 ) and S = U Dα , we have
S −1 AS = D 1 T Dα = F (say), an upper triangular. Furthermore, note that for i < j, we have
α
|fij | = |tij |αj−i ≤ . Thus, the required result follows.
8.3. COMMUTING MATRICES AND SIMULTANEOUS DIAGONALIZATION 217
8.3 Commuting Matrices and Simultaneous Diagonalization
Definition 8.3.1. [Simultaneously Diagonalizable] Let A, B ∈ Mn (C). Then, they are said
to be simultaneously diagonalizable if there exists an invertible matrix S such that S −1 AS
and S −1 BS are both diagonal matrices.
Since diagonal matrices commute, we have our next result.
Proposition 8.3.2. Let A, B ∈ Mn (C). If A and B are simultaneously diagonalizable then

AB = BA.
Proof. By definition, there exists an invertible matrix S such that S −1 AS = Λ1 and S −1 BS =

Λ2 . Hence,
AB = (SΛ1 S −1 ) · (SΛ2 S −1 ) = SΛ1 Λ2 S −1 = SΛ2 Λ1 S −1 = SΛ2 S −1 SΛ1 S −1 = BA.
Thus, we have proved the required result.
Theorem 8.3.3. Let A, B ∈ Mn (C) be diagonalizable matrices. Then they are simultaneously
diagonalizable if and only if they commute.
Proof. One part of this theorem has already been proved in Proposition 8.3.2. For the other
T
part, let us assume that AB = BA. Since A is diagonalizable, there exists an invertible matrix
AF
S such that
DR
S −1 AS = Λ = λ1 I ⊕ · · · ⊕ λk I, (8.3.1)
where λ1 , . . . , λk are the distinct eigenvalues of

 A. We now use  the sub-matrix structure of
C11 · · · C1k
S −1 AS to decompose C = S −1 BS as C = 
 .. 
. Since AB = BA and S is
 . 
Ck1 · · · Ckk
invertible, we have ΛC = CΛ. Thus,
   
λ1 C11 · · · λ1 C1k λ1 C11 · · · λk C1k
 ..  
= .. 
.

 .   . 
λk Ck1 · · · λk Ckk λ1 Ck1 · · · λk Ckk
Since λi 6= λj for 1 ≤ i 6= j ≤ k, we have Cij = 0, whenever i 6= j. Thus, the matrix

C = C11 ⊕ · · · ⊕ Ckk .
Since B is diagonalizable, the matrix C is also diagonalizable and hence the matrices Cii ,
for 1 ≤ i]lek, are diagonalizable. So, for 1 ≤ i ≤ k, there exists invertible matrices Ti ’s such
that Ti−1 Cii Ti = Λi . Put T = T1 ⊕ · · · ⊕ Tk . Then,
     
T1−1 λ1 I T1 λ1 I
.. .. .. ..
T −1 S −1 AST = 
     
.  .  . = . 
     
Tk−1 λk I Tk λk I
and
     
T1−1 C11 T1 Λ1
.. .. .. ..
T −1 S −1 BST = 
     
.  .  . = . .
     
Tk−1 Ckk Tk Λk
Thus A and B are simultaneously diagonalizable and the required result follows.
Definition 8.3.4. [Commuting Family of Matrices]

1. Let F ⊆ Mn (C). Then F is said to be a commuting family if each pair of matrices in
F commutes.
2. Let B ∈ Mn (C) and W be a subspace of Cn . Then, W is said to be a B-invariant
subspace if Bw ∈ W , for all w ∈ W (or equivalently, BW ⊆ W ).
3. A subspace W of Cn is said to be F-invariant if W is B-invariant for each B ∈ F.
Example 8.3.5. Let A ∈ Mn (C) with (λ, x) as an eigenpair. Then, W = {cx : c ∈ C} is an

A-invariant subspace. Furthermore, if W is an A-invariant subspace with dim(W ) = 1 then
verify that any non-zero vector in W is an eigenvector of A.
Theorem 8.3.6. [An A-invariant Subspace Contains an Eigenvector of A] Let A ∈ Mn (C)

and W ⊆ Cn be an A-invariant subspace of dimension at least 1. Then W contains an eigen-
vector of A.
T
AF
h 1 , . . . , fk } ⊆ C be ian ordered basis for W . Define T : W → W as T v = Av.

Proof. Let B = {f n
DR
Then T [B, B] = [T f1 ]B · · · [T fk ]B is a k × k matrix which satisfies [T w]B = T [B, B] [w]B ,

for all w ∈ W . As T [B, B] ∈ Mk (C), it has an eigenpair, say (λ, x̂) with x̂ ∈ Ck . That is,
T [B, B]x̂ = λx̂. (8.3.2)

k
(x̂)i fi ∈ Cn . Then, verify that x ∈ W and [x]B = x̂. Thus, T x ∈ W and now
P
Now, put x =
i=1
using Equation (8.3.2), we get
k
X k
X k
X k
X k
X
Tx = ([T x]B )i fi = (T [B, B][x]B )i fi = (T [B, B]x̂)i fi = (λx̂)i fi = λ (x̂)i fi = λx.
i=1 i=1 i=1 i=1 i=1
So, A has an eigenvector x ∈ W corresponding to the eigenvalue λ.
Theorem 8.3.7. Let F ⊆ Mn (C) be a commuting family of matrices. Then, all the matrices
in F have a common eigenvector.
Proof. Note that Cn is F-invariant. Let W ⊆ Cn be F-invariant with minimum positive

dimension. Let y ∈ W such that y 6= 0. We claim that y is an eigenvector, for each A ∈ F.
So, on the contrary assume y is not an eigenvector for some A ∈ F. Then, by Theorem 8.3.6,
W contains an eigenvector x of A for some eigenvalue, say λ. Define W0 = {z ∈ W : Az = λz}.
So W0 is a proper subspace of W as y ∈ W \ W0 . Also, for z ∈ W0 and C ∈ F, we note that
A(Cz) = CAz = λ(Cz), so that Cz ∈ W0 . So W0 is F-invariant and 1 ≤ dim W0 < dim W , a
contradiction.
Theorem 8.3.8. Let F ⊆ Mn (C) be a family of diagonalizable matrices. Then F is commuting

if and only if F is simultaneously diagonalizable.
Proof. We prove the result by induction on n. The result is clearly true for n = 1. So, let us
assume the result to be valid for all n < m. Now, let us assume that F ⊆ Mm (C) is a family of
diagonalizable matrices.
If F is simultaneously diagonalizable, then by Proposition 8.3.2, the family F is commuting.
Conversely, let F be a commuting family. If each A ∈ F is a scalar matrix then they are simul-
taneously diagonalizable via I. So, let A ∈ F be a non-scalar matrix. As A is diagonalizable,
there exists an invertible matrix S such that
S −1 AS = λ1 I ⊕ · · · ⊕ λk I, k ≥ 2,
where λi ’s are distinct. Now, consider the family G = {X̂ = S −1 XS | X ∈ F }. As F is

a commuting family, the set G is also a commuting family. So, each X̂ ∈ G has the form
X̂ = X1 ⊕ · · · ⊕ Xk . Note that Hi = {Xi | X̂ ∈ G} is a commuting family of diagonalizable
matrices of size < m. Thus, by induction hypothesis, Hi ’s are simultaneously diagonalizable,
say by the invertible matrices Ti ’s. That is, Ti−1 Xi Ti = Λi , a diagonal matrix, for 1 ≤ i ≤ k.
Thus, if T = T1 ⊕ · · · ⊕ Tk then
T −1 S −1 X̂ST = T −1 (X1 ⊕ · · · ⊕ Xk )T = T1−1 X1 T1 ⊕ · · · ⊕ Tk−1 Xk Tk = Λ1 ⊕ · · · ⊕ Λk ,

T
a diagonal matrix, for all X ∈ F. Thus the result holds by induction.

AF
We now give prove of some parts of Exercise 6.2.7.exe:eigen:1.

DR
Remark 8.3.9. [σ(AB) and σ(BA)] Let m ≤ n, A ∈ Mm×n (C), and B ∈ Mn×m (C). Then
σ(BA) = σ(AB) with n − m extra 0’s. In particular, if A, B ∈ Mn (C) then, PAB (t) = PBA (t).
Proof. Note that

" #" # " # " #" #
AB 0 Im A AB ABA Im A 0 0
= = .
B 0 0 In B BA 0 In B BA
" # " #
AB 0 0 0
Thus, the matrices and are similar. Hence, AB and BA have precisely
B 0 B BA
the same non-zero eigenvalues. Therefore, if they have the same size, they must have the same
characteristic polynomial.
Exercise 8.3.10. [Miscellaneous Exercises]

1. Let A be nonsingular. Then, verify that A−1 (AB)A = BA. Hence, AB and BA are
similar. Thus, PAB (t) = PBA (t).
2. Fix a positive integer k, 0 ≤ k ≤ n. Now, define the function fk : Mn (C) → C by f (A) =
coefficient of tk in PA (t). Prove that fk is a continuous function.
3. For any matrix A, prove that there exists an > 0 such that Aα = A + αI is invertible,
for all α ∈ (0, ). Thus, use the first part to conclude that for any given B, we have
PAα B (t) = PBAα (t), for all α ∈ (0, ).
4. Now, use continuity to argue that PAB (t) = lim PAα B (t) = lim PBAα (t) = PBA (t).
α→0+ α→0+
5. Let σ(A) = {λ1 , . . . , λn }, σ(B) = {µ1 , . . . , µn } and suppose that AB = BA. Then,
(a) prove that there is a permutation π such that σ(A+B) = {λ1 +µπ(1) , . . . , λn +µπ(n) }.
In particular, σ(A + B) ⊆ σ(A) + σ(B).
(b) if we further assume that σ(A) ∩ σ(−B) = ∅ then the matrix A + B is nonsingular.
6. Let A and B be two non-commuting matrices. Then, give an example to show that it is
difficult to relate σ(A + B) with σ(A) and σ(B).
   
