Prof. R.

Hiptmair AS 2015
ETH Zürich
G. Alberti,
D-MATH
F. Leonardi Numerical Methods for CSE

Problem Sheet 4

Problem 1. Order of convergence from error recursion (core prob-

lem)
In [1, Exp. 2.3.26] we have observed fractional orders of convergence (→ [1, Def. 2.1.17])
for both the secant method, see [1, Code 2.3.25], and the quadratic inverse interpolation
method. This is fairly typical for 2-point methods in 1D and arises from the underly-
ing recursions for error bounds. The analysis is elaborated for the secant method in [1,
Rem. 2.3.27], where a linearized error recursion is given in [1, Eq. (2.3.31)].
Now we suppose the recursive bound for the norms of the iteration errors
�
�e(n+1) � ≤ �e(n) � �e(n−1) � , (11)

where e(n) = x(n) − x∗ is the error of n-th iterate.

(1a) Guess the maximal order of convergence of the method from a numerical exper-
iment conducted in MATLAB.

H INT:[1, Rem. 2.1.19]

Solution: See Listing 29.

Listing 29: MATLAB script for Sub-problem (1a)

1 % MATLAB test code for homework problem on "order of
convergence from error
2 % recursion", outputs a table
3 e(1) = 1; e(2) = 0.8; % Any values ≤ 1 possible , also
random

1
4 f o r k=2:20, e(k+1) = e(k)* s q r t (e(k-1)); end
5 le = l o g (e); d i f f (le(2:end))./ d i f f (le(1:end-1)),

(1b) Find the maximal guaranteed order of convergence of this method through ana-
lytical considerations.
H INT: First of all note that we may assume equality in both the error recursion (11) and the
bound �e(n+1) � ≤ C�e(n) �p that defines convergence of order p > 1, because in both cases
equality corresponds to a worst case scenario. Then plug the two equations into each other
and obtain an equation of the type . . . = 1, where the left hand side involves an error norm
that can become arbitrarily small. This implies a condition on p and allows to determine
C > 0. A formal proof by induction (not required) can finally establish that these values
provide a correct choice.
Solution: Suppose �e(n) � = C�e(n−1) �p (p is the largest convergence order and C is some
constant).
Then

�e(n+1) � = C�e(n) �p = C(C�e(n−1) �p )p = C p+1 �e(n−1) �p

2
(12)

In (11) we may assume equality, because this is the worst case. Thus,

�e(n+1) � = �e(n) � ⋅ �e(n−1) � 2 = C�e(n−1) �p+ 2

1 1
(13)

Combine (12) and (13),

C p+1 �e(n−1) �p = C�e(n−1) �p+ 2
2 1

i.e.

C p �e(n−1) �p = 1.
2 −p− 1
2 (14)

Since (14) holds for each n ≥ 1, we have

p2 − p − =0
1

√ √
2
1+ 3 1− 3
p= or p= (dropped).
2 2
For C we find the maximal value 1.

2
Problem 2. Convergent Newton iteration (core problem)
As explained in [1, Section 2.3.2.1], the convergence of Newton’s method in 1D may only
be local. This problem investigates a particular setting, in which global convergence can
be expected.
We recall the notion of a convex function and its geometric definition. A differentiable
function f ∶ [a, b] � R is convex, if and only if its graph lies on or above its tangent at
any point. Equivalently, differentiable function f ∶ [a, b] � R is convex, if and only if its
derivative is non-decreasing.
Give a “graphical proof” of the following statement:
If F (x) belongs to C 2 (R), is strictly increasing, is convex, and has a unique zero, then the
Newton iteration [1, (2.3.4)] for F (x) = 0 is well defined and will converge to the zero of
F (x) for any initial guess x(0) ∈ R.
Solution: The sketches in Figure 8 discuss the different cases.

Problem 3. The order of convergence of an iterative scheme (core

problem)
[1, Rem. 2.1.19] shows how to detect the order of convergence of an iterative method
from a numerical experiment. In this problem we study the so-called Steffensen’s method,
which is a derivative-free iterative method for finding zeros of functions in 1D.
Let f ∶ [a, b] � R be twice continuously differentiable with f (x∗ ) = 0 and f ′ (x∗ ) =� 0.
Consider the iteration defined by

f (x(n) ) f (x + f (x)) − f (x)

x(n+1) = x(n) − where g(x) =
g(x(n) ) f (x)
, .

