Part IB — Numerical Analysis

Based on lectures by G. Moore

Notes taken by Dexter Chua

Lent 2016

These notes are not endorsed by the lecturers, and I have modified them (often
significantly) after lectures. They are nowhere near accurate representations of what
was actually lectured, and in particular, all errors are almost surely mine.

Polynomial approximation
Interpolation by polynomials. Divided differences of functions and relations to deriva-
tives. Orthogonal polynomials and their recurrence relations. Least squares approx-
imation by polynomials. Gaussian quadrature formulae. Peano kernel theorem and
applications. [6]

Computation of ordinary differential equations

Euler’s method and proof of convergence. Multistep methods, including order, the root
condition and the concept of convergence. Runge-Kutta schemes. Stiff equations and
A-stability. [5]

Systems of equations and least squares calculations

LU triangular factorization of matrices. Relation to Gaussian elimination. Column
pivoting. Factorizations of symmetric and band matrices. The Newton-Raphson
method for systems of non-linear algebraic equations. QR factorization of rectangular
matrices by Gram-Schmidt, Givens and Householder techniques. Application to linear
least squares calculations. [5]

1
Contents IB Numerical Analysis

Contents
0 Introduction 3

1 Polynomial interpolation 4
1.1 The interpolation problem . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Lagrange formula . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 The Newton formula . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 A useful property of divided differences . . . . . . . . . . . . . . 7
1.5 Error bounds for polynomial interpolation . . . . . . . . . . . . . 8

2 Orthogonal polynomials 12
2.1 Scalar product . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Orthogonal polynomials . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Three-term recurrence relation . . . . . . . . . . . . . . . . . . . 14
2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Least-squares polynomial approximation . . . . . . . . . . . . . . 16

3 Approximation of linear functionals 18

3.1 Linear functionals . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Gaussian quadrature . . . . . . . . . . . . . . . . . . . . . . . . . 19

4 Expressing errors in terms of derivatives 22

5 Ordinary differential equations 26

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.2 One-step methods . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3 Multi-step methods . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4 Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . . . . . 34

6 Stiff equations 37
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.2 Linear stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

7 Implementation of ODE methods 40

7.1 Local error estimation . . . . . . . . . . . . . . . . . . . . . . . . 40
7.2 Solving for implicit methods . . . . . . . . . . . . . . . . . . . . . 41

8 Numerical linear algebra 42

8.1 Triangular matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 42
8.2 LU factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
8.3 A = LU for special A . . . . . . . . . . . . . . . . . . . . . . . . 46

9 Linear least squares 50

2
0 Introduction IB Numerical Analysis

0 Introduction
Numerical analysis is the study of algorithms. There are many problems we
would like algorithms to solve. In this course, we will tackle the problems of
polynomial approximation, solving ODEs and solving linear equations. These
are all problems that frequently arise when we do (applied) maths.
In general, there are two things we are concerned with — accuracy and speed.
Accuracy is particularly important in the cases of polynomial approximation and
solving ODEs, since we are trying approximate things. We would like to make
good approximations of the solution with relatively little work. On the other
hand, in the case of solving linear equations, we are more concerned with speed
— our solutions will be exact (up to numerical errors due to finite precision of
calculations), but we would like to solve it quickly. We might have to deal with
huge systems, and we don’t want the computation time to grow too quickly.
In the past, this was an important subject, since they had no computers.
The algorithms had to be implemented by hand. It was thus very important to
find some practical and efficient methods of computing things, or else it would
take them forever to calculate what they wanted. So we wanted quick algorithms
that give reasonably accurate results.
Nowadays, this is still an important subject. While we have computers that
are much faster at computation, we still want our programs to be fast. We would
also want to get really accurate results, since we might be using them to, say,
send our rocket to the Moon. Moreover, with more computational power, we
might sacrifice efficiency for some other desirable properties. For example, if we
are solving for the trajectory of a particle, we might want the solution to satisfy
the conservation of energy. This would require some much more complicated and
slower algorithms that no one would have considered in the past. Nowadays, with
computers, these algorithms become more feasible, and are becoming increasingly
more popular.

3
1 Polynomial interpolation IB Numerical Analysis

1 Polynomial interpolation
Polynomials are nice. Writing down a polynomial of degree n involves only n + 1
numbers. They are easy to evaluate, integrate and differentiate. So it would be
nice if we can approximate things with polynomials. For simplicity, we will only
deal with real polynomials.
Notation. We write Pn [x] for the real linear vector space of polynomials (with
real coefficients) having degree n or less.
It is easy to show that dim(Pn [x]) = n + 1.

1.1 The interpolation problem

The idea of polynomial interpolation is that we are given n + 1 distinct interpo-
lation points {xi }ni=0 ⊆ R, and n + 1 data values {fi }ni=0 ⊆ R. The objective is
to find a p ∈ Pn [x] such that

p(xi ) = fi for i = 0, · · · , n.

In other words, we want to fit a polynomial through the points (xi , fi ).

There are many situations where this may come up. For example, we may
be given n + 1 actual data points, and we want to fit a polynomial through the
points. Alternatively, we might have a complicated function f , and want to
approximate it with a polynomial p such that p and f agree on at least that
n + 1 points.
The naive way of looking at this is that we try a polynomial

p(x) = an xn + an−1 xn−1 + · · · + a0 ,

and then solve the system of equations

fi = p(xi ) = an xni + an−1 xn−1

i + · · · + a0 .

This is a perfectly respectable system of n + 1 equations in n + 1 unknowns.

From linear algebra, we know that in general, such a system is not guaranteed
to have a solution, and if the solution exists, it is not guaranteed to be unique.
That was not helpful. So our first goal is to show that in the case of polynomial
interpolation, the solution exists and is unique.

1.2 The Lagrange formula

It turns out the problem is not too hard. You can probably figure it out yourself
if you lock yourself in a room for a few days (or hours). In fact, we will come up
with two solutions to the problem.
The first is via the Lagrange cardinal polynomials. These are simple to
state, and it is obvious that these work very well. However, practically, this is
not the best way to solve the problem, and we will not talk about them much.
Instead, we use this solution as a proof of existence of polynomial interpolations.
We will then develop our second method on the assumption that polynomial
interpolations exist, and find a better way of computing them.

4
1 Polynomial interpolation IB Numerical Analysis

Definition (Lagrange cardinal polynomials). The Lagrange cardinal polynomials

with respect to the interpolation points {xi }ni=0 are, for k = 0, · · · , n,
n
Y x − xi
`k (x) = .
xk − xi
i=0,i6=k

Note that these polynomials have degree exactly n. The significance of these
polynomials is we have `k (xi ) = 0 for i 6= k, and `k (xk ) = 1. In other words, we
have
`k (xj ) = δjk .
This is obvious from definition.
With these cardinal polynomials, we can immediately write down a solution
to the interpolation problem.
Theorem. The interpolation problem has exactly one solution.
Proof. We define p ∈ Pn [x] by
n
X
p(x) = fk `k (x).
k=0

Evaluating at xi gives
n
X n
X
p(xj ) = fk `k (xj ) = fk δjk = fj .
k=0 k=0

So we get existence.
For uniqueness, suppose p, q ∈ Pn [x] are solutions. Then the difference
r = p − q ∈ Pn [x] satisfies r(xj ) = 0 for all j, i.e. it has n + 1 roots. However, a
non-zero polynomial of degree n can have at most n roots. So in fact p − q is
zero, i.e. p = q.
While this works, it is not ideal. If we one day decide we should add one more
interpolation point, we would have to recompute all the cardinal polynomials,
and that is not fun. Ideally, we would like some way to reuse our previous
computations when we have new interpolation points.

1.3 The Newton formula

The idea of Newton’s formula is as follows — for k = 0, · · · , n, we write pk ∈ Pk [x]
for the polynomial that satisfies

pk (xi ) = fi for i = 0, · · · , k.

This is the unique degree-k polynomial that satisfies the first k + 1 conditions,
whose existence (and uniqueness) is guaranteed by the previous section. Then
we can write

p(x) = pn (x) = p0 (x) + (p1 (x) − p0 (x)) + · · · + (pn (x) − pn−1 (x)).

Hence we are done if we have an efficient way of finding the differences pk − pk−1 .

5
1 Polynomial interpolation IB Numerical Analysis

We know that pk and pk−1 agree on x0 , · · · , xk−1 . So pk − pk−1 evaluates to

0 at those points, and we must have
k−1
Y
pk (x) − pk−1 (x) = Ak (x − xi ),
i=0

for some Ak yet to be found out. Then we can write

n
X k−1
Y
p(x) = pn (x) = A0 + Ak (x − xi ).
k=1 i=0

This formula has the advantage that it is built up gradually from the interpo-
lation points one-by-one. If we stop the sum at any point, we have obtained
the polynomial that interpolates the data for the first k points (for some k).
Conversely, if we have a new data point, we just need to add a new term, instead
of re-computing everything.
All that remains is to find the coefficients Ak . For k = 0, we know A0 is the
unique constant polynomial that interpolates the point at x0 , i.e. A0 = f0 .
For the others, we note that in the formula for pk − pk−1 , we find that Ak
is the leading coefficient of xk . But pk−1 (x) has no degree k term. So Ak must
be the leading coefficient of pk . So we have reduced our problem to finding the
leading coefficients of pk .
The solution to this is known as the Newton divided differences. We first
invent a new notation:
Ak = f [x0 , · · · , xk ].
Note that these coefficients depend only on the first k interpolation points.
Moreover, since the labelling of the points x0 , · · · , xk is arbitrary, we don’t have
to start with x0 . In general, the coefficient
f [xj , · · · , xk ]
is the leading coefficient of the unique q ∈ Pk−j [x] such that q(xi ) = fi for
i = j, · · · , k.
While we do not have an explicit formula for what these coefficients are, we
can come up with a recurrence relation for these coefficients.
Theorem (Recurrence relation for Newton divided differences). For 0 ≤ j <
k ≤ n, we have
f [xj+1 , · · · , xk ] − f [xj , · · · , xk−1 ]
f [xj , · · · , xk ] = .
xk − xj
Proof. The key to proving this is to relate the interpolating polynomials. Let
q0 , q1 ∈ Pk−j−1 [x] and q2 ∈ Pk−j satisfy
q0 (xi ) = fi i = j, · · · , k − 1
q1 (xi ) = fi i = j + 1, · · · , k
q2 (xi ) = fi i = j, · · · , k
We now claim that
x − xj xk − x
q2 (x) = q1 (x) + q0 (x).
xk − xj xk − xj

6
1 Polynomial interpolation IB Numerical Analysis

We can check directly that the expression on the right correctly interpolates
the points xi for i = j, · · · , k. By uniqueness, the two expressions agree. Since
f [xj , · · · , xk ], f [xj+1 , · · · , xk ] and f [xj , · · · , xk−1 ] are the leading coefficients of
q2 , q1 , q0 respectively, the result follows.
Thus the famous Newton divided difference table can be constructed
xi fi f [∗, ∗] f [∗, ∗, ∗] ··· f [∗, · · · , ∗]
x0 f [x0 ]
f [x0 , x1 ]
x1 f [x1 ] f [x0 , x1 , x2 ]
..
f [x1 , x2 ] .
x2 f [x2 ] f [x2 , x3 , x4 ] ··· f [x0 , x1 , · · · , xn ]
.
f [x2 , x3 ] ..
.
x3 f [x3 ] ..
.. .. .
. . ..
xn f [xn ]
From the first n columns, we can find the n + 1th column using the recurrence
relation above. The values of Ak can then be found at the top diagonal, and
this is all we really need. However, to compute this diagonal, we will need to
compute everything in the table.
In practice, we often need not find the actual interpolating polynomial. If we
just want to evaluate p(x̂) at some new point x̂ using the divided table, we can
use Horner’s scheme, given by
S <- f[x0 ,..., xn ]
for k = n - 1,..., 0
S <- (^x - xk )S + f[x0 ,..., xk ]
end
This only takes O(n) operations.
If an extra data point {xn+1 , fn+1 } is added, then we only have to compute
an extra diagonal f [xk , · · · , xn+1 ] for k = n, · · · , 0 in the divided difference table
to obtain the new coefficient, and the old results can be reused. This requires
O(n) operations. This is less straightforward for Lagrange’s method.

1.4 A useful property of divided differences

In the next couple of sections, we are interested in the error of polynomial inter-
polation. Suppose the data points come from fi = f (xi ) for some complicated
f we want to approximate, and we interpolate f at n data points {xi }ni=1 with
pn . How does the error en (x) = f (x) − pn (x) depend on n and the choice of
interpolation points?
At the interpolation point, the error is necessarily 0, by definition of interpo-
lation. However, this does not tell us anything about the error elsewhere.
Our ultimate objective is to bound the error en by suitable multiples of the
derivatives of f . We will do this in two steps. We first relate the derivatives to
the divided differences, and then relate the error to the divided differences.
The first part is an easy result based on the following purely calculus lemma.