0 1 0 0 0 0
   
7. Are the matrices A =  0 0 −1 and B = 1 0 0 simultaneously triangularizable?
  
0 0 0 0 1 0
8. Let F ⊆ Mn (C) be a family of commuting normal matrices. Then, prove that each element
of F is simultaneously unitarily diagonalizable.
9. Let A ∈ Mn (C) with A∗ = A and x∗ Ax ≥ 0, for all x ∈ Cn . Then prove that σ(A) ⊆ R+
and if tr(A) = 0, then A = 0.
8.3.1 Diagonalization and Real Orthogonal Matrix
Proposition 8.3.11. [Triangularization: Real Matrix] Let A ∈ Mn (R). Then, there exists a
T
real orthogonal matrix Q such that QT AQ is block upper triangular, where each diagonal block
AF
is of size either 1 or 2.
DR
Proof. If all the eigenvalues of A are real then the corresponding eigenvectors have real entries
and hence, one can use induction to get the result in this case (see Lemma 6.4.1).
So, now let us assume that A has a complex eigenvalue, say λ = α + iβ with β 6= 0 and
x = u + iv as an eigenvector for λ. Thus, Ax = λx and hence Ax = λx. But, λ 6= λ as
β 6= 0. Thus, the eigenvectors x, x are linearly independent and therefore, {u, v} is a linearly
independent set. By Gram-Schmidt Orthonormalization process, we get an ordered basis, say
{w1 , w2 , . . . , wn } of Rn , where LS(w1 , w2 ) = LS(u, v). Also, using the eigen-condition Ax =
λx gives
Aw1 = aw1 + bβw2 , Aw2 = cw1 + dw2 ,
for some real numbers a, b, c and d.

Now, form a matrix X = [w1 , w2 , . . . , wn ]. Then, X is a real orthogonal matrix and
 
w∗
 1∗ 
 w2 
X ∗ AX = X ∗ [Aw1 , Aw2 , . . . , Awn ] =  .  [aw1 + bw2 , cw1 + dw2 , . . . , Awn ]
 
 .. 
 
wn∗
 
a b
 ∗
=  c d 
 (8.3.3)
0 B
where B ∈ Mn−2 (R). Now, by induction hypothesis the required result follows.
The next result is a direct application of Proposition 8.3.11 and hence the proof is omitted.
Corollary 8.3.12. [Simultaneous Triangularization: Real Matrices] Let F ⊆ Mn (R) be a

commuting family. Then, there exists a real orthogonal matrix Q such that QT AQ is a block
upper triangular matrix, where each diagonal block is of size either 1 or 2, for all A ∈ F.
Proposition 8.3.13. Let A ∈ Mn (R). Then the following statements are equivalent.
1. A is normal.
2. There exists a real orthogonal matrix Q such that QT AQ =
L
i Ai , where Ai ’s are real
normal matrices of size either 1 or 2.
Proof. 2 ⇒ 1 is trivial. To prove 1 ⇒ 2, recall that Proposition 8.3.11 gives the existence of
a real orthogonal matrix Q such that QT AQ is upper triangular with diagonal blocks of size
either 1 or 2. So, we can write
 
λ1 ∗ ∗ ∗ ∗ ∗
..
0 . ∗ ∗ ∗ ∗
 

  " #
0 · · · λp ∗ ∗ ∗  R C
QT AQ =  = (say).
 
0 ··· 0 A11 · · · A1k  0 B
 .. 
0 ··· 0 0 . ∗ 
T
 
0 ··· 0 0 · · · Akk
AF
DR
" #" # " #" #

R C RT 0 RT 0 R C
As A is normal, = . Thus, tr(CCT ) = tr(RRT − RT R) =
0 B C T BT C T BT 0 B
0. Now, using Exercise 8.3.10.9, we get C = 0. Hence, RRT = RT R and therefore, R is a
diagonal matrix.
k
P
As B T B = BB T , we have A1i AT1i = A11 AT11 . So tr A1i AT
P
1i = 0. Now, using Exer-
2
k
A1i AT1i = 0 and so A1i AT1i = 0, for all i = 2, . . . , k. Thus, A1i = 0,
P
cise 8.3.10.9 again, we have
2
for all i = 2, . . . , k. Hence, the required result follows.
Exercise 8.3.14. Let A ∈ Mn (R). Then the following are true.

1. A = −AT if and only if A is real orthogonally similar to [ j 0] ⊕ [ i Ai ], where Ai =
L L
" #
0 ai
, for some real numbers ai ’s.
−ai 0
2. AAT = I if and only if A is real orthogonally similar to [ i λi ] ⊕ [ j Aj ], where λi = ±1
L L
" #
cos θj sin θj
and Aj = , for some real numbers θi ’s.
− sin θj cos θj
8.3.2 Convergent and nilpotent matrices
Definition 8.3.15. [Convergent matrices] A matrix A is called a convergent matrix if

Am → 0 as m → ∞.
Remark 8.3.16. 1. Let A be a diagonalizable matrix with ρ(A) < 1. Then, A is a convergent
matrix.
Proof. Let A = U ∗ diag(λ1 , . . . , λn )U . As ρ(A) < 1, for each i, 1 ≤ i ≤ n, λm
i → 0 as
m → ∞. Thus, Am = U ∗ diag(λm m
1 , . . . , λn )U → 0.
2. Even if the matrix A is not diagonalizable, the above result holds. That is, whenever
ρ(A) < 1, the matrix A is convergent. The converse is also true.
Proof. Let Jk (λ) = λIk + Nk be a Jordan block of J = Jordan CFA. Then as Nkk = 0,
for each fixed k, we have
Jk (λ)m = λm + C(m, 1)λm−1 Nk + · · · + C(m, k − 1)λm−k+1 Nkk−1 → 0, as m → ∞.
As λm → 0 as m → ∞, the matrix Jk (λ)m → 0 and hence J is convergent. Thus, A is a

convergent matrix.
Conversely, if A is convergent, then J must be convergent. Thus each Jordan block Jk (λ)
must be convergent. Hence |λ| < 1.
Theorem 8.3.17. [Decomposition into Diagonalizable and Nilpotent Parts] Let A ∈ Mn (C).
Then A = B + C, where B is diagonalizable matrix and C is nilpotent such that BC = CB.
T
Proof. Let J = Jordan CFA. Then, J = D + N , where D = diag(J) and N is clearly a

AF
nilpotent matrix.
Now, note that DN = N D as for each Jordan block Jk (λ) = Dk + Nk , we have Dk = λI and
DR
Nk = Jk (0) so that Dk Nk = Nk Dk . As J = Jordan CFA, there exists an invertible matrix

S, such that S −1 AS = J. Hence, A = SJS −1 = SDS −1 + SN S −1 = B + C, which satisfy the
required conditions.
Chapter 9
Appendix
9.1 Uniqueness of RREF

Definition 9.1.1. Fix n ∈ N. Then, for each f ∈ Sn , we associate an n × n matrix, denoted
P f = [pij ], such that pij = 1, whenver f (j) = i and 0, otherwise. The matrix P f is called the
Permutation " matrix
# corresponding to the permutation f . For example, I2 , corresponding
0 1
to Id2 , and = E12 , corresponding to the permutation (1, 2), are the two permutation
1 0
matrices of order 2 × 2.
T
Remark 9.1.2. Recall that in Remark 9.2.16.1, it was observed that each permutation is a
AF
product of n transpositions, (1, 2), . . . , (1, n).

DR
1. Verify that the elementary matrix Eij is the permutation matrix corresponding to the
transposition (i, j) .
2. Thus, every permutation matrix is a product of elementary matrices E1j , 1 ≤ j ≤ n.
   
1 0 0 0 1 0
   
3. For n = 3, the permutation matrices are I3 , 0 0 1 = E23 = E12 E13 E12 , 1
   0 0=
0 1 0 0 0 1
     
0 1 0 0 0 1 0 0 1
     
E12 , 
0 0 1 = E12 E13 , 1 0 0 = E13 E12 and 0 1 0 = E13 .
    
1 0 0 0 1 0 1 0 0
4. Let f ∈ Sn and P f = [pij ] be the corresponding permutation matrix. Since pij = δi,j and
{f (1), . . . , f (n)} = [n], each entry of P f is either 0 or 1. Furthermore, every row and
column of P f has exactly one nonzero entry. This nonzero entry is a 1 and appears at
the position pi,f (i) .
5. By the previous paragraph, we see that when a permutation matrix is multiplied to A
(a) from left then it permutes the rows of A.
(b) from right then it permutes the columns of A.
6. P is a permutation matrix if and only if P has exactly one 1 in each row and column.
Solution: If P has exactly one 1 in each row and column, then P is a square matrix, say
223
224 CHAPTER 9. APPENDIX
n × n. Now, apply GJE to P . The occurrence of exactly one 1 in each row and column
implies that these 1’s are the pivots in each column. We just need to interchange rows to
get it in RREF. So, we need to multiply by Eij . Thus, GJE of P is In and P is indeed a
product of Eij ’s. The other part has already been explained earlier.
We are now ready to prove Theorem 2.4.6.
Theorem 9.1.3. Let A and B be two matrices in RREF. If they are row equivalent then A = B.
Proof. Note that the matrix A = 0 if and only if B = 0. So, let us assume that the matrices
A, B 6= 0. Also, the row-equivalence of A and B implies that there exists an invertible matrix
C such that A = CB, where C is product of elementary matrices.
Since B is in RREF, either B[:, 1] = 0T or B[:, 1] = (1, 0, . . . , 0)T . If B[:, 1] = 0T then
A[:, 1] = CB[:, 1] = C0 = 0. If B[:, 1] = (1, 0, . . . , 0)T then A[:, 1] = CB[:, 1] = C[:, 1]. As C is
invertible, the first column of C cannot be the zero vector. So, A[:, 1] cannot be the zero vector.
Further, A is in RREF implies that A[:, 1] = (1, 0, . . . , 0)T . So, we have shown that if A and B
are row-equivalent then their first columns must be the same.
Now, let us assume that the first k − 1 columns of A and B are equal and it contains r
pivotal columns. We will now show that the k-th column is also the same.
Define Ak = [A[:, 1], . . . , A[:, k]] and Bk = [B[:, 1], . . . , B[:, k]]. Then, our assumption implies
that A[:, i] = B[:, i], for 1 ≤ i ≤ k − 1. Since, the first k − 1 columns contain r pivotal columns,
there exists a permutation matrix P such that
T
AF
" # " #
Ir W A[:, k] Ir W B[:, k]
Ak P = and Bk P = .
DR
0 0 0 0
" # If the k-th columns of A and B are pivotal columns then by definition of RREF, A[:, k] =
0
= B[:, k], where 0 is a vector of size r and e1 = (1, 0, . . . , 0)T . So, we need to consider two
e1
cases depending on whether both are non-pivotal or one is pivotal and the other is not.
As A = CB, we get Ak = CBk and
" # " #" # " #
Ir W A[:, k] C1 C2 Ir W B[:, k] C1 C1 W CB[:, k]
= Ak P = CBk P = = .
0 0 C3 C4 0 0 C3 C3 W
" #
I r C2
So, we see that C1 = Ir , C3 = 0 and A[:, k] = B[:, k].
0 C4
Case 1: Neither A[:, k] nor B[:, k] are pivotal. Then
" # " # " #" # " #
X I r C2 I r C2 Y Y
= A[:, k] = B[:, k] = = .
0 0 C4 0 C4 0 0
Thus, X = Y and in this case the k-th columns are equal.
Case 2: A[:, k] is pivotal but B[:, k] in non-pivotal. Then
" # " # " #" # " #
0 I r C2 Ir C2 Y Y
= A[:, k] = B[:, k] = = ,
e1 0 C4 0 C4 0 0
a contradiction as e1 6= 0. Thus, this case cannot arise.
Therefore, combining both the cases, we get the required result.
9.2. PERMUTATION/SYMMETRIC GROUPS 225
9.2 Permutation/Symmetric Groups

