Chapter 1

Getting Started

1.1 Introduction
This book is the second in a series of Computer Illustrated Texts on Nu-
merical Methods; the first was called A Simple Introduction to Numerical
Analysis. A Computer Illustrated Text (CIT) consists of a software package
which is integrated with a textbook. It is intended that the reader should,
ideally, use the software frequently whilst working through the book, as the
textbook and software components have been designed to complement and
reinforce each other. However, it is also intended that the textbook compo-
nent can be used without access to a computer, and that the software can
be used on its own, for example, to generate demonstration programs or to
solve problems beyond the scope of this introductory text. It is the com-
bination of text and programs which gives an ideal teaching environment.
Each CIT is self-contained in that all the programs which are required are
either included in the text or supplied on the disc. However, the nature of
the subject means that no short introductory textbook can include all the
material required and so explicit references are given to other sources that
may be needed. No ordering in terms of difficulty is implied between this
CIT and the first, but some of the exercises in the text can be performed
more easily if the disc associated with A Simple Introduction to Numerical
Analysis is available. This first chapter describes how to run the programs,
input data, store results and display graphics.
It is assumed that the reader has a basic knowledge of how to use a
microcomputer but no knowledge of programming languages is needed, only

1
2 CHAPTER 1. GETTING STARTED

an ability to enter data when required.

1.2 The software: an example

The software associated with this book was originally intended for the BBC
microcomputer, but, other versions of the software will be available, and all
instructions in the main part of the book are machine independent. However,
in this section the features of the software will be introduced using the BBC
version of the program INTERP. You may prefer to skip now to Chapter 2
as it is always possible to refer back to this chapter.
Users of machines other than the BBC microcomputer can usefully follow
this section through, making only minor obvious variations. They will also
find that their disc contains some introductory remarks which point out the
most important of such variations.

Figure no longer available

Figure 1.1: Initial screen setting of INTERP

For BBC microcomputer users with the disc version of the software, load-
ing of the program INTERP is accomplished by inserting the disc in drive 0,
holding down SHIFT and then pressing and releasing BREAK. This should
put up a menu on the screen giving a list of all the programs which are
available. (On a network system it will be necessary to ask for advice re-
garding the appropriate loading instructions.) On some screens you may find
that either the top line or the bottom line is obscured; this can be overcome
by selecting the menu option Screen adjustment and then following the
screen instructions to obtain the best setting. Once the menu is visible use
the cursor keys to select the required program. Now select the program IN-
TERP. The screen should clear and after a short delay it should look like
figure 1.1. The screen has been split into three windows. The top window
is used for dialogue, the bottom window contains current settings and the
middle window is used either for graphic or tabular display. Initially, the
option prompt should read
Select data options
where the data option is in inverse colours and a flashing cursor appears
under the ‘d’of data. The data option enables you to input, edit, plot
1.2. THE SOFTWARE: AN EXAMPLE 3

and store data. Other options can be selected by pressing normal character
keys: the SPACE bar is very convenient for this purpose. Alternatively, the
key marked ↑ can be used to cycle forwards and the key marked ↓ to cycle
backwards through the options available. An option is selected by pressing
RETURN, but do not press RETURN just yet.
Whenever the program is loaded a default set of data is supplied and
plotted; the five default points are shown in the display window by crosses.
These values can be tabulated and a method for doing so will be described
later. In the bottom window the current ranges of plotting are given together
with the number of data points (N). It is also possible to input data using
the cursor, the current position of which is also given in the bottom window.
It is a uniform convention, used throughout the software, that all data which
may be changed by the user will have its existing setting left unchanged if the
user presses RETURN or ENTER. Where options need to be selected, the
alternative may be viewed by pressing the SPACE bar repeatedly until the
desired option is showing, then selected by pressing RETURN or ENTER.
The purpose of the program INTERP is to construct smooth curves
through given data points. However, in this chapter we will only be interested
in how data can be input, examined, edited, plotted and stored. Therefore,
ensure that the Select option reads data options and then press RETURN
to select this option. A new prompt should appear:
Data: from keys
Cycle through the options available at this point by pressing the SPACE bar
repeatedly. If at any time you wish to quit press ESCAPE which will take
you back to the following prompt:
Action: Continue
Try it and see. Since we wish to continue, accept the default setting by
pressing RETURN; the program should then advance to the Select data
options prompt as above. Press RETURN again and we arrive back at the
Data prompt.
Select the option Data: from keys by pressing RETURN. Next we
have the option of clearing any existing data points. For the moment let us
just add to the existing points by using the default setting, i.e. by pressing
RETURN. Additional points can either be input in tabular or graphical form.
Pressing RETURN again will select the option
Data input format: tabular
4 CHAPTER 1. GETTING STARTED