(3a) Write a M ATLAB script that computes the order of convergence to the point x∗
of this iteration for the function f (x) = xex − 1 (see [1, Exp. 2.2.3]). Use x(0) = 1.
Solution:

Listing 30: Order of convergence

1 clear all;
2 % different definitions of f with the same zero:
3 f = @(x) x.* exp(x)-1;
4 % f = @(x) exp(x) -1./x; % f = @(x) x-exp(-x);

3
 5 x0 = 1;
6 x_star = f z e r o (f,x0);
7

8 x = x0; upd = 1;
9 w h i l e (abs(upd) > e p s )
10 fx = f(x(end)); % only 2 evaluations of f at each
step
11 i f fx ≠ 0;
12 upd = fx^2 / (f(x(end)+fx)-fx);
13 x = [x, x(end)-upd];
14 e l s e upd = 0;
15 end
16 end
17 residual = f(x);
18 err = abs(x-x_star);
19 log_err = l o g (err);
20 ratios = (log_err(3:end)-log_err(2:end-1))...
21 ./(log_err(2:end-1)-log_err(1:end-2));
22

23 d i s p ('x, err, residuals,ratios')

24 [x',err',residual',[0;0;ratios']]

The output is

log(en+1 )−log(en )
x error en log(en )−log(en −1)
1.000000000000000 0.432856709590216
0.923262600967822 0.356119310558038
0.830705934728425 0.263562644318641 1.542345498206531
0.727518499997190 0.160375209587406 1.650553641703975
0.633710518522047 0.066567228112263 1.770024323911885
0.579846053882820 0.012702763473036 1.883754995643305
0.567633791946526 0.000490501536742 1.964598248590593
0.567144031581974 0.000000741172191 1.995899954235929
0.567143290411477 0.000000000001693 1.999927865685712
0.567143290409784 0.000000000000000 0.741551601040667

The convergence is obviously quadratic. As in the experiments shown in class, roundoff

affects the estimated order, once the iteration error approaches the machine precision.

4
(3b)
x
The function g(x) contains a term like exe , thus it grows very fast in x and the
method can not be started for a large x(0) . How can you modify the function f (keeping
the same zero) in order to allow the choice of a larger initial guess?
H INT: If f is a function and h ∶ [a, b] → R with h(x) ≠ 0, ∀x ∈ [a, b], then (f h)(x) = 0 ⇔
f (x) = 0.
Solution: The choice f˜(x) = e−x f (x) = x − e−x prevents the blow up of the function g
and allows to use a larger set of positive initial points. Of course, f˜(x) = 0 exactly when
f (x) = 0.

Problem 4. Newton’s method for F (x) ∶= arctan x

The merely local convergence of Newton’s method is notorious, see[1, Section 2.4.2] and
[1, Ex. 2.4.46]. The failure of the convergence is often caused by the overshooting of
Newton correction. In this problem we try to understand the observations made in [1,
Ex. 2.4.46].

(4a) Find an equation satisfied by the smallest positive initial guess x(0) for which
Newton’s method does not converge when it is applied to F (x) = arctan x.
H INT: Find out when the Newton method oscillates between two values.
H INT: Graphical considerations may help you to find the solutions. See Figure 9: you
should find an expression for the function g.
Solution: The function arctan(x) is positive, increasing and concave for positive x,
therefore the first iterations of Newton’s method with initial points 0 < x(0) < y (0) sat-
isfy y (1) < x(1) < 0 (draw a sketch to see it). The function is odd, i.e., arctan(−x) =
− arctan(x) for every x ∈ R, therefore the analogous holds for initial negative values
(y (0) < x(0) < 0 gives 0 < x(1) < y (1) ). Moreover, opposite initial values give opposite
iterations: if y (0) = −x(0) then y (n) = −x(n) for every n ∈ N.
All these facts imply that, if �x(1) � < �x(0) �, then the absolute values of the following iter-
ations will converge monotonically to zero. Vice versa, if �x(1) � > �x(0) �, then the absolute
values of the Newton’s iterations will diverge monotonically. Moreover, the iterations
change sign at each step, i.e., x(n) ⋅ x(n+1) < 0.
It follows that the smallest positive initial guess x(0) for which Newton’s method does not
converge satisfies x(1) = −x(0) . This can be written as

f (x(0) )
x(1) = x(0) − = x(0) − (1 + (x(0) )2 ) arctan x(0) = −x(0) .
f ′ (x(0) )