Definition 9.2.1. For a positive integer n, denote [n] = {1, 2, . . . , n}. A function f : A → B is
called
1. one-one/injective if f (x) = f (y) for some x, y ∈ A necessarily implies that x = y.
2. onto/surjective if for each b ∈ B there exists a ∈ A such that f (a) = b.
3. a bijection if f is both one-one and onto.
Example 9.2.2. Let A = {1, 2, 3}, B = {a, b, c, d} and C = {α, β, γ}. Then, the function
1. j : A → B defined by j(1) = a, j(2) = c and j(3) = c is neither one-one nor onto.
2. f : A → B defined by f (1) = a, f (2) = c and f (3) = d is one-one but not onto.
3. g : B → C defined by g(a) = α, g(b) = β, g(c) = α and g(d) = γ is onto but not one-one.
4. h : B → A defined by h(a) = 2, h(b) = 2, h(c) = 3 and h(d) = 1 is onto.
5. h ◦ f : A → A is a bijection.
6. g ◦ f : A → C is neither one-one not onto.
Remark 9.2.3. Let f : A → B and g : B → C be functions. Then, the composition of

functions, denoted g ◦ f , is a function from A to C defined by (g ◦ f )(a) = g(f (a)). Also, if
1. f and g are one-one then g ◦ f is one-one.
T
AF
2. f and g are onto then g ◦ f is onto.

DR
Thus, if f and g are bijections then so is g ◦ f .
Definition 9.2.4. A function f : [n] → [n] is called a permutation on n elements if f is a

bijection. For example, f, g : [2] → [2] defined by f (1) = 1, f (2) = 2 and g(1) = 2, g(2) = 1 are
permutations.
Exercise 9.2.5. Let S3 be the set consisting of all permutation on 3 elements. Then, prove
that S3 has 6 elements. Moreover, they are one of the 6 functions given below.
1. f1 (1) = 1, f1 (2) = 2 and f1 (3) = 3.
2. f2 (1) = 1, f2 (2) = 3 and f2 (3) = 2.
3. f3 (1) = 2, f3 (2) = 1 and f3 (3) = 3.
4. f4 (1) = 2, f4 (2) = 3 and f4 (3) = 1.
5. f5 (1) = 3, f5 (2) = 1 and f5 (3) = 2.
6. f6 (1) = 3, f6 (2) = 2 and f6 (3) = 1.
Remark 9.2.6. Let f : [n] → [n] be a bijection. Then, the inverse of f , denote f −1 , is defined
by f −1 (m) = ` whenever f (`) = m for m ∈ [n] is well defined and f −1 is a bijection. For
example, in Exercise 9.2.5, note that fi−1 = fi , for i = 1, 2, 3, 6 and f4−1 = f5 .
Remark 9.2.7. Let Sn = {f : [n] → [n] : σ is a permutation}. Then, Sn has n! elements and
forms a group with respect to composition of functions, called product, due to the following.
1. Let f ∈ Sn . Then,
!
1 2 ··· n
(a) f can be written as f = , called a two row notation.
f (1) f (2) · · · f (n)
(b) f is one-one. Hence, {f (1), f (2), . . . , f (n)} = [n] and thus, f (1) ∈ [n], f (2) ∈ [n] \
{f (1)}, . . . and finally f (n) = [n]\{f (1), . . . , f (n−1)}. Therefore, there are n choices
for f (1), n − 1 choices for f (2) and so on. Hence, the number of elements in Sn
equals n(n − 1) · · · 2 · 1 = n!.
2. By Remark 9.2.3, f ◦ g ∈ Sn , for any f, g ∈ Sn .
3. Also associativity holds as f ◦ (g ◦ h) = (f ◦ g) ◦ h for all functions f, g and h.
4. Sn has a special permutation called the identity permutation, denoted Idn , such that
Idn (i) = i, for 1 ≤ i ≤ n.
5. For each f ∈ Sn , f −1 ∈ Sn and is called the inverse of f as f ◦ f −1 = f −1 ◦ f = Idn .
Lemma 9.2.8. Fix a positive integer n. Then, the group Sn satisfies the following:
1. Fix an element f ∈ Sn . Then, Sn = {f ◦ g : g ∈ Sn } = {g ◦ f : g ∈ Sn }.
2. Sn = {g −1 : g ∈ Sn }.
Proof. Part 1: Note that for each α ∈ Sn the functions f −1 ◦α, α◦f −1 ∈ Sn and α = f ◦(f −1 ◦α)
T
as well as α = (α ◦ f −1 ) ◦ f .
AF
Part 2: Note that for each f ∈ Sn , by definition, (f −1 )−1 = f . Hence the result holds.
DR
Definition 9.2.9. Let f ∈ Sn . Then, the number of inversions of f , denoted n(f ), equals
n(f ) = | {(i, j) : i < j, f (i) > f (j) } |

= | {j : i + 1 ≤ j ≤ n, f (j) < f (i)} | using two row notation. (9.2.1)
!
1 2 3 4
Example 9.2.10. 1. For f = , n(f ) = | {(1, 2), (1, 3), (2, 3)} | = 3.
3 2 1 4
2. In Exercise 9.2.5, n(f5 ) = 2 + 0 = 2.

!
1 2 3 4 5 6 7 8 9
3. Let f = . Then, n(f ) = 3 + 1 + 1 + 1 + 0 + 3 + 2 + 1 = 12.
4 2 3 5 1 9 8 7 6
Definition 9.2.11. [Cycle Notation] Let f ∈ Sn . Suppose there exist r, 2 ≤ r ≤ n and
i1 , . . . , ir ∈ [n] such that f (i1 ) = i2 , f (i2 ) = i3 , . . . , f (ir ) = i1 and f (j) = j for all j 6= i1 , . . . , ir .
Then, we represent such a permutation
! by f = (i1 , i2 , . . . , ir! ) and call it an r-cycle. For
1 2 3 4 5 1 2 3 4 5
example, f = = (1, 4, 5) and = (2, 3).
4 2 3 5 1 1 3 2 4 5
Remark 9.2.12. 1. One also write the r-cycle (i1 , i2 , . . . , ir ) as (i2 , i3 , . . . , ir , i1 ) and so on.
For example, (1, 4, 5) = (4, 5, 1) = (5, 1, 4).
!
1 2 3 4 5
2. The permutation f = is not a cycle.
4 3 2 5 1
3. Let f = (1, 3, 5, 4) and g = (2, 4, 1) be two cycles. Then, their product, denoted f ◦ g or
(1, 3, 5, 4)(2, 4, 1) equals (1, 2)(3, 5, 4). The calculation proceeds as (the arrows indicate the
images):
1 → 2. Note (f ◦ g)(1) = f (g(1)) = f (2) = 2.
2 → 4 → 1 as (f ◦ g)(2) = f (g(2)) = f (4) = 1. So, (1, 2) forms a cycle.
3 → 5 as (f ◦ g)(3) = f (g(3)) = f (3) = 5.
5 → 4 as (f ◦ g)(5) = f (g(5)) = f (5) = 4.
4 → 1 → 3 as (f ◦ g)(4) = f (g(4)) = f (1) = 3. So, the other cycle is (3, 5, 4).
4. Let f = (1, 4, 5) and g = (2, 4, 1) be two permutations. Then, (1, 4, 5)(2, 4, 1) = (1, 2, 5)(4) =
(1, 2, 5) as 1 → 2, 2 → 4 → 5, 5 → 1, 4 → 1 → 4 and
(2, 4, 1)(1, 4, 5) = (1)(2, 4, 5) = (2, 4, 5) as 1 → 4 → 1, 2 → 4, 4 → 5, 5 → 1 → 2.
!
1 2 3 4 5
5. Even though is not a cycle, verify that it is a product of the cycles
4 3 2 5 1
(1, 4, 5) and (2, 3).
Definition 9.2.13. A permutation f ∈ Sn is called a transposition if there exist m, r ∈ [n]

such that f = (m, r).
Remark 9.2.14. Verify that

1. (2, 4, 5) = (2, 5)(2, 4) = (4, 2)(4, 5) = (5, 4)(5, 2) = (1, 2)(1, 5)(1, 4)(1, 2).
T
2. in general, the r-cycle (i1 , . . . , ir ) = (1, i1 )(1, ir )(1, ir−1 ) · · · (1, i2 )(1, i1 ).
AF
3. So, every r-cycle can be written as product of transpositions. Furthermore, they can be
DR
written using the n transpositions (1, 2), (1, 3), . . . , (1, n).
With the above definitions, we state and prove two important results.
Theorem 9.2.15. Let f ∈ Sn . Then, f can be written as product of transpositions.
Proof. Note that using use Remark 9.2.14, we just need to show that f can be written as
product of disjoint cycles.
Consider the set S = {1, f (1), f (2) (1) = (f ◦ f )(1), f (3) (1) = (f ◦ (f ◦ f ))(1), . . .}. As S is an
infinite set and each f (i) (1) ∈ [n], there exist i, j with 0 ≤ i < j ≤ n such that f (i) (1) = f (j) (1).
Now, let j1 be the least positive integer such that f (i) (1) = f (j1 ) (1), for some i, 0 ≤ i < j1 .
Then, we claim that i = 0.
For if, i − 1 ≥ 0 then j1 − 1 ≥ 1 and the condition that f is one-one gives

f (i−1) (1) = (f −1 ◦ f (i) )(1) = f −1 f (i) (1) = f −1 f (j1 ) (1) = (f −1 ◦ f (j1 ) )(1) = f (j1 −1) (1).
Thus, we see that the repetition has occurred at the (j1 − 1)-th instant, contradicting the
assumption that j1 was the least such positive integer. Hence, we conclude that i = 0. Thus,
(1, f (1), f (2) (1), . . . , f (j1 −1) (1)) is one of the cycles in f .
Now, choose i1 ∈ [n] \ {1, f (1), f (2) (1), . . . , f (j1 −1) (1)} and proceed as above to get another
cycle. Let the new cycle by (i1 , f (i1 ), . . . , f (j2 −1) (i1 )). Then, using f is one-one follows that
1, f (1), f (2) (1), . . . , f (j1 −1) (1) ∩ i1 , f (i1 ), . . . , f (j2 −1) (i1 ) = ∅.