and display the current points. A flashing cursor should now appear at
the bottom of the table and instructions concerning the input of data are
given in the top window. Try adding the point (0.95, 0.95) by typing in
each coordinate followed by RETURN. If you make a mistake you can delete
characters using the DELETE key before correcting them but each coordinate
is terminated by RETURN. To quit press Q. This process will return the
program to the option Data: from keys after sorting the tabular points
into ascending order. Now press RETURN to accept the current option of
supplying data from the keyboard and then press RETURN again to retain
the existing points. The program now advances to the option Data input
format: tabular. Press the SPACE bar to change this to Data input
format: graphical and then press RETURN. All the current data points
are displayed on the default graphical settings. These can be changed at this
point using the Graph data option. Cycle through the options available,
superimpose, clear - same axes and clear - new axes. For example, the
last of these options allows the user to set up different sets of axes. When
this option is shown press RETURN. The flashing cursor now appears in the
bottom window and it is possible to alter the range of each axis in turn. For
the moment accept the default settings by pressing RETURN each time the
flashing cursor moves to a new position in the bottom window. When all the
required values have been input the axes will be redrawn.
We will next investigate the graphical editing options. Firstly, select the
option Add point. It is now possible to add additional points either by (i)
specifying the x and y coordinates or (ii) using the cursor. For (i) first press
‘X’ and a flashing cursor will appear in the bottom window as the x coordi-
nate of the cursor. Try inputting a value 0.05 followed by RETURN. Now
press ‘Y’ and input a y coordinate of 0.0 followed by RETURN. The flashing
cursor in the middle window is now at the position (0.05, 0.0). In order to
add this point to the table press RETURN again, this will be confirmed by
an increase in N in the bottom window. To try out method (ii), first select
the Add point option again by pressing RETURN. Additional points can
now be added by moving the cursor with the arrow keys. The increment by
which the cursor moves is controlled by the ‘<’ (increase) or ‘>’ (decrease)
keys and is displayed in the top window. Note that the two methods can
be combined; for example, you may set the x coordinate and then adjust it
using the cursor control keys. The x and y values are open to change up to
the moment when you press RETURN.
The other graphical editing options are Change point, Delete point
1.2. THE SOFTWARE: AN EXAMPLE 5

and Quit editor. The Change point option first asks you to select a point
by using the cursor control keys; at this stage each keystroke makes the cursor
jump to the next point in ascending order (right or up arrow), or descending
order (left or down arrow). When the cursor flashes on the point you wish to
change press RETURN, and the situation is exactly as it was with the Add
point option. The Delete point option also begins by asking you to select
a point; once you have selected it you are asked to confirm that you wish to
delete the current point. The Quit editor option takes the program back
to the option level from which the editor was invoked.
At any stage of data input it is possible to view the current set of data
points by escaping from the current option and selecting the option
Data: tabulate
An option to send any tabulation to a connected printer is available which
reminds you to ensure that the printer is switched on. The option
Data: plot &/or edit
is available to display or modify any existing data set.
It is possible to store data in a file using the option
Data: store on file
and retrieve suitable data using
Data: from file
Wherever possible data files created by one program can be used as input for
any other when it is appropriate, but if an inappropriate data file is requested
then the program will display a suitable error message. Five simple data files
are provided on the program disc1 , as follows:
DATA
SIN
SPLINE
RAMP
FE
When loading a file you will need to type the whole file name at every occa-
sion, however, you need not distinguish between upper and lower case. You
will not be able to store any data on the program disc which is supplied be-
cause it is write protected and so you will need an unprotected disc of your
1
With BBC discs data files have a prefix D.; MS-DOS files will have a .DAT extension.
6 CHAPTER 1. GETTING STARTED