7
1 Polynomial interpolation IB Numerical Analysis

Lemma. Let g ∈ C m [a, b] have a continuous mth derivative. Suppose g is zero

at m + ` distinct points. Then g (m) has at least ` distinct zeros in [a, b].
Proof. This is a repeated application of Rolle’s theorem. We know that between
every two zeros of g, there is at least one zero of g 0 ∈ C m−1 [a, b]. So by
differentiating once, we have lost at most 1 zeros. So after differentiating m
times, g (m) has lost at most m zeros. So it still has at least ` zeros.
Theorem. Let {xi }ni=0 ∈ [a, b] and f ∈ C n [a, b]. Then there exists some
ξ ∈ (a, b) such that
1
f [x0 , · · · , xn ] = f (n) (ξ).
n!
Proof. Consider e = f − pn ∈ C n [a, b]. This has at least n + 1 distinct zeros in
(n)
[a, b]. So by the lemma, e(n) = f (n) − pn must vanish at some ξ ∈ (a, b). But
(n)
then pn = n!f [x0 , · · · , xn ] constantly. So the result follows.

1.5 Error bounds for polynomial interpolation

The actual error bound is not too difficult as well. It turns out the error e = f −pn
is “like the next term in the Newton’s formula”. This vague sentence is made
precise in the following theorem:
Theorem. Assume {xi }ni=0 ⊆ [a, b] and f ∈ C[a, b]. Let x̄ ∈ [a, b] be a non-
interpolation point. Then

en (x̄) = f [x0 , x1 , · · · , xn , x̄]ω(x̄),

where
n
Y
ω(x) = (x − xi ).
i=0

Note that we forbid the case where x̄ is an interpolation point, since it is

not clear what the expression f [x0 , x1 , · · · , xn , x̄] means. However, if x̄ is an
interpolation point, then both en (x̄) and ω(x̄) are zero, so there isn’t much to
say.
Proof. We think of x̄ = xn+1 as a new interpolation point so that

pn+1 (x) − pn (x) = f [x0 , · · · , xn , x̄]ω(x)

for all x ∈ R. In particular, putting x = x̄, we have pn+1 (x̄) = f (x̄), and we get
the result.
Combining the two results, we find
Theorem. If in addition f ∈ C n+1 [a, b], then for each x ∈ [a, b], we can find
ξx ∈ (a, b) such that
1
en (x) = f (n+1) (ξx )ω(x)
(n + 1)!
Proof. The statement is trivial if x is an interpolation point — pick arbitrary
ξx , and both sides are zero. Otherwise, this follows directly from the last two
theorems.

8
1 Polynomial interpolation IB Numerical Analysis

This is an exact result, which is not too useful, since there is no easy
constructive way of finding what ξx should be. Instead, we usually go for a
bound. We introduce the max norm

kgk∞ = max |g(t)|.

t∈[a,b]

This gives the more useful bound

Corollary. For all x ∈ [a, b], we have
1
|f (x) − pn (x)| ≤ kf (n+1) k∞ |ω(x)|
(n + 1)!

Assuming our function f is fixed, this error bound depends only on ω(x),
which depends on our choice of interpolation points. So can we minimize ω(x)
in some sense by picking some clever interpolation points ∆ = {xi }ni=0 ? Here
we will have n fixed. So instead, we put ∆ as the subscript. We can write our
bound as
1
kf − p∆ k∞ ≤ kf (n+1) k∞ kω∆ k∞ .
(n + 1)!
So the objective is to find a ∆ that minimizes kω∆ k∞ .
For the moment, we focus on the special case where the interval is [−1, 1].
The general solution can be obtained by an easy change of variable.
For some magical reasons that hopefully will become clear soon, the optimal
choice of ∆ comes from the Chebyshev polynomials.
Definition (Chebyshev polynomial). The Chebyshev polynomial of degree n on
[−1, 1] is defined by
Tn (x) = cos(nθ),
where x = cos θ with θ ∈ [0, π].
So given an x, we find the unique θ that satisfies x = cos θ, and then find
cos(nθ). This is in fact a polynomial in disguise, since from trigonometric
identities, we know cos(nθ) can be expanded as a polynomial in cos θ up to
degree n.
Two key properties of Tn on [−1, 1] are
(i) The maximum absolute value is obtained at
 
πk
Xk = cos
n
for k = 0, · · · , n with
Tn (Xk ) = (−1)k .

(ii) This has n distinct zeros at

 
2k − 1
xk = cos π .
2n
for k = 1, · · · , n.
When plotted out, the polynomials look like this:

9
1 Polynomial interpolation IB Numerical Analysis

T4 (x)

x
−1 1

−1

All that really matters about the Chebyshev polynomials is that the maximum
is obtained at n + 1 distinct points with alternating sign. The exact form of the
polynomial is not really important.
Notice there is an intentional clash between the use of xk as the zeros and
xk as the interpolation points — we will show these are indeed the optimal
interpolation points.
We first prove a convenient recurrence relation for the Chebyshev polynomials:
Lemma (3-term recurrence relation). The Chebyshev polynomials satisfy the
recurrence relations
Tn+1 (x) = 2xTn (x) − Tn−1 (x)
with initial conditions
T0 (x) = 1, T1 (x) = x.
Proof.
cos((n + 1)θ) + cos((n − 1)θ) = 2 cos θ cos(nθ).
This recurrence relation can be useful for many things, but for our purposes,
we only use it to show that the leading coefficient of Tn is 2n−1 (for n ≥ 1).
Theorem (Minimal property for n ≥ 1). On [−1, 1], among all polynomials
1
p ∈ Pn [x] with leading coefficient 1, 2n−1 kTn k minimizes kpk∞ . Thus, the
1
minimum value is 2n−1 .
Proof. We proceed by contradiction. Suppose there is a polynomial qn ∈ Pn
1
with leading coefficient 1 such that kqn k∞ < 2n−1 . Define a new polynomial
1
r= Tn − qn .
2n−1
This is, by assumption, non-zero.
Since both the polynomials have leading coefficient 1, the difference must
1 1
have degree at most n − 1, i.e. r ∈ Pn−1 [x]. Since 2n−1 Tn (Xk ) = ± 2n−1 , and
1
|qn (Xn )| < 2n−1 by assumption, r alternates in sign between these n + 1 points.
But then by the intermediate value theorem, r has to have at least n zeros. This
is a contradiction, since r has degree n − 1, and cannot be zero.

10
1 Polynomial interpolation IB Numerical Analysis

Corollary. Consider
n
Y
w∆ = (x − xi ) ∈ Pn+1 [x]
i=0

for any distinct points ∆ = {xi }ni=0 ⊆ [−1, 1]. Then

1
min kω∆ k∞ = .
∆ 2n
This minimum is achieved by picking the interpolation points to be the zeros of
Tn+1 , namely  
2k + 1
xk = cos π , k = 0, · · · , n.
2n + 2
Theorem. For f ∈ C n+1 [−1, 1], the Chebyshev choice of interpolation points
gives
1 1
kf − pn k∞ ≤ n kf (n+1) k∞ .
2 (n + 1)!
Suppose f has as many continuous derivatives as we want. Then as we
increase n, what happens to the error bounds? The coefficients involve dividing
by an exponential and a factorial. Hence as long as the higher derivatives of f
don’t blow up too badly, in general, the error will tend to zero as n → ∞, which
makes sense.
The last two results can be easily generalized to arbitrary intervals [a, b], and
this is left as an exercise for the reader.

11
2 Orthogonal polynomials IB Numerical Analysis

2 Orthogonal polynomials
It turns out the Chebyshev polynomials is just an example of a more general
class of polynomials, known as orthogonal polynomials. As in linear algebra, we
can define a scalar product on the space of polynomials, and then find a basis
of orthogonal polynomials of the vector space under this scalar product. We
shall show that each set of orthogonal polynomials has a three-term recurrence
relation, just like the Chebyshev polynomials.

2.1 Scalar product

The scalar products we are interested in would be generalization of the usual
scalar product on Euclidean space,
n
X
hx, yi = xi yi .
i=1

We want to generalize this to vector spaces of functions and polynomials. We

will not provide a formal definition of vector spaces and scalar products on an
abstract vector space. Instead, we will just provide some examples of commonly
used ones.
Example.
(i) Let V = C s [a, b], where [a, b] is a finite interval and s ≥ 0. Pick a weight
function w(x) ∈ C(a, b) such that w(x) > 0 for all x ∈ (a, b), and w is
integrable over [a, b]. In particular, we allow w to vanish at the end points,
or blow up mildly such that it is still integrable.
We then define the inner product to be
Z b
hf, gi = w(x)f (x)d(x) dx.
a

(ii) We can allow [a, b] to be infinite, e.g. [0, ∞) or even (−∞, ∞), but we have
to be more careful. We first define
Z b
hf, gi = w(x)f (x)g(x) dx
a
Rb
as before, but we now need more conditions. We require that a w(x)xn dx
to exist for all n ≥ 0, since we want to allow polynomials in our vector
2
space. For example, w(x) = e−x on [0, ∞), works, or w(x) = e−x on
(−∞, ∞). These are scalar products for Pn [x] for n ≥ 0, but we cannot
extend this definition to all smooth functions since they might blow up too
fast at infinity. We will not go into the technical details, since we are only
interested in polynomials, and knowing it works for polynomials suffices.
(iii) We can also have a discrete inner product, defined by
m
X
hf, gi = wj f (ξj )g(ξj )
j=1

12
2 Orthogonal polynomials IB Numerical Analysis

with {ξj }m m
j=1 distinct points and {wj }j=1 > 0. Now we have to restrict
ourselves a lot. This is a scalar product for V = Pm−1 [x], but not for
higher degrees, since a scalar product should satisfy hf, f i > 0 for f 6= 0.
In particular, we cannot extend this to all smooth functions.
With an inner product, we can define orthogonality.
Definition (Orthogonalilty). Given a vector space V and an inner product
h · , · i, two vectors f, g ∈ V are orthogonal if hf, gi = 0.

2.2 Orthogonal polynomials

Definition (Orthogonal polynomial). Given a vector space V of polynomials
and inner product h · , · i, we say pn ∈ Pn [x] is the nth orthogonal polynomial if

hpn , qi = 0 for all q ∈ Pn−1 [x].

In particular, hpn , pm i = 0 for n 6= m.

We said the orthogonal polynomial, but we need to make sure such a polyno-
mial has to be unique. It is clearly not unique, since if pn satisfies these relations,
then so does λpn for all λ =6 0. For uniqueness, we need to impose some scaling.
We usually do so by requiring the leading polynomial to be 1, i.e. it is monic.
Definition (Monic polynomial). A polynomial p ∈ Pn [x] is monic if the coeffi-
cient of xn is 1.
In practice, most famous traditional polynomials are not monic. They have
a different scaling imposed. Still, as long as we have some scaling, we will have
uniqueness.
We will not mess with other scalings, and stick to requiring them to be monic
since this is useful for proving things.
Theorem. Given a vector space V of functions and an inner product h · , · i,
there exists a unique monic orthogonal polynomial for each degree n ≥ 0. In
addition, {pk }nk=0 form a basis for Pn [x].
Proof. This is a big induction proof over both parts of the theorem. We induct
over n. For the base case, we pick p0 (x) = 1, which is the only degree-zero monic
polynomial.
Now suppose we already have {pn }nk=0 satisfying the induction hypothesis.
Now pick any monic qn+1 ∈ Pn+1 [x], e.g. xn+1 . We now construct pn+1 from
qn+1 by the Gram-Schmidt process. We define
n
X hqn+1 , pk i
pn+1 = qn+1 − pk .
hpk , pk i
k=0

This is again monic since qn+1 is, and we have

hpn+1 , pm i = 0

for all m ≤ n, and hence hpn+1 , pi = 0 for all p ∈ Pn [x] = hp0 , · · · , pn i.

13
2 Orthogonal polynomials IB Numerical Analysis

To obtain uniqueness, assume both pn+1 , p̂n+1 ∈ Pn+1 [x] are both monic
orthogonal polynomials. Then r = pn+1 − p̂n+1 ∈ Pn [x]. So

hr, ri = hr, pn+1 − p̂n+1 i = hr, pn+1 i − hr, p̂n+1 i = 0 − 0 = 0.

So r = 0. So pn+1 = p̂n−1 .
Finally, we have to show that p0 , · · · , pn+1 form a basis for Pn+1 [x]. Now
note that every p ∈ Pn+1 [x] can be written uniquely as

p = cpn+1 + q,

where q ∈ Pn [x]. But {pk }nk=0 is a basis for Pn [x]. So q can be uniquely
decomposed as a linear combination of p0 , · · · , pn .
Alternatively, this follows from the fact that any set of orthogonal vectors
must be linearly independent, and since there are n + 2 of these vectors and
Pn+1 [x] has dimension n + 2, they must be a basis.
In practice, following the proof naively is not the best way of producing the
new pn+1 . Instead, we can reduce a lot of our work by making a clever choice of
qn+1 .

2.3 Three-term recurrence relation

Recall that for the Chebyshev polynomials, we obtained a three-term recurrence
relation for them using special properties of the cosine. It turns out these
recurrence relations exist in general for orthogonal polynomials.
We start by picking qn+1 = xpn in the previous proof. We now use the fact
that
hxf, gi = hf, xgi.
This is not necessarily true for arbitrary inner products, but for most sensible
inner products we will meet in this course, this is true. In particular, it is clearly
true for inner products of the form
Z
hf, gi = w(x)f (x)g(x) dx.

Assuming this, we obtain the following theorem.

Theorem. Monic orthogonal polynomials are generated by

pk+1 (x) = (x − αk )pk (x) − βk pk−1 (x)

with initial conditions

p0 = 1, p1 (x) = (x − α0 )p0 ,

where
hxpk , pk i hpk , pk i
αk = , βk = .
hpk , pk i hpk−1 , pk−1 i
Proof. By inspection, the p1 given is monic and satisfies

hp1 , p0 i = 0.