So, the above process needs to be repeated at most n times to get all the disjoint cycles. Thus,
Remark 9.2.16. Note that when one writes a permutation as product of disjoint cycles, cycles
of length 1 are suppressed so as to match Definition 9.2.11. For example, the algorithm in the
proof of Theorem 9.2.15 implies
1. Using Remark 9.2.14.3, we see that every permutation can be written as product of the n
transpositions (1, 2), (1, 3), . . . , (1, n).
!
1 2 3 4 5
2. = (1)(2, 4, 5)(3) = (2, 4, 5).
1 4 3 5 2
!
1 2 3 4 5 6 7 8 9
3. = (1, 4, 5)(2)(3)(6, 9)(7, 8) = (1, 4, 5)(6, 9)(7, 8).
4 2 3 5 1 9 8 7 6
Note that Id3 = (1, 2)(1, 2) = (1, 2)(2, 3)(1, 2)(1, 3), as well. The question arises, is it
possible to write Idn as a product of odd number of transpositions? The next lemma answers
this question in negative.
Lemma 9.2.17. Suppose there exist transpositions fi , 1 ≤ i ≤ t, such that
Idn = f1 ◦ f2 ◦ · · · ◦ ft ,
then t is even.
T
Proof. We will prove the result by mathematical induction. Observe that t 6= 1 as Idn is not a
AF
transposition. Hence, t ≥ 2. If t = 2, we are done. So, let us assume that the result holds for
DR
all expressions in which the number of transpositions t ≤ k. Now, let t = k + 1.

Suppose f1 = (m, r) and let `, s ∈ [n] \ {m, r}. Then, the possible choices for the com-
position f1 ◦ f2 are (m, r)(m, r) = Idn , (m, r)(m, `) = (r, `)(r, m), (m, r)(r, `) = (`, r)(`, m)
and (m, r)(`, s) = (`, s)(m, r). In the first case, f1 and f2 can be removed to obtain Idn =
f3 ◦ f4 ◦ · · · ◦ ft , where the number of transpositions is t − 2 = k − 1 < k. So, by mathematical
induction, t − 2 is even and hence t is also even.
In the remaining cases, the expression for f1 ◦ f2 is replaced by their counterparts to obtain
another expression for Idn . But in the new expression for Idn , m doesn’t appear in the first
transposition, but appears in the second transposition. The shifting of m to the right can
continue till the number of transpositions reduces by 2 (which in turn gives the result by
mathematical induction). For if, the shifting of m to the right doesn’t reduce the number
of transpositions then m will get shifted to the right and will appear only in the right most
transposition. Then, this expression for Idn does not fix m whereas Idn (m) = m. So, the later
case leads us to a contradiction. Hence, the shifting of m to the right will surely lead to an
expression in which the number of transpositions at some stage is t − 2 = k − 1. At this stage,
one applies mathematical induction to get the required result.
Theorem 9.2.18. Let f ∈ Sn . If there exist transpositions g1 , . . . , gk and h1 , . . . , h` with
f = g1 ◦ g2 ◦ · · · ◦ gk = h1 ◦ h2 ◦ · · · ◦ h`
then, either k and ` are both even or both odd.

Proof. As g1 ◦ · · · ◦ gk = h1 ◦ · · · ◦ h` and h−1 = h for any transposition h ∈ Sn , we have
Idn = g1 ◦ g2 ◦ · · · ◦ gk ◦ h` ◦ h`−1 ◦ · · · ◦ h1 .
Hence by Lemma 9.2.17, k + ` is even. Thus, either k and ` are both even or both odd.
Definition 9.2.19. [Even and Odd Permutation] A permutation f ∈ Sn is called an

1. even permutation if f can be written as product of even number of transpositions.
2. odd permutation if f can be written as a product of odd number of transpositions.
Definition 9.2.20. Observe that if f and g are both even or both odd permutations, then f ◦ g
and g ◦ f are both even. Whereas, if one of them is odd and the other even then f ◦ g and g ◦ f
are both odd. We use this to define a function sgn : Sn → {1, −1}, called the signature of a
permutation, by (
1 if f is an even permutation
sgn(f ) = .
−1 if f is an odd permutation
Example 9.2.21. Consider the set Sn . Then,
1. by Lemma 9.2.17, Idn is an even permutation and sgn(Idn ) = 1.
2. a transposition, say f , is an odd permutation and hence sgn(f ) = −1
3. using Remark 9.2.20, sgn(f ◦ g) = sgn(f ) · sgn(g) for any two permutations f, g ∈ Sn .
T
We are now ready to define determinant of a square matrix A.

AF
Definition 9.2.22. Let A = [aij ] be an n × n matrix with complex entries. Then, the deter-
DR
minant of A, denoted det(A), is defined as

X X n
Y
det(A) = sgn(g)a1g(1) a2g(2) . . . ang(n) = sgn(g) aig(i) . (9.2.2)
g∈Sn σ∈Sn i=1
" #
1 2
For example, if S2 = {Id, f = (1, 2)} then for A = , det(A) = sgn(Id) · a1Id(1) a2Id(2) +
2 1
sgn(f ) · a1f (1) a2f (2) = 1 · a11 a22 + (−1)a12 a21 = 1 − 4 = −3.
Observe that det(A) is a scalar quantity. Even though the expression for det(A) seems
complicated at first glance, it is very helpful in proving the results related with “properties of
determinant”. We will do so in the next section. As another examples, we verify that this
definition also matches for 3 × 3 matrices. So, let A = [aij ] be a 3 × 3 matrix. Then, using
Equation (9.2.2),
X 3
Y
det(A) = sgn(σ) aiσ(i)
σ∈Sn i=1
3
Y 3
Y 3
Y
= sgn(f1 ) aif1 (i) + sgn(f2 ) aif2 (i) + sgn(f3 ) aif3 (i) +
i=1 i=1 i=1
3
Y 3
Y 3
Y
sgn(f4 ) aif4 (i) + sgn(f5 ) aif5 (i) + sgn(f6 ) aif6 (i)
i=1 i=1 i=1
= a11 a22 a33 − a11 a23 a32 − a12 a21 a33 + a12 a23 a31 + a13 a21 a32 − a13 a22 a31 .
9.3 Properties of Determinant

Theorem 9.3.1 (Properties of Determinant). Let A = [aij ] be an n × n matrix.
1. If A[i, :] = 0T for some i then det(A) = 0.
2. If B = Ei (c)A, for some c 6= 0 and i ∈ [n] then det(B) = c det(A).
3. If B = Eij A, for some i 6= j then det(B) = − det(A).
4. If A[i, :] = A[j, :] for some i 6= j then det(A) = 0.
5. Let B and C be two n×n matrices. If there exists m ∈ [n] such that B[i, :] = C[i, :] = A[i, :]
for all i 6= m and C[m, :] = A[m, :] + B[m, :] then det(C) = det(A) + det(B).
6. If B = Eij (c), for c 6= 0 then det(B) = det(A).
7. If A is a triangular matrix then det(A) = a11 · · · ann , the product of the diagonal entries.
8. If E is an n × n elementary matrix then det(EA) = det(E) det(A).
9. A is invertible if and only if det(A) 6= 0.
10. If B is an n × n matrix then det(AB) = det(A) det(B).

T
AF
11. If AT denotes the transpose of the matrix A then det(A) = det(AT ).

DR
Proof. Part 1: Note that each sum in det(A) contains one entry from each row. So, each sum
has an entry from A[i, :] = 0T . Hence, each sum in itself is zero. Thus, det(A) = 0.
Part 2: By assumption, B[k, :] = A[k, :] for k 6= i and B[i, :] = cA[i, :]. So,
   
X Y X Y
det(B) = sgn(σ)  bkσ(k)  biσ(i) = sgn(σ)  akσ(k)  caiσ(i)
σ∈Sn k6=i σ∈Sn k6=i
X n
Y
= c sgn(σ) akσ(k) = c det(A).
σ∈Sn k=1
Part 3: Let τ = (i, j). Then, sgn(τ ) = −1, by Lemma 9.2.8, Sn = {σ ◦ τ : σ ∈ Sn } and
X n
Y X n
Y
det(B) = sgn(σ) biσ(i) = sgn(σ ◦ τ ) bi,(σ◦τ )(i)
σ∈Sn i=1 σ◦τ ∈Sn i=1
 
X Y
= sgn(τ ) · sgn(σ)  bkσ(k)  bi(σ◦τ )(i) bj(σ◦τ )(j)
σ◦τ ∈Sn k6=i,j
 
X Y X n
Y
= sgn(τ ) sgn(σ)  bkσ(k)  biσ(j) bjσ(i) = − sgn(σ) akσ(k)
σ∈Sn k6=i,j σ∈Sn k=1
= − det(A).
Part 4: As A[i, :] = A[j, :], A = Eij A. Hence, by Part 3, det(A) = − det(A). Thus, det(A) = 0.
9.3. PROPERTIES OF DETERMINANT 231
Part 5: By assumption, C[i, :] = B[i, :] = A[i, :] for i 6= m and C[m, :] = B[m, :] + A[m, :]. So,
 
X Yn X Y
det(C) = sgn(σ) ciσ(i) = sgn(σ)  ciσ(i)  cmσ(m)
σ∈Sn i=1 σ∈Sn i6=m
 
X Y
= sgn(σ)  ciσ(i)  (amσ(m) + bmσ(m) )
σ∈Sn i6=m
X n
Y X n
Y
= sgn(σ) aiσ(i) + sgn(σ) biσ(i) = det(A) + det(B).
σ∈Sn i=1 σ∈Sn i=1
Part 6: By assumption, B[k, :] = A[k, :] for k 6= i and B[i, :] = A[i, :] + cA[j, :]. So,
 