own. You can give the data files any name which is consistent with the rules
of the filing system you are using. Try loading the file DATA, examining it,
editing it and then storing the result.

1.3 The software: general remarks

The previous section demonstrated how to supply simple numerical data and
also the use of cyclical prompts. Although the actual prompts and options
will vary from program to program the repeated use of the RETURN key
will always accept the current default setting and consequently the first time
a program is run results from the default parameters should agree with those
given in the text. Throughout the programs the use of ESCAPE will always
return the program to the Action prompt from which the program can be
restarted or terminated as required. The ultimate panic button on the BBC
model B is the BREAK key; this should only be pressed in emergencies. For
MS-DOS versions BREAK and CTRL/C have the same effect as ESC.
An alternative to numeric input is to supply data by means of expressions.
For example, at any point where the exponential number ‘e’ is required it is
possible to type the string EXP(1). However, if the string supplied cannot be
evaluated then a message such as ‘bad input’ or ‘error evaluating function’
will be given. Such an error will not necessarily be fatal to the program: a
suitable error message will be given and usually the user will be invited to
type in the item again. Some errors cannot be detected until later in the
calculation: these will also give rise to a suitable message, but it will act like
an ESC in that the program will jump to the Action prompt.

1.4 Screen dumps

It is frequently useful to obtain a permanent copy of the screen display for
later examination. Programs running under MS-DOS may use the key la-
belled Print Screen, or an equivalent such as Prnt Sc, to obtain hard
copies of the screen, provided that a suitable printer has been attached and
has been selected by using the GRAPHICS command (consult your manual).
For BBC microcomputer users, there is, unfortunately, no universal standard
for printers or the software to drive them. Therefore, rather than supply a
program which gives a direct screen-to-printer dump we have provided a fa-
1.5. FINAL REMARKS 7

cility for copying the screen image to memory, usually a disc file. The dump
file can then be used to recreate the screen image and users can then provide
their own software to make a hardcopy on any printer available. The details
of this process will depend on the hardware available and the system being
used. For the BBC microcomputer to which a disc drive is connected it is
possible to obtain a screen dump as follows. Whenever a dump option is
shown press RETURN at which point the top window should show
Dump to filename:
Make sure that the disc in the drive is not write-protected and then type in
a suitable file name, which must satisfy the usual BBC filename conventions,
followed by RETURN.
If an Epson printer, or any other device which is software compatible,
is to be used then the screen dump can be copied to the printer using the
program PRNTDMP which is available from the main menu. Selecting
this program gives a self-explanatory dialogue for the user to follow.
One final advantage to this method of obtaining screen dumps is that it
is possible to keep a permanent copy of screen images and recall them to the
screen for future demonstration. For BBC users this can be carried out from
within the program or else in intermediate command mode. In the latter
case if the file DMPF were the name given to a dumpfile created in mode 4
then it can be recalled to the screen, but not printed, using the command
MODE 4:*LOAD DMPF

1.5 Final remarks

If you want to use numerical methods in more complicated programs, as
opposed to learning about them, and you wish to write programs in BBC
BASIC, then A Mathematical Toolkit: Numerical Routines with Applications
in Engineering, Mathematics and the Sciences (Harding 1986) contains sub-
routines which can be built into your own programs. For graphical output,
Graphs and Charts on the BBC Microcomputer (Harding 1982) contains a
library of subroutines for this purpose. It is intended that in the near fu-
ture the CIT series will contain a Mathematical and Graphical Toolkit for
MS-DOS machines.
8 CHAPTER 1. GETTING STARTED
Chapter 2

Interpolation

‘Interpolation - the art of reading between the lines.’