14
2 Orthogonal polynomials IB Numerical Analysis

Using qn+1 = xpn in the Gram-Schmidt process gives

n
X hxpn , pk i
pn+1 = xpn − pk
hpk , pk i
k=0
n
X hpn , xpk i
pn+1 = xpn − pk
hpk , pk i
k=0

We notice that hpn , xpk i and vanishes whenever xpk has degree less than n. So
we are left with
hxpn , pn i hpn , xpn−1 i
= xpn − pn − pn−1
hpn , pn i hpn−1 , pn−1 i
hpn , xpn−1 i
= (x − αn )pn − pn−1 .
hpn−1 , pn−1 i

Now we notice that xpn−1 is a monic polynomial of degree n so we can write

this as xpn−1 = pn + q. Thus

hpn , xpn−1 i = hpn , pn + qi = hpn , pn i.

Hence the coefficient of pn−1 is indeed the β we defined.

2.4 Examples
The four famous examples are the Legendre polynomials, Chebyshev polynomials,
Laguerre polynomials and Hermite polynomials. We first look at how the
Chebyshev polynomials fit into this framework.
Chebyshev is based on the scalar product defined by
Z 1
1
hf, gi = √ f (x)g(x) dx.
−1 1 − x2
Note that the weight function blows up mildly at the end, but this is fine since
it is still integrable.
This links up with
Tn (x) = cos(nθ)
for x = cos θ via the usual trigonometric substitution. We have
Z π
1
hTn , Tm i = √ cos(nθ) cos(mθ) sin θ dθ
0 1 − cos2 θ
Z π
= cos(nθ) cos(mθ) dθ
0
= 0 if m 6= n.

The other orthogonal polynomials come from scalar products of the form
Z b
hf, gi = w(x)f (x)g(x) dx,
a

as described in the table below:

15
2 Orthogonal polynomials IB Numerical Analysis

Type Range Weight

Legendre [−1, 1] 1
Chebyshev [−1, 1] √ 1
1−x2
−x
Laguerre [0, ∞) e
2
Hermite (−∞, ∞) e−x

2.5 Least-squares polynomial approximation

If we want to approximate a function with a polynomial, polynomial interpolation
might not be the best idea, since all we do is make sure the polynomial agrees
with f at certain points, and then hope it is a good approximation elsewhere.
Instead, the idea is to choose a polynomial p in Pn [x] that “minimizes the error”.
What exactly do we mean by minimizing the error? The error is defined as
the function f − p. So given an appropriate inner product on the vector space of
continuous functions, we want to minimize

kf − pk2 = hf − p, f − pi.

This is usually of the form

Z b
hf − p, f − pi = w(x)[f (x) − p(x)]2 dx,
a

but we can also use alternative inner products such as

m
X
hf − p, f − pi = wj [f (ξi ) − p(ξi )]2 .
j=1

Unlike polynomial interpolation, there is no guarantee that the approximation

agrees with the function anywhere. Unlike polynomial interpolation, there is
some guarantee that the total error is small (or as small as we can get, by
definition). In particular, if f is continuous, then the Weierstrass approximation
theorem tells us the total error must eventually vanish.
Unsurprisingly, the solution involves the use of the orthogonal polynomials
with respect to the corresponding inner products.
Theorem. If {pn }nk=0 are orthogonal polynomials with respect to h · , · i, then
the choice of ck such that
X n
p= ck pk
k=0
2
minimizes kf − pk is given by

hf, pk i
ck = ,
kpk k2

and the formula for the error is

n
X hf, pk i2
kf − pk2 = kf k2 − .
kpk k2
k=0

16
2 Orthogonal polynomials IB Numerical Analysis

Note that the solution decouples, in the sense that ck depends only on f and
pk . If we want to take one more term, we just need to compute an extra term,
and not redo our previous work.
Also, we notice that the formula for the error is a positive term kf k2 sub-
tracting a lot of squares. As we increase n, we subtract more squares, and the
error decreases. If we are lucky, the error tends to 0 as we take n → ∞. Even
though we might not know how many terms we need in order to get the error to
be sufficiently small, we can just keep adding terms until the computed error
small enough (which is something we have to do anyway even if we knew what
n to take).
Proof. We consider a general polynomial
n
X
p= ck pk .
k=0

We substitute this in to obtain

n
X n
X
hf − p, f − pi = hf, f i − 2 ck hf, pk i + c2k kpk k2 .
k=0 k=0

Note that there are no cross terms between the different coefficients. We minimize
this quadratic by setting the partial derivatives to zero:
∂
0= hf − p, f − pi = −2hf, pk i + 2ck kpk k2 .
∂ck
To check this is indeed a minimum, note that the Hessian matrix is simply 2I,
which is positive definite. So this is really a minimum. So we get the formula for
the ck ’s as claimed, and putting the formula for ck gives the error formula.
Note that our constructed p ∈ Pn [x] has a nice property: for k ≤ n, we have

hf, pk i
hf − p, pk i = hf, pk i − hp, pk i = hf, pk i − hpk , pk i = 0.
kpk k2

Thus for all q ∈ Pn [x], we have

hf − p, qi = 0.

In particular, this is true when q = p, and tells us hf, pi = hp, pi. Using this to
expand hf − p, f − pi gives

kf − pk2 + kpk2 = kf k2 ,

which is just a glorified Pythagoras theorem.

17
3 Approximation of linear functionals IB Numerical Analysis

3 Approximation of linear functionals

3.1 Linear functionals
In this chapter, we are going to study approximations of linear functions. Before
we start, it is helpful to define what a linear functional is, and look at certain
examples of these.

Definition (Linear functional). A linear functional is a linear mapping L : V →

R, where V is a real vector space of functions.
In generally, a linear functional is a linear mapping from a vector space to
its underlying field of scalars, but for the purposes of this course, we will restrict
to this special case.
We usually don’t put so much emphasis on the actual vector space V . Instead,
we provide a formula for L, and take V to be the vector space of functions for
which the formula makes sense.
Example.
(i) We can choose some fixed ξ ∈ R, and define a linear functional by

L(f ) = f (ξ).

(ii) Alternatively, for fixed η ∈ R we can define our functional by

L(f ) = f 0 (η).

In this case, we need to pick a vector space in which this makes sense, e.g.
the space of continuously differentiable functions.
(iii) We can define
Z b
L(f ) = f (x) dx.
a

The set of continuous (or even just integrable) functions defined on [a, b]
will be a sensible domain for this linear functional.
(iv) Any linear combination of these linear functions are also linear functionals.
For example, we can pick some fixed α, β ∈ R, and define
β−α 0
L(f ) = f (β) − f (α) − [f (β) + f 0 (α)].
2

The objective of this chapter is to construct approximations to more compli-

cated linear functionals (usually integrals, possibly derivatives point values) in
terms of simpler linear functionals (usually point values of f itself).
For example, we might produce an approximation of the form
N
X
L(f ) ≈ ai f (xi ),
i=0

where V = C p [a, b], p ≥ 0, and {xi }N

i=0 ⊆ [a, b] are distinct points.

18
3 Approximation of linear functionals IB Numerical Analysis

How can we choose the coefficients ai and the points xi so that our approxi-
mation is “good”?
We notice that most of our functionals can be easily evaluated exactly when
f is a polynomial. So we might approximate our function f by a polynomial,
and then do it exactly for polynomials.
More precisely, we let {xi }N
i=0 ⊆ [a, b] be arbitrary points. Then using the
Lagrange cardinal polynomials `i , we have
N
X
f (x) ≈ f (xi )`i (x).
i=0

Then using linearity, we can approximate

N
! N
X X
L(f ) ≈ L f (xi )`i (x) = L(`i )f (xi ).
i=0 i=0

So we can pick
ai = L(`i ).
Similar to polynomial interpolation, this formula is exact for f ∈ PN [x]. But we
could do better. If we can freely choose {ai }N N
i=0 and {xi }i=0 , then since we now
have 2n + 2 free parameters, we might expect to find an approximation that is
exact for f ∈ P2N +1 [x]. This is not always possible, but there are cases when
we can. The most famous example is Gaussian quadrature.

3.2 Gaussian quadrature

The objective of Gaussian quadrature is to approximate integrals of the form
Z b
L(f ) = w(x)f (x) dx,
a

where w(x) is a weight function that determines a scalar product.

Traditionally, we have a different set of notations used for Gaussian quadra-
ture. So just in this section, we will use some funny notation that is inconsistent
with the rest of the course.
We let Z b
hf, gi = w(x)f (x)g(x) dx
a

be a scalar product for Pν [x]. We will show that we can find weights, written
{bn }νk=1 , and nodes, written {ck }νk=1 ⊆ [a, b], such that the approximation
Z b ν
X
w(x)f (x) dx ≈ bk f (ck )
a k=1

is exact for f ∈ P2ν−1 [x]. The nodes {ck }νk=1 will turn out to be the zeros of
the orthogonal polynomial pν with respect to the scalar product. The aim of
this section is to work this thing out.
We start by showing that this is the best we can achieve.

19
3 Approximation of linear functionals IB Numerical Analysis

Proposition. There is no choice of ν weights and nodes such that the approxi-
Rb
mation of a w(x)f (x) dx is exact for all f ∈ P2ν [x].
Proof. Define
ν
Y
q(x) = (x − ck ) ∈ Pν [x].
k=1

Then we know Z b
w(x)q 2 (x) dx > 0,
a

since q 2 is always non-negative and has finitely many zeros. However,

ν
X
bk q 2 (cn ) = 0.
k=1

So this cannot be exact for q 2 .

Recall that we initially had the idea of doing the approximation by interpo-
lating f at some arbitrary points in [a, b]. We call this ordinary quadrature.
Theorem (Ordinary quadrature). For any distinct {ck }νk=1 ⊆ [a, b], let {`k }νk=1
be the Lagrange cardinal polynomials with respect to {ck }νk=1 . Then by choosing
Z b
bk = w(x)`k (x) dx,
a

the approximation
Z b ν
X
L(f ) = w(x)f (x) dx ≈ bk f (ck )
a k=1

is exact for f ∈ Pν−1 [x].

We call this method ordinary quadrature.

This simple idea is how we generate many classical numerical integration

techniques, such as the trapezoidal rule. But those are quite inaccurate. It turns
out a clever choice of {ck } does much better — take them to be the zeros of
the orthogonal polynomials. However, to do this, we must make sure the roots
indeed lie in [a, b]. This is what we will prove now — given any inner product,
the roots of the orthogonal polynomials must lie in [a, b].
Theorem. For ν ≥ 1, the zeros of the orthogonal polynomial pν are real, distinct
and lie in (a, b).
We have in fact proved this for a particular case in IB Methods, and the
same argument applies.
Proof. First we show there is at least one root. Notice that p0 = 1. Thus for
ν ≥ 1, by orthogonality, we know
Z b Z b
w(x)pν (x)p1 (x) dx = w(x)pν (x) dx = 0.
a a

20
3 Approximation of linear functionals IB Numerical Analysis

So there is at least one sign change in (a, b). We have already got the result we
need for ν = 1, since we only need one zero in (a, b).
Now for ν > 1, suppose {ξj }m j=1 are the places where the sign of pν changes
in (a, b) (which is a subset of the roots of pν ). We define
m
Y
q(x) = (x − ξj ) ∈ Pm [x].
j=1

Since this changes sign at the same place as pν , we know qpν maintains the same
sign in (a, b). Now if we had m < ν, then orthogonality gives
Z b
hq, pν i = w(x)q(x)pν (x) dx = 0,
a

which is impossible, since qpν does not change sign. Hence we must have
m = ν.
Theorem. In the ordinary quadrature, if we pick {ck }νk=1 to be the roots of
pν (x), then get we exactness for f ∈ P2ν−1 [x]. In addition, {bn }νk=1 are all
positive.
Proof. Let f ∈ P2ν−1 [x]. Then by polynomial division, we get

f = qpν + r,

where q, r are polynomials of degree at most ν − 1. We apply orthogonality to

get
Z b Z b Z b
w(x)f (x) dx = w(x)(q(x)pν (x) + r(x)) dx = w(x)r(x) dx.
a a a

Also, since each ck is a root of pν , we get

ν
X ν
X ν
X
bk f (ck ) = bk (q(ck )pν (ck ) + r(ck )) = bk r(ck ).
k=1 k=1 k=1

But r has degree at most ν − 1, and this formula is exact for polynomials in
Pν−1 [x]. Hence we know
Z b Z b ν
X ν
X
w(x)f (x) dx = w(x)r(x) dx = bk r(ck ) = bk f (ck ).
a a k=1 k=1

To show the weights are positive, we simply pick as special f . Consider f ∈

{`2k }νk=1 ⊆ P2ν−2 [x], for `k the Lagrange cardinal polynomials for {ck }νk=1 . Since
the quadrature is exact for these, we get
Z b ν
X ν
X
0< w(x)`2k (x) dx = bj `2k (cj ) = bj δjk = bk .
a j=1 j=1

Since this is true for all k = 1, · · · , ν, we get the desired result.

21
4 Expressing errors in terms of derivatives IB Numerical Analysis

4 Expressing errors in terms of derivatives

As before, we approximate a linear functional L by
n
X
L(f ) ≈ ai Li (f ),
i=0

where Li are some simpler linear functionals, and suppose this is exact for
f ∈ Pk [x] for some k ≥ 0.
Hence we know the error
n
X
eL (f ) = L(f ) − ai Li (f ) = 0
i=0

whenever f ∈ Pk [x]. We say the error annihilates for polynomials of degree less
than k.
How can we use this property to generate formulae for the error and error
bounds? We first start with a rather simple example.