X Yn X Y
det(B) = sgn(σ) bkσ(k) = sgn(σ)  bkσ(k)  biσ(i)
σ∈Sn k=1 σ∈Sn k6=i
 
X Y
= sgn(σ)  akσ(k)  (aiσ(i) + cajσ(j) )
σ∈Sn k6=i
   
X Y X Y
= sgn(σ)  akσ(k)  aiσ(i) + c sgn(σ)  akσ(k)  ajσ(j) )
σ∈Sn k6=i σ∈Sn k6=i
X n
Y
= sgn(σ) akσ(k) + c · 0 = det(A). U seP art 4
T
σ∈Sn k=1
AF
Part 7: Observe that if σ ∈ Sn and σ 6= Idn then n(σ) ≥ 1. Thus, for every σ 6= Idn , there
DR
exists m ∈ [n] (depending on σ) such that m > σ(m) or m < σ(m). So, if A is triangular,
amσ(m) = 0. So, for each σ 6= Idn , ni=1 aiσ(i) = 0. Hence, det(A) = ni=1 aii . the result follows.
Q Q
Part 8: Using Part 7, det(In ) = 1. By definition Eij = Eij In and Ei (c) = Ei (c)In and
Eij (c) = Eij (c)In , for c 6= 0. Thus, using Parts 2, 3 and 6, we get det(Ei (c)) = c, det(Eij ) = −1
and det(Eij (k)) = 1. Also, again using Parts 2, 3 and 6, we get det(EA) = det(E) det(A).
Part 9: Suppose A is invertible. Then, by Theorem 2.7.1, A = E1 · · · Ek , for some elementary
matrices E1 , . . . , Ek . So, a repeated application of Part 8 implies det(A) = det(E1 ) · · · det(Ek ) 6=
0 as det(Ei ) 6= 0 for 1 ≤ i ≤ k.
Now, suppose that det(A) 6= 0. We need to show that A is invertible. On the contrary, as-
sume that A is not invertible. Then, by Theorem 2.7.1, Rank(A) <"n. So, # by Proposition 2.4.9,
B
there exist elementary matrices E1 , . . . , Ek such that E1 · · · Ek A = . Therefore, by Part 1
0
and a repeated application of Part 8 gives
" #!
B
det(E1 ) · · · det(Ek ) det(A) = det(E1 · · · Ek A) = det = 0.
0
As det(Ei ) 6= 0, for 1 ≤ i ≤ k, we have det(A) = 0, a contradiction. Thus, A is invertible.

Part 10: Let A be invertible. Then, by Theorem 2.7.1, A = E1 · · · Ek , for some elementary
matrices E1 , . . . , Ek . So, applying Part 8 repeatedly gives det(A) = det(E1 ) · · · det(Ek ) and
det(AB) = det(E1 · · · Ek B) = det(E1 ) · · · det(Ek ) det(B) = det(A) det(B).

In case A is not invertible, by Part 9, det(A) = 0. Also, AB is not invertible (AB is invertible
will imply A is invertible using the rank argument). So, again by Part 9, det(AB) = 0. Thus,
det(AB) = det(A) det(B).
Part 11: Let B = [bij ] = AT . Then, bij = aji , for 1 ≤ i, j ≤ n. By Lemma 9.2.8, we know that
Sn = {σ −1 : σ ∈ Sn }. As σ ◦ σ −1 = Idn , sgn(σ) = sgn(σ −1 ). Hence,
X n
Y X n
Y X n
Y
det(B) = sgn(σ) biσ(i) = sgn(σ) aσ(i),i = sgn(σ −1 ) aiσ−1 (i)
σ∈Sn i=1 σ∈Sn i=1 σ −1 ∈S n i=1
= det(A).
Remark 9.3.2. 1. As det(A) = det(AT ), we observe that in Theorem 9.3.1, the condition
on “row” can be replaced by the condition on “column”.
2. Let A = [aij ] be a matrix satisfying a1j = 0, for 2 ≤ j ≤ n. Let B = A(1 | 1), the submatrix
of A obtained by removing the first row and the first column. Then det(A) = a11 det(B).
Proof: Let σ ∈ Sn with σ(1) = 1. Then, σ has a cycle (1). So, a disjoint cycle represen-
tation of σ only has numbers {2, 3, . . . , n}. That is, we can think of σ as an element of
Sn−1 . Hence,
X n
Y X n
Y
det(A) = sgn(σ) aiσ(i) = sgn(σ) aiσ(i)
T
AF
σ∈Sn i=1 σ∈Sn ,σ(1)=1 i=1

n n−1
DR
X Y X Y
= a11 sgn(σ) aiσ(i) = a11 sgn(σ) biσ(i) = a11 det(B).
σ∈Sn ,σ(1)=1 i=2 σ∈Sn−1 i=1
We now relate this definition of determinant with the one given in Definition 2.8.1.
n
(−1)1+j a1j det A(1 | j) , where
P
Theorem 9.3.3. Let A be an n × n matrix. Then, det(A) =
j=1
recall that A(1 | j) is the submatrix of A obtained by removing the 1st row and the j th column.
 
0 0 · · · a1j · · · 0
 
 a21 a22 · · · a2j · · · a2n 
Proof. For 1 ≤ j ≤ n, define an n × n matrix Bj =  . . Also, for
 
 .. .. .. .. .. 
 . . . . 

an1 an2 · · · anj · · · ann
each matrix Bj , we define the n × n matrix Cj by
1. Cj [:, 1] = Bj [:, j],
2. Cj [:, i] = Bj [:, i − 1], for 2 ≤ i ≤ j and
3. Cj [:, k] = Bj [:, k] for k ≥ j + 1.
Also, observe that Bj ’s have been defined to satisfy B1 [1, :] + · · · + Bn [1, :] = A[1, :] and
Bj [i, :] = A[i, :] for all i ≥ 2 and 1 ≤ j ≤ n. Thus, by Theorem 9.3.1.5,
n
X
det(A) = det(Bj ). (9.3.1)
j=1
9.4. DIMENSION OF W1 + W2 233
Let us now compute det(Bj ), for 1 ≤ j ≤ n. Note that Cj = E12 E23 · · · Ej−1,j Bj , for 1 ≤ j ≤ n.
Then, by Theorem 9.3.1.3, we get det(Bj ) = (−1)j−1 det(Cj ). So, using Remark 9.3.2.2 and
Theorem 9.3.1.2 and Equation (9.3.1), we have
n
X n
X
(−1)j−1 det(Cj ) = (−1)j+1 a1j det A(1 | j) .

det(A) =
j=1 j=1
Thus, we have shown that the determinant defined in Definition 2.8.1 is valid.
9.4 Dimension of W1 + W2
Theorem 9.4.1. Let V be a finite dimensional vector space over F and let W1 and W2 be two
subspaces of V. Then,
dim(W1 ) + dim(W2 ) = dim(W1 + W2 ) + dim(W1 ∩ W2 ). (9.4.1)
Proof. Since W1 ∩ W2 is a vector subspace of V , let B = {u1 , . . . , ur } be a basis of W1 ∩ W2 .

As, W1 ∩ W2 is a subspace of both W1 and W2 , let us extend the basis B to form a basis
B1 = {u1 , . . . , ur , v1 , . . . , vs } of W1 and a basis B2 = {u1 , . . . , ur , w1 , . . . , wt } of W2 .
We now prove that D = {u1 , . . . , ur , w1 , . . . , ws , v1 , v2 , . . . , vt } is a basis of W1 + W2 . To
do this, we show that
1. D is linearly independent subset of V and

T
AF
2. LS(D) = W1 + W2 .
DR
The second part can be easily verified. For the first part, consider the linear system
α1 u1 + · · · + αr ur + β1 w1 + · · · + βs ws + γ1 v1 + · · · + γt vt = 0 (9.4.2)
in the variables αi ’s, βj ’s and γk ’s. We re-write the system as
α1 u1 + · · · + αr ur + β1 w1 + · · · + βs ws = −(γ1 v1 + · · · + γt vt ).
s r T
γi vi ∈ LS(B1 ) = W1 . Also, v = βk wk . So, v ∈ LS(B2 ) = W2 .
P P P
Then, v = − αr ur +
i=1 j=1 k=1
r
Hence, v ∈ W1 ∩ W2 and therefore, there exists scalars δ1 , . . . , δk such that v =
P
δ j uj .
j=1
Substituting this representation of v in Equation (9.4.2), we get
(α1 − δ1 )u1 + · · · + (αr − δr )ur + β1 w1 + · · · + βt wt = 0.
So, using Exercise 3.4.16.1, αi − δi = 0, for 1 ≤ i ≤ r and βj = 0, for 1 ≤ j ≤ t. Thus, the

system (9.4.2) reduces to
α1 u1 + · · · + αk uk + γ1 v1 + · · · + γr vr = 0
which has αi = 0 for 1 ≤ i ≤ r and γj = 0 for 1 ≤ j ≤ s as the only solution. Hence, we see that
the linear system of Equations (9.4.2) has no nonzero solution. Therefore, the set D is linearly
independent and the set D is indeed a basis of W1 + W2 . We now count the vectors in the sets
B, B1 , B2 and D to get the required result.
9.5 When does Norm imply Inner Product

In this section, we prove the following result. A generalization of this result to complex vector
space is left as an exercise for the reader as it requires similar ideas.
Theorem 9.5.1. Let V be a real vector space. A norm k · k is induced by an inner product if
and only if, for all x, y ∈ V, the norm satisfies
kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 (parallelogram law). (9.5.1)
Proof. Suppose that k · k is indeed induced by an inner product. Then, by Exercise 4.2.7.3 the
result follows.
So, let us assume that k · k satisfies the parallelogram law. So, we need to define an inner
product. We claim that the function f : V × V → R defined by
1
kx + yk2 − kx − yk2 , for all x, y ∈ V

f (x, y) =
4
satisfies the required conditions for an inner product. So, let us proceed to do so.
1
Step 1: Clearly, for each x ∈ V, f (x, 0) = 0 and f (x, x) = kx + xk2 = kxk2 . Thus,
4
f (x, x) ≥ 0. Further, f (x, x) = 0 if and only if x = 0.
Step 2: By definition f (x, y) = f (y, x) for all x, y ∈ V.