2.1 Polynomial interpolation

We shall begin our discussion of interpolation by considering the following
problem.

Example 2.1.1
In an experiment the yield of a chemical species is recorded as a function
of the concentration of another reagent which is present. As it is impossible
to vary the concentration of the reagent continuously, only a finite set of
different values can be tabulated. Let us assume that the following results
have been obtained.

Table 2.1: Data for example 2.1.1

Concentration of reagent Yield
0.10 0.09956
0.30 0.39646
0.50 0.58813
0.70 0.77210
0.90 0.89608

9
10 CHAPTER 2. INTERPOLATION

We can now ask questions like:

1. Is it possible to estimate the yield when the reagent has a concentration

of 0.51?

2. How accurate is such an approximation going to be?

3. Are the results which are produced sensitive to perturbations or errors

in the tabular values? Clearly, we would like small changes in the data
to produce small changes in the results. If this is the case then the
problem is said to be well-conditioned.

4. Is it possible to determine a reagent concentration which gives a pre-

scribed yield, for example, find a concentration such that the yield is
0.5? (This is the inverse problem to (1)).

The problem of attempting to estimate intermediate values from a given

table is called interpolation and is the basis for the following sections.
In order to provide a background for understanding the methods described
in this chapter we shall begin by considering the problem of obtaining non-
tabular values in elementary ways. We assume that we are given a table
of values (xi , yi), i = 0, . . . , n, and that we need to determine a non-tabular
point (x, y). For example, let us suppose that we wish to find an approximate
value which corresponds to x = 0.51 in table 2.1. Perhaps the simplest way
is to join together successive points as shown in figure 2.1. In particular,
between x = 0.5 and x = 0.7 we have the straight line shown in figure 2.2
and we can find the coordinates of the point C by linear interpolation.
From figure 2.2 the point C has as its y coordinate y(0.5)+CD. However,
by similar triangles,
CD BE
=
AD AE
therefore,

BE (0.77210 − 0.58813)
CD = AD = 0.01 = 0.00920,
AE (0.7 − 0.5)

to five significant figures. (Only five significant figures are quoted here since
the original tabular values are only given to this accuracy.) Hence, y(0.51)
is approximately 0.58813 + 0.00920 = 0.59733.
2.1. POLYNOMIAL INTERPOLATION 11

1.5

1.0

0.5

-0.5 0.5 1.0 1.5

-0.5

Figure 2.1: A simple linear interpolant for table 2.1.1

C
A E
D

0.5 x 0.7

Figure 2.2: Linear interpolation between two points

12 CHAPTER 2. INTERPOLATION

Figure not available

Figure 2.3: The derived function for figure 2.1.1

Instead of using the similar triangle argument above we could have de-
termined an explicit expression for the equation of the straight line joining
the points A and B in figure 2.2. This line will be of the form

y = c0 + c1 x,

and passes through the points (0.5, 0.58813) and (0.7, 0.77210). Therefore,

0.58813 = c0 + 0.5c1

and
0.77210 = c0 + 0.7c1 .
Eliminating c0 gives c1 = (0.77210 − 0.58813)/(0.7 − 0.5) = 0.91985, which
is the gradient of the line. The intercept c0 is then found to be at 0.12821,
hence the equation of the line joining the points A and B is

