Nonlinear Systems

by Peter J. Olver
University of Minnesota

1. Introduction.
Nonlinearity is ubiquitous in physical phenomena. Fluid and plasma mechanics, gas
dynamics, elasticity, relativity, chemical reactions, combustion, ecology, biomechanics, and
many, many other phenomena are all governed by inherently nonlinear equations. (The one
notable exception is quantum mechanics, which is a fundamentally linear theory, although
recent attempts at grand unification of all fundamental physical theories, such as string
theory and conformal field theory, [8], are nonlinear.) For this reason, an ever increasing
proportion of modern mathematical research is devoted to the analysis of nonlinear systems
and nonlinear phenomena.
Why, then, does one devote so much time studying linear mathematics? The facile
answer is that nonlinear systems are vastly more difficult to analyze. In the nonlinear
regime, many of the most basic questions remain unanswered: existence and uniqueness of
solutions are not guaranteed; explicit formulae are difficult to come by; linear superposition
is no longer available; numerical approximations are not always sufficiently accurate; etc.,
etc. A more intelligent answer is that a thorough understanding of linear phenomena
and linear mathematics is an essential prerequisite for progress in the nonlinear arena.
Therefore, one must first develop the proper linear foundations in sufficient depth before
we can realistically confront the untamed nonlinear wilderness. Moreover, many important
physical systems are “weakly nonlinear”, in the sense that, while nonlinear effects do
play an essential role, the linear terms tend to dominate the physics, and so, to a first
approximation, the system is essentially linear. As a result, such nonlinear phenomena are
best understood as some form of perturbation of their linear approximations. The truly
nonlinear regime is, even today, only sporadically modeled and even less well understood.
The advent of powerful computers has fomented a veritable revolution in our under-
standing of nonlinear mathematics. Indeed, many of the most important modern analytical
techniques drew their inspiration from early computer-aided investigations of nonlinear sys-
tems. However, despite dramatic advances in both hardware capabilities and sophisticated
mathematical algorithms, many nonlinear systems — for instance, fully general Einsteinian
gravitation, or the Navier–Stokes equations of fluid mechanics at high Reynolds numbers
— still remain beyond the capabilities of today’s computers.
The goal of these lecture notes is to provide a brief overview of some of the most
important ideas, mathematical techniques, and new physical phenomena in the nonlinear
realm. We start with iteration of nonlinear functions, also known as discrete dynamical
systems. Building on our experience with iterative linear systems, as developed in Chap-
ter 10 of [15], we will discover that functional iteration, when it converges, provides a

10/23/22 1 c 2022 Peter J. Olver

powerful mechanism for solving equations and for optimization. On the other hand, even
very simple non-convergent nonlinear iterative systems may admit remarkably complex,
chaotic behavior. The third section is devoted to basic solution techniques for nonlinear
equations and nonlinear systems, and includes bisection, general iteration, and the very
powerful Newton Method. The fourth section is devoted to finite-dimensional optimiza-
tion principles, i.e., the minimization or maximization of nonlinear functions, and including
systems with constraints leading to the method of Lagrange multipliers. Numerical opti-
mization procedures rely on iterative procedures, and we concentrate on those associated
with gradient descent.

2. Iteration of Functions.
Iteration, meaning repeated application of a function, can be viewed as a discrete
dynamical system in which the continuous time variable has been “quantized” to assume
integer values. Even iterating a very simple quadratic scalar function can lead to an amaz-
ing variety of dynamical phenomena, including multiply-periodic solutions and genuine
chaos. Nonlinear iterative systems arise not just in mathematics, but also underlie the
growth and decay of biological populations, predator-prey interactions, spread of commu-
nicable diseases such as Aids, and host of other natural phenomena. Moreover, many
numerical solution methods — for systems of algebraic equations, ordinary differential
equations, partial differential equations, and so on — rely on iteration, and so the the-
ory underlies the analysis of convergence and efficiency of such numerical approximation
schemes.
In general, an iterative system has the form

u(k+1) = g(u(k) ), (2.1)

where g: R n → R n is a real vector-valued function. (One can similarly treat iteration of
complex-valued functions g: C n → C n , but, for simplicity, we only deal with real systems
here.) A solution is a discrete collection of points† u(k) ∈ R n , in which the index k =
0, 1, 2, 3, . . . takes on non-negative integer values. Chapter 10 of [15] dealt with the case
when g(u) = A u is a linear function, necessarily given by multiplication by an n × n
matrix A. In this chapter, we enlarge our scope to the nonlinear case.
Once we specify the initial iterate,

u(0) = c, (2.2)
then the resulting solution to the discrete dynamical system (2.1) is easily computed:

u(1) = g(u(0) ) = g(c), u(2) = g(u(1) ) = g(g(c)), u(3) = g(u(2) ) = g(g(g(c))), . . .

and so on. Thus, unlike continuous dynamical systems, the existence and uniqueness of
solutions is not an issue. As long as each successive iterate u(k) lies in the domain of

†
The superscripts on u(k) refer to the iteration number, and do not denote derivatives.

10/23/22 2 c 2022 Peter J. Olver

definition of g one merely repeats the process to produce the solution,
k times
z }| { (2.3)
u(k) = g◦ ··· ◦g (c), k = 0, 1, 2, . . . ,
which is obtained by composing the function g with itself a total of k times. In other
words, the solution to a discrete dynamical system corresponds to repeatedly pushing the
g key on your calculator. For example, entering 0 and then repeatedly hitting the cos key
corresponds to solving the iterative system

u(k+1) = cos u(k) , u(0) = 0. (2.4)

The first 10 iterates are displayed in the following table:
k 0 1 2 3 4 5 6 7 8 9
(k)
u 0 1 .540302 .857553 .65429 .79348 .701369 .76396 .722102 .750418
For simplicity, we shall always assume that the vector-valued function g: R n → R n is
defined on all of R n ; otherwise, we must always be careful that the successive iterates u(k)
never leave its domain of definition, thereby causing the iteration to break down. To avoid
technical complications, we will also assume that g is at least continuous; later results rely
on additional smoothness requirements, e.g., continuity of its first and second order partial
derivatives.
While the solution to a discrete dynamical system is essentially trivial, understanding
its behavior is definitely not. Sometimes the solution converges to a particular value —
the key requirement for numerical solution methods. Sometimes it goes off to ∞, or, more
precisely, the norms† of the iterates are unbounded: k u(k) k → ∞ as k → ∞. Sometimes
the solution repeats itself after a while. And sometimes the iterates behave in a seemingly
random, chaotic manner — all depending on the function g and, at times, the initial
condition c. Although all of these cases may arise in real-world applications, we shall
mostly concentrate upon understanding convergence.
Definition 2.1. A fixed point or equilibrium of a discrete dynamical system (2.1) is
a vector u⋆ ∈ R n such that
g(u⋆ ) = u⋆ . (2.5)
We easily see that every fixed point provides a constant solution to the discrete dy-
namical system, namely u(k) = u⋆ for all k. Moreover, it is not hard to prove that any
convergent solution necessarily converges to a fixed point.
Proposition 2.2. If a solution to a discrete dynamical system converges,
lim u(k) = u⋆ ,
k→∞

then the limit u⋆ is a fixed point.

†
In view of the equivalence of norms on finite-dimensional vector spaces, cf. [ 15 ], any norm
will do here.

10/23/22 3 c 2022 Peter J. Olver

 Proof : This is a simple consequence of the continuity of g. We have
 
⋆ (k+1) (k) (k)
u = lim u = lim g(u ) = g lim u = g(u⋆ ),
k→∞ k→∞ k→∞

the last two equalities following from the continuity of g. Q.E.D.

For example, continuing the cosine iteration (2.4), we find that the iterates gradually
converge to the value u⋆ ≈ .739085, which is the unique solution to the fixed point equation

cos u = u.
Later we will see how to rigorously prove this observed behavior.
Of course, not every solution to a discrete dynamical system will necessarily converge,
but Proposition 2.2 says that if it does, then it must converge to a fixed point. Thus, a
key goal is to understand when a solution converges, and, if so, to which fixed point —
if there is more than one. (In the linear case, only the actual convergence is a significant
issues since most linear systems admit exactly one fixed point, namely u⋆ = 0.)
Fixed points are roughly divided into three classes:
• asymptotically stable, with the property that all nearby solutions converge to it,
• stable, with the property that all nearby solutions stay nearby, and
• unstable, almost all of whose nearby solutions diverge away from the fixed point.
Thus, from a practical standpoint, convergence of the iterates of a discrete dynamical
system requires asymptotic stability of the fixed point. Examples will appear in abundance
in the following sections.
Scalar Functions
As always, the first step is to thoroughly understand the scalar case, and so we begin
with a discrete dynamical system

u(k+1) = g(u(k) ), u(0) = c, (2.6)

in which g: R → R is a continuous, scalar-valued function. As noted above, we will assume,
for simplicity, that g is defined everywhere, and so we do not need to worry about whether
the iterates u(0) , u(1) , u(2) , . . . are all well-defined.
The elementary linear case g(u) = a u is treated in [15; Chapter 10]. The simplest
“nonlinear” case is that of an affine function

g(u) = a u + b, (2.7)
leading to an affine discrete dynamical system

u(k+1) = a u(k) + b. (2.8)

The only fixed point is the solution to
b
u⋆ = g(u⋆ ) = a u⋆ + b, namely, u⋆ = . (2.9)
1−a

10/23/22 4 c 2022 Peter J. Olver

 u⋆3

u⋆2

u⋆1

Figure 1. Fixed Points.

The formula for u⋆ requires that a 6= 1, and, indeed, the case a = 1 has no fixed point, as
the reader can easily confirm.
Since we already know the value of u⋆ , we can readily analyze the differences

e(k) = u(k) − u⋆ , (2.10)

between successive iterates and the fixed point. Observe that, the smaller e(k) is, the closer
u(k) is to the desired fixed point. In many applications, the iterate u(k) is viewed as an
approximation to the fixed point u⋆ , and so e(k) is interpreted as the error in the k th
iterate. Subtracting the fixed point equation (2.9) from the iteration equation (2.8), we
find
u(k+1) − u⋆ = a (u(k) − u⋆ ).
Therefore the errors e(k) are related by a linear iteration

e(k+1) = a e(k) , and hence e(k) = ak e(0) . (2.11)

Therefore, the solutions to this scalar linear iteration converge:

e(k) −→ 0 and hence u(k) −→ u⋆ , if and only if | a | < 1.

This is the criterion for asymptotic stability of the fixed point, or, equivalently, convergence
of the affine iterative system (2.8). The magnitude of a determines the rate of convergence,
and the closer it is to 0, the faster the iterates approach the fixed point.
Example 2.3. The affine function
1
g(u) = 4 u+2
leads to the iterative scheme
u(k+1) = 1
4 u(k) + 2.

10/23/22 5 c 2022 Peter J. Olver

 Figure 2. Tangent Line Approximation.

Starting with the initial condition u(0) = 0, the ensuing values are
k 1 2 3 4 5 6 7 8
(k)
u 2.0 2.5 2.625 2.6562 2.6641 2.6660 2.6665 2.6666
8
Thus, after 8 iterations, the iterates have produced the fixed point u⋆ = 3 to 4 decimal
places. The rate of convergence is 14 , and indeed


1 k


1 k
| e(k) | = | u(k) − u⋆ | = 4
| u(0) − u⋆ | = 8
3 4
−→ 0 as k −→ ∞.
Let us now turn to the fully nonlinear case. First note that the fixed points of g(u)
correspond to the intersections of its graph with the graph of the function i(u) = u. For
instance Figure 1 shows the graph of a function that has 3 fixed points, labeled u⋆1 , u⋆2 , u⋆3 .
In general, near any point in its domain, a (smooth) nonlinear function can be well
approximated by its tangent line, which represents the graph of an affine function; see
Figure 2. Therefore, if we are close to a fixed point u⋆ , then we might expect the iterative
system based on the nonlinear function g(u) to behave very much like that of its affine
tangent line approximation. And, indeed, this intuition turns out to be essentially correct.
This result forms our first concrete example of linearization, in which the analysis of a
nonlinear system is based on its linear (or, more precisely, affine) approximation.
The explicit formula for the tangent line to g(u) near the fixed point u = u⋆ = g(u⋆ )
is
g(u) ≈ g(u⋆ ) + g ′ (u⋆ )(u − u⋆ ) ≡ a u + b, (2.12)
where

a = g ′ (u⋆ ), b = g(u⋆ ) − g ′ (u⋆ ) u⋆ = 1 − g ′ (u⋆ ) u⋆ .
Note that u⋆ = b /(1 − a) remains a fixed point for the affine approximation: a u⋆ + b =
u⋆ . According to the preceding discussion, the convergence of the iterates for the affine
approximation is governed by the size of the coefficient a = g ′ (u⋆ ). This observation
inspires the basic stability criterion for fixed points of scalar iterative systems.

10/23/22 6 c 2022 Peter J. Olver

 Theorem 2.4. Let g(u) be a continuously differentiable scalar function. Suppose
u⋆ = g(u⋆ ) is a fixed point. If | g ′ (u⋆ ) | < 1, then u⋆ is an asymptotically stable fixed
point, and hence any sequence of iterates u(k) which starts out sufficiently close to u⋆ will
converge to u⋆ . On the other hand, if | g ′ (u⋆ ) | > 1, then u⋆ is an unstable fixed point, and
the only iterates which converge to it are those that land exactly on it, i.e., u(k) = u⋆ for
some k ≥ 0.
Proof : The goal is to prove that the errors e(k) = u(k) − u⋆ between the iterates and
the fixed point tend to 0 as k → ∞. To this end, we try to estimate e(k+1) in terms of
e(k) . According to (2.6) and the Mean Value Theorem from calculus,

e(k+1) = u(k+1) − u⋆ = g(u(k) ) − g(u⋆ ) = g ′ (v) (u(k) − u⋆ ) = g ′ (v) e(k) , (2.13)

for some v lying between u(k) and u⋆ . By continuity, if | g ′ (u⋆ ) | < 1 at the fixed point,
then we can choose δ > 0 and | g ′ (u⋆ ) | < σ < 1 such that the estimate

| g ′ (v) | ≤ σ < 1 whenever | v − u⋆ | < δ (2.14)

holds in a (perhaps small) interval surrounding the fixed point. Suppose

| e(k) | = | u(k) − u⋆ | < δ.

Then the point v in (2.13), which is closer to u⋆ than u(k) , satisfies (2.14). Therefore,

| u(k+1) − u⋆ | ≤ σ | u(k) − u⋆ |, and hence | e(k+1) | ≤ σ | e(k) |. (2.15)

In particular, since σ < 1, we have | u(k+1) − u⋆ | < δ, and hence the subsequent iterate
u(k+1) also lies in the interval where (2.14) holds. Repeating the argument, we conclude
that, provided the initial iterate satisfies

| e(0) | = | u(0) − u⋆ | < δ,

the subsequent errors are bounded by

e(k) ≤ σ k e(0) , and hence e(k) = | u(k) − u⋆ | −→ 0 as k → ∞,

which completes the proof of the theorem in the stable case.
The proof in unstable case is left as an exercise for the reader. Q.E.D.
Remark : The constant σ governs the rate of convergence of the iterates to the fixed
point. The closer the iterates are to the fixed point, the smaller we can choose δ in (2.14),
and hence the closer we can choose σ to | g ′ (u⋆ ) |. Thus, roughly speaking, | g ′ (u⋆ ) | governs
the speed of convergence, once the iterates get close to the fixed point. This observation
will be developed more fully in the following subsection.
Remark : The cases when g ′ (u⋆ ) = ± 1 are not covered by the theorem. For a linear
system, such fixed points are stable, but not asymptotically stable. For nonlinear systems,
more detailed knowledge of the nonlinear terms is required in order to resolve the status —
stable or unstable — of the fixed point. Despite their importance in certain applications,
we will not try to analyze such borderline cases any further here.

10/23/22 7 c 2022 Peter J. Olver

 m

Figure 3. Planetary Orbit.

Example 2.5. Given constants ǫ, m, the trigonometric equation

u = m + ǫ sin u (2.16)
is known as Kepler’s equation. It arises in the study of planetary motion, in which 0 < ǫ < 1
represents the eccentricity of an elliptical planetary orbit, u is the eccentric anomaly,
defined as the angle formed at the center of the ellipse by the planet and the major axis,
and m = 2 π t / T is its mean anomaly, which is the time, measured in units of T /(2 π)
where T is the period of the orbit, i.e., the length of the planet’s year, since perihelion or
point of closest approach to the sun; see Figure 3.
The solutions to Kepler’s equation are the fixed points of the discrete dynamical
system based on the function
g(u) = m + ǫ sin u.
Note that
| g ′ (u) | = | ǫ cos u | ≤ | ǫ | < 1, (2.17)
which automatically implies that the as yet unknown fixed point is stable. Indeed, condi-
tion (2.17) is enough to prove the existence of a unique stable fixed point; see Theorem 2.18
below. In the particular case m = ǫ = 12 , the result of iterating u(k+1) = 12 + 21 sin u(k)
starting with u(0) = 12 is

k 0 1 2 3 4 5 6 7 8
u(k) .5 .7397 .8370 .8713 .8826 .8862 .8873 .8877 .8878
After 12 iterations, we have converged sufficiently close to the solution (fixed point) u⋆ =
.887862 to have computed its value to 6 decimal places.

10/23/22 8 c 2022 Peter J. Olver

 L+ (u)

g(u)

L− (u)
u
⋆
u

Figure 4. Graph of a Contraction.