Example. Let L(f ) = f (β). We decide to be silly and approximate L(f ) by

β−α 0
L(f ) ≈ f (α) + (f (β) + f 0 (α)),
2
where α 6= β. This is clearly much easier to evaluate. The error is given by
β−α 0
eL (f ) = f (β) − f (α) − (f (β) + f 0 (α)),
2
and this vanishes for f ∈ P2 [x].
How can we get a more useful error formula? We can’t just use the fact that
it annihilates polynomials of degree k. We need to introduce something beyond
this — the k + 1th derivative. We now assume f ∈ C k+1 [a, b].
Note that so far, everything we’ve done works if the interval is infinite, as long
as the weight function vanishes sufficiently quickly as we go far away. However,
for this little bit, we will need to require [a, b] to be finite, since we want to make
sure we can take the supremum of our functions.
We now seek an exact error formula in terms of f (k+1) , and bounds of the
form
|eL (f )| ≤ cL kf (k+1) k∞
for some constant cL . Moreover, we want to make cL as small as possible. We
don’t want to give a constant of 10 million, while in reality we can just use 2.
Definition (Sharp error bound). The constant cL is said to be sharp if for any
ε > 0, there is some fε ∈ C k+1 [a, b] such that

|eL (f )| ≥ (cL − ε)kfε(k+1) k∞ .

This makes it precise what we mean by cL “cannot be better”. This doesn’t

say anything about whether cL can actually be achieved. This depends on the
particular form of the question.

22
4 Expressing errors in terms of derivatives IB Numerical Analysis

To proceed, we need Taylor’s theorem with the integral remainder, i.e.

(x − a)k (k) 1 x
Z
f (x) = f (a) + (x − a)f 0 (a) + · · · + f (a) + (x − θ)k f (k+1) (θ) dθ.
k! k! a
This is not really good, since there is an x in the upper limit of the integral.
Instead, we write the integral as
Z b
(x − θ)k+ f (k+1) (θ) dθ,
a

where we define (x − θ)k+ is a new function defined by

(
k (x − θ)k x ≥ θ
(x − θ)+ =
0 x < θ.

Then if λ is a linear functional that annihilates Pk [x], then we have

!
1 b
Z
k (k+1)
λ(f ) = λ (x − θ)+ f (θ) dθ
k! a

for all f ∈ C k+1 [a, b].

For our linear functionals, we can simplify by taking the λ inside the integral
sign and obtain
1 b
Z
λ(f ) = λ((x − θ)k+ )f (k+1) (θ) dθ,
k! a
noting that λ acts on (x − θ)k+ ∈ C k−1 [a, b] as a function of x, and think of θ as
being held constant.
Of course, pure mathematicians will come up with linear functionals for
which we cannot move the λ inside, but for our linear functionals (point values,
derivative point values, integrals etc.), this is valid, as we can verify directly.
Hence we arrive at
Theorem (Peano kernel theorem). If λ annihilates polynomials of degree k or
less, then
1 b
Z
λ(f ) = K(θ)f (k+1) (θ) dθ
k! a
for all f ∈ C k+1 [a, b], where
Definition (Peano kernel). The Peano kernel is
K(θ) = λ((x − θ)k+ ).
The important thing is that the kernel K is independent of f . Taking suprema
in different ways, we obtain different forms of bounds:
Z b
|K(θ)| dθkf (k+1) k∞




 a


1  Z b ! 12
|λ(f )| ≤ .
k! 
 |K(θ)|2 dθ kf (k+1) k2
 a




kK(θ)k∞ kf (k+1) k1


23
4 Expressing errors in terms of derivatives IB Numerical Analysis

Hence we can find the constant cL for different choices of the norm. When
1
computing cL , don’t forget the factor of k! !
By fiddling with functions a bit, we can show these bounds are indeed sharp.
Example. Consider our previous example where
β−α 0
eL (f ) = f (β) − f (α) − (f (β) + f 0 (α)),
2
with exactness up to polynomials of degree 2. We wlog assume α < β. Then

K(θ) = eL ((x − θ)2+ ) = (β − θ)2+ − (α − θ)2+ − (β − α)((β − θ)+ + (α − θ)+ ).

Hence we get 
0
 a≤θ≤α
K(θ) = (α − θ)(β − θ) α ≤ θ ≤ β

0 β ≤ θ ≤ b.


Hence we know Z β
1
eL (f ) = (α − θ)(β − θ)f 000 (θ) dθ
2 α

for all f ∈ C 3 [a, b].

Note that in this particular case, our function K(θ) does not change sign on
[a, b]. Under this extra assumption, we can say a bit more.
First, we note that the bound
Z b
1
|λ(f )| ≤ K(θ) dθ kf (k+1) k∞
k! a

can be achieved by xk+1 , since this has constant k + 1th derivative. Also, we
can use the integral mean value theorem to get the bound
Z b !
1
λ(f ) = K(θ) dθ f (k+1) (ξ),
k! a

where ξ ∈ (a, b) depends on f . These are occasionally useful.

Example. Continuing our previous example, we see that K(θ) ≤ 0 on [a, b],
and Z b
1
K(θ) dθ = − (β − α)3 .
a 6
Hence we have the bound
1
|eL (f )| ≤ (β − α)3 kf 000 k∞ ,
12
and this bound is achieved for x3 . We also have
1
eL (f ) = − (β − α)3 f 000 (ξ)
12
for some f -dependent value of some ξ ∈ (a, b).

24
4 Expressing errors in terms of derivatives IB Numerical Analysis

Finally, note that Peano’s kernel theorem says if eL (f ) = 0 for all f ∈ Pk [x],
then we have
1 b
Z
eL (f ) = K(θ)f (k+1) (θ) dθ
k! a
for all f ∈ C k+1 [a, b].
But for any other fixed j = 0, · · · , k − 1, we also have eL (f ) = 0 for all
f ∈ Pj [x]. So we also know
Z b
1
eL (f ) = Kj (θ)f (j+1) (θ) dθ
j! a

for all f ∈ C j+1 [a, b]. Note that we have a different kernel.
In general, this might not be a good idea, since we are throwing information
away. Yet, this can be helpful if we get some less smooth functions that don’t
have k + 1 derivatives.

25
5 Ordinary differential equations IB Numerical Analysis

5 Ordinary differential equations

5.1 Introduction
Our next big goal is to solve ordinary differential equations numerically. We will
focus on differential equations of the form

y0 (t) = f (t, y(t))

for 0 ≤ t ≤ T , with initial conditions

y(0) = y0 .

The data we are provided is the function f : R×RN → RN , the ending time T > 0,
and the initial condition y0 ∈ Rn . What we seek is the function y : [0, T ] → RN .
When solving the differential equation numerically, our goal would be to
make our numerical solution as close to the true solution as possible. This makes
sense only if a “true” solution actually exists, and is unique. From IB Analysis
II, we know a unique solution to the ODE exists if f is Lipschitz.
Definition (Lipschitz function). A function f : R × RN → RN is Lipschitz with
Lipschitz constant λ ≥ 0 if

kf (t, x) − f (t, x̂)k ≤ λkx − x̂k

for all t ∈ [0, T ] and x, x̂ ∈ RN .

A function is Lipschitz if it is Lipschitz with Lipschitz constant λ for some λ.

It doesn’t really matter what norm we pick. It will just change the λ. The
importance is the existence of a λ.
A special case is when λ = 0, i.e. f does not depend on x. In this case, this
is just an integration problem, and is usually easy. This is a convenient test case
— if our numerical approximation does not even work for these easy problems,
then it’s pretty useless.
Being Lipschitz is sufficient for existence and uniqueness of a solution to
the differential equation, and hence we can ask if our solution converges to
this unique solution. An extra assumption we will often make is that f can
be expanded in a Taylor series to as many degrees as we want, since this is
convenient for our analysis.
What exactly does a numerical solution to the ODE consist of? We first
choose a small time step h > 0, and then construct approximations

yn ≈ y(tn ), n = 1, 2, · · · ,

with tn = nh. In particular, tn − tn−1 = h and is always constant. In practice,

we don’t fix the step size tn − tn−1 , and allow it to vary in each step. However,
this makes the analysis much more complicated, and we will not consider varying
time steps in this course.
If we make h smaller, then we will (probably) make better approximations.
However, this is more computationally demanding. So we want to study the
behaviour of numerical methods in order to figure out what h we should pick.

26
5 Ordinary differential equations IB Numerical Analysis

5.2 One-step methods

There are many ways we can classify numerical methods. One important classifi-
cation is one-step versus multi-step methods. In one-step methods, the value of
yn+1 depends only on the previous iteration tn and yn . In multi-step methods,
we are allowed to look back further in time and use further results.
Definition ((Explicit) one-step method). A numerical method is (explicit)
one-step if yn+1 depends only on tn and yn , i.e.

yn+1 = φh (tn , yn )

for some function φh : R × RN → RN .

We will later see what “explicit” means.
The simplest one-step method one can imagine is Euler’s method.

Definition (Euler’s method). Euler’s method uses the formula

yn+1 = yn + hf (tn , yn ).

We want to show that this method “converges”. First of all, we need to make
precise the notion of “convergence”. The Lipschitz condition means there is a
unique solution to the differential equation. So we would want the numerical
solution to be able to approximate the actual solution to arbitrary accuracy as
long as we take a small enough h.
Definition (Convergence of numerical method). For each h > 0, we can produce
a sequence of discrete values yn for n = 0, · · · , [T /h], where [T /h] is the integer
part of T /h. A method converges if, as h → 0 and nh → t (hence n → ∞), we
get
yn → y(t),
where y is the true solution to the differential equation. Moreover, we require
the convergence to be uniform in t.
We now prove that Euler’s method converges. We will only do this properly for
Euler’s method, since the algebra quickly becomes tedious and incomprehensible.
However, the proof strategy is sufficiently general that it can be adapted to most
other methods.
Theorem (Convergence of Euler’s method).
(i) For all t ∈ [0, T ], we have

lim yn − y(t) = 0.
h→0
nh→t

(ii) Let λ be the Lipschitz constant of f . Then there exists a c ≥ 0 such that

eλT − 1
ken k ≤ ch
λ
for all 0 ≤ n ≤ [T /h], where en = yn − y(tn ).

27
5 Ordinary differential equations IB Numerical Analysis

Note that the bound in the second part is uniform. So this immediately gives
the first part of the theorem.
Proof. There are two parts to proving this. We first look at the local truncation
error. This is the error we would get at each step assuming we got the previous
steps right. More precisely, we write
y(tn+1 ) = y(tn ) + h(f , tn , y(tn )) + Rn ,
and Rn is the local truncation error. For the Euler’s method, it is easy to get
Rn , since f (tn , y(tn )) = y0 (tn ), by definition. So this is just the Taylor series
expansion of y. We can write Rn as the integral remainder of the Taylor series,
Z tn+1
Rn = (tn+1 − θ)y 00 (θ) dθ.
tn

By some careful analysis, we get

kRn k∞ ≤ ch2 ,
where
1 00
c=ky k∞ .
2
This is the easy part, and tends to go rather smoothly even for more complicated
methods.
Once we have bounded the local truncation error, we patch them together to
get the actual error. We can write
en+1 = yn+1 − y(tn+1 )
 
= yn + hf (tn , yn ) − y(tn ) + hf (tn , y(tn )) + Rn
 
= (yn − y(tn )) + h f (tn , yn ) − f (tn , y(tn )) − Rn

Taking the infinity norm, we get

ken+1 k∞ ≤ kyn − y(tn )k∞ + hkf (tn , yn ) − f (tn , y(tn ))k∞ + kRn k∞
≤ ken k∞ + hλken k∞ + ch2
= (1 + λh)ken k∞ + ch2 .
This is valid for all n ≥ 0. We also know ke0 k = 0. Doing some algebra, we get
n−1
X ch
ken k∞ ≤ ch2 (1 + hλ)j ≤ ((1 + hλ)n − 1) .
j=0
λ

Finally, we have
(1 + hλ) ≤ eλh ,
since 1 + λh is the first two terms of the Taylor series, and the other terms are
positive. So
(1 + hλ)n ≤ eλhn ≤ eλT .
So we obtain the bound
eλT − 1
ken k∞ ≤ ch .
λ
Then this tends to 0 as we take h → 0. So the method converges.

28
5 Ordinary differential equations IB Numerical Analysis

6 0. However, λ = 0 is the easy case, since it is

This works as long as λ =
just integration. We can either check this case directly, or use the fact that
eλT −1
λ → T as λ → 0.
The same proof strategy works for most numerical methods, but the algebra
will be much messier. We will not do those in full. We, however, take note of
some useful terminology:
Definition (Local truncation error). For a general (multi-step) numerical
method
yn+1 = φ(tn , y0 , y1 , · · · , yn ),
the local truncation error is
η n+1 = y(tn+1 ) − φn (tn , y(t0 ), y(t1 ), · · · , y(tn )).
This is the error we will make at the (n + 1)th step if we had accurate values
for the first n steps.
For Euler’s method, the local truncation error is just the Taylor series
remainder term.
Definition (Order). The order of a numerical method is the largest p ≥ 1 such
that η n+1 = O(hp+1 ).
The Euler method has order 1. Notice that this is one less than the power
of the local truncation error, since when we look at the global error, we drop a
power, and only have en ∼ h.
Let’s try to get a little bit beyond Euler’s method.
Definition (θ-method). For θ ∈ [0, 1], the θ-method is
 
yn+1 = yn + h θf (tn , yn ) + (1 − θ)f (tn+1 , yn+1 ) .