T
Step 3: Now note that kx + yk2 − kx − yk2 = 2 kx + yk2 − kxk2 − kyk2 . Or equivalently,

AF
2f (x, y) = kx + yk2 − kxk2 − kyk2 , for x, y ∈ V. (9.5.2)

DR
Thus, for x, y, z ∈ V, we have
4 (f (x, y) + f (z, y)) = kx + yk2 − kx − yk2 + kz + yk2 − kz − yk2

= 2 kx + yk2 + kz + yk2 − kxk2 − kzk2 − 2kyk2

= kx + z + 2yk2 + kx − zk2 − kx + zk2 + kx − zk2 − 4kyk2

= kx + z + 2yk2 − kx + zk2 − k2yk2

= 2f (x + z, 2y) using Equation (9.5.2). (9.5.3)
Now, substituting z = 0 in Equation (9.5.3) and using Equation (9.5.2), we get 2f (x, y) =
f (x, 2y) and hence 4f (x + z, y) = 2f (x + z, 2y) = 4 (f (x, y) + f (z, y)). Thus,
f (x + z, y) = f (x, y) + f (z, y), for all x, y ∈ V. (9.5.4)
Step 4: Using Equation (9.5.4), f (x, y) = f (y, x) and the principle of mathematical induction,
it follows that nf (x, y) = f (nx, y), for all x, y ∈ V and n ∈ N. Another application of
Equation (9.5.4) with f (0, y) = 0 implies that nf (x, y) = f (nx, y), for all x, y ∈ V and
n ∈ Z. Also, for m 6= 0,
n n
mf x, y = f (m x, y) = f (nx, y) = nf (x, y).
m m
Hence, we see that for all x, y ∈ V and a ∈ Q, f (ax, y) = af (x, y).
9.6. ROOTS OF A POLYNOMIALS 235
Step 5: Fix u, v ∈ V and define a function g : R → R by
g(x) = f (xu, v) − xf (u, v)

1 x
kxu + vk2 − kxuk2 − kvk2 − ku + vk2 − kuk2 − kvk2 .

=
2 2
Then, by previous step g(x) = 0, for all x ∈ Q. So, if g is a continuous function then
continuity implies g(x) = 0, for all x ∈ R. Hence, f (xu, v) = xf (u, v), for all x ∈ R.
Note that the second term of g(x) is a constant multiple of x and hence continuous. Using
a similar reason, it is enough to show that g1 (x) = kxu + vk, for certain fixed vectors
u, v ∈ V, is continuous. To do so, note that
kx1 u + vk = k(x1 − x2 )u + x2 u + vk ≤ k(x1 − x2 )uk + kx2 u + vk.

Thus, kx1 u + vk − kx2 u + vk ≤ k(x1 − x2 )uk. Hence, taking the limit as x1 → x2 , we

get lim kx1 u + vk = kx2 u + vk.

x1 →x2
Thus, we have proved the continuity of g and hence the prove of the required result.
9.6 Roots of a Polynomials

The main aim of this subsection is to prove the continuous dependence of the zeros of a poly-
nomial on its coefficients and to recall Descartes’s rule of signs.
T
Definition 9.6.1. [Jordan Curves] A curve in C is a continuous function f : [a, b] → C,

AF
where [a, b] ⊆ R.
DR
1. If the function f is one-one on [a, b) and also on (a, b], then it is called a simple curve.
2. If f (b) = f (a), then it is called a closed curve.
3. A closed simple curve is called a Jordan curve.
4. The derivative (integral) of a curve f = u+iv is defined component wise. If f 0 is continuous
on [a, b], we say f is a C 1 -curve (at end points we consider one sided derivatives and
continuity).
5. A C 1 -curve on [a, b] is called a smooth curve, if f 0 is never zero on (a, b).
6. A piecewise smooth curve is called a contour.
7. A positively oriented simple closed curve is called a simple closed curve such that
while traveling on it the interior of the curve always stays to the left. (Camille Jordan
has proved that such a curve always divides the plane into two connected regions, one of
which is called the bounded region and the other is called the unbounded region. The
one which is bounded is considered as the interior of the curve.)
We state the famous Rouche Theorem of complex analysis without proof.
Theorem 9.6.2. [Rouche’s Theorem] Let C be a positively oriented simple closed contour.
Also, let f and g be two analytic functions on RC , the union of the interior of C and the curve
C itself. Assume also that |f (x)| > |g(x)|, for all x ∈ C. Then, f and f + g have the same
number of zeros in the interior of C.
Corollary 9.6.3. [Alen Alexanderian, The University of Texas at Austin, USA.] Let P (t) =
tn +an−1 tn−1 +· · ·+a0 have distinct roots λ1 , . . . , λm with multiplicities α1 , . . . , αm , respectively.
Take any > 0 for which the balls B (λi ) are disjoint. Then, there exists a δ > 0 such that the
polynomial q(t) = tn + a0n−1 tn−1 + · · · + a00 has exactly αi roots (counting with multiplicities) in
B (λi ), whenever |aj − a0j | < δ.
Proof. For an > 0 and 1 ≤ i ≤ m, let Ci = {z ∈ C : |z − λi | = }. Now, for each i, 1 ≤ i ≤ m,

take νi = min |p(z)|, ρi = max[1 + |z| + · · · + |z|n−1 ] and choose δ > 0 such that ρi δ < νi . Then,
z∈Ci z∈Ci
for a fixed j and z ∈ Cj , we have
|q(z) − P (z)| = |(a0n−1 − an−1 )z n−1 + · · · + (a00 − a0 )| ≤ δρj < νj ≤ |p(z)|.
Hence, by Rouche’s theorem, p(z) and q(z) have the same number of zeros inside Cj , for each
j = 1, . . . , m. That is, the zeros of q(t) are within the -neighborhood of the zeros of P (t).
As a direct application, we obtain the following corollary.
Corollary 9.6.4. Eigenvalues of a matrix are continuous functions of its entries.
Proof. Follows from Corollary 9.6.3.

Pn
Remark 9.6.5. 1. [Sign changes in a polynomial] Let P (x) = 0 ai xn−i be a real polyno-
mial, with a0 6= 0. Read the coefficient from the left a0 , a1 , . . .. We say the sign changes
of ai occur at m1 < m2 < · · · < mk to mean that am1 is the first after a0 with sign
T
opposite to a0 ; am2 is the first after am1 with sign opposite to am1 ; and so on.
AF
Pn n−i be a real polynomial. Then, the

2. [Descartes’ Rule of Signs] Let P (x) = 0 ai x
DR
maximum number of positive roots of P (x) = 0 is the number of changes in sign of the
coefficients and that the maximum number of negative roots is the number of sign changes
in P (−x) = 0.
Proof. Assume that a0 , a1 , · · · , an has k > 0 sign changes. Let b > 0. Then, the coeffi-
cients of (x − b)P (x) are
a0 , a1 − ba0 , a2 − ba1 , · · · , an − ban−1 , −ban .
This list has at least k + 1 changes of signs. To see this, assume that a0 > 0 and an 6= 0.
Let the sign changes of ai occur at m1 < m2 < · · · < mk . Then, setting
c0 = a0 , c1 = am1 − bam1 −1 , c2 = am2 − bam2 −1 , · · · , ck = amk − bamk −1 , ck+1 = −ban ,
we see that ci > 0 when i is even and ci < 0, when i is odd. That proves the claim.
Now, assume that P (x) = 0 has k positive roots b1 , b2 , · · · , bk . Then,
P (x) = (x − b1 )(x − b2 ) · · · (x − bk )Q(x),
where Q(x) is a real polynomial. By the previous observation, the coefficients of (x −

bk )Q(x) has at least one change of signs, coefficients of (x − bk−1 )(x − bk )Q(x) has at least
two, and so on. Thus coefficients of P (x) has at least k change of signs. The rest of the
proof is similar.
9.7. VARIATIONAL CHARACTERIZATIONS OF HERMITIAN MATRICES 237
9.7 Variational characterizations of Hermitian Matrices

Let A ∈ Mn (C) be a Hermitian matrix. Then, by Theorem 6.4.10, we know that all the
eigenvalues of A are real. So, we write λi (A) to mean the i-th smallest eigenvalue of A. That
is, the i-th from the left in the list λ1 (A) ≤ λ2 (A) ≤ · · · ≤ λn (A).
Lemma 9.7.1. [Rayleigh-Ritz Ratio] Let A ∈ Mn (C) be a Hermitian matrix. Then,

1. λ1 (A)x∗ x ≤ x∗ Ax ≤ λn (A)x∗ x, for each x ∈ Cn .
∗
2. λ1 (A) = min xx∗Ax ∗
x = min x Ax.
x6=0 kxk=1
x∗ Ax
3. λn (A) = max x∗ x = max x∗ Ax.
x6=0 kxk=1
Proof. Proof of Part 1: By spectral theorem (see Theorem 6.4.10, there exists a unitary matrix
U such that A = U DU ∗ , where D = diag(λ1 (A), . . . , λn (A)) is a real diagonal matrix. Thus,
the set {U [:, 1], . . . , U [:, n]} is a basis of Cn . Hence, for each x ∈ Cn , there exists Ans :i ’s
(scalar) such that x = αi U [:, i]. So, note that x∗ x = |αi |2 and
P
X X X
λ1 (A)x∗ x = λ1 (A) |αi |2 ≤ |αi |2 λi (A) = x∗ Ax ≤ λn |αi |2 = λn x∗ x.
For Part 2 and Part 3, take x = U [:, 1] and x = U (:, n), respectively.
As an immediate corollary, we state the following result.
T
x∗ Ax
AF
Corollary 9.7.2. Let A ∈ Mn (C) be a Hermitian matrix and α = . Then, A has an

x∗ x
eigenvalue in the interval (−∞, α] and has an eigenvalue in the interval [α, ∞).
DR
We now generalize the second and third parts of Theorem 9.7.2.
Proposition 9.7.3. Let A ∈ Mn (C) be a Hermitian matrix with A = U DU ∗ , where U is a

unitary matrix and D is a diagonal matrix consisting of the eigenvalues λ1 ≤ λ2 ≤ · · · ≤ λn .
Then, for any positive integer k, 1 ≤ k ≤ n,
λk = min x∗ Ax = max x∗ Ax.

kxk=1 kxk=1
x⊥U [:,1],...,U [:,k−1] x⊥U [:,n],...,U [:,k+1]
Proof. Let x ∈ Cn such that x is orthogonal to U [, 1], . . . , U [:, k − 1]. Then, we can write
Pn
x= αi U [:, i], for some scalars αi ’s. In that case,
i=k
n
X n
X
λk x∗ x = λk |αi |2 ≤ |αi |2 λi = x∗ Ax
i=k i=k
and the equality occurs for x = U [:, k]. Thus, the required result follows.
Theorem 9.7.4. [Courant-Fischer] Let A ∈ Mn (C) be a Hermitian matrix with eigenvalues

λ1 ≤ λ2 ≤ · · · ≤ λn . Then,
λk = max min x∗ Ax = min max x∗ Ax.

w1 ,...,wk−1 kxk=1 wn ,...,wk+1 kxk=1
x⊥w1 ,...,wk−1 x⊥wn ,...,wk+1
Proof. Let A = U DU ∗ , where U is a unitary matrix and D = diag(λ1 , . . . , λn ). Now, choose a

set of k − 1 linearly independent vectors from Cn , say S = {w1 , . . . , wk−1 }. Then, some of the
eigenvectors {U [:, 1], . . . , U [:, k − 1]} may be an element of S ⊥ . Thus, using Proposition 9.7.3,
we see that
λk = min x∗ Ax ≥ min x∗ Ax.
kxk=1, kxk=1
x⊥U [:,1],...,U [:,k−1] x∈S ⊥
Hence, λk ≥ max min x∗ Ax, for each choice of k − 1 linearly independent vectors.
w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1
But, by Proposition 9.7.3, the equality holds for the linearly independent set {U [:, 1], . . . , U [:
, k − 1]} which proves the first equality. A similar argument gives the second equality and hence
the proof is omitted.
Theorem 9.7.5. [Weyl Interlacing Theorem] Let A, B ∈ Mn (C) be a Hermitian matrices.