Inspection of the proof of Theorem 2.4 reveals that we never really used the differen-
tiability of g, except to verify the inequality
| g(u) − g(u⋆ ) | ≤ σ | u − u⋆ | for some fixed σ < 1. (2.18)
A function that satisfies (2.18) for all nearby u is called a contraction at the point u⋆ . Any
function g(u) whose graph lies between the two lines
L± (u) = g(u⋆ ) ± σ (u − u⋆ ) for some σ < 1,
for all u sufficiently close to u⋆ , i.e., such that | u − u⋆ | < δ for some δ > 0, defines a
contraction, and hence fixed point iteration starting with | u(0) − u⋆ | < δ will converge to
u⋆ ; see Figure 4. In particular, any function that is differentiable at u⋆ with | g ′ (u⋆ ) | < 1
defines a contraction at u⋆ .
Example 2.6. The simplest truly nonlinear example is a quadratic polynomial. The
most important case is the so-called logistic map
g(u) = λ u(1 − u), (2.19)
where λ 6= 0 is a fixed non-zero parameter. (The case λ = 0 is completely trivial. Why?)
In fact, an elementary change of variables can make any quadratic iterative system into
one involving a logistic map.
The fixed points of the logistic map are the solutions to the quadratic equation
u = λ u(1 − u), or λ u2 − λ u + 1 = 0.
Using the quadratic formula, we conclude that g(u) has two fixed points:
1
u⋆1 = 0, u⋆2 = 1 − .
λ
10/23/22 9 c 2022 Peter J. Olver
 1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

20 40 60 80 100 20 40 60 80 100 20 40 60 80 100

λ = 1.0 λ = 2.0 λ = 3.0

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

20 40 60 80 100 20 40 60 80 100 20 40 60 80 100

λ = 3.4 λ = 3.5 λ = 3.55

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

20 40 60 80 100 20 40 60 80 100 20 40 60 80 100

λ = 3.6 λ = 3.7 λ = 3.8

Figure 5. Logistic Iterates.

Let us apply Theorem 2.4 to determine their stability. The derivative is

g ′ (u) = λ − 2 λ u, and so g ′ (u⋆1 ) = λ, g ′ (u⋆2 ) = 2 − λ.

Therefore, if | λ | < 1, the first fixed point is stable, while if 1 < λ < 3, the second fixed
point is stable. For λ < −1 or λ > 3 neither fixed point is stable, and we expect the
iterates to not converge at all.
Numerical experiments with this example show that it is the source of an amazingly
diverse range of behavior, depending upon the value of the parameter λ. In the accompa-
nying Figure 5, we display the results of iteration starting with initial point u(0) = .5 for
several different values of λ; in each plot, the horizontal axis indicates the iterate number
k and the vertical axis the iterate value u(k) for k = 0, . . . , 100. As expected from Theo-
rem 2.4, the iterates converge to one of the fixed points in the range −1 < λ < 3, except
when λ = 1. For λ a little bit larger than λ1 = 3, the iterates do not converge to a fixed
point. But it does not take long for them to settle down, switching back and forth between
two particular values. This behavior indicates the existence of a (stable) period 2 orbit for
the discrete dynamical system, in accordance with the following definition.

10/23/22 10 c 2022 Peter J. Olver

 Definition 2.7. A period k orbit of a discrete dynamical system is a solution that
satisfies u(n+k) = u(n) for all n = 0, 1, 2, . . . . The (minimal) period is the smallest positive
value of k for which this condition holds.
Thus, a fixed point
u(0) = u(1) = u(2) = · · ·
is a period 1 orbit. A period 2 orbit satisfies
u(0) = u(2) = u(4) = · · · and u(1) = u(3) = u(5) = · · · ,
but u(0) 6= u(1) , as otherwise the minimal period would be 1. Similarly, a period 3 orbit
has
u(0) = u(3) = u(6) = · · · , u(1) = u(4) = u(7) = · · · , u(2) = u(5) = u(8) = · · · ,
with u(0) , u(1) , u(2) distinct. Stability of a period k orbit implies that nearby iterates
converge to this periodic solution.
For the logistic map, the period 2 orbit persists until λ = λ2 ≈ 3.4495, after which
the iterates alternate between four values — a period 4 orbit. This again changes at
λ = λ3 ≈ 3.5441, after which the iterates end up alternating between eight values. In fact,
there is an increasing sequence of values
3 = λ1 < λ2 < λ3 < λ4 < · · · ,
where, for any λn < λ ≤ λn+1 , the iterates eventually follow a period 2n orbit. Thus, as λ
passes through each value λn the period of the orbit goes from 2n to 2 · 2n = 2n+1 , and the
discrete dynamical system experiences a bifurcation. The bifurcation values λn are packed
closer and closer together as n increases, piling up on an eventual limiting value
λ⋆ = lim λn ≈ 3.5699,
n→∞
at which point the orbit’s period has, so to speak, become infinitely large. The entire
phenomena is known as a period doubling cascade.
Interestingly, the ratios of the distances between successive bifurcation points ap-
proaches a well-defined limit,
λn+2 − λn+1
−→ 4.6692 . . . , (2.20)
λn+1 − λn
known as Feigenbaum’s constant. In the 1970’s, the American physicist Mitchell Feigen-
baum, [6], discovered that similar period doubling cascades appear in a broad range of
discrete dynamical systems. Even more remarkably, in almost all cases, the corresponding
ratios of distances between bifurcation points has the same limiting value. Feigenbaum’s
experimental observations were rigorously proved by Oscar Lanford in 1982, [10].
After λ passes the limiting value λ⋆ , all hell breaks loose. The iterates become com-
pletely chaotic† , moving at random over the interval [ 0, 1 ]. But this is not the end of the

†
The term “chaotic” does have a precise mathematical definition, [ 5 ], but the reader can take
it more figuratively for the purposes of this elementary exposition.

10/23/22 11 c 2022 Peter J. Olver

 1

0.8

0.6

0.4

0.2

2.5 3 3.5 4

Figure 6. The Logistic Map.

story. Embedded within this chaotic regime are certain small ranges of λ where the system
settles down to a stable orbit, whose period is no longer necessarily a power of 2. In fact,
there exist values of λ for which the iterates settle down to a stable orbit of period k for any
positive integer k. For instance, as λ increases past λ3,⋆ ≈ 3.83, a period 3 orbit appears
over a small range of values, after which, as λ increses slightly further, there is a period
doubling cascade where period 6, 12, 24, . . . orbits successively appear, each persisting on
a shorter and shorter range of parameter values, until λ passes yet another critical value
where chaos breaks out yet again. There is a well-prescribed order in which the periodic
orbits make their successive appearance, and each odd period k orbit is followed by a very
closely spaced sequence of period doubling bifurcations, of periods 2n k for n = 1, 2, 3, . . . ,
after which the iterates revert to completely chaotic behavior until the next periodic case
emerges. The ratios of distances between bifurcation points always have the same Feigen-
baum limit (2.20). Finally, these periodic and chaotic windows all pile up on the ultimate
parameter value λ⋆⋆ = 4. And then, when λ > 4, all the iterates go off to ∞, and the
system ceases to be interesting.
The reader is encouraged to write a simple computer program and perform some
numerical experiments. In particular, Figure 6 shows the asymptotic behavior of the
iterates for values of the parameter in the interesting range 2 < λ < 4. The horizontal
axis is λ, and the marked points show the ultimate fate of the iteration for the given
value of λ. For instance, each point the single curve lying above the smaller values of
λ represents a stable fixed point; this bifurcates into a pair of curves representing stable
period 2 orbits, which then bifurcates into 4 curves representing period 4 orbits, and so
on. Chaotic behavior is indicated by a somewhat random pattern of points lying above the
value of λ. To plot this figure, we ran the logistic iteration u(n) for 0 ≤ n ≤ 100, discarded
the first 50 points, and then plotted the next 50 iterates u(51) , . . . , u(100) . Investigation of
the fine detailed structure of the logistic map requires yet more iterations with increased
numerical accuracy. In addition one should discard more of the initial iterates so as to give

10/23/22 12 c 2022 Peter J. Olver

the system enough time to settle down to a stable periodic orbit or, alternatively, continue
in a chaotic manner.
Remark : So far, we have only looked at real scalar iterative systems. Complex discrete
dynamical systems display yet more remarkable and fascinating behavior. The complex
version of the logistic iteration equation leads to the justly famous Julia and Mandelbrot
sets, [11], with their stunning, psychedelic fractal structure, [16].
The rich range of phenomena in evidence, even in such extremely simple nonlinear
iterative systems, is astounding. While intimations first appeared in the late nineteenth
century research of the influential French mathematician Henri Poincaré, serious investiga-
tions were delayed until the advent of the computer era, which precipitated an explosion of
research activity in the area of dynamical systems. Similar period doubling cascades and
chaos are found in a broad range of nonlinear systems, [1], and are often encountered in
physical applications, [13]. A modern explanation of fluid turbulence is that it is a (very
complicated) form of chaos, [1].
Quadratic Convergence
Let us now return to the more mundane case when the iterates converge to a stable
fixed point of the discrete dynamical system. In applications, we use the iterates to compute
a precise† numerical value for the fixed point, and hence the efficiency of the algorithm
depends on the speed of convergence of the iterates.
According to the remark following the proof Theorem 2.4, the convergence rate of an
iterative system is essentially governed by the magnitude of the derivative | g ′ (u⋆ ) | at the
fixed point. The basic inequality (2.15) for the errors e(k) = u(k) − u⋆ , namely
| e(k+1) | ≤ σ | e(k) |,
is known as a linear convergence estimate. It means that, once the iterates are close to
the fixed point, the error decreases by a factor of (at least) σ ≈ | g ′ (u⋆ ) | at each step. If
the k th iterate u(k) approximates the fixed point u⋆ correctly to m decimal places, so its
error is bounded by
| e(k) | < .5 × 10− m ,
then the (k + 1)st iterate satisfies the error bound
| e(k+1) | ≤ σ | e(k) | < .5 × 10− m σ = .5 × 10− m+log10 σ .
More generally, for any j > 0,
| e(k+j) | ≤ σ j | e(k) | < .5 × 10− m σ j = .5 × 10− m+j log10 σ ,
which means that the (k + j)th iterate u(k+j) has at least‡
m − j log10 σ = m + j log10 σ −1

†
The degree of precision is to be specified by the user and the application.
‡
Note that since σ < 1, the logarithm log10 σ −1 = − log10 σ > 0 is positive.

10/23/22 13 c 2022 Peter J. Olver

correct decimal places. For instance, if σ = .1 then each new iterate produces one new
decimal place of accuracy (at least), while if σ = .9 then it typically takes 22 ≈ −1/ log10 .9
iterates to produce just one additional accurate digit!
This means that there is a huge advantage — particularly in the application of iterative
methods to the numerical solution of equations — to arrange that | g ′ (u⋆ ) | be as small as
possible. The fastest convergence rate of all will occur when g ′ (u⋆ ) = 0. In fact, in such a
happy situation, the rate of convergence is not just slightly, but dramatically faster than
linear.

Theorem 2.8. Suppose that† g ∈ C2 , and u⋆ = g(u⋆ ) is a fixed point such that
g ′ (u⋆ ) = 0. Then, for all iterates u(k) sufficiently close to u⋆ , the errors e(k) = u(k) − u⋆
satisfy the quadratic convergence estimate

| e(k+1) | ≤ τ | e(k) |2 (2.21)

for some constant τ > 0.

Proof : Just as that of the linear convergence estimate (2.15), the proof relies on
approximating g(u) by a simpler function near the fixed point. For linear convergence, an
affine approximation sufficed, but here we require a higher order approximation. Thus, we
replace the mean value formula (2.13) by the first order Taylor expansion

g(u) = g(u⋆ ) + g ′ (u⋆ ) (u − u⋆ ) + 12 g ′′ (w) (u − u⋆ )2 , (2.22)

where the final error term depends on an (unknown) point w that lies between u and u⋆ .
At a fixed point, the constant term is g(u⋆ ) = u⋆ . Furthermore, under our hypothesis
g ′ (u⋆ ) = 0, and so (2.22) reduces to

g(u) − u⋆ = 1
2 g ′′ (w) (u − u⋆ )2 .

Therefore,
| g(u) − u⋆ | ≤ τ | u − u⋆ |2 , (2.23)

where τ is chosen so that

1
2 | g ′′ (w) | ≤ τ (2.24)

for all w sufficiently close to u⋆ . Therefore, the magnitude of τ is governed by the size
of the second derivative of the iterative function g(u) near the fixed point. We use the
inequality (2.23) to estimate the error

| e(k+1) | = | u(k+1) − u⋆ | = | g(u(k) ) − g(u⋆ ) | ≤ τ | u(k) − u⋆ |2 = τ | e(k) |2 ,

which establishes the quadratic convergence estimate (2.21). Q.E.D.

†
The notation means that g(u) is twice continuously differentiable, i.e., g, g ′ , g ′′ are all defined
and continuous near the fixed point u⋆ .

10/23/22 14 c 2022 Peter J. Olver

 Let us see how the quadratic estimate (2.21) speeds up the convergence rate. Following
our earlier argument, suppose u(k) is correct to m decimal places, so

| e(k) | < .5 × 10− m .

Then (2.21) implies that

| e(k+1) | < .5 × (10− m )2 τ = .5 × 10− 2 m+log10 τ ,

and so u(k+1) has 2 m − log10 τ accurate decimal places. If τ ≈ | g ′′ (u⋆ ) | is of moderate

size, we have essentially doubled the number of accurate decimal places in just a single
iterate! A second iteration will double the number of accurate digits yet again. Thus,
the convergence of a quadratic iteration scheme is extremely rapid, and, barring round-off
errors, one can produce any desired number of digits of accuracy in a very short time. For
example, if we start with an initial guess that is accurate in the first decimal digit, then a
linear iteration with σ = .1 will require 49 iterations to obtain 50 decimal place accuracy,
whereas a quadratic iteration (with τ = 1) will only require 6 iterations to obtain 26 = 64
decimal places of accuracy!

Example 2.9. Consider the function

2 u3 + 3
g(u) = .
3 u2 + 3
There is a unique (real) fixed point u⋆ = g(u⋆ ), which is the real solution to the cubic
equation
1 3
3
u + u − 1 = 0.

Note that

′ 2 u4 + 6 u2 − 6 u 6 u 31 u3 + u − 1
g (u) = = ,
3 (u2 + 1)2 3 (u2 + 1)2

and hence g ′ (u⋆ ) = 0 vanishes at the fixed point. Theorem 2.8 implies that the iterations
will exhibit quadratic convergence to the root. Indeed, we find, starting with u(0) = 0, the
following values:

k 1 2 3
(k)
u 1.00000000000000 .833333333333333 .817850637522769

4 5 6
.817731680821982 .817731673886824 .817731673886824
The convergence rate is dramatic: after only 5 iterations, we have produced the first 15
decimal places of the fixed point. In contrast, the linearly convergent scheme based on
ge(u) = 1 − 31 u3 takes 29 iterations just to produce the first 5 decimal places of the same
solution.

10/23/22 15 c 2022 Peter J. Olver

 In practice, the appearance of a quadratically convergent fixed point is a matter of
luck. The construction of quadratically convergent iterative methods for solving equations
will be the focus of the following Section 3.

Vector–Valued Iteration

Extending the preceding analysis to vector-valued iterative systems is not especially

difficult. We will build on our experience with linear iterative systems, and so the reader is
advised to review the basic concepts and results from Chapter 10 of [15] before proceeding
to the nonlinear systems presented here.
We begin by fixing a norm k · k on R n . Since we will also be computing the associated
matrix norm k A k, as defined in [15], it may be more convenient for computations to adopt
either the 1 or the ∞ norms rather than the standard Euclidean norm.
We begin by defining the vector-valued counterpart of the basic linear convergence
condition (2.18).

Definition 2.10. A function g: R n → R n is a contraction at a point u⋆ ∈ R n if

there exists a constant 0 ≤ σ < 1 such that

k g(u) − g(u⋆ ) k ≤ σ k u − u⋆ k (2.25)

for all u sufficiently close to u⋆ , i.e., k u − u⋆ k < δ for some fixed δ > 0.

Remark : The notion of a contraction depends on the underlying choice of matrix

norm. Indeed, the linear function g(u) = A u if and only if k A k < 1, which implies
that A is a convergent matrix. While every convergent matrix satisfies k A k < 1 in some
matrix norm, and hence defines a contraction relative to that norm, it may very well have
k A k > 1 in a particular norm, violating the contraction condition.

Theorem 2.11. If u⋆ = g(u⋆ ) is a fixed point for the discrete dynamical system
(2.1) and g is a contraction at u⋆ , then u⋆ is an asymptotically stable fixed point.

Proof : The proof is a copy of the last part of the proof of Theorem 2.4. We write

k u(k+1) − u⋆ k = k g(u(k) ) − g(u⋆ ) k ≤ σ k u(k) − u⋆ k,

using the assumed estimate (2.25). Iterating this basic inequality immediately demon-
strates that

k u(k) − u⋆ k ≤ σ k k u(0) − u⋆ k for k = 0, 1, 2, 3, . . . . (2.26)

Since σ < 1, the right hand side tends to 0 as k → ∞, and hence u(k) → u⋆ . Q.E.D.