If we put θ = 1, then we just get Euler’s method. The other two most
common choices of θ are θ = 0 (backward Euler) and θ = 12 (trapezoidal rule).
Note that for θ =6 1, we get an implicit method. This is since yn+1 doesn’t
just appear simply on the left hand side of the equality. Our formula for yn+1
involves yn+1 itself! This means, in general, unlike the Euler method, we can’t
just write down the value of yn+1 given the value of yn . Instead, we have to
treat the formula as N (in general) non-linear equations, and solve them to find
yn+1 !
In the past, people did not like to use this, because they didn’t have computers,
or computers were too slow. It is tedious to have to solve these equations in
every step of the method. Nowadays, these are becoming more and more popular
because it is getting easier to solve equations, and θ-methods have some huge
theoretical advantages (which we do not have time to get into).
We now look at the error of the θ-method. We have
 
η = y(tn+1 ) − y(tn ) − h θy0 (tn ) + (1 − θ)y0 (tn+1 )

We expand all terms about tn with Taylor’s series to obtain

   
1 2 00 1 1
= θ− h y (tn ) + θ− h3 y000 (tn ) + O(h4 ).
2 2 3
1
We see that θ = 2 gives us an order 2 method. Otherwise, we get an order 1
method.

29
5 Ordinary differential equations IB Numerical Analysis

5.3 Multi-step methods

We can try to make our methods more efficient by making use of previous values
of yn instead of the most recent one. One common method is the AB2 method:
Definition (2-step Adams-Bashforth method). The 2-step Adams-Bashforth
(AB2) method has
1
yn+2 = yn+1 + h (3f (tn+1 , yn+1 ) − f (tn , yn )) .
2
This is a particular example of the general class of Adams-Bashforth methods.
In general, a multi-step method is defined as follows:
Definition (Multi-step method). A s-step numerical method is given by
s
X s
X
ρ` yn+` = h σ` f (tn+` , yn+` ).
`=0 `=0

This formula is used to find the value of yn+s given the others.
One point to note is that we get the same method if we multiply all the
constants ρ` , σ` by a non-zero constant. By convention, we normalize this by
setting ρs = 1. Then we can alternatively write this as
s
X s−1
X
yn+s = h σ` f (tn+` , yn+` ) − ρ` yn+` .
`=0 `=0

This method is an implicit method if σs 6= 0. Otherwise, it is explicit.

Note that this method is linear in the sense that the coefficients ρ` and σ`
appear linearly, outside the f s. Later we will see more complicated numerical
methods where these coefficients appear inside the arguments of f .
For multi-step methods, we have a slight problem to solve. In a one-step
method, we are given y0 , and this allows us to immediately apply the one-step
method to get higher values of yn . However, for an s-step method, we need
to use other (possibly 1-step) method to obtain y1 , · · · , ys−1 before we can get
started.
Fortunately, we only need to apply the one-step method a fixed, small number
of times, even as h → 0. So the accuracy of the one-step method at the start
does not matter too much.
We now study the properties of a general multi-step method. The first thing
we can talk about is the order:
Theorem. An s-step method has order p (p ≥ 1) if and only if
s
X
ρ` = 0
`=0

and
s
X s
X
ρ` `k = k σ` `k−1
`=0 `=0
0
for k = 1, · · · , p, where 0 = 1.

30
5 Ordinary differential equations IB Numerical Analysis

This is just a rather technical result coming directly from definition.

Proof. The local truncation error is
s
X s
X
ρ` y(tn+` ) − h σ` y0 (tn+` ).
`=0 `=0
0
We now expand the y and y about tn , and obtain
∞
s
! s s
!
X X hk X k X
k−1
ρ` y(tn ) + ρ` ` − k σ` ` y(k) (tn ).
k!
`=0 k=1 `=0 `=0
p+1
This is O(h ) under the given conditions.
Example (AB2). In the two-step Adams-Bashforth method, We see that the
conditions hold for p = 2 but not p = 3. So the order is 2.
Instead of dealing with the coefficients ρ` and σ` directly, it is often convenient
to pack them together into two polynomials. This uses two polynomials associated
with the numerical method. Unsurprisingly, they are
s
X s
X
ρ(w) = ρ` w ` , σ(w) = σ` w ` .
`=0 `=0

We can then use this to restate the above theorem.

Theorem. A multi-step method has order p (with p ≥ 1) if and only if
ρ(ex ) − xσ(ex ) = O(xp+1 )
as x → 0.
Proof. We expand
s
X s
X
ρ(ex ) − xσ(ex ) = ρ` e`x − x σ` e`x .
`=0 `=0

We now expand the e`x in Taylor series about x = 0. This comes out as
∞
s s s
!
X X 1 X k X
ρ` + ρ` ` − k σ` `k−1 xk .
k!
`=0 k=1 `=0 `=0

So the result follows.

Ps
Note that `=0 ρ` = 0, which is the condition required for the method to
even have an order at all, can be expressed as ρ(1) = 0.
Example (AB2). In the two-step Adams-Bashforth method, we get
3 1
ρ(w) = w2 − w, σ(w) = w− .
2 2
We can immediately check that ρ(1) = 0. We also have
5 3
ρ(ex ) − xσ(ex ) = x + O(x4 ).
12
So the order is 2.

31
5 Ordinary differential equations IB Numerical Analysis

We’ve sorted out the order of multi-step methods. The next thing to check
is convergence. This is where the difference between one-step and multi-step
methods come in. For one-step methods, we only needed the order to understand
convergence. It is a fact that a one step method converges whenever it has an
order p ≥ 1. For multi-step methods, we need an extra condition.
Definition (Root condition). We say ρ(w) satisfies the root condition if all its
zeros are bounded by 1 in size, i.e. all roots w satisfy |w| ≤ 1. Moreover any
zero with |w| = 1 must be simple.
We can imagine this as saying large roots are bad — they cannot get past 1,
and we cannot have too many with modulus 1.
We saw any sensible multi-step method must have ρ(1) = 0. So in particular,
1 must be a simple zero.
Theorem (Dahlquist equivalence theorem). A multi-step method is convergent
if and only if
(i) The order p is at least 1; and
(ii) The root condition holds.
The proof is too difficult to include in this course, or even the Part II version.
This is only done in Part III.
Example (AB2). Again consider the two-step Adams-Bashforth method. We
have seen it has order p = 2 ≥ 1. So we need to check the root condition. So
ρ(w) = w2 − w = w(w − 1). So it satisfies the root condition.
Let’s now come up with a sensible strategy for constructing convergent s-step
methods:
(i) Choose a ρ so that ρ(1) = 0 and the root condition holds.
(ii) Choose σ to maximize the order, i.e.
(
ρ(w) O(|w − 1|s+1 ) if implicit
σ= +
log w O(|w − 1|s ) if explicit

We have the two different conditions since for implicit methods, we have
one more coefficient to fiddle with, so we can get a higher order.
Where does the log1 w come from? We try to substitute w = ex (noting that
ex − 1 ∼ x). Then the formula says
(
x 1 x O(xs+1 ) if implicit
σ(e ) = ρ(e ) + .
x O(xs ) if explicit

Rearranging gives
(
O(xs+2 ) if implicit
ρ(ex ) − xσ(ex ) = ,
O(xs+1 ) if explicit

which is our order condition. So given any ρ, there is only one sensible way to
pick σ. So the key is in picking a good enough ρ.

32
5 Ordinary differential equations IB Numerical Analysis

The root conditions is “best” satisfied if ρ(w) = ws−1 (w − 1), i.e. all but one
roots are 0. Then we have
s
X
yn+s − yn+s−1 = h σ` f (tn+` , yn+` ),
`=0

where α is chosen to maximize order.

Definition (Adams method). An Adams method is a multi-step numerical
method with ρ(w) = ws−1 (w − 1).
These can be either explicit or implicit. In different cases, we get different
names.
Definition (Adams-Bashforth method). An Adams-Bashforth method is an
explicit Adams method.
Definition (Adams-Moulton method). An Adams-Moulton method is an im-
plicit Adams method.
Example. We look at the two-step third-order Adams-Moulton method. This
is given by
 
1 2 5
yn+2 − yn+1 = h − f (tn , yn ) + f (tn+1 , yn+1 ) + f (tn+1 , yn+2 ) .
2 3 12
ρ(w)
Where do these coefficients come from? We have to first expand our log w about
w − 1:
ρ(w) w(w − 1) 3 5
= = 1 + (w − 1) + (w − 1)2 + O(|w − 1|3 ).
log w log w 2 12
These aren’t our coefficients of σ, since what we need to do is to rearrange the
first three terms to be expressed in terms of w. So we have
ρ(w) 1 2 5
= − + w + w2 + O(|w − 1|3 ).
log w 12 3 12
Another important class of multi-step methods is constructed in the opposite
way — we choose a particular σ, and then find the most optimal ρ.
Definition (Backward differentiation method). A backward differentiation
method has σ(w) = σs ws , i.e.
s
X
ρ` yn+` = σs f (tn+s , yn+s ).
`=0

This is a generalization of the one-step backwards Euler method.

Given this σ, we need to choose the appropriate ρ. Fortunately, this can be
done quite easily.
Lemma. An s-step backward differentiation method of order s is obtained by
choosing
s
X 1 s−`
ρ(w) = σs w (w − 1)` ,
`
`=1

33
5 Ordinary differential equations IB Numerical Analysis

with σs chosen such that ρs = 1, namely

s
!−1
X 1
σs = .
`
`=1

Proof. We need to construct ρ so that

ρ(w) = σs ws log w + O(|w − 1|s+1 ).

This is easy, if we write

 
1
log w = − log
w
 
w−1
= − log 1 −
w
∞  `
X 1 w−1
= .
` w
`=1

s
Multiplying by σs w gives the desired result.
For this method to be convergent, we need to make sure it does satisfy the
root condition. It turns out the root condition is satisfied only for s ≤ 6. This is
not obvious by first sight, but we can certainly verify this manually.

5.4 Runge-Kutta methods

Finally, we look at Runge-Kutta methods. These methods are very complicated,
and are rather tedious to analyse. They have been largely ignored for a long
time, until more powerful computers came along and made these methods much
more practical. These are used quite a lot nowadays since they have many nice
properties.
Runge-Kutta methods can be motivated by Gaussian quadrature, but we
will not go into that connection here. Instead, we’ll go straight in and work with
the method.
Definition (Runge-Kutta method). General (implicit) ν-stage Runge-Kutta
(RK) methods have the form
ν
X
yn+1 = yn + h b` k` ,
`=1

where  
ν
X
k` = f tn + cn h, yn + h a`j kj 
j=1

for ` = 1, · · · , ν.
There are a lot of parameters we have to choose. We need to pick

{b` }ν`=1 , {c}ν`=1 , {a`j }ν`,j=1 .

34
5 Ordinary differential equations IB Numerical Analysis

Note that in general, {k` }ν`=1 have to be solved for, since they are defined in
terms of one another. However, for certain choices of parameters, we can make
this an explicit method. This makes it easier to compute, but we would have
lost some accuracy and flexibility.
Unlike all the other methods we’ve seen so far, the parameters appear inside
f . They appear non-linearly inside the functions. This makes the method much
more complicated and difficult to analyse using Taylor series. Yet, once we
manage to do this properly, these have lots of nice properties. Unfortunately, we
will not have time to go into what these properties actually are.
Notice this is a one-step method. So once we get order p ≥ 1, we will have
convergence. So what conditions do we need for a decent order?
This is in general very complicated. However, we can quickly obtain some
necessary conditions. We can consider the case where f is a constant. Then k`
is always that constant. So we must have
ν
X
b` = 1.
`=1

It turns out we also need, for ` = 1, · · · , ν,

ν
X
c` = a`j .
j=1

While these are necessary conditions, they are not sufficient. We need other
conditions as well, which we shall get to later. It is a fact that the best possible
order of a ν-stage Runge-Kutta method is 2ν.
To describe a Runge-Kutta method, a standard notation is to put the
coefficients in the Butcher table:
c1 a11 ··· a1ν
.. .. .. ..
. . . .
cν aν1 ··· aνν
b1 ··· bν
We sometimes more concisely write it as

c A
vT
This table allows for a general implicit method. Initially, explicit methods came
out first, since they are much easier to compute. In this case, the matrix A is
strictly lower triangular, i.e. a`j whenever ` ≤ j.