Then, λk (A) + λ1 (B) ≤ λk (A + B) ≤ λk (A) + λn (B). In particular, if B = P ∗ P , for some
matrix P , then λk (A + B) ≥ λk (A). In particular, for z ∈ Cn , λk (A + zz∗ ) ≤ λk+1 (A).
Proof. As A and B are Hermitian matrices, the matrix A + B is also Hermitian. Hence, by
Courant-Fischer theorem and Lemma 9.7.1.1,
λk (A + B) = max min x∗ (A + B)x

w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1
T
≤ max min [x∗ Ax + λn (B)] = λk (A) + λn (B)

AF
w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1
DR
and
λk (A + B) = max min x∗ (A + B)x

w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1
≥ max min [x∗ Ax + λ1 (B)] = λk (A) + λ1 (B).

w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1
If B = P ∗ P , then λ1 (B) = min x∗ (P ∗ P )x = min kP xk2 ≥ 0. Thus,

kxk=1 kxk=1
λk (A + B) ≥ λk (A) + λ1 (B) ≥ λk (A).
In particular, for z ∈ Cn , we have
λk (A + zz∗ ) = max min [x∗ Ax + |x∗ z|2 ]

w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1
≤ max min [x∗ Ax + |x∗ z|2 ]

w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1 ,z
= max min x∗ Ax
w1 ,...,wk−1 kxk=1
x⊥w1 ,...,wk−1 ,z
≤ max min x∗ Ax = λk+1 (A).

w1 ,...,wk−1 ,wk kxk=1
x⊥w1 ,...,wk−1 ,wk
9.7. VARIATIONAL CHARACTERIZATIONS OF HERMITIAN MATRICES 239
Theorem 9.7.6.
" [Cauchy
# Interlacing Theorem] Let A ∈ Mn (C) be a Hermitian matrix.
A y
Define Â = ∗ , for some a ∈ R and y ∈ Cn . Then,
y a
λk (Â) ≤ λk (A) ≤ λk+1 (Â).
Proof. Note that
λk+1 (Â) = max min x∗ Âx ≤ max min x∗ Âx

w1 ,...,wk ∈Cn+1 kxk=1 w1 ,...,wk ∈Cn+1 kxk=1
xn+1 =0
x⊥w1 ,...,wk x⊥w1 ,...,wk
∗
= max min x Ax = λk+1 (A)
w1 ,...,wk ∈Cn kxk=1
x⊥w1 ,...,wk
and
λk+1 (Â) = min max x∗ Âx ≥ min max x∗ Âx

w1 ,...,wn−k ∈Cn+1 kxk=1 w1 ,...,wn−k ∈Cn+1 kxk=1
xn+1 =0
x⊥w1 ,...,wn−k x⊥w1 ,...,wn−k
∗
= min max x Ax = λk (A)
w1 ,...,wn−k ∈Cn kxk=1
x⊥w1 ,...,wn−k
As an immediate corollary, one has the following result.
Corollary 9.7.7. [Inclusion principle] Let A ∈ Mn (C) be a Hermitian matrix and r be a

T
positive integer with 1 ≤ r ≤ n. If Br×r is a principal submatrix of A then, λk (A) ≤ λk (B) ≤

AF
λk+n−r (A).
DR
Theorem 9.7.8. [Poincare Separation Theorem] Let A ∈ Mn (C) be a Hermitian matrix and
{u1 , . . . , ur } ⊆ Cn be an orthonormal set for some positive integer r, 1 ≤ r ≤ n. If further
B = [bij ] is an r × r matrix with bij = u∗i Auj , 1 ≤ i, j ≤ r then, λk (A) ≤ λk (B) ≤ λk+n−r (A).
Proof. Let us extend the i set {u1 , . . . , ur } to an orthonormal basis, say {u1 , . . . , un }
h orthonormal
of Cn and write U = u1 · · · un . Then, B is a r × r principal submatrix of U ∗ AU . Thus, by
inclusion principle, λk (U ∗ AU ) ≤ λk (B) ≤ λk+n−r (U ∗ AU ). But, we know that σ(U ∗ AU ) = σ(A)
and hence the required result follows.
The proof of the next result is left for the reader.
Corollary 9.7.9. Let A ∈ Mn (C) be a Hermitian matrix and r be a positive integer with
1 ≤ r ≤ n. Then,
λ1 (A) + · · · + λr (A) = ∗min trU∗ AU and λn−r+1 (A) + · · · + λn (A) = max

∗
trU∗ AU.
U U =Ir U U=Ir
Corollary 9.7.10. Let A ∈ Mn (C) be a Hermitian matrix and W be a k-dimensional subspace

of Cn . Suppose, there exists a real number c such that x∗ Ax ≥ cx∗ x, for each x ∈ W .
Then,
∗
λn−k+1 (A) ≥ c. In particular, if x Ax > 0, for each nonzero x ∈ W , then λn−k+1 > 0. Note
that, a k-dimensional subspace need not contain
" # an eigenvector of A. For example, the line
1 0

y = 2x does not contain an eigenvector of .
0 2
Proof. Let {x1 , . . . , xn−k } be a basis of W ⊥ . Then,
λn−k+1 (A) = max min x∗ Ax ≥ min x∗ Ax ≥ c.

w1 ,...,wn−k kxk=1 kxk=1
x⊥w1 ,...,wn−k x⊥x1 ,...,xn−k
Now assume that x∗ Ax > 0 holds for each nonzero x ∈ W and that λn−k+1 = 0. Then, it
follows that min x∗ Ax = 0. Now, define f : Cn → C by f (x) = x∗ Ax.
kxk=1
x⊥x1 ,...,xn−k
Then, f is a continuous function and min f (x) = 0. Thus, f must attain its bound on the
kxk=1
x∈W
unit sphere. That is, there exists y ∈ W with kyk = 1 such that y∗ Ay = 0, a contradiction.
T
AF
DR
Index
Jk (λ), 190 Equality of two Matrices, 10

k-th Moment, 208 Matrix, 9
Transpose of a Matrix, 11
Adjugate of a Matrix, 57
Descartes’ Rule of Signs, 236
Algebraic Multiplicity, 161
Determinant
Annihilating Polynomial, 197
Properties, 230
Basic Variables, 49 Determinant of a Square Matrix, 55, 229
Basis Matrix, 131 Diagonal matrix, 10
Basis of a Vector Space, 85 Diagonalizable Matrices, 164
Double Dual Space, 149
Cauchy-Schwartz Inequality, 99
Dual Space, 149
Cayley Hamilton Theorem, 173
Change of Basis Matrix, 140 Eigen-Condition, 156
Characteristic Equation, 156 Eigen-pair, 156
T
AF
Characteristic Polynomial, 156 Eigen-space, 156

Derivative, 206 Eigenvalue, 156
DR
Characteristic Root, 156 Eigenvalues

Characteristic Value, 156 Linear Operator, 159
Characteristic Vector, 156 Eigenvector, 156
Characteristic-pair, 156 Left, 159
Closed Curve, 235 Elementary Matrix, 33
Cofactor Matrix, 57 Elementary Row Operations, 33
Column vector, 9 Elementary Symmetric Functions, 205
Commuting Family of Matrices, 218 Equality of Linear Transformations, 117
Companion Matrix, 197
Finite Dimensional Vector Space, 73
Complex Vector Space, 66
Form
Contour, 235
Hermitian, 177
Convergent Matrix, 221
Linear, 178
convex hull, 141
Quadratic, 177
Coordinate Matrix, 134
Sesquilinear, 177
Coordinate Vector, 132
Free Variables, 49
Curves, 235
Function
Defective Matrix, 166 bijective, 225
Definition injective, 225
Conjugate Transpose of a Matrix, 11 Inverse, 128
241
242 INDEX
Left Inverse, 128 Linear forms, 178

one-one, 225 Linear Functional, 148
onto, 225 linear Independence, 77
Right Inverse, 128 Linear Operator, 117
surjective, 225 Eigenvalues, 159
Fundamental Theorem of Linear Algebra, 92 Linear Space
Norm, 101
Generalized Schur’s Theorem, 189 Linear Span of Vectors, 73
Geometric Multiplicity, 161 Linear System, 28
Given’s Method, 212 Associated Homogeneous System, 29
Gram-Schmidt Orthogonalization Process, 109 Augmented Matrix, 29
Coefficient Matrix, 29
Hamel basis, 85
Equivalent Systems, 35
Hermitian Form, 177
Homogeneous, 28
Inertial degree, 179
Non-Homogeneous, 28
Householder matrix, 24, 213
Non-trivial Solution, 29
Identity matrix, 10 Solution, 29
Identity Operator, 117 Solution Set, 29
Identity Transformation, 117 Trivial Solution, 29
Linear Transformation, 117
T
Index of an Eigenvalue, 191

AF
Inertia of a matrix, 177 Composition, 138

Equality, 117
DR
Inertial degree, 179

Inner Product, 97 Inverse, 129, 139
Inner Product Space, 97 Isomorphism, 130
Invariant Subspace, 218 Kernel, 124
Inverse of a Linear Transformation, 129 Matrix, 134
Inverse of a Matrix, 17 Matrix Product, 138
Isometry, 146 Non-Singular, 129
Null Space, 124
Jacobi Method, 213 Range Space, 124
Jordan Block, 190 Scalar Product, 126
Jordan Canonical Form, 194 Singular, 129
Jordan Curve, 235 Sum, 126
Jordan Matrix, 190 Lower Triangular Matrix, 10
Latent Roots, 156 Matrix, 9