5
Therefore, x(0) is a zero of the function
g(x) = 2x − (1 + x2 ) arctan x with g ′ (x) = 1 − 2x arctan x.

(4b) Use Newton’s method to find an approximation of such x(0) , and implement it
with Matlab.
Solution: Newton’s iteration to find the smallest positive initial guess reads
2x(n) − (1 + (x(n) )2 ) arctan x(n) −x(n) + �1 − (x(n) )2 � arctan x(n)
(n+1)
=x (n)
− =
1 − 2x arctan x(n) 1 − 2x(n) arctan x(n)
x .

The implementation in Matlab is given in Listing 31 (see also Figure 9).

Listing 31: Matlab Code for NewtonArctan.m

1 clear all;
2 x0 = 2; % initial guess
3 r = 1;
4 w h i l e (r > e p s )
5 x1 = x0-( 2*x0-(1+x0^2)* a t a n (x0) ) /
(1-2*x0* a t a n (x0));
6 % x1 = (-x0+(1-x0^2) * atan(x0))/(1-2 * x0 * atan(x0));
7 r = abs ( (x1 - x0) / x1 );
8 x0 = x1;
9 f p r i n t f ( 'x0 = %16.14e , accuracy = %16.14e \n ' ,
x1, r );
10 end
11

12 figure;
13 x1 = x0- a t a n (x0)*(1+x0^2); x2 = x1- a t a n (x1)*(1+x1^2);
14 X=[-2:0.01:2];
15 p l o t (X, a t a n (X),'k',...
16 X, 2*(X)-(1+(X).^2).* a t a n ((X)),'r--',...
17 [x0, x1, x1, x2, x2], [ a t a n (x0), 0, a t a n (x1), 0,
a t a n (x2)],...
18 [x0,x1],[0,0],'ro',[-2,2], [0,0],'k','linewidth',2);
19 l e g e n d ('arctan', 'g', 'Newton critical iteration'); a x i s
equal;

6
20 p r i n t -depsc2 'ex_NewtonArctan.eps'

Problem 5. Order-p convergent iterations

In [1, Section 2.1.1] we investigated the speed of convergence of iterative methods for the
solution of a general non-linear problem F (x) = 0 and introduced the notion of conver-
gence of order p ≥ 1, see [1, Def. 2.1.17]. This problem highlights the fact that for p > 1
convergence may not be guaranteed, even if the error norm estimate of [1, Def. 2.1.17]
may hold for some x∗ ∈ Rn and all iterates x(k) ∈ Rn .
Suppose that, given x∗ ∈ Rn , a sequence x(k) satisfies

�x(k+1) − x∗ � ≤ C�x(k) − x∗ �p ∀k and some p > 1 .

(5a) Determine ✏0 > 0 such that

�x(0) − x∗ � ≤ ✏0 �⇒ lim x(k) = x∗ .

k→∞

In other words, ✏0 tells us which distance of the initial guess from x∗ still guarantees local
convergence.
Solution:
lim x(k) = x∗ ⇐⇒ lim �x(k) − x∗ � = 0
k→∞ k→∞

Thus we seek an upper bound B(k) for �x(k) − x∗ � and claim that: lim B(k) = 0.
k→∞

�x(k) − x∗ � ≤ C�x(k−1) − x∗ �p
≤ C ⋅ C p �x(k−2) − x∗ �p
2

≤ C ⋅ C p ⋅ C p �x(k−3) − x∗ �p
2 3

⋮
≤ �x(0) − x∗ �p
k−1 k
C�C p
∑ pi
k−1

= C i=0 �x(0) − x∗ �p
k

pk −1
= �x(0) − x∗ �p
geom. series k
C p−1

pk −1
≤ ✏p0 = C 1−p ⋅�C p−1 ✏0 � = B(k)
k 1 1 pk

�
C p−1

const.

lim B(k) = 0 ⇐⇒ C p−1 ✏0 < 1

1

k→∞

7
 �⇒ 0 < ✏0 < C 1−p
1

(5b) Provided that �x(0) − x∗ � < ✏0 is satisfied, determine the minimal kmin =
kmin (✏0 , C, p, ⌧ ) such that
�x(k) − x∗ � < ⌧.