y = 0.12821 + 0.91985x,

from which y(0.51) = 0.59733 as before. Clearly, this very naive approach is
simple to use but ignores most of the information given in the table. Fur-
thermore, it fits an unnatural function to the data supplied since the fitted
function has a strange derived function. Between each pair of consecutive
data points the fitted function is linear and therefore has constant slope. The
derivative of the piecewise linear function shown in figure 2.1 will therefore be
piecewise constant and look like the function shown in figure 2.3. In order to
overcome this problem we could use higher order functions rather than linear
ones. For example, through any three points we can fit a unique parabola.
Exercise 2.1.1 (Analytical/Computational) Find the parabola through the
points (0.3, 0.39646), (0.5, 0.58813) and (0.7, 0.77210). (Hint: the general
parabolic function is y = a + bx + cx2 , and the three coefficients can be
determined by substituting the three points into this equation to produce
three linear equations which must be solved to find values for a, b and c.)
If the idea of fitting a polynomial through a number of points is unfamiliar,
then it is important to work through exercise 2.1.1. In particular, check that
2.1. POLYNOMIAL INTERPOLATION 13

the resulting quadratic function takes the correct values at x = 0.3, x = 0.5
and x = 0.7. It is not difficult to extend this process to work for arbitrary
polynomials, and it is to this problem that we next turn our attention.
A polynomial of degree n has n + 1 coefficients and so we can fit a unique
polynomial of degree n through n + 1 distinct tabular points, xi , i = 0, . . . , n.
Let us assume that we are given a table of values (xi , yi ), i = 0, . . . , n, and
try to fit such a polynomial. Let Pn be the interpolating polynomial
Pn (x) = c0 + c1 x + c2 x2 + . . . + cn xn .
Now we select the coefficients cj , j = 0, . . . , n, so that the equations
Pn (xi ) = yi, i = 0, . . . , n, (2.1)
are satisfied. This gives a system of n + 1 linear equations in the n + 1
unknown coefficients cj , j = 0, . . . , n. These linear equations may then be
solved using, for example, Gaussian elimination with partial pivoting. (See
Harding and Quinney (1986).)
Example 2.1.2
For the five values in table 2.1 we can fit a unique polynomial of degree 4,
i.e.
P4 (x) = c0 + c1 x + c2 x2 + c3 x3 + c4 x4 .
Evaluating this polynomial at the five data points in table 2.1 gives the
following system of linear equations:

0.12 0.13 0.14

    
1 0.1 c0 0.09956
1 0.3 0.32 0.33 0.34 c1 0.39646
    
    
0.52 0.53 0.54
    

 1 0.5 
 c2 
 = 
 0.58813 .
 (2.2)

 1 0.7 0.72 0.73 0.74 
 c3 


 0.77210 

1 0.9 0.92 0.93 0.94 c4 0.89608
Using Gaussian elimination, the solution of this system of equations is
[−0.159796, 3.164656, −6.499406, 8.274375, −3.901563]T
which produces the interpolating quartic
P4 (x) = −0.159796 + 3.164656x − 6.499406x2 + 8.274375x3 − 3.901563x4.
(You can easily check these results using the program GAUSS from Harding
and Quinney (1986).) Evaluating P4 at the given points produces the values
in table 2.2. (Later we shall provide an alternative simpler method.)
14 CHAPTER 2. INTERPOLATION

Table 2.2: Polynomial interpolation applied to table 2.1.1

x y P4 (x) P4 (x) − y
0.1 0.09956 0.0995598 2.4 × 10−7
0.3 0.39646 0.3964597 2.8 × 10−7
0.5 0.58813 0.5881297 3.1 × 10−7
0.7 0.77210 0.7720996 3.9 × 10−7
0.9 0.89608 0.8960794 5.7 × 10−7