In most interesting situations, the function g is differentiable, and so can be approxi-

mated by its first order Taylor polynomial:

g(u) ≈ g(u⋆ ) + g ′ (u⋆ ) (u − u⋆ ) = u⋆ + g ′ (u⋆ ) (u − u⋆ ). (2.27)

10/23/22 16 c 2022 Peter J. Olver

Here  
∂g1 ∂g1 ∂g1
...
 ∂u1 ∂u2 ∂un 
 
 
 ∂g2 ∂g2 ∂g2 
 ... 
 ∂u ∂u2 ∂un 
 1 
g ′ (u) =  , (2.28)
 . .. .. .. 
 .. . . . 
 
 
 ∂g ∂gn ∂gn 
 n
... 
∂u1 ∂u2 ∂un

denotes the n × n Jacobian matrix of the vector-valued function g, whose entries are the
partial derivatives of its individual components. Since u⋆ is fixed, the the right hand
side of (2.27) is an affine function of u. Moreover, u⋆ remains a fixed point of the affine
approximation. Proposition 10.44 of [15] tells us that iteration of the affine function
will converge to the fixed point if and only if its coefficient matrix, namely g ′ (u⋆ ), is
a convergent matrix, meaning that its spectral radius ρ(g ′ (u⋆ )) < 1. This observation
motivates the following theorem and corollary.
Theorem 2.12. Let u⋆ be a fixed point for the discrete dynamical system u(k+1) =
g(u ). If the Jacobian matrix norm k g ′ (u⋆ ) k < 1, then g is a contraction at u⋆ , and
(k)

hence the fixed point u⋆ is asymptotically stable.

Proof : The first order Taylor expansion of g(u) at the fixed point u⋆ takes the form

g(u) = g(u⋆ ) + g ′ (u⋆ ) (u − u⋆ ) + R(u − u⋆ ), (2.29)

where the remainder term satisfies
R(u − u⋆ )
lim ⋆ = 0.
u→u k u − u⋆ k
Let ε > 0 be such that
σ = k g ′ (u⋆ ) k + ε < 1.
Choose 0 < δ < 1 such that k R(u − u⋆ ) k ≤ ε k u − u⋆ k whenever k u − u⋆ k ≤ δ. For
such u, we have, by the Triangle Inequality,

k g(u) − g(u⋆ ) k ≤ k g ′ (u⋆ ) (u − u⋆ ) k + k R(u − u⋆ ) k

≤ k g ′ (u⋆ ) k + ε k u − u⋆ k = σ k u − u⋆ k,
which establishes the contraction inequality (2.25). Q.E.D.
Corollary 2.13. If the Jacobian
matrix g ′ (u⋆ ) is a convergent matrix, meaning that
its spectral radius satisfies ρ g ′ (u⋆ ) < 1, then u⋆ is an asymptotically stable fixed point.
Proof : Corollary 10.32 in [15] assures us that k g ′ (u⋆ ) k < 1 in some matrix norm.
Using this norm, the result immediately follows from the theorem. Q.E.D.

10/23/22 17 c 2022 Peter J. Olver

 Theorem 2.12 tells us that initial values u(0) that are sufficiently near a stable fixed
point u⋆ are guaranteed to converge to it. In the linear case, closeness of the initial data
to the fixed point was not, in fact, an issue; all stable fixed points are, in fact, globally
stable. For nonlinear iteration, it is of critical importance, and one does not typically
expect iteration starting with far away initial data to converge to the desired fixed point.
An interesting (and difficult) problem is to determine the so-called basin of attraction of a
stable fixed point, defined as the set of all initial data that ends up converging to it. As in
the elementary logistic map (2.19), initial values that lie outside a basin of attraction can
lead to divergent iterates, periodic orbits, or even exhibit chaotic behavior. The full range
of possible phenomena is a topic of contemporary research in dynamical systems theory,
[1, 5, 16, 18], and in numerical analysis, [3, 4].
Alternatively, a fixed point is unstable when the spectral radius of its Jacobian matrix
is strictly greater than one, meaning that it has at least one eigenvalue of modulus | λ | > 1.
More precisely, the following result holds, cf. [9; Theorem 5.1.5].
Theorem 2.14. Let u⋆ be a fixed point for the discrete
dynamical system u(k+1) =
g(u(k) ). If the Jacobian matrix has spectral radius ρ g ′ (u⋆ ) > 1, then u⋆ is an unstable
fixed point, in the sense that there exists r > 0 with the property that one can find initial
points u(0) lying arbitrarily close to the fixed point u⋆ that produce some eventual iterate
u(k) that lies a distance at least r away from u⋆ ; that is, k u(k) − u⋆ k ≥ r for some k ≥ 1,
depending upon the initial point u(0) .
In fact, most initial points near u⋆ exhibit such instability. However, keep in mind
that, while the theorem states that, for suitable initial data, some subsequent iterate will
lie outside the ball of radius r centered at u⋆ , it is not claimed that this is necessarily true
for all subsequent iterates, since it is possible that later iterates re-enter the ball and even
(albeit highly unlikely), if they happen to then land on the so-called stable manifold † of
the fixed point, eventually converging to it.
Example 2.15. Consider the function
 1 3 9 
− 4 u + 8 u + 41 v 3
g(u, v) = 3 1 .
4 v − 2 uv

There are four (real) fixed points; stability is determined by the size of the eigenvalues of
the Jacobian matrix !
9 3 2 1
− u − v
g′ (u, v) = 8 4 2
3 2 3
4
v 4
− 12 u

†
For linear iterative systems, the stable manifold of the origin coincides with the stable sub-
space spanned by the real and imaginary parts of eigenvectors corresponding to the stable eigen-
values of modulus | λ | < 1. In the nonlinear case, the stable manifold at u⋆ is tangent to the
stable subspace of its Jacobian matrix. Practically detecting such convergent points whose iter-
ates eventually lie on the stable manifold is rather challenging, since (almost) any small numerical
error can dislodge the iterate off the stable submanifold, causing it to eventually go away from
the fixed point again.

10/23/22 18 c 2022 Peter J. Olver

 Figure 7. Basins of Attraction.

at each of the fixed points. The results are summarized in the following table:

  ! ! !
0 √1 − √12 − 21
⋆
fixed point u1 = u⋆2 = 2 u⋆3 = u⋆4 =
0 0 0 1
2

! ! ! !
9 3 3 15
8 0 4 0 4 0 16 − 14
Jacobian matrix 3 1 3 1
3 0 − 0 + 3
0 4 4
√
2 2 4
√
2 2 16
1

eigenvalues 1.125, .75 .75, .396447 1.10355, .75 .96875 ± .214239 i

spectral radius 1.125 .75 1.10355 .992157

Thus, u⋆2 and u⋆4 are stable fixed points, whereas u⋆1 and u⋆3 are both unstable. Indeed,
T
starting with u(0) = ( .5, .5 ) , it takes 24 iterates to converge to u⋆2 with 4 significant
T
decimal digits, whereas starting with u(0) = ( −.7, .7 ) , it takes 1049 iterates to converge
to within 4 digits of u⋆4 ; the slower convergence rate is predicted by the larger Jacobian
spectral radius. The two basins of attraction are plotted in Figure 7. The stable fixed
points are indicated by black dots. The light gray region contains u⋆2 and indicates all the
points that converge to it; the darker gray indicates points converging, more slowly, to u⋆4 .
All other initial points, except u⋆1 and u⋆3 , have rapidly unbounded iterates: k u(k) k → ∞.
The smaller the spectral radius or matrix norm of the Jacobian matrix at the fixed
point, the faster the nearby iterates will converge to it. As in the scalar case, quadratic
convergence will occur when the Jacobian matrix g ′ (u⋆ ) = O is the zero matrix, i.e.,
all first order partial derivatives of the components of g vanish at the fixed point. The
quadratic convergence estimate
k u(k+1) − u⋆ k ≤ τ k u(k) − u⋆ k2 (2.30)

10/23/22 19 c 2022 Peter J. Olver

 g

Ω Ω

Figure 8. A Contraction Mapping.

is a consequence of the second order Taylor expansion at the fixed point. Details of the
proof are left as an exercise.
Of course, in practice we don’t know the norm or spectral radius of the Jacobian
matrix g ′ (u⋆ ) because we don’t know where the fixed point is. This apparent difficulty
can be easily circumvented by requiring that k g ′ (u) k < 1 for all u — or, at least, for all
u in a domain Ω containing the fixed point. In fact, this hypothesis can be used to prove
the existence and uniqueness of asymptotically stable fixed points. Rather than work with
the Jacobian matrix, let us return to the contraction condition (2.25), but now imposed
uniformly on an entire domain.
Definition 2.16. A function g: R n → R n is called a contraction mapping on a
domain Ω ⊂ R n if
(a) it maps Ω to itself, so g(u) ∈ Ω whenever u ∈ Ω, and
(b) there exists a constant 0 ≤ σ < 1 such that

k g(u) − g(v) k ≤ σ k u − v k for all u, v ∈ Ω. (2.31)

In other words, applying a contraction mapping reduces the mutual distance between
points. In view of Theorem 2.12, we can detect contraction mappings by looking at the
norm of their Jacobian matrix.
Lemma 2.17. If g: Ω → Ω and k g ′ (u) k < 1 for all u ∈ Ω, then g is a contraction
mapping.
So, as its name indicates, a contraction mapping effectively shrinks the size of its
domain; see Figure 8. As the iterations proceed, the successive image domains become
smaller and smaller. If the original domain is closed and bounded, then it is forced to
shrink down to a single point, which is the unique fixed point of the iterative system,
leading to the Contraction Mapping Theorem.
Theorem 2.18. If g is a contraction mapping on a closed bounded domain Ω ⊂ R n ,
then g admits a unique fixed point u⋆ ∈ Ω. Moreover, starting with any initial point
u(0) ∈ Ω, the iterates u(k+1) = g(u(k) ) necessarily converge to the fixed point: u(k) → u⋆ .
Proof : The basic contraction estimate (2.31) implies that, for any positive integers
k, l,
k u(k+1) − u(l+1) k = k g(u(k) ) − g(u(l) ) k ≤ σ k u(k) − u(l) k.

10/23/22 20 c 2022 Peter J. Olver

Iterating, we find

k u(k+n) − u(l+n) k ≤ σ n k u(k) − u(l) k ≤ σ n D,

where
D = max { k u − v k | u, v ∈ Ω } < ∞,
the diameter of Ω, is bounded since Ω is a bounded domain. Thus, to prove u(k) forms
a Cauchy sequence, given ε > 0, choose N such that σ N D < ε, which is possible since
0 ≤ σ < 1. Then the preceding estimate show that

k u(k) − u(l) k ≤ σ N k u(k−N) − u(l−N) k ≤ σ N D < ε,

whenever k, l ≥ N , proving the result. Since we know u(k) → u⋆ converges, the usual
argument proves that u⋆ is a fixed point.
To prove uniqueness of the fixed point, suppose that there are two fixed points, so
g(u⋆ ) = u⋆ and g(e
u⋆ ) = u
e ⋆ . Then, by (2.25)

e ⋆ − u⋆ k = k g(e
0 ≤ ku u⋆ ) − g(u⋆ ) k ≤ σ k u
e ⋆ − u⋆ k < k u
e ⋆ − u⋆ k,
e ⋆ − u⋆ k = 0 and hence u
which implies k u e ⋆ = u⋆ , proving the result. Q.E.D.
Example 2.19. The function

1
g(u) = u + 12 π − tan−1 u satisfies | g ′ (u) | = 1 − <1
1 + u2
for all u ∈ R, and hence defines a contraction mapping. However, g(u) has no fixed point.
Why does this not contradict Theorem 2.18?

3. Solution of Equations and Systems.

Solving nonlinear equations and systems of equations is, of course, a problem of utmost
importance in mathematics and its manifold applications. It can also be extremely difficult.
Indeed, finding a complete set of (numerical) solutions to a complicated nonlinear system
can be an almost insurmountable challenge. In its most general version, we are given a
collection of m functions f1 , . . . , fm depending upon n variables u1 , . . . , un , and are asked
T
to determine all possible solutions u = ( u1 , u2 , . . . , un ) to the system

f1 (u1 , . . . , un ) = 0, ... fm (u1 , . . . , un ) = 0. (3.1)

In many applications, the number of equations equals the number of unknowns, m = n,
in which case one expects both existence and uniqueness of solutions. This point will be
discussed in further detail below.
Here are some prototypical examples:
(a) Find the roots of the quintic polynomial equation

u5 + u + 1 = 0. (3.2)

10/23/22 21 c 2022 Peter J. Olver

 3

-1 -0.5 0.5 1

-1

-2

-3

-4

Figure 9. Graph of u5 + u + 1.

Graphing the left hand side of the equation, as in Figure 9, convinces us that there is just
one real root, lying somewhere between −1 and −.5. While there are explicit algebraic
formulas for the roots of quadratic, cubic, and quartic polynomials, a famous theorem† due
to the Norwegian mathematician Nils Henrik Abel in the early 1800’s states that there is
no such formula for generic fifth order polynomial equations.
(b) Any fixed point equation u = g(u) has the form (3.4) where f (u) = u − g(u). For
example, the trigonometric Kepler equation

u − ǫ sin u = m
arises in the study of planetary motion, cf. Example 2.5. Here ǫ, m are fixed constants,
and we seek a corresponding solution u.
(c) Suppose we are given chemical compounds A, B, C that react to produce a fourth
compound D according to

2 A + B ←→ D, A + 3 C ←→ D.
Let a, b, c be the initial concentrations of the reagents A, B, C injected into the reaction
chamber. If u denotes the concentration of D produced by the first reaction, and v that
by the second reaction, then the final equilibrium concentrations

a⋆ = a − 2 u − v, b⋆ = b − u, c⋆ = c − 3 v, d⋆ = u + v,
of the reagents will be determined by solving the nonlinear system

(a − 2 u − v)2 (b − u) = α (u + v), (a − 2 u − v)(c − 3 v)3 = β (u + v), (3.3)

where α, β are the known equilibrium constants of the two reactions.
Our immediate goal is to develop numerical algorithms for solving such nonlinear
equations. Unfortunately, there is no direct universal solution method for nonlinear sys-
tems comparable to Gaussian elimination. As a result, numerical solution techniques rely
almost exclusively on iterative algorithms. This section presents the principal methods for

†
A modern proof of this fact relies on Galois theory, [ 7 ].

10/23/22 22 c 2022 Peter J. Olver

 f (u)

a b
u⋆

Figure 10. Intermediate Value Theorem.

numerically approximating the solution(s) to a nonlinear system. We shall only discuss

general purpose algorithms; specialized methods for solving particular classes of equations,
e.g., polynomial equations, can be found in numerical analysis texts, e.g., [3, 4, 17]. Of
course, the most important specialized methods — those designed for solving linear systems
— will continue to play a critical role, even in the nonlinear regime.
The Bisection Method
We begin, as always, with the scalar case. Thus, we are given a real-valued function
f : R → R, and seek its roots, i.e., the real† solution(s) to the scalar equation

f (u) = 0. (3.4)
Our immediate goal is to develop numerical algorithms for solving such nonlinear scalar
equations. The most primitive algorithm, and the only one that is guaranteed to work in
all cases, is the Bisection Method. While it has an iterative flavor, it cannot be properly
classed as a method governed by functional iteration as defined in the preceding section,
and so must be studied directly in its own right.
The starting point is the Intermediate Value Theorem, which we state in simplified
form. See Figure 10 for an illustration, and [14] for a proof.
Lemma 3.1. Let f (u) be a continuous scalar function. Suppose we can find two
points a < b where the values of f (a) and f (b) take opposite signs, so either f (a) < 0 and
f (b) > 0, or f (a) > 0 and f (b) < 0. Then there exists at least one point a < u⋆ < b where
f (u⋆ ) = 0.
Note that if f (a) = 0 or f (b) = 0, then finding a root is trivial. If f (a) and f (b)
have the same sign, then there may or may not be a root in between. Figure 11 plots
the functions u2 + 1, u2 and u2 − 1, on the interval −2 ≤ u ≤ 2. The first has two simple
roots; the second has a single double root, while the third has no root. We also note

†
Complex roots to complex equations will be discussed later.