Solution: Using the previous upper bound and the condition ⌧ , we obtain:

�x(k) − x∗ � ≤ C 1−p ⋅ �C p−1 ✏0 � <⌧

1 1 pk

Solving for the minimal k (and calling the solution kmin ), with the additional requirement
that k ∈ N, we obtain:

ln �C 1−p � + pk ⋅ ln �C p−1 ✏0 � < ln ⌧

1 1

��� � � � � � � � � � � �� � � � � � � � � � � � �
<0

� ln �⌧ ) + p−1 ln (C) �
1
k > ln ⋅ kmin ∈ N
1
� � ln (p)
,
ln (C ✏0 )
1
p−1

� �
� � ln �⌧ ) + p−1 ln (C) � 1 ��
�
1
kmin = � ln ⋅
� � (C p−1 ✏ ) � ln (p) ��
� �
1
ln 0

(5c) Write a M ATLAB function

k_min = @(epsilon,C,p,tau) ...

and plot kmin = kmin (✏0 , ⌧ ) for the values p = 1.5, C = 2. Test you implementation for
every (✏0 , ⌧ ) ∈ linspace(0, C 1−p )2 ∩ (0, 1)2 ∩ {(i, j) � i ≥ j}
1

H INT: Use a M ATLAB pcolor plot and the commands linspace and meshgrid.
Solution: See k_min_plot.m.

Listing 32: Matlab Code for k_min_plot.m, Problem 5.

1 C = 2;
2 p = 1.5;

8
 3 eps_max = C^(1/(1-p));
4

5 ngp = 100; % number of grid points

6 eps_lin = l i n s p a c e (0, eps_max, ngp);
7 tau_lin = l i n s p a c e (0, eps_max, ngp);
8 [eps_msh,tau_msh] =
meshgrid(eps_lin(2:(end-1)),tau_lin(2:(end-1)));
9

10 kmin = @(eps , C, p, tau) c e i l ( l o g ( ( l o g (tau) +

(1/(p-1)).* l o g (C) ) ./ l o g (C^(1/(p-1)) .* e p s ) ) ./
l o g (p) );
11 k = kmin(eps_msh, C, p, tau_msh);
12

13 % Consider only gridpoints where: eps larger as tau

14 f o r ne = 1:ngp-2
15 f o r nt = 1:ngp-2
16 i f (ne > nt)
17 k(ne,nt) = 0;
18 end
19 end
20 end
21

22 % Plotting
23 p c o l o r (eps_msh,tau_msh,k)
24 c o l o r b a r ()
25 t i t l e ('Minimal number of iterations for error < \tau')
26 x l a b e l ('\epsilon_0')
27 y l a b e l ('\tau')
28 xlim([0,eps_max])
29 ylim([0,eps_max])
30 s h a d i n g flat

Problem 6. Code quiz

A frequently encountered drudgery in scientific computing is the use and modification of
poorly documented code. This makes it necessary to understand the ideas behind the code
first. Now we practice this in the case of a simple iterative method.

9
(6a) What is the purpose of the following MATLAB code?

1 f u n c t i o n y = myfn(x)
2 l o g 2 = 0.693147180559945;
3

4 y = 0;
5 w h i l e (x > s q r t (2)), x = x/2; y = y + l o g 2 ; end
6 w h i l e (x < 1/ s q r t (2)), x = x*2; y = y - l o g 2 ; end
7 z = x-1;
8 dz = x* exp(-z)-1;
9 w h i l e (abs(dz/z) > e p s )
10 z = z+dz;
11 dz = x* exp(-z)-1;
12 end
13 y = y+z+dz;

Solution: The MATLAB code computes y = log (x), for a given x. The program can be
regarded as Newton iterations for finding the zero of

f (z) = ez − x (15)

where z is unknown and x is given.

(6b) Explain the rationale behind the two while loops in lines #5, 6.
Solution: The purpose of the two while loops is to shift the function values of (15) and
modify the initial z0 = x − 1 in such a way that good convergence is reached (according to
the function derivative).

(6c) Explain the loop body of lines #10, 11.