Leading Entry, 30 Addition, 11
Left Eigenvector, 159 Additive Identity, 12
Linear Algebra Additive Inverse, 12
Fundamental Theorem, 92 Adjoint, 57
Linear Combination of Vectors, 72 Adjugate, 57
Linear Dependence, 77 Change of Basis, 133, 140
INDEX 243
Cofactor, 57 Row, 9
Column, 9 Row Echelon Form, 30
Commuting Family, 218 Row Equivalence, 34
Companion, 197 Row-Reduced Echelon Form, 38
Defective, 166 Scalar, 10
Determinant, 55 Scalar Multiplication, 11
Diagonal, 10 Singular, 56
Diagonalization, 164 Size, 9
Eigen-pair, 156 Skew-Hermitian, 19
Eigenvalue, 156 Skew-Symmetric, 19
Eigenvector, 156 Spectral Radius, 156
Elementary, 33 Spectrum, 156
Generalized Inverse, 24 Square, 10
Hermitian, 19 Submatrix, 21
Householder, 24, 213 Symmetric, 19
Idempotent, 20 Toeplitz, 203
Identity, 10 Trace, 23
Inverse, 17 triangular, 10
Jordan, 190 Unitary, 19
T
Linear Transformation, 134 Upper triangular, 10

AF
Lower triangular, 10 Zero, 10

DR
Minimal Polynomial, 198 Matrix Equality, 10

Negative definite, 175 Matrix Multiplication, 13
Negative semi-definite, 175 Matrix of a Linear Transformation, 134
Nilpotent, 20 Maximal linearly independent set, 83
Non-derogatory, 166 Maximal subset, 83
Non-Singular, 56 Minimal Polynomial, 198
Normal, 19
Non-derogatory Matrix, 166
Order, 9
Non-Singular Matrix, 56
Orthogonal, 19
Notation
Permutation, 19, 223
A(S | T ), 21
Positive definite, 175
A[S, T ], 21
Positive semi-definite, 175
[k], k a positive integer, 21
Principal Diagonal, 10
Principal Minor, 205 Operator
Principal Submatrix, 21 Identity, 117
Product of Matrices, 13 Zero, 117
Projection, 20 Order of Nilpotency, 20
Pseudo, 24 Ordered Basis, 131
Quadratic Form, 180 Orthogonal Complement, 104
Reflection, 20, 145 Orthogonal Operators, 145
244 INDEX
Orthogonal Projection, 142 Similar Matrices, 140

Orthogonal Vectors, 103 Simple Closed Curve, 235
Orthogonally Congruent, 147 Simple Curve, 235
Orthonormal Basis, 105 Simultaneously Diagonalizable, 217
Orthonormal Vectors, 105 Singular Matrix, 56
singular values, 187
Parseval’s formula, 107
Solution Set of a Linear System, 29
Permutation, 225
Space
Cycle notation, 226
Column Space, 88
Even, 229
Linear, 66
Odd, 229
Normed Linear, 101
Sgn function, 229
Null Space, 88
Signature, 229
Range, 88
Permutation Matrix, 19
Row Space, 88
Pivot Entry, 30
Spectrum of a matrix, 156
Pivotal Column, 30
Square Matrix
Plane Rotations, 210
Determinant, 229
Polynomial
Square matrix, 10
Sign Changes, 236
Standard basis
Principal Minor, 205
Cn , 11
T
Principal Submatrix, 21
AF
Mm,n (C), 19
Properties of Determinant, 230
Mn,1 (C), 11
DR
QR Decomposition, 112 Standard unit vectors, 11

Generalized, 113 Star congruence, 179
Quadratic Form, 177, 180 Submatrix of a Matrix, 21
Subset
Rank-Nullity Theorem, 125 Maximal, 83
Real Vector Space, 66 Maximal linear independent, 83
Reflection Operator, 145 Sum, 75
Reisz Representation Theorem, 121 Subspace
Rigid Motion, 146 Linear Span, 75
Row Equivalent Matrices, 34 Orthogonal Complement, 103
Row Operations Sum, 75
Elementary, 33 Sum of Principal Minors, 205
Row vector, 9 Sum of two Matrices, 11
Row-Reduced Echelon Form, 38 Sylvester’s law of inertia, 179
System of Linear Equations, 28
Scalar matrix, 10
Schur’s generalized theorem, 189 Toeplitz Matrix, 203
Selection Principle of Unitary Matrices, 209 Trace of a Matrix, 23
Sesquilinear Form, 177 Transformation
Signature of a matrix, 177 Zero, 118
INDEX 245
Triangular Matrix, 10 Word in Symbols s and t, 215

Trivial Subspace, 69
Zero matrix, 10
Unit Vector, 19, 99 Zero Operator, 117
Unitarily Congruent, 147 Zero Transformation, 118
Unitary Equivalence, 168
Unitary Group, 209
Unitary Similar, 168
Upper triangular form, 11
Upper triangular matrix, 10
Vector
Column, 9
Coordinate, 132
Row, 9
Unit, 19
Vector Space, 66
Basis, 85
Complex, 66
Complex n-tuple, 67
T
Dimension of M + N , 233
AF
Dual Basis, 149

DR
Finite Dimensional, 73
Infinite Dimensional, 73
Inner Product, 97
Isomorphic, 130
Minimal spanning set, 85
Real, 66
Real n-tuple, 67
Subspace, 69
Vector Subspace, 69
Vectors
Angle, 100
Length, 99
Linear Combination, 72
Linear Dependence, 77
Linear Independence, 77
Linear Span, 73
Mutually Orthogonal, 105
Norm, 99
Orthogonal, 103
Orthonormal, 105

Linear Algebra

Uploaded by

Linear Algebra

Uploaded by

Linear Algebra through Matrices

Arbind K Lal Sukant Pati

July 10, 2018

1.5 Submatrix of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 System of Linear Equations 27

4 Inner Product Spaces 97

4.4.1 Properties of Orthonormal Vectors . . . . . . . . . . . . . . . . . . . . . . 105

4.6 QR Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

5 Linear Transformations 117

6 Eigenvalues, Eigenvectors and Diagonalizability 153

6.4.1 Diagonalizability of some Special Matrices . . . . . . . . . . . . . . . . . . 170

7 Jordan canonical form 189

8 Advanced Topics on Diagonalizability and Triangularization∗ 205

8.2.3 Schur’s Upper Triangularization Revisited . . . . . . . . . . . . . . . . . . 215

8.3 Commuting Matrices and Simultaneous Diagonalization . . . . . . . . . . . . . . 217

8.3.1 Diagonalization and Real Orthogonal Matrix . . . . . . . . . . . . . . . . 220

Example 1.1.1. 1. Consider the linear system

given lines 2x + 5y = 7 and 2x + 4y = 6. But, we also see that

which corresponds to the solution of

2. Consider the linear system 

where the second

3. As a last example, consider the linear system

1.2 Definition of a Matrix

Definition 1.2.1. A rectangular array of numbers is called a matrix.

1.2.1 Special Matrices

Definition 1.2.4. Let A = [aij ] be an m × n matrix with aij ∈ F.

1. Then A is called a zero-matrix, denoted 0"(order # is mostly clear

2. Then A is called a square matrix if m = n and is denoted by A ∈ Mn (F).

4. An m × n matrix A = [aij ] is said to have an uppertriangular form if aij = 0 for all

1.3 Matrix Operations

1.3.1 Transpose and Conjugate Transpose of Matrices

Definition 1.3.1. Let A = [aij ] ∈ Mm,n (C). Then

Theorem 1.3.2. For any matrix A, (A∗ )∗ = A and (AT )T = A.

1.3.2 Sum and Scalar Multiplication of Matrices

Definition 1.3.3. Let A = [aij ], B = [bij ] ∈ Mm,n (C) and k ∈ C.

Theorem 1.3.5. Let A, B, C ∈ Mm,n (C) and let k, ` ∈ C. Then

Proof. (1). Let A = [aij ] and B = [bij ]. Then by definition

A + B = [aij ] + [bij ] = [aij + bij ] = [bij + aij ] = [bij ] + [aij ] = B + A

Definition 1.3.6. Let A ∈ Mm,n (C). Then

1. the matrix 0m×n satisfying A + 0 = 0 + A = A is called the additive identity.

2. the matrix B with A + B = 0 is called the additive inverse of A, denoted −A = (−1)A.

Exercise 1.3.7. 1. Find non zero, non-identity matrices A satisfying

(a) A∗ = A (such matrices are called Hermitian matrices).

2. Suppose A = [aij ], B = [bij ] ∈ Mm,n (C).

5. Write the 3 × 3 matrices A = [aij ] satisfying

1.3.3 Multiplication of Matrices

1. Entry-wise Method: (AB)  11 = 1 · 3 +

A(B[:, 1]) is a 3 × 1 matrix. So, matrix multiplication is defined. Thus,

defined and is a 3 × 3 matrix. Hence,

1. AB corresponds to operating (combining certain multiples of rows) on the rows of B. This

(AB)[i, :] = A[i, :]B = ai1 B[1, :] + · · · + ain B[n, :], for 1 ≤ i ≤ m.

2. AB also corresponds to operating (combining certain multiples of columns) on the columns

(AB)[:, j] = AB[:, j] = A[:, 1]b1j + · · · + A[:, n]bnj , for 1 ≤ j ≤ p.

AB = A[:, 1]B[1, :] + A[:, 2]B[2, :] + · · · + A[:, n]B[n, :].

5. If m 6= p then the product BA is NOT defined.

7. If m = n = p, then the orders

1. Then (AB)C = A(BC), i.e., the matrix multiplication is associative.

2. For any k ∈ C, (kA)B = k(AB) = A(kB).

3. Then A(B + C) = AB + AC, i.e., multiplication distributes over addition.

4. If A ∈ Mn (C) then AIn = In A = A.

2. Let L1 , L2 ∈ Mn (C) be lower triangular matrices. If D ∈ Mn (C) is a diagonal matrix then

(a) L1 L2 is a lower triangular matrix.

The same holds for upper triangular matrices.

3. Let A ∈ Mm,n (C) and B ∈ Mn,p (C).

(a) Prove that (AB)∗ = B ∗ A∗ .

4. Construct matrices A and B that satisfy the following statements.

(a) The product AB is defined but BA is not defined.

8. Let A be an upper triangular matrix. If A∗ A = AA∗ then prove that A is a diagonal

9. Let A be a 3 × 3 upper triangular matrix with diagonal entries a, b, c. Then

(A − aI3 )(A − bI3 )(A − cI3 ) = 0.

10. Let A and B be two m × n matrices. Then, prove that (A + B)∗ = A∗ + B ∗ .

11. Find A, B, C ∈ M2 (C) such that AB = AC but B 6= C (Cancellation laws don’t

1.3.4 Inverse of a Matrix

Definition 1.3.14. Let A ∈ Mn (C). Then

1. B ∈ Mn (C) is said to be a left inverse of A if BA = In .

2. C ∈ Mn (C) is called a right inverse of A if AC = In .