Example. The most famous explicit Runge-Kutta method is the 4-stage 4th
order one, often called the classical Runge-Kutta method. The formula can be
given explicitly by
h
yn+1 = yn + (k1 + 2k2 + 2k3 + k4 ),
6

35
5 Ordinary differential equations IB Numerical Analysis

where
k1 = f (xn , yn )
 
1 1
k2 = f xn + h, yn + hk1
2 2
 
1 1
k3 = f xn + h, yn + hk2
2 2
k4 = f (xn + h, yn + hk3 ) .
We see that this is an explicit method. We don’t need to solve any equations.
Choosing the parameters for the Runge-Kutta method to maximize order
is hard. Consider the simplest case, the 2-stage explicit method. The general
formula is
yn+1 = yn + h(b1 k1 + b2 k2 ),
where
k1 = f (xn , yn )
k2 = f (xn + c2 h, yn + r2 hk1 ).
To analyse this, we insert the true solution into the method. First, we need to
insert the true solution of the ODE into the k’s. We get
k1 = y0 (tn )
k2 = f (tn + c2 h, y(tn ) + c2 hy0 (tn ))
 
0 ∂f 0
= y (tn ) + c2 h (tn .y(tn )) + ∇f (tn , y(tn ))y (tn ) + O(h2 )
∂t
Fortunately, we notice that the thing inside the huge brackets is just y00 (tn ). So
this is
= y0 (tn ) + c2 hy00 (tn ) + O(h2 ).
Hence, our local truncation method for the Runge-Kutta method is

y(tn+1 ) − y(tn ) − h(b1 k1 + b2 k2 )

 
1
= (1 − b1 − b2 )hy0 (tn ) + − b2 c2 h2 y00 (tn ) + O(h3 ).
2
Now we see why Runge-Kutta methods are hard to analyse. The coefficients
appear non-linearly in this expression. It is still solvable in this case, in the
obvious way, but for higher stage methods, this becomes much more complicated.
In this case, we have a 1-parameter family of order 2 methods, satisfying
1
b1 + b2 = 1, b2 c2 = .
2
It is easy to check using a simple equation y 0 = λh that it is not possible to get a
higher order method. So as long as our choice of b1 and b2 satisfy this equation,
we get a decent order 2 method.
As we have seen, Runge-Kutta methods are really complicated, even in the
simplest case. However, they have too many good properties that they are
becoming very popular nowadays.

36
6 Stiff equations IB Numerical Analysis

6 Stiff equations
6.1 Introduction
Initially, when people were developing numerical methods, people focused mostly
on quantitative properties like order and accuracy. We then develop many
different methods like multi-step methods and Runge-Kutta methods.
More recently, people started to look at structural properties. Often, equations
come with some special properties. For example, a differential equation describing
the motion of a particle would most probably conserve energy. When we
approximate it numerically, we would like the numerical approximation to satisfy
conservation of energy as well. This is what recent developments are looking at —
we want to look at whether numerical methods preserve certain nice properties.
We are not going to look at conservation of energy — this is too complicated
for a first course. Instead, we look at the following problem. Suppose we have a
system of ODEs for 0 ≤ t ≤ T :
y0 (t) = f (t, y(t))
y(0) = y0 .
Suppose T > 0 is arbitrary, and
lim y(t) = 0.
t→∞

What restriction on h is necessary for a numerical method to satisfy lim yn = 0?

n→∞
This question is still too complicated for us to tackle. It can only be easily
solved for linear problems, namely ODEs of the form
y0 (t) = Ay(t),
for A ∈ RN ×N .
Firstly, for what A do we have y(t) → 0 as t → ∞? By some basic linear
algebra, we know this holds only if Re(λ) < 0 for all eigenvalues A. To simplify
further, we consider the case where
y 0 (t) = λy(t), Re(λ) < 0.
It should be clear that if A is diagonalizable, then it can be reduced to multiple
instances of this case. Otherwise, we need to do some more work, but we’ll not
do that in this course.
And that, at last, is enough simplification.

6.2 Linear stability

This is the easy part of the course, where we are just considering the problem
y 0 = λy. No matter how complicated are numerical method is, when applied to
this problem, it usually becomes really simple.
Definition (Linear stability domain). If we apply a numerical method to
y 0 (t) = λy(t)
with y(0) = 1, λ ∈ C, then its linear stability domain is
n o
D = z = hλ : lim yn = 0 .
n→∞

37
6 Stiff equations IB Numerical Analysis

Example. Consider the Euler method. The discrete solution is

yn = (1 + hλ)n .
Thus we get
D = {z ∈ C : |1 + z| < 1}.
We can visualize this on the complex plane as the open unit ball
Im

Re
−1

Example. Consider instead the backward Euler method. This method is

implicit, but since our problem is simple, we can find what nth step is. It turns
out it is
yn = (1 − λh)−n .
Then we get
D = {z ∈ C : |1 − z| > 1}.
We can visualize this as the region:
Im

Re
1

We make the following definition:

Definition (A-stability). A numerical method is A-stable if
C− = {z ∈ C : Re(z) < 0} ⊆ D.
In particular, for Re(z) < λ, A-stability means that yn will tend to 0 regardless
of how large h is.
Hence the backwards Euler method is A-stable, but the Euler method is not.
A-stability is a very strong requirement. It is hard to get A-stability. In
particular, Dahlquist proved that no multi-step method of order p ≥ 3 is A-stable.
Moreover, no explicit Runge-Kutta method can be A-stable.
So let’s look at some other implicit methods.

38
6 Stiff equations IB Numerical Analysis

Example (Trapezoidal rule). Again consider y 0 (t) = λy, with the trapezoidal
rule. Then we can find  n
1 + hλ/2
yn = .
1 − hλ/2
So the linear stability domain is
 
2+z
D= z∈C: <1 .
2−z

What this says is that z has to be closer to −2 than to 2. In other words, D is

exactly C− .

When testing a numerical method for A-stability in general, complex analysis

is helpful. Usually, when applying a numerical method to the problem y 0 = λy,
we get
yn = [r(hλ)]n ,
where r is some rational function. So

D = {z ∈ C : |r(z)| < 1}.

We want to know if D contains the left half-plane. For more complicated

expressions of r, like the case of the trapezoidal rule, this is not so obvious.
Fortunately, we have the maximum principle:
Theorem (Maximum principle). Let g be analytic and non-constant in an open
set Ω ⊆ C. Then |g| has no maximum in Ω.

Since |g| needs to have a maximum in the closure of Ω, the maximum must
occur on the boundary. So to show |g| ≤ 1 on the region Ω, we only need to
show the inequality holds on the boundary ∂Ω.
We try Ω = C− . The trick is to first check that g is analytic in Ω, and
then check what happens at the boundary. This technique is made clear in the
following example:

Example. Consider
6 − 2z
r(z) = .
6 − 4z + z 2
This is still pretty simple, but can illustrate how we can use the maximum
principle.
We
√ first check if it is analytic. This certainly has some poles, but they are
2 ± 2i, and are in the right-half plane. So this is analytic in C− .
Next, what happens at the boundary of the left-half plane? Firstly, as
|z| → ∞, we find r(z) → 0, since we have a z 2 at the denominator. The next
part is checking when z is on the imaginary axis, say z = it with t ∈ R. Then
we can check by some messy algebra that

|r(it)| ≤ 1

for t ∈ R. Therefore, by the maximum principle, we must have |r(z)| ≤ 1 for all
z ∈ C− .

39
7 Implementation of ODE methods IB Numerical Analysis

7 Implementation of ODE methods

We just did quite a lot of theory about numerical methods. To end this section,
we will look at some more practical sides of ODE methods.

7.1 Local error estimation

The first problem we want to tackle is what h we should pick. Usually, when
using numerical analysis software, you will be asked for an error tolerance, and
then the software will automatically compute the required h we need. How is
this done?
Milne’s device is a method for estimating the local truncation error of a
multi-step method, and hence changing the step-length h (there are similar
techniques for Runge-Kutta methods, but they are more complicated). This uses
two multi-step methods of the same order.
To keep things simple, we consider the two-step Adams-Bashforth method.
Recall that this is given by
h
yn+1 = yn + (2f (tn , yn ) − f (tn−1 , yn−1 )).
2
This is an order 2 error with
5 3 000
η n+1 = h y (tn ) + O(h4 ).
12
The other multi-step method of order 2 is our old friend, the trapezoidal rule.
This is an implicit method
h
yn+1 = yn + (f (tn , yn ) + f (tn+1 , yn+1 )).
2
This has a local truncation error of
1
η n+1 = − h3 y000 (tn ) + O(h4 ).
12
The key to Milne’s device is the coefficients of h3 y000 (yn ), namely
5 1
cAB = , cTR = − .
12 12
Since these are two different methods, we get different yn+1 ’s. We distinguish
these by superscripts, and have
AB
y(tn+1 ) − yn+1 ' cAB h3 y000 (tn )
TR
y(tn+1 ) − yn+1 ' cTR h3 y000 (tn )
We can now eliminate some terms to obtain
TR −cTR
y(tn+1 ) − yn+1 ' (yAB − yn+1
TR
).
cAB − cTR n+1
In this case, the constant we have is 16 . So we can estimate the local truncation
error for the trapezoidal rule, without knowing the value of y000 . We can then
use this to adjust h accordingly.
The extra work we need for this is to compute numerical approximations
with two methods. Usually, we will use a simple, explicit method such as the
Adams-Bashforth method as the second method, when we want to approximate
the error of a more complicated but nicer method, like the trapezoidal rule.

40
7 Implementation of ODE methods IB Numerical Analysis

7.2 Solving for implicit methods

Implicit methods are often more likely to preserve nice properties like conservation
of energy. Since we have more computational power nowadays, it is often
preferable to use these more complicated methods. When using these implicit
methods, we have to come up with some way to solve the equations involved.
As an example, we consider the backward Euler method

yn+1 = yn + hf (tn+1 , yn+1 ).

There are two ways to solve this equation for yn+1 . The simplest method is
functional iteration. As the name suggests, this method is iterative. So we use
superscripts to denote the iterates. In this case, we use the formula
(k+1) k
yn+1 = yn + hf (tn+1 , yn+1 ).
(0) (0)
To do this, we need to find a yn+1 . Usually, we start yn+1 = yn . Even better,
(0)
we can use some simpler explicit method to obtain our first guess of yn+1 .
The question, of course, is whether this converges. Fortunately, this converges
to a locally unique solution if λh is sufficiently small, where λ is the Lipschitz
constant of f . For the backward Euler, we will require λh < 1. This relies on
the contraction mapping theorem, which you may have met in IB Analysis II.
Does this matter? Sometimes it does. Usually, we pick an h using accuracy
considerations, picking the largest possible h that still gives us the desired
accuracy. However, if we use this method, we might need to pick a much smaller
h in order for this to work. This will require us to compute much more iterations,
and can take a lot of time.
An alternative is Newton’s method. This is given by the formula
(k) (k)
(I − hJ (k) )z(k) = yn+1 − (yn + hf (tn+1 , yn+1 ))
(k+1)
yn+1 = yn(k) − z(k) ,

where J (k) is the Jacobian matrix

(k)
J (k) = ∇f (tn+1 , yn+1 ) ∈ RN ×N .

This requires us to solve for z in the first equation, but this is a linear system,
which we have some efficient methods for solving.
There are several variants to Newton’s method. This is the full Newton’s
method, where we re-compute the Jacobian in every iteration. It is also possible
to just use the same Jacobian over and over again. There are some speed gains
in solving the equation, but then we will need more iterations before we can get
our yn+1 .

41
8 Numerical linear algebra IB Numerical Analysis

8 Numerical linear algebra

In the last part of the course, we study numerical algorithms for solving certain
problems in linear algebra. Here we are not so much concerned about accuracy
— the solutions we obtained are theoretically exact. Instead, we will try to find
some efficient methods of solving these problems. However, we will occasionally
make some comments about accuracy problems that arise from the fact that we
only work with finite precision.
We start off with the simplest problem in linear algebra: given A ∈ Rn×n ,
b ∈ Rn , we want to find an x such that
Ax = b.
We all know about the theory — if A is non-singular, then there is a unique
solution for every possible b. Otherwise, there are no solutions for some b and
infinitely many solutions for other b’s.
Most of the time, we will only consider the case where A is non-singular.
However, we will sometimes comment on what happens when A is singular.

8.1 Triangular matrices

While matrices in general and big and scary, triangular matrices tend to be much
nicer.
Definition (Triangular matrix). A matrix A is upper triangular if Aij = 0
whenever i > j. It is lower triangular if Aij = 0 whenever i < j. We usually
denote upper triangular matrices as U and lower triangular matrices as L.
We can visualize these matrices as follows:

L U

Why are triangular matrices nice? First of all, it is easy to find the determinants
of triangular matrices. We have
n
Y n
Y
det(L) = Lii , det(U ) = U11 .
i=1 i=1

In particular, a triangular matrix is non-singular if and only if it has no zero

entries in its diagonal.
How do we solve equations involving triangular matrices? Suppose we have
a lower triangular matrix equation
    
L11 0 ··· 0 x1 b1
 L12 L22 · · · 0   x 2   b2 
..   ..  =  ..  ,
    
 .. .. ..
 . . . .  .   . 
L1n L2n ··· Lnn xn bn
or, more visually, we have

42
8 Numerical linear algebra IB Numerical Analysis

There is nothing to solve in the first line. We can immediately write down the
value of x1 . Substituting this into the second line, we can then solve for x2 . In
general, we have  
i−1
1  X
xi = bi − Lij xj  .
Lii j=1

This is known as forward substitution.

For upper triangular matrices, we can do a similar thing, but we have to
solve from the bottom instead. We have
 
n
1  X
xi = bi − Uij xj  .
Uii j=1+i

This is known as backwards substitution.

This can be used to find the inverse of matrices as well. The solution to
Lxj = ej is the jth column of L−1 . It is then easy to see from this that the
inverse of a lower triangular matrix is also lower triangular.
Similarly, the columns of U −1 are given by solving U xj = ei , and we see that
−1
U is also upper triangular.
It is helpful to analyse how many operations it takes to compute this. If we
look carefully, solving Lx = b or U x = b this way requires O(n2 ) operations.

8.2 LU factorization
In general, we don’t always have triangular matrices. The idea is, for every
matrix A, find some lower and triangular matrices L, U such that A = LU .