Solution: The while-loop computes the zero of (15) using Newton iterations
�
� z = x − 1,
� (0)
� (n+1) z (n)
�
� = z (n) − e z(n)−x , n ≥ 1
�
z
e

(6d) Explain the conditional expression in line #9.

Solution: This is a correction based termination criterium, see [1, § 2.1.26].

10
(6e) Replace the while-loop of lines #9 through #12 with a fixed number of itera-
tions that, nevertheless, guarantee that the result has a relative accuracy eps.
Solution: Denote the zero of f (z) with z ∗ , and e(n) = z (n) − z ∗ . Use Taylor expansion of
f (z), f ′ (z):

f (z (n) ) = ez e(n) + ez (e(n) )2 + O((e(n) )3 )

∗ 1 ∗
2
= xe(n) + x(e(n) )2 + O((e(n) )3 ),
1
2
f (z ) = e + e e + O((e(n) )2 )
′ (n) z∗ z ∗ (n)

= x + xe(n) + O((e(n) )2 )

Considering Newton’s Iteration z (n+1) = z (n) − ff′(z )

(n)

(z (n) )
,

f (z (n)
e(n+1) = e(n) −
f ′ (z (n)
xe(n) + 12 x(e(n) )2 + O((e(n) )3 )
= e(n) −
x + xe(n) + O((e(n) )2 )
� (e )
1 (n) 2

= �
2

� ( )1+2+�+2 (e0 )2
1 n n+1

2
= 2 ⋅ ( )2 (e0 )2
1 n+1 n+1

2
= 2 ⋅ ( e0 )2 ,
1 n+1

2
where e0 = z 0 − z ∗ = x − 1 − log (x). So it is enough for us to determine the number n of
iteration steps by � elog x � = eps. Thus
(n+1)

log (log (2) − log (log (x)) − log (eps)) − log (log (2) − log (�e0 �))
n = −1
log (2)
log (− log (eps)) − log (− log (�e0 �))
≈ −1
log (2)
The following code is for your reference.

11
 function y = myfn(x)
log2 = 0.693147180559945;
y = 0;
w h i l e (x > sqrt(2)), x = x/2; y = y + log2; end
w h i l e (x < 1/sqrt(2)), x = x*2; y = y - log2; end
z = x-1;
dz = x*exp(-z)-1;
e0=z-log(x);
k=(log(-log(eps))-log(-log(abs(e0))))/log(2);
f o r i=1:k
z = z+dz;
dz = x*exp(-z)-1;
end
y = y+z+dz;

Issue date: 08.10.2015

Hand-in: 15.10.2015 (in the boxes in front of HG G 53/54).

Version compiled on: October 15, 2015 (v. 1.0).

12
 If the starting value is chosen to be less than the zero
point, then xk > x⇤ for any k 1, and then f (xk ) > 0.

x⇤

x0
x1

The sequence {xk } generated by Newton Method is decreasing, since

f(xk )
xk+1 = xk f 0 (xk ) < xk , for any k 1,

as f (xk ) > 0 and f 0 (xk ) > 0.

xk+1 xk xk 1

Every tangent of the convex function

gk (x) = f (xk ) + f 0 (xk )(x xk ) is
below the function graph f (x). Thus xk > x⇤ for every k 1 and
then the limit of the sequence exists with lim xk = y ⇤ x⇤ . If
k!1
y ⇤ > x⇤ , then gk (y ⇤ ) = f (y ⇤ ) > f (x⇤ ) = 0 which means that the
method can not be stopped at this moment(as the stopping
criterion |xk+1 xk | = ff(x k)
0 (x ) is not small enough). Thus y
k
⇤
= x⇤ .

gk (xk )

gk (x) = f (xk ) + f 0 (xk )(x xk )

13

Figure 8: Graphical proof for Problem 2.

 Figure 9: Newton iterations with F (x) = arctan(x) for the critical initial value x(0)

1.5 arctan
g
Newton critical iteration

0.5

−0.5

−1

−1.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Minimal number of iterations for error < τ

0.25 16

14

0.2
12

10
0.15
τ

0.1
6

4
0.05

0 0
0 0.05 0.1 0.15 0.2 0.25
ϵ0

Figure 10: Minimum number of iterations given ⌧ and ✏0 , Problem 5.

14