10/23/22 23 c 2022 Peter J. Olver

 5 5 5

4 4 4

3 3 3

2 2 2

1 1 1

-2 -1 1 2 -2 -1 1 2 -2 -1 1 2

-1 -1 -1

-2 -2 -2

Figure 11. Roots of Quadratic Functions.

that continuity of the function on the entire interval [ a, b ] is an essential hypothesis. For
example, the function f (u) = 1/u satisfies f (−1) = −1 and f (1) = 1, but there is no root
to the equation 1/u = 0.
Note carefully that the Lemma 3.1 does not say there is a unique root between a and
b. There may be many roots, or even, in pathological examples, infinitely many. All the
theorem guarantees is that, under the stated hypotheses, there is at least one root.
Once we are assured that a root exists, bisection relies on a “divide and conquer”
strategy. The goal is to locate a root a < u⋆ < b between the endpoints. Lacking any
additional evidence, one tactic would be to try the midpoint c = 21 (a + b) as a first guess
for the root. If, by some miracle, f (c) = 0, then we are done, since we have found a
solution! Otherwise (and typically) we look at the sign of f (c). There are two possibilities.
If f (a) and f (c) are of opposite signs, then the Intermediate Value Theorem tells us that
there is a root u⋆ lying between a < u⋆ < c. Otherwise, f (c) and f (b) must have opposite
signs, and so there is a root c < u⋆ < b. In either event, we apply the same method to
the interval in which we are assured a root lies, and repeat the procedure. Each iteration
halves the length of the interval, and chooses the half in which a root is sure to lie. (There
may, of course, be a root in the other half interval, but as we cannot be sure, we discard
it from further consideration.) The root we home in on lies trapped in intervals of smaller
and smaller width, and so convergence of the method is guaranteed.
Example 3.2. The roots of the quadratic equation

f (u) = u2 + u − 3 = 0
can be computed exactly by the quadratic formula:
√ √
⋆ − 1 + 13 ⋆ − 1 − 13
u1 = ≈ 1.302775 . . . , u2 = ≈ − 2.302775 . . . .
2 2
Let us see how one might approximate them by applying the Bisection Algorithm. We start
the procedure by choosing the points a = u(0) = 1, b = v (0) = 2, noting that f (1) = −1
and f (2) = 3 have opposite signs and hence we are guaranteed that there is at least one
root between 1 and 2. In the first step we look at the midpoint of the interval [ 1, 2 ],
which is 1.5, and evaluate f (1.5) = .75. Since f (1) = −1 and f (1.5) = .75 have opposite
signs, we know that there is a root lying between 1 and 1.5. Thus, we take u(1) = 1 and
v (1) = 1.5 as the endpoints of the next interval, and continue. The next midpoint is at

10/23/22 24 c 2022 Peter J. Olver

 k u(k) v (k) w(k) = 12 (u(k) + v (k) ) f (w(k) )

0 1 2 1.5 .75
1 1 1.5 1.25 −.1875
2 1.25 1.5 1.375 .2656
3 1.25 1.375 1.3125 .0352
4 1.25 1.3125 1.2813 −.0771
5 1.2813 1.3125 1.2969 −.0212
6 1.2969 1.3125 1.3047 .0069
7 1.2969 1.3047 1.3008 −.0072
8 1.3008 1.3047 1.3027 −.0002
9 1.3027 1.3047 1.3037 .0034
10 1.3027 1.3037 1.3032 .0016
11 1.3027 1.3032 1.3030 .0007
12 1.3027 1.3030 1.3029 .0003
13 1.3027 1.3029 1.3028 .0001
14 1.3027 1.3028 1.3028 −.0000

1.25, where f (1.25) = −.1875 has the opposite sign to f (1.5) = .75, and so a root lies
between u(2) = 1.25 and v (2) = 1.5. The process is then iterated as long as desired — or,
more practically, as long as your computer’s precision does not become an issue.
The table displays the result of the algorithm, rounded off to four decimal places.
After 14 iterations, the Bisection Method has correctly computed the first four decimal
digits of the positive root u⋆1 . A similar bisection starting with the interval from u(1) = − 3
to v (1) = − 2 will produce the negative root.
A formal implementation of the Bisection Algorithm appears in the accompanying
pseudocode program. The endpoints
of the k th interval are denoted by u(k) and v (k) . The
midpoint is w(k) = 21 u(k) + v (k) , and the key decision is whether w(k) should be the
right or left hand endpoint of the next interval. The integer n, governing the number of
iterations, is to be prescribed in accordance with how accurately we wish to approximate
the root u⋆ .
The algorithm produces two sequences of approximations u(k) and v (k) that both
converge monotonically to u⋆ , one from below and the other from above:

a = u(0) ≤ u(1) ≤ u(2) ≤ · · · ≤ u(k) −→ u⋆ ←− v (k) ≤ · · · ≤ v (2) ≤ v (1) ≤ v (0) = b.

and u⋆ is trapped between the two. Thus, the root is trapped inside a sequence of intervals
[ u(k) , v (k) ] of progressively shorter and shorter length. Indeed, the length of each interval
is exactly half that of its predecessor:

v (k) − u(k) = 1
2 (v (k−1) − u(k−1) ).

10/23/22 25 c 2022 Peter J. Olver

 The Bisection Method

start
if f (a) f (b) < 0 set u(0) = a, v (0) = b
else print “Bisection Method not applicable”
for k = 0 to n − 1
set w(k) = 21 (u(k) + v (k) )
if f (w(k) ) = 0, stop; print u⋆ = w(k)
if f (u(k) ) f (w(k) ) < 0, set u(k+1) = u(k) , v (k+1) = w(k)
else set u(k+1) = w(k) , v (k+1) = v (k)
next k
print u⋆ = w(n) = 21 (u(n) + v (n) )
end

Iterating this formula, we conclude that

n

1 n
v (n) − u(n) = 12 (v (0) − u(0) ) = 2 (b − a) −→ 0 as n −→ ∞.

The midpoint
w(n) = 1
2 (u(n) + v (n) )

lies within a distance

1 n+1
| w(n) − u⋆ | ≤ 1
2
(v (n) − u(n) ) = 2
(b − a)

of the root. Consequently, if we desire to approximate the root within a prescribed tolerance
ε, we should choose the number of iterations n so that


1 n+1 b−a
2
(b − a) < ε, or n > log2 − 1. (3.5)
ε
Summarizing:

Theorem 3.3. If f (u) is a continuous function, with f (a) f (b) < 0, then the Bisec-
tion Method starting with u(0) = a, v (0) = b, will converge to a solution u⋆ to the equation
f (u) = 0 lying between a and b. After n steps, the midpoint w(n) = 12 (u(n) + v (n) ) will be
within a distance of ε = 2− n−1 (b − a) from the solution.

For example, in the case of the quadratic equation in Example 3.2, after 14 iterations,
we have approximated the positive root to within


1 15
ε= 2 (2 − 1) ≈ 3.052 × 10−5 ,

10/23/22 26 c 2022 Peter J. Olver

reconfirming our observation that we have accurately computed its first four decimal places.
If we are in need of 10 decimal places, we set our tolerance to ε = .5 × 10−10 , and so,
according to (3.5), must perform n = 34 > 33.22 ≈ log2 2 × 1010 − 1 successive bisections† .
Example 3.4. As noted at the beginning of this section, the quintic equation

f (u) = u5 + u + 1 = 0
has one real root, whose value can be readily computed by bisection. We start the algorithm
with the initial points u(0) = − 1, v (0) = 0, noting that f (− 1) = − 1 < 0 while f (0) = 1 > 0
are of opposite signs. In order to compute the root to 6 decimal places, we set ε = .5×10−6
in (3.5), and so need to perform n = 20 > 19.93 ≈ log2 2 × 106 − 1 bisections. Indeed,
the algorithm produces the approximation u⋆ ≈ − .754878 to the root, and the displayed
digits are guaranteed to be accurate.

Fixed Point Methods

The Bisection Method has an ironclad guarantee to converge to a root of the function
— provided it can be properly started by locating two points where the function takes
opposite signs. This may be tricky if the function has two very closely spaced roots and
is, say, negative only for a very small interval between them, and may be impossible
for multiple roots, e.g., the root u⋆ = 0 of the quadratic function f (u) = u2 . When
applicable, its convergence rate is completely predictable, but not especially fast. Worse,
it has no immediately apparent extension to systems of equations, since there is no obvious
counterpart to the Intermediate Value Theorem for vector-valued functions.
Most other numerical schemes for solving equations rely on some form of fixed point
iteration. Thus, we seek to replace the system of equations f (u) = 0 with a fixed point
system u = g(u), that leads to the iterative solution scheme u(k+1) = g(u(k) ). For this to
work, there are two key requirements:
(a) The solution u⋆ to the equation f (u) = 0 is also a fixed point for g(u), and
(b) u⋆ is, in fact a stable fixed point, meaning that the Jacobian g ′ (u⋆ ) is a convergent
matrix, or, slightly more restrictively, k g ′ (u⋆ ) k < 1 for a prescribed matrix norm.
If both conditions hold, then, provided we choose the initial iterate u(0) = c sufficiently
close to u⋆ , the iterates u(k) → u⋆ will converge to the desired solution. Thus, the key
to the practical use of functional iteration for solving equations is the proper design of an
iterative system — coupled with a reasonably good initial guess for the solution. Before
implementing general procedures, let us discuss an elementary example.
Example 3.5. To solve the cubic equation

f (u) = u3 − u − 1 = 0 (3.6)
we note that f (1) = −1 while f (2) = 5, and so there is a root between 1 and 2. Indeed,
the Bisection Method leads to the approximate value u⋆ ≈ 1.3247 after 17 iterations.

†
This assumes we have sufficient precision on the computer to avoid round-off errors.

10/23/22 27 c 2022 Peter J. Olver

 Let us try to find the same root by fixed point iteration. As a first, naı̈ve, guess, we
rewrite the cubic equation in fixed point form

u = u3 − 1 = ge(u).
Starting with the initial guess u(0) = 1.5, successive approximations to the solution are
found by iterating

u(k+1) = e
g (u(k) ) = (u(k) )3 − 1, k = 0, 1, 2, . . . .
However, their values

u(0) = 1.5, u(1) = 2.375, u(2) = 12.396,

u(3) = 1904, u(4) = 6.9024 × 109 , u(5) = 3.2886 × 1029 , ...
rapidly become unbounded, and so fail to converge. This could, in fact, have been predicted
by the convergence criterion in Theorem 2.4. Indeed, e g ′ (u) = − 3 u2 and so | e
g ′ (u) | > 3
for all u ≥ 1, including the root u⋆ . This means that u⋆ is an unstable fixed point, and
the iterates cannot converge to it.
On the other hand, we can rewrite the equation (3.6) in the alternative iterative form
√
u = 3 1 + u = g(u).
In this case
1 1
0 ≤ g ′ (u) = ≤ for u > 0.
3(1 + u)2/3 3
Thus, the stability condition (2.14) is satisfied, and we anticipate convergence at a rate of
at least 31 . (The Bisection Method converges more slowly, at rate 12 .) Indeed, the first few
√
iterates u(k+1) = 1 + u(k) are
3

1.5, 1.35721, 1.33086, 1.32588, 1.32494, 1.32476, 1.32473,

and we have converged to the root, correct to four decimal places, in only 6 iterations.

Newton’s Method
Our immediate goal is to design an efficient iterative scheme u(k+1) = g(u(k) ) whose
iterates converge rapidly to the solution of the given scalar equation f (u) = 0. As we
learned in Section 2, the convergence of the iteration is governed by the magnitude of its
derivative at the fixed point. At the very least, we should impose the stability criterion
| g ′ (u⋆ ) | < 1, and the smaller this quantity can be made, the faster the iterative scheme
converges. if we are able to arrange that g ′ (u⋆ ) = 0, then the iterates will converge
quadratically fast, leading, as noted in the discussion following Theorem 2.8, to a dramatic
improvement in speed and efficiency.
Now, the first condition requires that g(u) = u whenever f (u) = 0. A little thought
will convince you that the iterative function should take the form

g(u) = u − h(u) f (u), (3.7)

10/23/22 28 c 2022 Peter J. Olver

where h(u) is a reasonably nice function. If f (u⋆ ) = 0, then clearly u⋆ = g(u⋆ ), and so u⋆
is a fixed point. The converse holds provided h(u) 6= 0 is never zero.
For quadratic convergence, the key requirement is that the derivative of g(u) be zero
at the fixed point solutions. We compute

g ′ (u) = 1 − h′ (u) f (u) − h(u) f ′ (u).

Thus, g ′ (u⋆ ) = 0 at a solution to f (u⋆ ) = 0 if and only if

0 = 1 − h′ (u⋆ ) f (u⋆ ) − h(u⋆ ) f ′ (u⋆ ) = 1 − h(u⋆ ) f ′ (u⋆ ).

Consequently, we should require that

1
h(u⋆ ) = (3.8)
f ′ (u⋆ )
to ensure a quadratically convergent iterative scheme. This assumes that f ′ (u⋆ ) 6= 0,
which means that u⋆ is a simple root of f . For here on, we leave aside multiple roots,
which require a different approach.
Of course, there are many functions h(u) that satisfy (3.8), since we only need to
specify its value at a single point. The problem is that we do not know u⋆ — after all this
is what we are trying to compute — and so cannot compute the value of the derivative
of f there. However, we can circumvent this apparent difficulty by a simple device: we
impose equation (3.8) at all points, setting

1
h(u) = , (3.9)
f ′ (u)
which certainly guarantees that it holds at the solution u⋆ . The result is the function

f (u)
g(u) = u − , (3.10)
f ′ (u)
and the resulting iteration scheme is known as Newton’s Method , which, as the name
suggests, dates back to the founder of the calculus. To this day, Newton’s Method remains
the most important general purpose algorithm for solving equations. It starts with an
initial guess u(0) to be supplied by the user, and then successively computes

f (u(k) )
u(k+1) = u(k) − . (3.11)
f ′ (u(k) )

As long as the initial guess is sufficiently close, the iterates u(k) are guaranteed to converge,
quadratically fast, to the (simple) root u⋆ of the equation f (u) = 0.

Theorem 3.6. Suppose f (u) ∈ C2 is twice continuously differentiable. Let u⋆ be

a solution to the equation f (u⋆ ) = 0 such that f ′ (u⋆ ) 6= 0. Given an initial guess u(0)
sufficiently close to u⋆ , the Newton iteration scheme (3.11) converges at a quadratic rate
to the solution u⋆ .

10/23/22 29 c 2022 Peter J. Olver

 Proof : By continuity, if f ′ (u⋆ ) 6= 0, then f ′ (u) 6= 0 for all u sufficiently close to u⋆ , and
hence the Newton iterative function (3.10) is well defined and continuously differentiable
near u⋆ . Since g ′ (u) = f (u) f ′′ (u)/f ′ (u)2 , we have g ′ (u⋆ ) = 0 when f (u⋆ ) = 0, as promised
by our construction. The quadratic convergence of the resulting iterative scheme is an
immediate consequence of Theorem 2.8. Q.E.D.
Example 3.7. Consider the cubic equation
f (u) = u3 − u − 1 = 0,
that we already solved in Example 3.5. The function used in the Newton iteration is
f (u) u3 − u − 1
g(u) = u − = u − ,
f ′ (u) 3 u2 − 1
which is well-defined as long as u 6= ± √13 . We will try to avoid these singular points. The
iterative procedure
(u(k) )3 − u(k) − 1
u(k+1) = g(u(k) ) = u(k) −
3 (u(k) )2 − 1
with initial guess u(0) = 1.5 produces the following values:
1.5, 1.34783, 1.32520, 1.32472,
and we have computed the root to 5 decimal places after only three iterations. The
quadratic convergence of Newton’s Method implies that, roughly, each new iterate doubles
the number of correct decimal places. Thus, to compute the root accurately to 40 decimal
places would only require 3 further iterations† . This underscores the tremendous advantage
that the Newton algorithm offers over competing methods.
Example 3.8. Consider the cubic polynomial equation
f (u) = u3 − 32 u2 + 95 u − 1
27
= 0.
Since
1 1

1 2

1 1
f (0) = − 27 , f 3 = 54 , f 3 = − 27 , f (1) = 54 ,
the Intermediate Value Lemma 3.1 guarantees that there are three roots on the interval
[ 0, 1 ]: one between 0 and 31 , the second between 13 and 23 , and the third between 32 and
1. The graph in Figure 12 reconfirms this observation. Since we are dealing with a cubic
polynomial, there are no other roots. (Why?)
 1It takes
 1 2sixteen
 iterations
2
 of the Bisection Method starting with the three subintervals
0 , 3 , 3 , 3 and 3 , 1 , to produce the roots to six decimal places:
u⋆1 ≈ .085119, u⋆2 ≈ .451805, u⋆3 ≈ .963076.

†
This assumes we are working in a sufficiently high precision arithmetic so as to avoid round-off
errors.

10/23/22 30 c 2022 Peter J. Olver

 0.1

0.05

0.2 0.4 0.6 0.8 1

-0.05

-0.1

-0.15

-0.2

-0.25

Figure 12. The function f (u) = u3 − 32 u2 + 95 u − 1

27 .

Incidentally, if we start with the interval [ 0, 1 ] and apply bisection, we converge (perhaps
surprisingly) to the largest root u⋆3 in 17 iterations.
Fixed point iteration based on the formulation

u = g(u) = − u3 + 23 u2 + 94 u + 1
27

can be used to find the first and third roots, but not the second root. For instance, starting
with u(0) = 0 produces u⋆1 to 5 decimal places after 23 iterations, whereas starting with
u(0) = 1 produces u⋆3 to 5 decimal places after 14 iterations. The reason we cannot produce
u⋆2 is due to the magnitude of the derivative

g ′ (u) = − 3 u2 + 3 u + 4
9

at the roots, which is

g ′ (u⋆1 ) ≈ 0.678065, g ′ (u⋆2 ) ≈ 1.18748, g ′ (u⋆3 ) ≈ 0.551126.

Thus, u⋆1 and u⋆3 are stable fixed points, but u⋆2 is unstable. However, because g ′ (u⋆1 ) and
g ′ (u⋆3 ) are both bigger than .5, this iterative algorithm actually converges slower than
ordinary bisection!
Finally, Newton’s Method is based upon iteration of the rational function

f (u) u3 − 32 u2 + 95 u − 1
27
g(u) = u − ′ =u− .
f (u) 3 u2 − 3 u + 59

Starting with an initial guess of u(0) = 0, the method computes u⋆1 to 6 decimal places
after only 4 iterations; starting with u(0) = .5, it produces u⋆2 to similar accuracy after 2
iterations; while starting with u(0) = 1 produces u⋆3 after 3 iterations — a dramatic speed
up over the other two methods.