If we can do this, then we can solve Ax = b in two steps — we first find a y

such that Ly = b. Then we find an x such that U x = y. Then
Ax = LU x = Ly = b.
So what we want is such a factorization. To guarantee uniqueness of the
factorization, we require that L is unit, i.e. the diagonals are all 1. Otherwise,
given any such factorization, we can divide U by some (non-zero) constant and
multiply L by the same constant to get another factorization.
Definition (LU factorization). A = LU is an LU factorization if U is upper
triangular and L is unit lower triangular (i.e. the diagonals of L are all 1).
Note that since L has to be unit, it must be non-singular. However, we still
allow A and U to be singular. Note that
det(A) = det(L) det(U ) = det(U ).
So A is singular if and only if U is.
Unfortunately, even if A is non-singular, it may not have an LU factorization.
We take a simple example of  
0 1
1 1

43
8 Numerical linear algebra IB Numerical Analysis

This is clearly non-singular, with determinant −1. However, we can manually

check that there is no LU factorization of this.
On the other hand, while we don’t really like singular matrices, singular
matrices can still have LU factorizations. For example,
    
0 0 1 0 0 0
=
0 1 0 1 0 1

is trivially an LU factorization of a singular matrix.

Fortunately, if A is non-singular and has an LU factorization, then this
factorization is unique (this is not necessarily true if A is singular. For example,
we know     
0 0 1 0 0 0
=
0 1 a 1 0 1
for any real number a).
To understand when LU factorizations exist, we try to construct one, and
see when it could fail.
We write L in terms of its columns, and U in terms of its rows:
 T
u1

uT2 
L = l1 l2 · · · ln , U =  . 
 
 .. 
uTn

Clearly, these rows and columns cannot be arbitrary, since L and U are triangular.
In particular, li , ui must be zero in the first i − 1 entries.
Suppose this is an LU factorization of A. Then we can write

A = L · U = l1 uT1 + l2 uT2 + · · · + ln uTn .

What do these matrices look like? For each i, we know li and ui have the first
i − 1 entries zero. So the first i − 1 rows and columns of li uTi are zero. In
particular, the first row and columns only have contributions from l1 uT1 , the
second row/column only has contributions from l2 uT2 and l1 uT1 etc.
The plan is as follows:

(i) Obtain l1 and u1 from the first row and column of A. Since the first entry
of l1 is 1, uT1 is exactly the first row of A. We can then obtain l2 by taking
the first column of A and dividing by U11 = A11 .
(ii) Obtain l2 and u2 form the second row and column of A − l1 uT1 similarly.

(iii) · · ·
Pn−1
(iv) Obtain ln and uTn from the nth row and column of A − i=1 li uTi .
We can turn this into an algorithm. We define the intermediate matrices, starting
with
A(0) = A.

44
8 Numerical linear algebra IB Numerical Analysis

For k = 1, · · · , n, we let
(k−1)
Ukj = Akj j = k, · · · , n
Aik
Lik = (k−1)
i = k, · · · , n
Akk
(k) (k−1)
Aij = Aij − Lik Ukj i, j ≥ k

When k = n, we end up with a zero matrix, and then U and L are completely
filled.
We can now see when this will break down. A sufficient condition for A = LU
(k−1) (k−1)
to exist is that Akk 6= 0 for all k. Since Akk = Ukk , this sufficient condition
ensures U , and hence A is non-singular. Conversely, if A is non-singular and
an LU factorization exists, then this would always work, since we must have
(k−1)
Akk = Ukk 6= 0. Moreover, the LU factorization must be given by this
algorithm. So we get uniqueness.
The problem with this sufficient condition is that most of these coefficients
do not appear in the matrix A. They are constructed during the algorithm. We
don’t know easily what they are in terms of the coefficients of A. We will later
come up with an equivalent condition on our original A that is easier to check.
Note that as long as this method does not break down, we need O(n3 )
operations to perform this factorization. Recall we only needed O(n2 ) operations
to solve the equation after factorization. So the bulk of the work in solving
Ax = b is in doing the LU factorization.
As before, this allows us to find the inverse of A if it is non-singular. In
particular, solving Axj = ej gives the jth column of A−1 . Note that we are
solving the system for the same A for each j. So we only have to perform the
LU factorization once, and then solve n different equations. So in total we need
O(n3 ) operations.
However, we still have the problem that factorization is not always possible.
Requiring that we must factorize A as LU is too restrictive. The idea is to factor
something closely related, but not exactly A. Instead, we want a factorization

P A = LU,

where P is a permutation matrix. Recall a permutation matrix acting on a

column vector just permutes the elements of the vector, and P acting on A
would just permute the rows of A. So we want to factor A up to a permutation
of rows.
Our goal now is to extend the previous algorithm to allow permutations of
rows, and then we shall show that we will be able to perform this factorization
all the time.
(0)
Suppose our breakdown occurs at k = 1, i.e. A11 = A11 = 0. We find a
(0)
permutation matrix P1 and let it act via P1 A . The idea is to look down the
first column of A, and find a row starting with a non-zero element, say p. Then
we use P1 to interchange rows 1 and p such that P1 A(0) has a non-zero top-most
(0)
entry. For simplicity, we assume we always need a P1 , and if A11 is non-zero in
the first place, we just take P1 to be the identity.
After that, we can carry on. We construct l1 and u1 from P1 A(0) as before,
and set A(1) = P1 A(0) − l1 uT1 .

45
8 Numerical linear algebra IB Numerical Analysis

But what happens if the first column of A is completely zero? Then no

interchange will make the (1, 1) entry non-zero. However, in this case, we don’t
actually have to do anything. We can immediately find our l1 and u1 , namely
set l1 = e1 (or anything) and let uT1 be the first row of A(0) . Then this already
works. Note however that this corresponds to A (and hence U ) being singular,
and we are not too interested with these.
(k−1)
The later steps are exactly analogous. Suppose we have Akk = 0. Again
(k−1)
we find a Pk such that Pk A has a non-zero (k, k) entry. We then construct
lk and uk from Pk A(k−1) and set

A(k) = Pk A(n−1) − lk utk .

Again, if the kth column of A(k−1) is completely zero, we set lk = ek and uTk to
be the kth row of A(k−1) . But again this implies A and U will be singular.
However, as we do this, the permutation matrices appear all over the place
inside the algorithm. It is not immediately clear that we do get a factorization
of the form P A = LU . Fortunately, keeping track of the interchanges, we do
have an LU factorization
P A = L̃U,
where U is what we got from the algorithm,

P = Pn−1 · · · P2 P1 ,

while L̃ is given by


L̃ = l̃1 ··· l̃n , l̃k = Pn−1 · · · Pk−1 lk .

Note that in particular, we have

l̃n−1 = ln−1 , l̃n = ln .

One problem we have not considered is the problem of inexact arithmetic. While
these formula are correct mathematically, when we actually implement things,
we do them on computers with finite precision. As we go through the algorithm,
errors will accumulate, and the error might be amplified to a significant amount
when we get the reach the end. We want an algorithm that is insensitive to
errors. In order to work safely in inexact arithmetic, we will put the element of
largest modulus in the (k, k)th position, not just an arbitrary non-zero one, as
this minimizes the error when dividing.

8.3 A = LU for special A

There is one thing we haven’t figured out yet. Can we just look at the matrix A,
and then determine whether it has an LU factorization (that does not involve
permutations)? To answer this, we start with some definitions.
Definition (Leading principal submatrix). The leading principal submatrices
Ak ∈ Rk×k for k = 1, · · · , n of A ∈ Rn×n are

(Ak )ij = Aij , i, j = 1, · · · , k.

In other words, we take the first k rows and columns of A.

46
8 Numerical linear algebra IB Numerical Analysis

Theorem. A sufficient condition for the existence for both the existence and
uniqueness of A = LU is that det(Ak ) 6= 0 for k = 1, · · · , n − 1.
Note that we don’t need A to be non-singular. This is equivalent to the
(k−1)
restriction Akk 6= 0 for k = 1, · · · , n − 1. Also, this is a sufficient condition,
not a necessary one.
Proof. Straightforward induction.
We extend this result a bit:
Theorem. If det(Ak ) 6= 0 for all k = 1, · · · , n, then A ∈ Rn×n has a unique
factorization of the form
A = LDÛ ,
where D is non-singular diagonal matrix, and both L and Û are unit triangular.
Proof. From the previous theorem, A = LU exists. Since A is non-singular, U
is non-singular. So we can write this as

U = DÛ ,

where D consists of the diagonals of U and Û = D−1 U is unit.

This is not hard, but is rather convenient. In particular, it allows us to
factorize symmetric matrices in a nice way.
Theorem. Let A ∈ Rn×n be non-singular and det(Ak ) 6= 0 for all k = 1, · · · , n.
Then there is a unique “symmetric” factorization

A = LDLT ,

with L unit lower triangular and D diagonal and non-singular.

Proof. From the previous theorem, we can factorize A uniquely as

A = LDÛ .

We take the transpose to obtain

A = AT = Û T DLT .

This is a factorization of the form “unit lower”-“diagonal”-“unit upper”. By

uniqueness, we must have Û = LT . So done.
We just looked at a special type of matrices, the symmetric matrices. We
look at another type of matrices:
Definition (Positive definite matrix). A matrix A ∈ Rn×n is positive-definite if

xT Ax > 0

for x 6= 0 ∈ Rn .
Theorem. Let A ∈ Rn×n be a positive-definite matrix. Then det(Ak ) 6= 0 for
all k = 1, · · · , n.

47
8 Numerical linear algebra IB Numerical Analysis

Proof. First consider k = n. To show A is non-singular, it suffices to show that

Ax = 0 implies x = 0. But we can multiply the equation by xT to obtain
xT Ax = 0. By positive-definiteness, we must have x = 0. So done.
Now suppose Ak y = 0 for k < n and y = Rk . Then yT Ak y = 0. We
invent a new x ∈ Rn by taking y and pad it with zeros. Then xT Ax = 0. By
positive-definiteness, we know x = 0. Then in particular y = 0.
We are going to use this to prove the most important result in this section.
We are going to consider matrices that are symmetric and positive-definite.
Theorem. A symmetric matrix A ∈ Rn×n is positive-definite iff we can factor
it as
A = LDLT ,
where L is unit lower triangular, D is diagonal and Dkk > 0.
Proof. First suppose such a factorization exists, then
xT Ax = xT LDLT x = (LT x)T D(LT x).
We let y = LT x. Note that y = 0 if and only if x = 0, since L is invertible. So
n
X
xT Ax = yT Dy = yk2 Dkk > 0
k=1

if y 6= 0.
Now if A is positive definite, it has an LU factorization, and since A is
symmetric, we can write it as
A = LDLT ,
where L is unit lower triangular and D is diagonal. Now we have to show
Dkk > 0. We define yk such that LT yk = ek , which exist, since L is invertible.
Then clearly yk 6= 0. Then we have
Dkk = eTk Dek = ykT LDLT yk = ykT Ayk > 0.
So done.
This is a practical check for symmetric A being positive definite. We can
perform this LU factorization, and then check whether the diagonal has positive
entries.
Definition (Cholesky factorization). The Cholesky factorization of a symmetric
positive-definite matrix A is a factorization of the form
A = LDLT ,
with L unit lower triangular and D a positive-definite diagonal matrix.
There is another way of doing this. We let D1/2 be the “square root” of D,
by taking the positive square root of the diagonal entries of D. Then we have
A = LDLT = LD1/2 D1/2 LT = (LD1/2 )(LD1/2 )T = GGT ,
where G is lower triangular with Gkk > 0. This is another way of presenting this
result.
Finally, we look at the LU factorization of band matrices.

48
8 Numerical linear algebra IB Numerical Analysis

Definition (Band matrix). A band matrix of band width r is a matrix A such

that Aij 6= 0 implies |i − j| ≤ r.

For example, a band matrix of band width 0 is a diagonal matrix; a band

matrix of band width 1 is a tridiagonal matrix.

The result is
Proposition. If a band matrix A has band width r and an LU factorization
A = LU , then L and U are both band matrices of width r.

Proof. Straightforward verification.

49
9 Linear least squares IB Numerical Analysis

9 Linear least squares

Finally, we consider a slightly different kind of problem, whose solution involves
a different kind of factorization.
The question we are interested in is how we can “solve”

Ax = b,

where A ∈ Rm×n , with m > n, and b ∈ Rm and x ∈ Rn .

Of course, this is an over-determined system. In general, this has no solution.
So we want to find the “best” solution to this system of equations. More precisely,
we want to find an x∗ ∈ Rn that minimizes kAx − bk2 . Note that we use the
Euclidean norm. This is the least squares problem.
Why would one be interested in this? Often in, say, statistics, we have some
linear model that says the outcome B is related to the input variables A1 , · · · , An
by
B = A1 X1 + A2 X2 + · · · + An Xn + ε,
where the Xi are some unknown parameters, and ε is some random fluctuation.
The idea is to do lots of experiments, say m of them, collect the data in A ∈ Rm×n
and b ∈ Rm , and then find the Xi ’s that predict these results the most accurately,
i.e. find the x such that Ax is as close to b as possible.
So how can we find this x∗ ? First of all, we find a necessary and sufficient
condition for x∗ to be a solution to this.
Theorem. A vector x∗ ∈ Rn minimizes kAx − bk2 if and only if

AT (Ax∗ − b) = 0.

Proof. A solution, by definition, minimizes

f (x) = hAx − b, Ax − bi
= xT AAx − 2xT AT b + bT b.

Then as a function of x, the partial derivatives of this must vanish. We have

∇f (x) = 2AT (Ax − b).

So a necessary condition is
AT (Ax − b).
Now suppose our x∗ satisfies AT (Ax∗ − b) = 0. Then for all x ∈ Rn , we write
x = x∗ + y, and then we have

kAx − bk2 = kA(x∗ + y) − bk

= kAx∗ − bk + 2yT AT (Ax − b) + kAyk2
= kAx∗ − bk + kAyk2
≥ kAx∗ − bk.

So x∗ must minimize the Euclidean norm.

Corollary. If A ∈ Rm×n is a full-rank matrix, then there is a unique solution
to the least squares problem.

50
9 Linear least squares IB Numerical Analysis

Proof. We know all minimizers are solutions to

(AT A)x = AT b.

The matrix A being full rank means y = 6 0 ∈ Rn implies Ay 6= 0 ∈ Rm . Hence

T n×n
A A∈R is positive definite (and in particular non-singular), since

xT AT Ax = (Ax)T (Ax) = kAxk2 > 0

for x 6= 0. So we can invert AT A and find a unique solution x.

Now to find the minimizing x∗ , we need to solve the normal equations

AT Ax = AT b.

If A has full rank, then AT A is non-singular, and there is a unique solution. If

not, then the general theory of linear equations tells us there are either infinitely
many solution or no solutions. But for this particular form of equations, it turns
out we can show this is always consistent. So there will always be a solution.
However, solving these equations is not the recommended algorithm for finding
x∗ for accuracy reasons. Instead the solution of these is based on orthogonal
matrices.

Definition (Orthogonal matrix). A square matrix Q ∈ Rn×n is orthogonal if

QT Q = I.
A key property of the orthogonal matrices is that we have

kQxk = kxk

for all x ∈ Rn . This is helpful since what we are trying to do in the least-squares
problem involves minimizing the Euclidean norm of something.
The idea is that for any orthogonal matrix, minimizing kAx − bk is the same
as minimizing kQAx − Qbk. So Q should we pick? Recall that a simple kind of
matrices is triangular matrices. So we might want to get something, say, upper
triangular.
Definition (QR factorization). A QR factorization of an m × n matrix A is a
factorization of the form
A = QR,
where Q ∈ Rm×m is an orthogonal matrix, and R ∈ Rm×n is “upper triangular”
matrix, where “upper triangular” means
 
R11 R12 · · · R1m
 0 R22 · · · R2m 
 .. .. .. 
 
..
 .
 . . . 
R=  0 0 · · · Rmm 
 0
 0 · · · 0  
 . . . .. 
 .. .. .. . 
0 0 ··· 0

51
9 Linear least squares IB Numerical Analysis

Note that every A ∈ Rm×n has a QR factorization, as we will soon show,

but this is not unique (e.g. we can multiply both Q and R by −1).
Once we have the QR factorization, we can multiply kAx − bk by QT = Q−1 ,
and get an equivalent problem of minimizing kRx − QT bk. We will not go into
details, but it should be clear that this is not too hard to solve.
There is an alternative “skinny” QR factorization, with A = Q̃R̃, where we
remove redundant rows and columns such that Q̃ ∈ Rm×n and R̃ ∈ Rn×n . This
is good enough to solve the least squares problem.
We will see that if A has full rank, then the skinny QR is unique up to a
sign, i.e. if we require R̃kk > 0 for k = 1, · · · , n.
We shall demonstrate three standard algorithms for QR factorization:
(i) Gram-Schmidt factorization
(ii) Givens rotations
(iii) Householder reflections
The latter two are based on some geometric ideas, while the first is based on the
familiar Gram-Schmidt process.

Gram-Schmidt factorization
This targets the skinny version. So we stop writing the tildes above, and just
write Q ∈ Rm×n and R ∈ Rn×n .
As we all know, the Gram-Schmidt process orthogonalizes vectors. What we
are going to orthogonalize are the columns of A. We write



A = a1 · · · an , Q = q1 · · · qn .
By definition of the QR factorization, we need
j
X
aj = Rij qi .
i=1

This is just done in the usual Gram-Schmidt way.

(i) To construct column 1, if a1 6= 0, then we set
a1
q1 = , R11 = ka1 k.
ka1 k
Note that the only non-unique possibility here is the sign — we can let
R11 = −ka1 k and q1 = − kaa11 k instead. But if we require R11 > 0, then
this is fixed.
In the degenerate case, a1 = 0, then we can just set R11 = 0, and the pick
any q1 ∈ Rn with kq1 k = 1.
(ii) For columns 1 < k ≤ n, for i = 1, · · · , k − 1, we set
Rik = hqi , ak i,
and set
k−1
X
dk = ak − qi hqi , ak i.
i=1

52
9 Linear least squares IB Numerical Analysis

If dk 6= 0, then we set
dk
qk = ,
kdk k
and
Rkk = kdk k.
In the case where dk = 0, we again set Rkk = 0, and pick qk to be anything
orthonormal to q1 , · · · , qk−1 .
In practice, a slightly different algorithm (modified Gram-Schmidt process)
is used, which is (much) superior with inexact arithmetic. The modified Gram-
Schmidt process is in fact the same algorithm, but performed in a different order
in order to minimize errors.
However, this is often not an ideal algorithm for large matrices, since there
are many divisions and normalizations involved in computing the qi , and the
accumulation of errors will cause the resulting matrix Q to lose orthogonality.

Givens rotations
This works with the full QR factorization.
Recall that in R2 , a clockwise rotation of θ ∈ [−π, π] is performed by
    
cos θ sin θ α α cos θ + β sin θ
= .
− sin θ cos θ β −α sin θ + β cos θ
By choosing θ such that
α β
cos θ = p , sin θ = p ,
α + β2
2 α2 + β2
this then becomes p 
α2 + β 2
.
0
Of course, by choosing
pa slightly different θ, we can make the result zero in the
first component and α2 + β 2 in the second.
Definition (Givens rotation). In Rm , where m > 2, we define the Givens
rotation on 3 parameters 1 ≤ p < q ≤ m, θ ∈ [−π, π] by
 
1
 .. 

 . 


 1 


 cos θ sin θ 


 1 

[p,q]
Ωθ = 
 ..  ∈ Rm×m ,

 . 

 1 


 − sin θ cos θ 


 1 


 . ..


1
where the sin and cos appear at the p, qth rows and columns.

53
9 Linear least squares IB Numerical Analysis

[p,q]
Note that for y ∈ Rm , the effect of Ωθ only alters the p and q components.
[p,q]
In general, for B ∈ Rm×n , then Ωθ B only alters the p and q rows of B.
Moreover, just like the R case, given a particular z ∈ Rm , we can choose θ
2
[p,q]
such that the qth component (Ωθ z)q = 0.
Hence, A ∈ Rm×n can be transformed into an “upper triangular” form by
applying s = mn − n(n+1)
2 Givens rotations, since we need to introduce s many
zeros. Then
Qs · · · Q1 A = R.
How exactly do we do this? Instead of writing down a general algorithm, we
illustrate it with a matrix A ∈ R4×3 . The resultant R would look like
 
R11 R12 R13
 0 R22 R23 
R=  0
.
0 R33 
0 0 0

We will apply the Givens rotations in the following order:

[3,4] [3,4] [2,3] [1,4] [1,3] [1,2]
Ωθ6 Ωθ5 Ωθ4 Ωθ3 Ωθ2 Ωθ1 A = R.

The matrix A transforms as follows:

× × × Ω[1,2] × × × Ω[1,3] × × × Ω[1,3] ×××
××× θ1 0 ×× θ2 0 ×× θ3 0 ××
××× ××× 0 × × 0 ××
××× ××× ××× 0 ××

[2,3]
Ωθ4 × × × Ω[3,4] × × × Ω[3,4] ×××
0 ×× θ5 0 ×× θ6 0 ××
0 0× 0 0× 0 0×
0 ×× 0 0× 0 0 0
[2,3]
Note that when applying, say, Ωθ4 , the zeros of the first column get preserved,
[2,3]
since Ωθ4 only mixes together things in row 2 and 3, both of which are zero in
the first column. So we are safe.
Note that this gives us something of the form

Qs · · · Q1 A = R.

We can obtain a QR factorization by inverting to get

A = QT1 · · · QTs R.

However, we don’t really need to do this if we just want to do use this to solve
the least squares problem, since to do so, we need to multiply by QT , not Q,
which is exactly Qs · · · Q1 .

Householder reflections
6 0 ∈ Rm , we define the Householder
Definition (Householder reflection). For u =
reflection by
uuT
Hu = I − 2 T ∈ Rm×m .
u u

54
9 Linear least squares IB Numerical Analysis

Note that this is symmetric, and we can see Hu2 = I. So this is indeed
orthogonal.
To show this is a reflection, suppose we resolve x as the perpendicular and
parallel parts as x = αu + w ∈ Rm , where
uT x
α= , uT w = 0.
uT u
Then we have
Hu x = −αu + w.
So this is a reflection in the m − 1-dimensional hyperplane uT y = 0.
What is the cost of computing Hu z? This is evaluated as
uT z
Hu z = z − 2 u.
uT u
This only requires O(m) operations, which is nice.
Proposition. A matrix A ∈ Rm×n can be transformed into upper-triangular
form by applying n Householder reflections, namely

Hn · · · H1 A = R,

where each Hn introduces zero into column k and leaves the other zeroes alone.
This is much less multiplications needed than the number needed for Givens
rotations, in general.
To show this, we first prove some lemmas.
Lemma. Let a, b ∈ Rm , with a 6= b, but kak = kbk. Then if we pick u = a − b,
then
Hu a = b.
This is obvious if we draw some pictures in low dimensions.
Proof. We just do it:
2(kak − aT b)
Hu a = a − (a − b) = a − (a − b) = b,
kak2 − 2aT b + kbk2
where we used the fact that kak = kbk.
We will keep applying the lemma. Since we want to get an upper triangular
matrix, it makes sense to pick b to have many zeroes, and the best option would
be to pick b to be a unit vector.
So we begin our algorithm: we want to clear the first column of A. We let a
be the first column of A, and assume a ∈ Rm is not already in the correct form,
i.e. a is not a multiple of e1 . Then we define

u = a ∓ kake1 ,

where either choice of the sign is pure-mathematically valid. However, we will

later see that there is one choice that is better when we have to deal with inexact
arithmetic. Then we end up with

H1 a = Hu a = ±kake1 .

55
9 Linear least squares IB Numerical Analysis

Now we have
×××
××
H1 A = 0 × × .
0
0 ××
To do the next step, we need to be careful, since we don’t want to destroy the
previously created zeroes.
Lemma. If the first k − 1 components of u are zero, then
(i) For every x ∈ Rm , Hu x does not alter the first k − 1 components of x.
(ii) If the last (m − k + 1) components of y ∈ Rm are zero, then Hu y = y.
These are all obvious from definition. All these say is that reflections don’t
affect components perpendicular to u, and in particular fixes all vectors perpen-
dicular to u.
Lemma. Let a, b ∈ Rm , with
   
ak bk
 ..   .. 
 .  6=  .  ,
am bm
but
m
X m
X
a2j = b2j .
j=k j=k

Suppose we pick
u = (0, 0, · · · , 0, ak − bk , · · · , am − bm )T .
Then we have
Hu a = (a1 , · · · , ak−1 bk , · · · , bm ).
This is a generalization of what we’ve had before for k = 1. Again, the proof
is just straightforward verification.
Now we can mess with the second column of H1 A. We let a be the second
column of H1 A, and assume a3 , · · · , am are not all zero, i.e. (0, a2 , · · · , am )T is
not a multiple of e2 . We choose
u = (0, a2 ∓ γ, a3 , · · · , am )T ,
where v
um
uX
γ=t aj .
j=2

Then
Hu a = (a1 , ±γ, 0, · · · ).
Also, by the previous lemma, this does not affect anything in the first column
and the first row. Now we have
×××
××
H2 H1 A = 0 ×
0 0
0 0×

56
9 Linear least squares IB Numerical Analysis

Suppose we have reached Hk−1 · · · H1 A, where the first k − 1 rows are of the
correct form, i.e.

Hk−1 · · · H1 A =
0

k−1
We consider the kth column of Hk−1 · · · H1 A, say a. We assume (0, · · · , 0, ak , · · · , am )T
is not a multiple of ek . Choosing
v
um 2
uX
u = γ(0, · · · , 0, ak ∓ γ, ak+1 , · · · , am )t , γ = t aj ,
j=k

we find that
Hu a = (a1 , · · · , ak−1 , ±γ, 0, · · · , 0)T .
Now we have

Hk · · · H1 A = ,
0

k
and Hk did not alter the first k − 1 rows and columns of Hk−1 · · · H1 A.
There is one thing we have to decide on — which sign to pick. As mentioned,
these do not matter in pure mathematics, but with inexact arithmetic, we
should pick the sign in ak ∓ γ such that ak ∓ γ has maximum magnitude, i.e.
ak ∓ γ = ak + sgn(ak )γ. It takes some analysis to justify why this is the right
choice, but it is not too surprising that there is some choice that is better.
So how do Householder and Givens compare? We notice that the Givens
method generates zeros at one entry at a time, while Householder does it column
by column. So in general, the Householder method is superior.
However, for certain matrices with special structures, we might need the
extra delicacy of introducing a zero one at a time. For example, if A is a band
matrix, then it might be beneficial to just remove the few non-zero entries one
at a time.

57