Newton’s Method has a very pretty graphical interpretation, that helps us understand
what is going on and why it converges so fast. Given the equation f (u) = 0, suppose we

10/23/22 31 c 2022 Peter J. Olver

 f (u)

u(k+1) u(k)

Figure 13. Newton’s Method.

know an approximate value u = u(k) for a solution. Nearby u(k) , we can approximate the
nonlinear function f (u) by its tangent line

y = f (u(k) ) + f ′ (u(k) )(u − u(k) ). (3.12)

As long as the tangent line is not horizontal — which requires f ′ (u(k) ) 6= 0 — it crosses
the axis at
f (u(k) )
u(k+1) = u(k) − ′ (k) ,
f (u )
which represents a new, and, presumably more accurate, approximation to the desired
root. The procedure is illustrated pictorially in Figure 13. Note that the passage from
u(k) to u(k+1) is exactly the Newton iteration step (3.11). Thus, Newtonian iteration is
the same as the approximation of function’s root by those of its successive tangent lines.
Given a sufficiently accurate initial guess, Newton’s Method will rapidly produce
highly accurate values for the simple roots to the equation in question. In practice, barring
some kind of special exploitable structure, Newton’s Method is the root-finding algorithm
of choice. The one caveat is that we need to start the process reasonably close to the
root we are seeking. Otherwise, there is no guarantee that a particular set of iterates will
converge, although if they do, the limiting value is necessarily a root of our equation. The
behavior of Newton’s Method as we change parameters and vary the initial guess is very
similar to the simpler logistic map that we studied in Section 2, including period dou-
bling bifurcations and chaotic behavior. The reader is invited to experiment with simple
examples; further details can be found in [16].

Example 3.9. For fixed values of the eccentricity ǫ, Kepler’s equation

u − ǫ sin u = m (3.13)

10/23/22 32 c 2022 Peter J. Olver

 1.4

1.2

0.8

0.6

0.4

0.2

0.2 0.4 0.6 0.8 1

Figure 14. The Solution to the Kepler Equation for Eccentricity ǫ = .5.

can be viewed as a implicit equation defining the eccentric anomaly u as a function of

the mean anomaly m. To solve Kepler’s equation by Newton’s Method, we introduce the
iterative function
u − ǫ sin u − m
g(u) = u − .
1 − ǫ cos u
Notice that when | ǫ | < 1, the denominator never vanishes and so the iteration remains
well-defined everywhere. Starting with a sufficiently close initial guess u(0) , we are assured
that the method will quickly converge to the solution.
Fixing the eccentricity ǫ, we can employ the method of continuation to determine
how the solution u⋆ = h(m) depends upon the mean anomaly m. Namely, we start at
m = m0 = 0 with the obvious solution u⋆ = h(0) = 0. Then, to compute the solution
at successive closely spaced values 0 < m1 < m2 < m3 < · · · , we use the previously
computed value as an initial guess u(0) = h(mk ) for the value of the solution at the next
mesh point mk+1 , and run the Newton scheme until it converges to a sufficiently accurate
approximation to the value u⋆ = h(mk+1 ). As long as mk+1 is reasonably close to mk ,
Newton’s Method will converge to the solution quite quickly.
The continuation method will quickly produce the values of u at the sample points.
Intermediate values can either be determined by an interpolation scheme, e.g., a cubic
spline fit of the data, or by running the Newton scheme using the closest known value as
an initial condition. A plot for 0 ≤ m ≤ 1 using the value ǫ = .5 appears in Figure 14.

Systems of Equations
Let us now turn our attention to nonlinear systems of equations. We shall only
consider the case when there are the same number of equations as unknowns:
f1 (u1 , . . . , un ) = 0, ... fn (u1 , . . . , un ) = 0. (3.14)
We shall rewrite the system in vector form
f (u) = 0, (3.15)

10/23/22 33 c 2022 Peter J. Olver

where f : R n → R n is a vector-valued function of n variables. In practice, we do not
necessarily require that f be defined on all of R n , although this does simplify the exposition.
We shall only consider solutions that are separated from any others. More formally:
Definition 3.10. A solution u⋆ to a system f (u) = 0 is called isolated if there exists
δ > 0 such that f (u) 6= 0 for all u satisfying 0 < k u − u⋆ k < δ.

Example 3.11. Consider the planar equation

x2 + y 2 = (x2 + y 2 )2 .
Rewriting the equation in polar coordinates as

r = r2 or r(r − 1) = 0,
we immediately see that the solutions consist of the origin x = y = 0 and all points on the
unit circle r 2 = x2 + y 2 = 1. Only the origin is an isolated solution, since every solution
lying on the circle has plenty of other points on the circle that lie arbitrarily close to it.
Typically, solutions to a system of n equations in n unknowns are isolated, although
this is not always true. For example, if A is a singular n × n matrix, then the solutions to
the homogeneous linear system A u = 0 form a nontrivial subspace, and so are not isolated.
Nonlinear systems with non-isolated solutions can similarly be viewed as exhibiting some
form of degeneracy. In general, the numerical computation of non-isolated solutions, e.g.,
solving the implicit equations for a curve or surface, is a much more difficult problem, and
we will not attempt to discuss these issues in this introductory presentation. (However,
our continuation approach to the Kepler equation in Example 3.9 indicates how one might
proceed in such situations.)
In the case of a single scalar equation, the simple roots, meaning those for which
′ ⋆
f (u ) 6= 0, are the easiest to compute. In higher dimensions, the role of the derivative
of the function is played by the Jacobian matrix (2.28), and this motivates the following
definition.
Definition 3.12. A solution u⋆ to a system f (u) = 0 is called nonsingular if the
associated Jacobian matrix is nonsingular there: det f ′ (u⋆ ) 6= 0.

Note that the Jacobian matrix is square if and only if the system has the same number
of equations as unknowns, which is thus one of the requirements for a solution to be
nonsingular in our sense. Moreover, the Inverse Function Theorem from multivariable
calculus, [2, 12], implies that a nonsingular solution is necessarily isolated.

Theorem 3.13. Every nonsingular solution u⋆ to a system f (u) = 0 is isolated.

Being the multivariate counterparts of simple roots also means that nonsingular solu-
tions of systems are the most amenable to practical computation. Computing non-isolated
solutions, as well as isolated solutions with a singular Jacobian matrix, is a considerable
challenge, and practical algorithms remain much less well developed. For this reason, we
focus exclusively on numerical solution techniques for nonsingular solutions.

10/23/22 34 c 2022 Peter J. Olver

 Now, let us turn to numerical solution techniques. The first remark is that, unlike the
scalar case, proving existence of a solution to a system of equations is often a challenging
issue. There is no counterpart to the Intermediate Value Lemma 3.1 for vector-valued
functions. It is not hard to find vector-valued functions whose entries take on both positive
and negative values, but admit no solutions; a simple example follows. This precludes any
simple analog of the Bisection Method for nonlinear systems in more than one unknown.
T
Example 3.14. Let f (x, y) = ( x, 2 x y − 1 ) . Note that
               
1 1 1 1 −1 −1 −1 −1
f = , f = , f = , f = ,
1 1 0 −1 −1 1 0 −1
and hence the entries of f (x, y) take on all possible combinations of signs, i.e., both entries
positive, the first positive and the second negative, etc. On the other hand, there is no
solution to the system f (x, y) = 0. This implies that one cannot detect roots of a vector-
valued function by looking at sign changes in the entries.
On the other hand, Newton’s Method can be straightforwardly adapted to compute
nonsingular solutions to systems of equations, and is the most widely used method for this
purpose. The derivation proceeds in very similar manner to the scalar case. First, we
replace the system (3.15) by a fixed point system
u = g(u) (3.16)
having the same solutions. By direct analogy with (3.7), any (reasonable) fixed point
method will take the form
g(u) = u − L(u) f (u), (3.17)
where L(u) is an n × n matrix-valued function. Clearly, if f (u) = 0 then g(u) = u;
conversely, if g(u) = u, then L(u) f (u) = 0. If we further require that the matrix L(u)
be nonsingular, i.e., det L(u) 6= 0, then every fixed point of the iterator (3.17) will be a
solution to the system (3.15) and vice versa.
According to Theorem 2.12, the speed of convergence (if any) of the iterative method
u(k+1) = g(u(k) ) (3.18)
is governed by the matrix norm (or, more precisely, the spectral radius) of the Jacobian
matrix g ′ (u⋆ ) at the fixed point. In particular, if
g ′ (u⋆ ) = O (3.19)
is the zero matrix, then the method converges quadratically fast. Let’s figure out how this
can be arranged. Computing the derivative using the matrix version of the Leibniz rule
for the derivative of a matrix product, we find
g ′ (u⋆ ) = I − L(u⋆ ) f ′ (u⋆ ), (3.20)
where I is the n × n identity matrix. (Fortunately, all the terms that involve derivatives
of the entries of L(u) go away since f (u⋆ ) = 0 by assumption.) Therefore, the quadratic
convergence criterion (3.19) holds if and only if
−1
L(u⋆ ) f ′ (u⋆ ) = I , and hence L(u⋆ ) = f ′ (u⋆ ) (3.21)

10/23/22 35 c 2022 Peter J. Olver

 Figure 15. Computing the Cube Roots of Unity by Newton’s Method.

should be the inverse of the Jacobian matrix of f at the solution, which, fortuitously, was
already assumed to be nonsingular.
As in the scalar case, we don’t know the solution u⋆ , but we can arrange that condition
(3.21) holds by setting
−1
L(u) = f ′ (u)
everywhere — or at least everywhere that f has a nonsingular Jacobian matrix. The
resulting fixed point system
u = g(u) = u − f ′ (u)−1 f (u), (3.22)
leads to the quadratically convergent Newton iteration scheme
u(k+1) = u(k) − f ′ (u(k) )−1 f (u(k) ). (3.23)
All it requires is that we guess an initial value u(0) that is sufficiently close to the desired
solution u⋆ . We are then guaranteed that the iterates u(k) will converge quadratically fast
to u⋆ .
Theorem 3.15. Let u⋆ be a nonsingular solution to the system f (u) = 0. Then,
provided u(0) is sufficiently close to u⋆ , the Newton iteration scheme (3.23) converges at a
quadratic rate to the solution: u(k) → u⋆ .

Example 3.16. Consider the pair of simultaneous cubic equations

f1 (u, v) = u3 − 3 u v 2 − 1 = 0, f2 (u, v) = 3 u2 v − v 3 = 0. (3.24)
It is not difficult to prove that there are precisely three solutions:
     
⋆ 1 ⋆ −.5 ⋆ −.5
u1 = , u2 = , u3 = . (3.25)
0 .866025 . . . −.866025 . . .

10/23/22 36 c 2022 Peter J. Olver

The Newton scheme relies on the Jacobian matrix
 2 
′ 3 u − 3 v2 − 6uv
f (u) = .
6uv 3 u2 − 3 v 2
Since det f ′ (u) = 9(u2 + v 2 ) is non-zero except at the origin, all three solutions are non-
singular, and hence, for a sufficiently close initial value, Newton’s Method will converge to
the nearby solution. We explicitly compute the inverse Jacobian matrix:
 2 
′ −1 1 3 u − 3 v2 6uv
f (u) = .
9(u2 + v 2 ) − 6uv 3 u2 − 3 v 2
Hence, in this particular example, the Newton iterator (3.22) is
   2  3 
u 1 3 u − 3 v2 6uv u − 3 u v2 − 1
g(u) = − .
v 9(u2 + v 2 ) − 6uv 3 u2 − 3 v 2 3 u2 v − v 3
A complete diagram of the three basins of attraction, consisting of points whose New-
ton iterates converge to each of the three roots, has a remarkably complicated, fractal-like
structure, as illustrated in Figure 15. In this plot, the x and y coordinates run from −1.5
to 1.5. The points in the black region all converge to u⋆1 ; those in the light gray region
all converge to u⋆2 ; while those in the dark gray region all converge to u⋆3 . The closer one
is to the root, the sooner the iterates converge. On the interfaces between the basins of
attraction are points for which the Newton iterates fail to converge, but exhibit a random,
chaotic behavior. However, round-off errors will cause such iterates to fall into one of the
basins, making it extremely difficult to observe such behavior over the long run.
Remark : The alert reader may notice that in this example, we are in fact merely
computing the cube roots of unity, i.e., equations (3.24) are the real and imaginary parts
of the complex equation z 3 = 1 when z = u + i v.
Example 3.17. A robot arm consists of two rigid rods that are joined end-to-end to
a fixed point in the plane, which we take as the origin 0. The arms are free to rotate, and
the problem is to configure them so that the robot’s hand ends up at the prescribed position
T
a = ( a, b ) . The first rod has length ℓ and makes an angle α with the horizontal, so its
T
end is at position v1 = ( ℓ cos α, ℓ sin α ) . The second rod has length m and makes an
T
angle β with the horizontal, and so is represented by the vector v2 = ( m cos β, m sin β ) .
The hand at the end of the second arm is at position v1 + v2 , and the problem is to find
values for the angles α, β so that v1 + v2 = a; see Figure 16. To this end, we need to solve
the system of equations

ℓ cos α + m cos β = a, ℓ sin α + m sin β = b, (3.26)

for the angles α, β.
To find the solution, we shall apply Newton’s Method. First, we compute the Jacobian
matrix of the system with respect to α, β, which is
 
′ − ℓ sin α − m sin β
f (α, β) = ,
ℓ cos α m cos β

10/23/22 37 c 2022 Peter J. Olver

 v1 + v2

β
v1
ℓ

Figure 16. Robot Arm.

with inverse

 
′ −1 1 − ℓ sin α m sin β
f (α, β) = .
ℓ m sin(β − α) − ℓ cos α m cos β

As a result, the Newton iteration equation (3.23) has the explicit form

   
α(k+1) α(k)
= −
β (k+1) β (k)
  
1 − ℓ cos α(k) m sin β (k) ℓ cos α(k) + m cos β (k) − a
− .
ℓ m sin(β (k) − α(k) ) − ℓ cos α(k) m sin β (k) ℓ sin α(k) + m sin β (k) − b

when running the iteration, one must be careful to avoid points at which α(k) − β (k) = 0
or π, i.e., where the robot arm has straightened out.

As an example, let us assume that the rods have lengths ℓ = 2, m = 1, and the
T
desired location of the hand is at a = ( 1, 1 ) . We start with an initial guess of α(0) = 0,
β (0) = 21 π, so the first rod lies along the x–axis and the second is perpendicular. The first
few Newton iterates are given in the accompanying table. The first column is the iterate
number k; the second and third columns indicate the angles α(k) , β (k) of the rods. The
fourth and fifth give the position (x(k) , y (k) )T of the joint or elbow, while the final two
indicate the position (z (k) , w(k) )T of the robot’s hand.

10/23/22 38 c 2022 Peter J. Olver

 k α(k) β (k) x(k) y (k) z (k) w(k)

0 .0000 1.5708 2.0000 .0000 2.0000 1.0000

1 .0000 2.5708 2.0000 .0000 1.1585 .5403
2 .3533 2.8642 1.8765 .6920 .9147 .9658
3 .2917 2.7084 1.9155 .5751 1.0079 .9948
4 .2987 2.7176 1.9114 .5886 1.0000 1.0000
5 .2987 2.7176 1.9114 .5886 1.0000 1.0000

Observe that the robot has rapidly converged to one of the two possible configurations.
(Can you figure out what the second equilibrium is?) In general, convergence depends on
the choice of initial configuration, and the Newton iterates do not always settle down to
a fixed point. For instance, if k a k > ℓ + m, there is no possible solution, since the arms
are too short for the hand to reach to desired location; thus, no choice of initial conditions
will lead to a convergent scheme and the robot arm flaps around in a chaotic manner.
Now that we have gained a little experience with Newton’s Method for systems of
equations, some supplementary remarks are in order. As we know, [15], except perhaps
in very low-dimensional situations, one should not directly invert a matrix, but rather use
Gaussian elimination, or, in favorable situations, a linear iterative scheme, e.g., Jacobi,
Gauss–Seidel or even SOR. So a better strategy is to leave the Newton system (3.23) in
unsolved, implicit form

f ′ (u(k) ) v(k) = − f (u(k) ), u(k+1) = u(k) + v(k) . (3.27)

Given the iterate u(k) , we compute the Jacobian matrix f ′ (u(k) ) and the right hand side
− f (u(k) ), and then use our preferred linear systems solver to find v(k) . Adding u(k) to
the result immediately yields the updated approximation u(k+1) to the solution.
The main bottleneck in the implementation of the Newton scheme, particularly for
large systems, is solving the linear system in (3.27). The coefficient matrix f ′ (u(k) ) must be
recomputed at each step of the iteration, and hence knowing the solution to the k th linear
system does not appear to help us solve the subsequent system. Performing a complete
Gaussian elimination at every step will tend to slow down the algorithm, particularly in
high dimensional situations involving many equations in many unknowns.
One simple dodge for speeding up the computation is to note that, once we start
converging, u(k) will be very close to u(k−1) and so we will probably not go far wrong by
using f ′ (u(k−1) ) in place of the updated Jacobian matrix f ′ (u(k) ). Since we have already
solved the linear system with coefficient matrix f ′ (u(k−1) ), we know its L U factorization,
and hence can use Forward and Back Substitution to quickly solve the modified system

f ′ (u(k−1) ) v(k+1) = − f (u(k) ), u(k+1) = u(k) + v(k) . (3.28)

If u(k+1) is still close to u(k−1) , we can continue to use f ′ (u(k−1) ) as the coefficient matrix
when proceeding on to the next iterate u(k+2) . We proceed in this manner until there

10/23/22 39 c 2022 Peter J. Olver

has been a notable change in the iterates, at which stage we can revert to solving the
correct, unmodified linear system (3.27) by Gaussian Elimination. This strategy may
dramatically reduce the total amount of computation required to approximate the solution
to a prescribed accuracy. The down side is that this quasi-Newton scheme is usually
only linearly convergent, and so does not home in on the root as fast as the unmodified
implementation. The user needs to balance the trade-off between speed of convergence
versus amount of time needed to solve the linear system at each step in the process. See
[17] for further discussion.

4. Optimization.
We have already noted the importance of quadratic minimization principles for char-
acterizing the equilibrium solutions of linear systems of physical significance. In nonlinear
systems, optimization — either maximization or minimization — retains its centrality, and
the wealth of practical applications has spawned an entire sub-discipline of applied mathe-
matics. Physical systems naturally seek to minimize the potential energy function, and so
determination of the possible equilibrium configurations requires solving a nonlinear min-
imization principle. Engineering design is guided by a variety of optimization constraints,
such as performance, longevity, safety, and cost. Non-quadratic minimization principles
also arise in the fitting of data by schemes that go beyond the simple linear least squares
approximation method discussed in [15; Section 4.3]. Additional applications naturally
appear in economics and financial mathematics — one often wishes to minimize expenses
or maximize profits, in biological and ecological systems, in pattern recognition and signal
processing, in statistics, and so on. In this section, we will describe the basic mathematics
underlying simple nonlinear optimization problems along with basic numerical techniques.
The Objective Function
Throughout this section, the real-valued function F (u) = F (u1 , . . . , un ) to be op-
timized — the energy, cost, entropy, performance, etc. — will be called the objective
T
function. As such, it depends upon one or more variables u = ( u1 , u2 , . . . , un ) that
belong to a prescribed subset Ω ⊂ R n .
Definition 4.1. A point u⋆ ∈ Ω is a global minimum of the objective function F (u)
on the domain Ω if
F (u⋆ ) ≤ F (u) for all u ∈ Ω. (4.1)
The minimum is called strict if

F (u⋆ ) < F (u) for u⋆ 6= u ∈ Ω. (4.2)

The point u⋆ is called a (strict) local minimum if the relevant inequality holds just for
points u ∈ Ω nearby u⋆ , i.e., satisfying k u − u⋆ k < δ for some δ > 0.
The definition of a maximum — local or global — is the same, but with the reversed
inequality: F (u⋆ ) ≥ F (u) or, in the strict case, F (u⋆ ) > F (u). Alternatively, a maximum
of F (u) is the same as a minimum of the negative − F (u). Therefore, every result that
applies to minimization of a function can easily be translated into a result on maximization,

10/23/22 40 c 2022 Peter J. Olver

which allows us to concentrate exclusively on the minimization problem without any loss of
generality. We will use extremum as a shorthand term for either a minimum or a maximum.

Remark : In fact, any system of equations can be readily converted into a minimization
principle. Given a system f (u) = 0, we introduce the objective function

F (u) = k f (u) k2 , (4.3)

where k · k is any convenient norm on R n . By the basic properties of the norm, the
minimum value is F (u) = 0, and this is achieved if and only if f (u) = 0, i.e., at a solution
to the system. More generally, if there is no solution to the system, the minimizer(s) of
F (u) play the role of a least squares solution, at least for an inner product-based norm,
along with the extensions to more general norms.

In contrast to the rather difficult question of existence of solutions to systems of

equations, there is a general theorem that guarantees the existence of minima (and, hence,
maxima) for a broad class of optimization problems.

Theorem 4.2. If F : Ω → R is continuous, and Ω ⊂ R n is a compact, meaning closed

and bounded, subset, then F has at least one global minimum u⋆ ∈ Ω.

Proof : Let m⋆ = min{ F (u) | u ∈ Ω }, which may, a priori , be − ∞. Choose points

u , u(2), . . . ∈ Ω, such that F (u(k) ) → m as k → ∞. By the basic properties of compact
(1)

sets, [19], there is a convergent subsequence u(ki ) → u⋆ ∈ Ω. By continuity,

   
⋆ (ki )
F (u ) = F lim u = lim F u(ki ) = m⋆ ,
ki → ∞ ki → ∞

and hence u⋆ is a minimizer. Q.E.D.

Although Theorem 4.2 assures us of the existence of a global minimum of any contin-
uous function on a bounded domain, it does not guarantee uniqueness, nor does it indicate
how to go about finding it. Just as with the solution of nonlinear systems of equations, it
is quite rare that one can extract explicit formulae for the minima of non-quadratic func-
tions. Our goal, then, is to formulate practical algorithms that can accurately compute
the minima of general nonlinear functions.
The most naı̈ve algorithm, but one that is often successful in small scale problems,
[17], is to select a reasonably dense set of sample points u(k) in the domain and choose
the one that provides the smallest value for F (u(k) ). If the points are sufficiently densely
distributed and the function is not too wild, this will give a reasonable approximation to
the minimum. The algorithm can be speeded up by appealing to more sophisticated means
of selecting the sample points.
In the rest of this section, we will discuss optimization strategies that exploit the
differential calculus. Let us first review the basic procedure for optimizing functions that
you learned in first and second year calculus. As you no doubt remember, there are two
different possible types of minima. An interior minimum occurs at an interior point of the
domain of definition of the function, whereas a boundary minimum occurs on its boundary

10/23/22 41 c 2022 Peter J. Olver

 6

-1 -0.5 0.5 1

Figure 17. The function 8 u3 + 5 u2 − 6 u.

∂Ω. Interior local minima are easier to find, and, to keep the presentation simple, we shall
focus our efforts on them. Let us begin with a simple scalar example.
Example 4.3. Let us optimize the scalar objective function

F (u) = 8 u3 + 5 u2 − 6 u
on the domain −1 ≤ u ≤ 1. To locate the minimum, the first step is to look at the critical
points where the derivative vanishes:

F ′ (u) = 24 u2 + 10 u − 6 = 0, and hence u = 13 , − 43 .

To ascertain the local nature of the two critical points, we apply the second derivative test.
Since F ′′ (u) = 48 u + 10, we have



F ′′ 13 = 26 > 0, whereas F ′′ − 43 = − 26 < 0.
We conclude that 31 is a local minimum, while 43 is a local maximum.
To find the global minimum and maximum on the interval [ − 1, 1 ], we must also take
into account the boundary points ± 1. Comparing the function values at the four points,


63
F (− 1) = 3, F 31 = − 3127 ≈ − 1.148, F − 3
4 = 16 = 3.9375, F (1) = 7,
we see that 31 is the global minimum, whereas 1 is the global maximum — which occurs on
the boundary of the interval. This is borne out by the graph of the function, as displayed
in Figure 17.

The Gradient
As you first learn in multi-variable calculus, [12], the interior extrema — minima
and maxima — of a smooth function F (u) = F (u1 , . . . , un ) are necessarily critical points,
meaning places where the gradient of F vanishes. The standard gradient is the vector field
whose entries are its first order partial derivatives:
 T
∂F ∂F
∇F (u) = , ... , . (4.4)
∂u1 ∂un

10/23/22 42 c 2022 Peter J. Olver

Let us, in preparation for the more general minimization problems over infinite-dimensional
function spaces that arise in the calculus of variations, reformulate the definition of the
gradient in a more intrinsic manner. An important but subtle point is that the gradient
operator, in fact, relies upon the introduction of an inner product on the underlying vector
space. The version (4.4) is, in fact, based upon on the Euclidean dot product on R n .
Altering the inner product will change the formula for the gradient!

Definition 4.4. Let V be an inner product space. The gradient of a function

F : V → R at a point u ∈ V is the vector ∇F (u) ∈ V that satisfies

d
h ∇F (u) , v i = F (u + t v) for all v ∈ V. (4.5)
dt t=0

Remark : The function F does not have to be defined on all of the space V in order
for this definition to make sense.

The quantity displayed in the preceding formula is known as the directional derivative
of F with respect to v ∈ V , and typically denoted by ∂F/∂v. The directional derivative
measures the rate of change of F in the direction of the vector v, scaled in proportion to
its length.
In the Euclidean case, when F (u) = F (u1 , . . . , un ) is a function of n variables, defined
T
for u = ( u1 , u2 , . . . , un ) ∈ R n , we can use the chain rule to compute

d d
F (u + t v) = F (u1 + t v1 , . . . , un + t vn )
dt dt
(4.6)
∂F ∂F
= (u + t v) v1 + · · · + (u + t v) vn .
∂u1 ∂un
Setting t = 0, the right hand side of (4.6) reduces to

d ∂F ∂F
F (u + t v) = (u) v1 + · · · + (u) vn = ∇F (u) · v.
dt t=0 ∂u1 ∂un
Therefore, the directional derivative equals the Euclidean dot product between the usual
gradient of the function (4.4) and the direction vector v, justifying (4.5) in the Euclidean
case.

Remark : In this chapter, we will only deal with the standard Euclidean dot product,
which results in the usual gradient formula (4.4). If we introduce an alternative inner
product on R n , then the notion of gradient, as defined in (4.5) will change.

A function F (u) is continuously differentiable if and only if its gradient ∇F (u) is a

continuously varying vector-valued function of u. This is equivalent to the requirement
that its first order partial derivatives ∂F/∂ui are all continuous. As usual, we use C1 (Ω)
to denote the vector space of all continuously differentiable scalar-valued functions defined
on a domain Ω ⊂ R n . From now on, all objective functions are assumed to be continuously
differentiable on their domain of definition.

10/23/22 43 c 2022 Peter J. Olver

 If u(t) represents a parametrized curve contained within the domain of definition of
F (u), then the instantaneous rate of change in the scalar quantity F as we move along the
curve is given by
d du
F (u(t)) = ∇F (u) · , (4.7)
dt dt
which is the directional derivative of F with respect to the velocity or tangent vector

v = u to the curve. For instance, suppose F (u1 , u2 ) represents the elevation of a mountain
T
range at position u = ( u1 , u2 ) . If we travel through the mountains along the path
T
u(t) = ( u1 (t), u2 (t) ) , then our instantaneous rate of ascent or descent (4.7) is equal to

the dot product of our velocity vector u(t) with the gradient of the elevation function.
This observation leads to an important interpretation of the gradient vector.

Theorem 4.5. The gradient ∇F (u) of a scalar function F (u) points in the direction
of its steepest increase at the point u. The negative gradient, − ∇F (u), which points in
the opposite direction, indicates the direction of steepest decrease.

Thus, when F represents elevation, ∇F tells us the direction that is steepest uphill,
while − ∇F points directly downhill — the direction water will flow. Similarly, if F
represents the temperature of a solid body, then ∇F tells us the direction in which it
is heating up the quickest. Heat energy (like water) will flow in the opposite, coldest
direction, namely that of the negative gradient vector − ∇F .
But you need to be careful in how you interpret Theorem 4.5. Clearly, the faster you
move along a curve, the faster the function F (u) will vary, and one needs to take this
into account when comparing the rates of change along different curves. The easiest way

to effect the comparison is to assume that the tangent vector a = u has unit norm, so
k a k = 1, which means that we are passing through the point u(t) with unit speed. Once
this is done, Theorem 4.5 is an immediate consequence of the Cauchy–Schwarz inequality,
cf. [15]. Indeed,

∂F
= | a · ∇F | ≤ k a k k ∇F k = k ∇F k, when k a k = 1,
∂a
with equality if and only if a points in the same direction as the gradient. Therefore,
the maximum rate of change is when a = ∇F/k ∇F k is the unit vector in the gradient
direction, while the minimum is achieved when a = − ∇F/k ∇F k points in the opposite
direction.

Critical Points

Thus, the only points at which the gradient fails to indicate directions of increase/de-
crease of the objective function are where it vanishes. Such points play a critical role in
the analysis, whence the following definition.

Definition 4.6. A point u⋆ is called a critical point of the objective function F (u)
if
∇F (u⋆ ) = 0. (4.8)

10/23/22 44 c 2022 Peter J. Olver

 Let us prove that all local minima are indeed critical points. The most important thing
about this proof is that it only relies on the intrinsic definition of gradient, and therefore
applies to any function on any inner product space. Moreover, even though the gradient
will change if we alter the underlying inner product, the requirement that it vanish at a
local minimum does not.
Theorem 4.7. Every local (interior) minimum u⋆ of a continuously differentiable
function F (u) is a critical point: ∇F (u⋆ ) = 0.
Proof : Let 0 6= v ∈ R n be any vector. Consider the scalar function

g(t) = F (u⋆ + t v) = F (u⋆1 + t v1 , . . . , u⋆n + t vn ), (4.9)

where t ∈ R is sufficiently small to ensure that u⋆ + t v remains strictly inside the domain
of F . Note that g measures the values of F along a straight line passing through u⋆ in the
direction prescribed by v. Since u⋆ is a local minimum,

F (u⋆ ) ≤ F (u⋆ + t v), and hence g(0) ≤ g(t)

for all t sufficiently close to zero. In other words, g(t), as a function of the single variable
t, has a local minimum at t = 0. By the basic calculus result on minima of functions of
one variable, the derivative of g(t) must vanish at t = 0. Therefore, by the definition (4.5)
of gradient,
d
0 = g ′ (0) = F (u⋆ + t v) = h ∇F (u⋆ ) , v i.
dt t=0

We conclude that the gradient vector ∇F (u⋆ ) at the critical point must be orthogonal to
every vector v ∈ R n , which is only possible if ∇F (u⋆ ) = 0. Q.E.D.
Thus, provided the objective function is continuously differentiable, every interior
minimum, both local and global, is necessarily a critical point. The converse is not true;
critical points can be maxima; they can also be saddle points or of some degenerate form.
The basic analytical method† for determining the (interior) minima of a given function is
to first find all its critical points by solving the system of equations (4.8). Each critical
point then needs to be examined more closely — as it could be either a minimum, or a
maximum, or neither.
Example 4.8. Consider the function

F (u, v) = u4 − 2 u2 + v 2 ,

T
which is defined and continuously differentiable on all of R 2 . Since ∇F = 4 u3 − 4 u, 2 v ,
its critical points are obtained by solving the pair of equations

4 u3 − 4 u = 0, 2 v = 0.

†
Numerical methods are discussed below.

10/23/22 45 c 2022 Peter J. Olver

 Figure 18. The Function u4 − 2 u2 + v 2 .

The solutions to the first equation are u = 0, ± 1, while the second equation requires v = 0.
Therefore, F has three critical points:
     
0 1 −1
u⋆1 = , u⋆2 = , u⋆3 = . (4.10)
0 0 0

Inspecting its graph in Figure 18, we suspect that the first critical point u⋆1 is a saddle
point, whereas the other two appear to be local minima, having the same value F (u⋆2 ) =
F (u⋆3 ) = − 1. This will be confirmed once we learn how to rigorously distinguish critical
points.

The student should also pay attention to the distinction between local minima and
global minima. In the absence of theoretical justification, the only practical way to deter-
mine whether or not a minimum is global is to find all the different local minima, including
those on the boundary, and see which one gives the smallest value. If the domain is un-
bounded, one must also worry about the asymptotic behavior of the objective function for
large u.

The Second Derivative Test

The status of critical point — minimum, maximum, or neither — can often be resolved
by analyzing the second derivative of the objective function at the critical point. Let us
first review the one variable second derivative test you learned in first year calculus.

Proposition 4.9. Let g(t) ∈ C2 be a scalar function, and suppose that t⋆ a critical
point: g ′ (t⋆ ) = 0. If t⋆ is a local minimum, then g ′′ (t⋆ ) ≥ 0. Conversely, if g ′′ (t⋆ ) > 0,
then t⋆ is a strict local minimum. Similarly, g ′′ (t⋆ ) ≤ 0 is required at a local maximum,
while g ′′ (t⋆ ) < 0 implies that t⋆ is a strict local maximum.

10/23/22 46 c 2022 Peter J. Olver

 The proof of this result relies on the fact that we can approximate the function by its
quadratic Taylor polynomial near the critical point:

g(t) ≈ g(t⋆ ) + 1
2
(t − t⋆ )2 g ′′ (t⋆ ),
since g ′ (t⋆ ) = 0, and so the linear terms in the Taylor polynomial vanish. If g ′′ (t⋆ ) 6= 0,
then the quadratic Taylor polynomial has a minimum or maximum at t⋆ according to the
sign of the second derivative, and this provides the key to the proof. In the borderline
case, when g ′′ (t⋆ ) = 0, the second derivative test is inconclusive, and the point could be
either maximum or minimum or neither. One must analyze the higher order terms in the
Taylor expansion to resolve the status of the critical point.
In multi-variate calculus, the “second derivative” of a function F (u) = F (u1 , . . . , un )
is represented by the n × n Hessian † matrix , whose entries are its second order partial
derivatives:  
∂ 2F ∂ 2F ∂ 2F
 ∂u2 ...
 1 ∂u1 ∂u2 ∂u1 ∂un  
 
 ∂ F 2 2
∂ F ∂ F 
2
 ... 
 ∂u ∂u ∂u 2 ∂u ∂u 
 2 1 2 2 n 
∇ F (u) = 
2

,
 (4.11)
 .. .. .. .. 
 . . . . 
 
 
 ∂ 2F ∂ 2
F ∂ 2
F 
 ... 
∂un ∂u1 ∂un ∂u2 ∂u2n

We will always assume that F (u) ∈ C2 has continuous second order partial derivatives. In
this case, its mixed partial derivatives are equal: ∂ 2 F/∂ui ∂uj = ∂ 2 F/∂uj ∂ui , cf. [2, 12].
As a result, the Hessian is a symmetric matrix: ∇2 F (u) = ∇2 F (u)T .
The second derivative test for a local minimum of scalar function relies on the positiv-
ity of its second derivative. For a function of several variables, the corresponding condition
is that the Hessian matrix be positive definite. See [15] for a detailed discussion of positive
definite matrices. More specifically:
Theorem 4.10. Let F (u) = F (u1 , . . . , un ) ∈ C2 (Ω) be a real-valued, twice contin-
uously differentiable function defined on an open domain Ω ⊂ R n . If u⋆ ∈ Ω is a (local,
interior) minimum for F , then it is necessarily a critical point, so ∇F (u⋆ ) = 0. Moreover,
the Hessian matrix (4.11) must be positive semi-definite at the minimum, so ∇2 F (u⋆ ) ≥ 0.
Conversely, if u⋆ is a critical point with positive definite Hessian matrix ∇2 F (u⋆ ) > 0,
then u⋆ is a strict local minimum of F .
A maximum requires a negative semi-definite Hessian matrix. If, moreover, the Hes-
sian at the critical point is negative definite, then the critical point is a strict local maxi-
mum. If the Hessian matrix is indefinite, then the critical point is a saddle point — neither

†
Named after the early eighteenth century German mathematician Ludwig Otto Hesse.

10/23/22 47 c 2022 Peter J. Olver

minimum nor maximum. In the borderline case, when the Hessian is only positive or nega-
tive semi-definite at the critical point, the second derivative test is inconclusive. Resolving
the nature of the critical point requires more detailed knowledge of the objective function,
e.g., its higher order derivatives.
We defer the proof of Theorem 4.10 until the end of this section.
Example 4.11. As a first, elementary example, consider the quadratic function

F (u, v) = u2 − 2 u v + 3 v 2 .
To minimize F , we begin by computing its gradient
 
2u − 2v
∇F (u, v) = .
−2u + 6v
Solving the pair of equations ∇F = 0, namely

2 u − 2 v = 0, − 2 u + 6 v = 0,
we see that the only critical point is the origin u = v = 0. To test whether the origin is a
maximum or minimum, we further compute the Hessian matrix
   
2 Fuu Fuv 2 −2
H = ∇ F (u, v) = = .
Fuv Fvv −2 6
Using the methods of [15; Section 3.5], we easily prove that the Hessian matrix is positive
definite. Therefore, by Theorem 4.10, u⋆ = 0 is a strict local minimum of F .
Indeed, we recognize F (u, v) to be, in fact, a homogeneous positive definite quadratic
form, which can be written in the form
   
T 1 −1 1 u
F (u, v) = u K u, where K= = 2 H, u= .
−1 3 v
Positive definiteness of the coefficient matrix K implies that F (u, v) > 0 for all u =
T
( u, v ) 6= 0, and hence 0 is, in fact, a global minimum.
In general, any quadratic function Q(u) = Q(u1 , . . . , un ) can be written in the form
m
X n
X
T T
Q(u) = u K u − 2 b u + c = kij ui uj − 2 bi ui + c, (4.12)
i,j = 1 i=1

where K = K T is a symmetric n × n matrix, b ∈ R n is a fixed vector, and c ∈ R is a

scalar. A straightforward computation produces the formula for its gradient and Hessian
matrix:
∇Q(u) = 2 K u − 2 b, ∇2 Q(u) = 2 K. (4.13)
As a result, the critical points of the quadratic function are the solutions to the linear
system K u = b. If K is nonsingular, there is a unique critical point u⋆ , which is a strict
local minimum if and only if K > 0 is positive definite. In fact, [15; Theorem 4.1] tells

10/23/22 48 c 2022 Peter J. Olver

 u2 + v 2 − v 3 u2 + v 4 u2 + v 3
Figure 19. Critical Points.

us that, in the positive definite case, u⋆ is a strict global minimum for Q(u). Thus, the
algebraic approach of [15; Chapter 4] provides additional, global information that cannot
be gleaned directly from the local, multivariable calculus Theorem 4.10. But algebra is only
able to handle quadratic minimization problems with ease. The analytical classification of
minima and maxima of more complicated objective functions necessarily relies the gradient
and Hessian criteria of Theorem 4.10.
Example 4.12. The function
 
2 2 3 2u
F (u, v) = u + v − v has gradient ∇F (u, v) = .
2 v − 3 v2
   
⋆ 0 ⋆ 0
The critical point equation ∇F = 0 has two solutions: u1 = and u2 = 2 . The
0 3
Hessian matrix of the objective function is
 
2 2 0
∇ F (u, v) = .
0 2 − 6v
 
2 2 0
At the first critical point, the Hessian ∇ F (0, 0) = is positive definite. Therefore,
0 2  
2

2 0
the origin is a strict local minimum. On the other hand, ∇2 F 0, 3 = is
  0 −2
0
indefinite, and hence u⋆2 = 2 a saddle point. The function is graphed in Figure 19,
3
with the critical points indicated by the small solid balls. The origin is, in fact, only a
local minimum, since F (0, 0) = 0, whereas F (0, v) < 0 for all v > 1. Thus, this particular
function has no global minimum or maximum on R 2 .
Next, consider the function
 
2 4 2u
F (u, v) = u + v , with gradient ∇F (u, v) = .
4 v3
The only critical point is the origin u = v = 0. The origin is a strict global minimum
T
because F (u, v) > 0 = F (0, 0) for all (u, v) 6= ( 0, 0 ) . However, its Hessian matrix
 
2 2 0
∇ F (u, v) =
0 12 v 2

10/23/22 49 c 2022 Peter J. Olver

  
2 2 0
is only positive semi-definite at the origin, ∇ F (0, 0) = .
0 0
On the other hand, the origin u = v = 0 is also the only critical point for the function
 
2 3 2u
F (u, v) = u + v with ∇F (u, v) = .
3 v2
The Hessian matrix is
   
2 2 0 2 2 0
∇ F (u, v) = , and so ∇ F (0, 0) =
0 6v 0 0
is the same positive semi-definite matrix at the critical point. However, in this case (0, 0)
is not a local minimum; indeed

F (0, v) < 0 = F (0, 0) whenever v < 0,

and so there exist points arbitrarily close to the origin where F takes on smaller values.
As illustrated in Figure 19, the origin is, in fact, a degenerate saddle point.
Finally, the function
 
2 2 2u − 2v
F (u, v) = u − 2 u v + v has gradient ∇F (u, v) = ,
−2u + 2v
and so every point u = v is a critical point. The Hessian matrix
   
2 Fuu Fuv 2 −2
∇ F (u, v) = =
Fuv Fvv −2 2
is positive semi-definite everywhere. Since F (u, u) = 0, while F (u, v) = (u − v)2 > 0
when u 6= v, each of these critical points is a non-isolated (and hence non-strict) local
minimum. Thus, comparing the three preceding examples, we see that a semi-definite
Hessian is unable to distinguish between different types of degenerate critical points.
Finally, the reader should always keep in mind that first and second derivative tests
only determine the local behavior of the function near the critical point. They cannot
be used to determine whether or not we are at a global minimum. This requires some
additional analysis, and, often, a fair amount of ingenuity.
Proof of Theorem 4.10 : We return to the proof of Theorem 4.7. Given a local
minimum u⋆ , the scalar function g(t) = F (u⋆ + t v) in (4.9) has a local minimum at t = 0.
As noted above, basic calculus tells us that its derivatives at t = 0 must satisfy

g ′ (0) = 0, g ′′ (0) ≥ 0. (4.14)

The first condition leads to the critical point equation ∇F (u⋆ ) = 0. A straightforward
chain rule calculation produces the formula
n
X ∂ 2F
′′
g (0) = (u⋆ ) vi vj = vT ∇2 F (u⋆ ) v.
i,j = 1
∂ui ∂uj

10/23/22 50 c 2022 Peter J. Olver

As a result, the second condition in (4.14) requires that

vT ∇2 F (u⋆ ) v ≥ 0.
Since this condition is required for every direction v ∈ R n , the Hessian matrix ∇2 F (u⋆ ) ≥ 0
satisfies the criterion for positive semi-definiteness, proving the first part of the theorem.
The proof of the converse relies† on the second order Taylor expansion of the function:

F (u) = F (u⋆ ) + ∇F (u⋆ ) · v + 1

2 vT ∇2 F (u⋆ ) v + S(v, u⋆ )
⋆ 1 T 2 ⋆ ⋆
where v = u − u⋆ ,
= F (u ) + 2 v ∇ F (u ) v + S(v, u ),
(4.15)
⋆
at the critical point, whence ∇F (u ) = 0. The remainder term in the Taylor formula goes
to 0 as u → u⋆ at a rate faster than quadratic:
S(v, u⋆ )
−→ 0 as v −→ 0. (4.16)
k v k2
Assuming ∇2 F (u⋆ ), there is a constant C > 0 such that

vT ∇2 F (u⋆ ) v ≥ C k v k2 for all v ∈ Rn.

This is a consequence of Theorem 3.7 of [15] on the equivalence of norms, coupled with
the fact that every positive definite matrix defines a norm. By (4.16), we can find δ > 0
such that

| S(v, u⋆ ) | < 1
2 C k v k2 whenever 0 < k v k = k u − u⋆ k < δ.
But then the Taylor formula (4.15) implies that, for all u satisfying the preceding inequality,

0< 1
2 vT ∇2 F (u⋆ ) v + S(v, u⋆ ) = F (u) − F (u⋆ ),
which implies u⋆ is a strict local minimum of F (u). Q.E.D.

Constrained Optimization and Lagrange Multipliers

In many applications, the function to be minimized is subject to constraints. For
instance, finding boundary minima requires constraining the minima to the boundary of
the domain. Another example would be to find the minimum temperature on the surface
of the earth. Assuming the earth is a perfect sphere of radius R, the temperature function
T (u, v, w) is then to be minimized subject to the constraint u2 + v 2 + w2 = R2 .
Let us focus on finding the minimum value of an objective function F (u) = F (u, v, w)
when its arguments (u, v, w) are constrained to lie on a regular surface S ⊂ R 3 . Suppose

†
Actually, it is not hard to prove the first part using the first order Taylor expansion without
resorting to the scalar function g. On the other hand, when we look at infinite-dimensional
minimization problems arising in the calculus of variations, we will no longer have the luxury of
appealing to the finite-dimensional Taylor expansion, whereas the previous argument continues
to apply in general contexts.

10/23/22 51 c 2022 Peter J. Olver

 T
u⋆ = ( u⋆ , v ⋆ , w⋆ ) ∈ S is a (local) minimum for the constrained objective function. Let
T
u(t) = ( u(t), v(t), w(t) ) ⊂ S be any curve contained within the surface that passes
through the minimum, with u(0) = u⋆ . Then the scalar function g(t) = F (u(t)) must
have a local minimum at t = 0, and hence, in view of (4.7),

d
0 = g ′ (0) = = ∇F (u(0)) · u(0) = ∇F (u⋆ ) · u(0) .
 
F (u(t)) (4.17)
dt t=0

Thus, the gradient of the objective function at the surface minimum must be orthogonal
to the tangent vector to the curve. Since the curve was constrained to lies entirely in S, its
tangent vector u(0) is tangent to the surface at the point u⋆ . Since every tangent vector to


the surface is tangent to some curve contained in the surface, ∇F (u⋆ ) must be orthogonal
to every tangent vector, and hence point in the normal direction to the surface. Thus, a
constrained critical point u⋆ ∈ S of a function on a surface is defined so that

∇F (u⋆ ) = λ n, (4.18)

where n denotes the normal to the surface at the point u⋆ . The scalar factor λ is known
as the Lagrange multiplier in honor of Joseph-Louis Lagrange, one of the pioneers of
constrained optimization. The value of the Lagrange multiplier is not fixed a priori, but
must be determined by solving the critical point system (4.18). The same reasoning applies
to local maxima, which are also constrained critical points. The nature of a constrained
critical point — local minimum, local maximum, local saddle point, etc. — is fixed by a
constrained second derivative test.

Example 4.13. Our problem is to find the minimum value of the objective function
F (u, v, w) = u2 −2 w3 when u, v, w are restricted to the unit sphere S = (u2 + v 2 + w2 = 1).
T
The radial vector n = ( u, v, w ) is normal to the sphere, and so the critical point condition
(4.18) is
   
2u u
∇F =  0  = λ v .

2
−6w w

Thus, we must solve the system of equations

2 u = λ u, 0 = λ v, − 6 w2 = λ w, subject to u2 + v 2 + w2 = 1,

for the unknowns u, v, w and λ. This needs to be done carefully to avoid missing any cases.
First, if u 6= 0, then λ = 2, v = 0, and either w = 0 whence u = ±1, or w = − 31 and so
√ √
u = ± 1 − w2 = ± 2 3 2 . On the other hand, if u = 0, then either λ = 0, w = 0 and so
v = ±1, or v = 0, w = ±1, and λ = ∓6. Collecting these together, we discover that there

10/23/22 52 c 2022 Peter J. Olver

are a total of 8 critical points on the unit sphere:
     √   √ 
2 2
1 −1 3
− 232
   
u⋆1 =  0 , u⋆2 =  0 , u⋆3 =  0 , u⋆4 =  0 ,
0 0 −1 − 13
     3  
0 0 0 0
⋆ ⋆ ⋆

u5 = 1 ,  
u6 = −1 ,  u7 = 0 ,
 u⋆8 =  0 .
0 0 1 −1
Since the unit sphere is closed and bounded, we are assured that F has a global maximum
and a global minimum when restricted to S, which are both to be found among our
candidate critical points. Thus, we merely compute the value of the objective function at
each critical point,
F (u⋆1 ) = 1, F (u⋆2 ) = 1, F (u⋆3 ) = 22
27 , F (u⋆4 ) = 26
27 ,
F (u⋆5 ) = 0, F (u⋆6 ) = 0, F (u⋆7 ) = −2, F (u⋆8 ) = 2.
Therefore, u⋆7 must be the global minimum and u⋆8 the global maximum of the objective
function restricted to the unit sphere. The status of the other six critical points — con-
strained local maximum, minimum, or neither — is less evident, and a full classification
requires the second derivative test for constrained extrema.
If the surface is given as the level set of a function

G(u, v, w) = c, (4.19)
then at any point u⋆ ∈ S, the gradient vector ∇G(u⋆ ) points in the normal direction to
the surface, and hence, provided n = ∇G(u⋆ ) 6= 0, the surface critical point condition can
be rewritten as
∇F (u⋆ ) = λ ∇G(u⋆ ), (4.20)
T
or, in full detail, the critical point ( u⋆ , v ⋆ , w⋆ ) must satisfy
∂F ∂G
(u, v, w) = λ (u, v, w),
∂u ∂u
∂F ∂G
(u, v, w) = λ (u, v, w), (4.21)
∂v ∂v
∂F ∂G
(u, v, w) = λ (u, v, w).
∂w ∂w
Thus, to find the constrained critical points, one needs to solve the combined system (4.19,
21) of 4 equations for the four unknowns u, v, w and the Lagrange multiplier λ.
Formally, one can reformulate the problem as an unconstrained optimization problem
by introducing the augmented objective function


E(u, v, w, λ) = F (u, v, w) − λ G(u, v, w) − c . (4.22)
The critical points of the augmented function are where its gradient, with respect to all four
arguments, vanishes. Setting the partial derivatives with respect to u, v, w to 0 reproduces

10/23/22 53 c 2022 Peter J. Olver

the system (4.21), while its partial derivative with respect to λ reproduces the constraint
(4.19).
If F (u) is defined on a closed subdomain Ω ⊂ R n , then its minima may also occur
at boundary points u ∈ ∂Ω. When the boundary is smooth, there is an analogous critical
point condition for local boundary extrema.
Theorem 4.14. Let Ω ⊂ R n be a domain with smooth boundary ∂Ω. Suppose F (u)
is continuously differentiable at all points in Ω = Ω ∪ ∂Ω. If the boundary point u0 ∈ ∂Ω is
a (local) minimum for F when restricted to the closed domain Ω, then the gradient vector
∇F (u0 ) is either 0 or points inside the domain in the normal direction to ∂Ω.
Proof : Let u(t) ⊂ ∂Ω be any curve that is entirely contained in the boundary, with
u(0) = u0 . Then the scalar function g(t) = F (u(t)) must have a local minimum at t = 0,
and hence, in view of (4.7),
d
0 = g ′ (0) =

F (u(t)) = ∇F (u(0)) , u(0) .
dt t=0

Since the curve lies entirely in ∂Ω, its tangent vector u(0) is tangent to the boundary at
the point u0 ; moreover, we can realize any such tangent vector by a suitable choice of
curve. We conclude that ∇F (u0 ) is orthogonal to every tangent vector to ∂Ω at the point
u0 , and hence must point in the normal direction to the boundary. Moreover, if non-zero,
it cannot point outside Ω since then − ∇F (u0 ), which is the direction of decrease in the
objective function, would point inside the domain, which would preclude u0 from being a
local minimum. Q.E.D.
The same ideas can be applied to optimization problems involving functions of several
variables subject to one or more constraints. Suppose the objective function F (u) =
F (u1 , . . . , un ) is subject to the constraints
G1 (u) = c1 , ··· Gk (u) = ck . (4.23)
A point u satisfying the constraints is termed regular if the corresponding gradient vectors
∇G1 (u), . . . , ∇Gk (u) are linearly independent. (Singular points are more tricky, and must
be handled separately.) A regular constrained critical point u⋆ necessarily satisfies the
vector equation
∇F (u⋆ ) = λ1 ∇G1 (u⋆ ) + · · · + λk ∇Gk (u⋆ ), (4.24)
where the unspecified scalars λ1 , . . . , λk are called the Lagrange multipliers for the con-
strained optimization problem. (See below for the proof details.) The critical points are
thus found by solving the combined system (4.23–24) for the n + k variables u1 , . . . , un and
λ1 , . . . , λk . As in (4.22) we can reformulate this as an unconstrained optimization problem
for the augmented objective function
k
X

E(u, λ) = F (u) − λi Gi (u) − ci . (4.25)
i=1

The gradient with respect to u reproduces the critical point system (4.24), while its gradient
with respect to λ = (λ1 , . . . , λk ) recovers the constraints (4.23).

10/23/22 54 c 2022 Peter J. Olver

 Theorem 4.15. Every regular constrained local minimum and maximum is a con-
strained critical point.
Example 4.16. The problem is to find the point or points on the intersection of the
elliptical cylinders
u2 + 4 v 2 = 1, u2 + 9 w2 = 4, (4.26)
that is the closest to the origin. Thus, we seek to minimize the squared† distance function

F (u, v, w) = u2 + v 2 + w2
subject to the constraints

G(u, v, w) = u2 + 4 v 2 = 1, H(u, v, w) = u2 + 9 w2 = 4,
The augmented objective function (4.22) is



E(u, v, w, λ, µ) = u2 + v 2 + w2 − λ u2 + 4 v 2 − 1 + µ u2 + 9 w2 − 4 .
To find its critical points, we set all its partial derivatives to zero:
∂E ∂E ∂E
= 2 u + 2 λ u + 2 µ u = 0, = 2 v + 8 λ v = 0, = 2 w + 18 λ w = 0,
∂u ∂v ∂w
while the partial derivatives with respect to the Lagrange multipliers λ, µ reproduce the
two constraints (4.26). Thus,
1
either u = 0 or λ + µ = −1, either v = 0 or λ = − 14 , and either w = 0 or µ = − 18 .
Thus, at least one of u, v, w must be zero. If u = 0, then v = ± 12 , w = ± 32 ; if v = 0,
then u = ± 1, w = ± √13 ; while there are no real solutions to the constraints when w = 0.

T
The first four critical points, 0, ± 12 , ± 23 , all lie a distance 65 ≈ .8333 from the origin,

T √
while the second four, ±1, 0, ± 32 , are further away, at distance 32 ≈ 1.1547. Thus,
the closest points on intersection of the cylinders are the first four, while the furthest
points from the origin are the last four. (The latter comment relies on the fact that the
intersection is a compact subset of R 3 .)
The second derivative test for regular constrained minima and maxima can be found
as follows. Let u(t) be a parametrized curve satisfying the constraints (4.23) for all t, with

u(0) = u⋆ ,
 
u(0) = v, u(0) = w.
The second order Taylor expansions of the functions evaluated on the curve take the form

 
F u(t) = F (u⋆ ) + t ∇F (u⋆ ) · v + 12 t2 ∇F (u⋆ ) · w + vT ∇2 F (u⋆ ) v + · · · ,

  (4.27)
Gi u(t) = Gi (u⋆ ) + t ∇Gi (u⋆ ) · v + 21 t2 ∇Gi (u⋆ ) · w + vT ∇2 Gi (u⋆ ) v + · · · .

†
Any distance minimizer also minimizes the squared distance; we work with the latter in order
to avoid square roots in the computation.

10/23/22 55 c 2022 Peter J. Olver

Since we are assuming Gi (u(t)) ≡ 0, its individual Taylor coefficients vanish

Gi (u⋆ ) = 0, ∇Gi (u⋆ ) · v = 0, ∇Gi (u⋆ ) · w + vT ∇2 Gi (u⋆ ) v = 0, i = 1, . . . , k. (4.28)

At a regular point, this implies that v must be orthogonal to each of the gradient vectors
∇G1 (u⋆ ), . . . , ∇Gk (u⋆ ), and hence must belong to the tangent space to the constraint
set (4.23), which can be identified as the orthogonal complement to the normal subspace
spanned by the gradient vectors. Indeed, the Implicit Function Theorem, [2, 12], implies
that the constraint set is, locally, a submanifold of dimension n − k, and hence every
tangent vector v can be realized as the tangent to some curve u(t) contained therein.


Furthermore, F u(t) has a minimum at t = 0, which implies the necessary conditions

∇F (u⋆ ) · v = 0, ∇F (u⋆ ) · w + vT ∇2 F (u⋆ ) v ≥ 0, (4.29)

which must hold whenever the vectors v, w satisfy (4.28). Thus, ∇F (u⋆ ) must be orthog-
onal to the tangent space to the constraint set, and hence a linear combination of the
normal gradient vectors, which implies the Lagrange multiplier equation (4.24). Thus, in
view of (4.28),

k
X k
X
∇F (u⋆ ) · w = λi ∇Gi (u⋆ ) · w = − λi ∇2 Gi (u⋆ ) v.
i=1 i=1

Substituting back into (4.29) produces the necessary condition

k
!
X
vT ∇2 F (u⋆ ) − λi ∇2 Gi (u⋆ ) v ≥ 0, (4.30)
i=1

which must hold whenever v satisfies the constraints in (4.28). In other words, the Hessian
matrix of the augmented objective function (4.25) with respect to u, evaluated at the
critical point u⋆ , must be positive semi-definite when restricted to the tangent space of the
constraint set. As in the unconstrained minimization problem, if this restricted augmented
Hessian matrix is positive definite, then the critical point is a strict local minimum to the
constrained optimization problem.
To convert these conditions into a practical criterion, let v1 , . . . , vk be a basis for
the tangent space, i.e., a basis for the solution space to the system of linear equations
∇Gi (u⋆ ) · v = 0, i = 1, . . . , k. Let V = (v1 , . . . , vk ) be the corresponding n × k matrix
with the indicated columns. Then a necessary condition for minimality of u⋆ is that the
restricted augmented Hessian matrix

k
!
X
H =VT ∇2 F (u⋆ ) − λi ∇2 Gi (u⋆ ) V (4.31)
i=1

be positive semi-definite, H ≥ 0, while when it is positive definite, H > 0, the critical

point u⋆ is a strict local minimum.

10/23/22 56 c 2022 Peter J. Olver

 Numerical Minimization of Scalar Functions
In practical optimization, one typically bypasses the preliminary characterization of
minima as critical points, and instead implements a direct iterative procedure that con-
structs a sequence of successively better approximations to the desired minimum. As the
computation progresses, the approximations are adjusted so that the objective function
is made smaller and smaller, which, we hope, will ensure that we are converging to some
form of minimum.
As always, to understand the issues involved, it is essential to consider the simplest
scalar situation. Thus, we are given the problem of minimizing a scalar function F (u) on
a bounded interval a ≤ u ≤ b. The minimum value can either be at an endpoint or an
interior minimum. Let us first state a result that plays a similar role to the Intermediate
Value Lemma 3.1 that formed the basis of the Bisection Method for locating roots.
Lemma 4.17. Suppose that F (u) is defined and continuous for all a ≤ u ≤ b.
Suppose that we can find a point a < c < b such that F (c) < F (a) and F (c) < F (b). Then
F (u) has a minimum at some point a < u⋆ < b.
The proof is an easy consequence of Theorem 4.2. Therefore, if we find three points
a < c < b satisfying the conditions of the lemma, we are assured of the existence of a local
minimum for the function between the two endpoints. Once this is done, we can design
an algorithm to home in on the minimum u⋆ . We choose another point, say d between
a and c and evaluate F (d). If F (d) < F (c), then F (d) < F (a) also, and so the points
a < d < c satisfy the hypotheses of Lemma 4.17. Otherwise, if F (d) > F (c) then the
points d < c < b satisfy the hypotheses of the lemma. In either case, a local minimum
has been narrowed down to a smaller interval, either [ a, c ] or [ d, b ]. In the unlikely even
that F (d) = F (c), one can try another point instead — unless the objective function is
constant, one will eventually find a suitable value of d. Iterating the method will produce
a sequence of progressively smaller and smaller intervals in which the minimum is trapped,
and, just like the Bisection Method, the endpoints of the intervals get closer and closer to
the local minimum u⋆ .
The one question is how to choose the point d. We described the algorithm when
it was selected to lie between a and c, but one could equally well try a point between
c and b. To speed up the algorithm, it makes sense to place d in the larger of the two
subintervals [ a, c ] and [ c, b ]. One could try placing d in the midpoint of the interval, but
a more inspired choice is to place it at a fraction θ = √52 − 12 ≈ .61803 of the way along
the interval, i.e., at θ a + (1 − θ) c if [ a, c ] is the longer interval. (One could equally well
take the point (1 − θ) a + θ c.) The result is the Golden Section Method . At each stage, the
length of the interval has been reduced by a factor of θ, so the convergence rate is linear,
although a bit slower than bisection.
Another strategy is to use an interpolating polynomial passing through the three
points on the graph of F (u) and use its minimum value as the next approximation to the
minimum. The minimizing value occurs at
ms − nt
d= ,
s−t

10/23/22 57 c 2022 Peter J. Olver

where
F (c) − F (a) F (b) − F (c) a+c c+b
s= , t= , m= , n= .
c−a b−c 2 2
As long as a < c < b satisfy the hypothesis of Lemma 4.17, we are assured that the
quadratic interpolant has a minimum (and not a maximum!), and that the minimum
remains between the endpoints of the interval: a < d < b. If the length of the interval is
small, the minimum value should be a good approximation to the minimizer u⋆ of F (u)
itself. Once d is determined, the algorithm proceeds as before. In this case, convergence is
not quite guaranteed, or, in unfavorable situations, could be much slower than the Golden
Section Method. One can even try using the method when the function values do not
satisfy the hypothesis of Lemma 4.17, although now the new point d will not necessarily
lie between a and b. Worse, the quadratic interpolant may have a maximum at d, and one
ends up going in the wrong direction, which can even happen in the minimizing case due
to the discrepancy between it and the objective function F (u). Thus, this method must
be handled with care.
A final idea is to focus not on the objective function F (u) but rather its derivative
f (u) = F ′ (u). The critical points of F are the roots of f (u) = 0, and so one can use one
of the solution methods, e.g., Bisection or Newton’s Method, to find the critical points.
Of course, one must then take care that the critical point u⋆ is indeed a minimum, as it
could equally well be a maximum of the original objective function. (It will probably not
be an inflection point, as these do not correspond to simple roots of f (u).) The status of
the critical point can be checked by looking at the sign of F ′′ (u⋆ ) = f ′ (u⋆ ); indeed, if we
use Newton’s Method we will be computing the derivative at each stage of the algorithm,
and can stop looking if the derivative turns out to be of the wrong sign.
Gradient Descent
Now, let us turn our attention to multi-dimensional optimization problems. We are
seeking to minimize a (smooth) scalar objective function F (u) = F (u1 , . . . , un ). According
to Theorem 4.5, at any given point u in the domain of definition of F , the negative gradient
vector − ∇F (u), if nonzero, points in the direction of the steepest decrease in F . Thus, to
minimize F , an evident strategy is to “walk downhill”, and, to be efficient, walk downhill
as fast as possible, namely in the direction − ∇F (u). After walking in this direction for a
little while, we recompute the gradient, and this tells us the new direction to head downhill.
With luck, we will eventually end up at the bottom of the valley, i.e., at a (local) minimum
value of the objective function.
This simple idea forms the basis of the Gradient Descent Method for minimizing the
objective function F (u). In a numerical implementation, we start the iterative procedure
with an initial guess u(0) , and let u(k) denote the k th approximation to the minimum u⋆ .
To compute the next approximation, we set out from u(k) in the direction of the negative
gradient, and set
u(k+1) = u(k) − tk ∇F (u(k) ). (4.32)
for some positive scalar tk > 0. We are free to adjust tk so as to optimize our descent
path, and this is the key to the success of the method.

10/23/22 58 c 2022 Peter J. Olver

 If ∇F (u(k) ) 6= 0, then, at least when tk > 0 is sufficiently small,

F (u(k+1) ) < F (u(k) ), (4.33)

and so u(k+1) is, presumably, a better approximation to the desired minimum. Clearly, we
cannot choose tk too large or we run the risk of overshooting the minimum and reversing
the inequality (4.33). Think of walking downhill in the Swiss Alps. If you walk too far
in a straight line, which is what happens as tk increases, then you might very well miss
the valley and end up higher than you began — not a good strategy for descending to the
bottom! On the other hand, if we choose tk too small, taking very tiny steps, then the
method may end up converging to the minimum much too slowly to be of practical use.
How should we choose an optimal value for the factor tk ? Keep in mind that the goal
is to minimize F (u). Thus, a good strategy would be to set tk equal to the value of t > 0
that minimizes the scalar objective function


g(t) = F u(k) − t ∇F (u(k) ) (4.34)
obtained by restricting F (u) to the ray emanating from u(k) that lies in the negative
gradient direction. Physically, this corresponds to setting off in a straight line in the
direction of steepest decrease, and continuing on until we cannot go down any further.
Barring luck, we will not have reached the actual bottom of the valley, but must then
readjust our direction and continue on down the hill in a series of straight line paths.
In practice, one can rarely compute the minimizing value t⋆ of (4.34) exactly. Instead,
we employ one of the scalar minimization algorithms presented in the previous subsection.
Note that we only need to look for a minimum among positive values of t > 0, since our
choice of the negative gradient direction assures us that, at least for t sufficiently small
and positive, g(t) < g(0).
A more sophisticated approach is to employ the second order Taylor polynomial to
approximate the function and then use its minimum (assuming such exists) as the next
approximation to the desired minimizer. Specifically, if xk is the current approximation
to the minimum, then we approximate
F (u) ≈ c(k) + (u − u(k) )T g(k) + 21 (u − u(k) )T H (k) (u − u(k) ) (4.35)
near u(k) , where

c(k) = F (u(k) ), g(k) = ∇F (u(k) ), H (k) = ∇2 F (u(k) ), (4.36)

are, respectively, the function value, the gradient and the Hessian at the current iterate.
If ∇2 F (x⋆ ) is positive definite, then, assuming xk is close, so is Hk = ∇2 F (xk ). Thus,
the quadratic Taylor approximation has a unique minimum value xk+1 which satisfies the
linear system
H (k) (u(k+1) − u(k) ) = − g(k) . (4.37)
The solution serves to define the next approximation u(k+1) . While not guaranteed to
converge, the method does perform well in all reasonable situations.

10/23/22 59 c 2022 Peter J. Olver

 References

[1] Alligood, K.T., Sauer, T.D., and Yorke, J.A., Chaos. An Introduction to Dynamical
Systems, Springer-Verlag, New York, 1997.
[2] Apostol, T.M., Mathematical Analysis, 2nd ed., Addison–Wesley Publ. Co.,
Reading, Mass., 1974.
[3] Bradie, B., A Friendly Introduction to Numerical Analysis, Prentice–Hall, Inc.,
Upper Saddle River, N.J., 2006.
[4] Burden, R.L., and Faires, J.D., Numerical Analysis, Seventh Edition, Brooks/Cole,
Pacific Grove, CA, 2001.
[5] Devaney, R.L., An Introduction to Chaotic Dynamical Systems, Addison–Wesley,
Redwood City, Calif., 1989.
[6] Feigenbaum, M.J., Qualitative universality for a class of nonlinear transformations,
J. Stat. Phys. 19 (1978), 25–52.
[7] Gaal, L., Classical Galois theory, 4th ed., Chelsea Publ. Co., New York, 1988.
[8] Greene, B., The Elegant Universe: Superstrings, Hidden Dimensions, and the Quest
for the Ultimate Theory, W. W. Norton, New York, 1999..
[9] Henry, D., Geometric Theory of Semilinear Parabolic Equations, Lecture Notes in
Math., vol. 840, Springer–Verlag, Berlin, 1981.
[10] Lanford, O., A computer-assisted proof of the Feigenbaum conjecture, Bull. Amer.
Math. Soc. 6 (1982), 427–434.
[11] Mandelbrot, B.B., The Fractal Geometry of Nature, W.H. Freeman, New York,
1983.
[12] Marsden, J.E., and Tromba, A.J., Vector Calculus, 4th ed., W.H. Freeman, New
York, 1996.
[13] Moon, F.C., Chaotic Vibrations, John Wiley & Sons, New York, 1987.
[14] Olver, P.J., Continuous Calculus, Lecture Notes, University of Minnesota, 2020.
[15] Olver, P.J., and Shakiban, C., Applied Linear Algebra, Prentice–Hall, Inc., Upper
Saddle River, N.J., 2005.
[16] Peitgen, H.-O., and Richter, P.H., The Beauty of Fractals: Images of Complex
Dynamical Systems,Springer–Verlag, New York, 1986.
[17] Press, W.H., Teukolsky, S.A., Vetterling, W.T., and Flannery, B.P., Numerical
Recipes: The Art of Scientific Computing, 3rd ed., Cambridge University Press,
Cambridge, 2007.
[18] Robinson, R.C., An Introduction to Dynamical Systems: Continuous and Discrete,
2nd ed., Pure and Applied Undergraduate Texts, vol. 19, Amer. Math. Soc.,
Providence, R.I., 2012.
[19] Royden, H.L., Real Analysis, Macmillan Co., New York, 1988.

10/23/22 60 c 2022 Peter J. Olver