1.5

1.0

0.5

-0.5 0.5 1.0 1.5

-0.5

Figure 2.4: Quartic polynomial for table 2.1.1

Notice that the values in the third column of table 2.2 are displayed to
seven decimal places even though the data are only given to five. This is
to emphasize the fact that whenever this process is carried out it will be
prone to numerical errors. For this example, if we round the results in col-
umn 3 to five decimal places we recover the initial data. Whether this is
always the case will be examined later. The function obtained, P4 (x), is
shown in figure 2.4. Notice that between the tabulated points the interpo-
lating polynomial is quite smooth and so we are justified in evaluating it at
non-tabular points. For example, we can determine a value for y(0.51) by
P4 (0.51) = 0.59733892. However, this is not always the case as the next ex-
ample will demonstrate. Outside the range of the tabular points it would be
unwise to use the interpolating polynomial to extrapolate the given values.
This can be seen by examining the curve in figure 2.4 for values of x in excess
of 0.9 where one might expect a monotonic increasing function.
2.1. POLYNOMIAL INTERPOLATION 15

Example 2.1.3
The interpolating polynomial which passes through the points (−2, e−4 ),
(−1, e−1 ), (0, 1), (1, e−1 ), and (2, e−4 ) is shown in figure 2.5. These val-
ues are obtained from the function exp(−x2 ) which is always positive but
as figure 2.5 shows the same is not true of the interpolating polynomial for
values of x in the intervals (−2, −1) and (1, 2)! Furthermore, as the number
of points increases the interpolating polynomial behaves less like the func-
tion exp(−x2 ). This is called Runge’s problem and will be discussed in more
detail in the next chapter. (See also exercises 2.1.4 and 3.4.4.)

1.5

1.0

0.5

-2 -1 1 2

-0.5

Figure 2.5: Runge’s problem: see example 2.1.3

Exercise 2.1.2 (Computational) The program INTERP is an interactive

program which can be used to construct interpolating polynomials through a
set of data points. The values in table 2.1 are set up as the default tabulated
points. Therefore, load the program INTERP and when the screen shows
Select data options
press the space bar which will change the option to Select method. Press
RETURN and the option prompt will change to
Method: Construct interpolating polyn.
You can cycle through the various methods by repeatedly pressing any char-
acter key except RETURN; now select the option above by pressing RE-
TURN, which will calculate the required interpolating polynomial passing
through the points in table 2.1. First of all, you are allowed the option of
16 CHAPTER 2. INTERPOLATION

selecting the graphics parameters which are used to plot the resulting poly-
nomial. For the moment select the option
Graph data: superimpose
The interpolating polynomial is now drawn in the graphics window and the
actual polynomial is given in the bottom window. (The coefficients are lim-
ited to five significant figures by the space available to display them.) We
can now evaluate the required interpolating polynomial using the option
Option: Evaluate interpolant
Entering the value for x = 0.51 produces the same value as in example
2.1.2. All computation is carried out to full accuracy; only the coefficients
are displayed to five significant figures.
Exercise 2.1.3 (Computational) Use the program INTERP and any four
values from table 2.1 to determine a cubic interpolant and compare the esti-
mate for y(0.51) with the value determined above. This can be carried out by
quitting the current option by pressing Q and then selecting the option Edit
data. Press RETURN in order to accept the current plotting parameters
and then cycle through the edit options available by pressing any key but
RETURN. (Recall that you can cycle backwards using the key marked ↓.)
When the current option is Delete point press RETURN and use the cursor
keys to select the point you wish to remove and then press RETURN again.
Cycle through the options now available until the option is Quit editor and
then press RETURN. Once again cycle through the options available until
the current option becomes Select method and press RETURN yet again.
Accepting the option Method: Construct interpolating polyn., the in-
terpolating polynomial through the remaining points will be determined and
plotted. The effect of deleting another point can be examined as follows.
Cycle until the option Edit data is current and then press RETURN twice.
This will take you into the data editor; the current edit option should be Add
point. If this is the case press RETURN, if not cycle until it is and then
press RETURN. The cursor in the graphics window should now be flashing
at the point which was previously deleted; this point can be added back to
the table simply by pressing RETURN. Other points can then be deleted to
consider the effect of fewer points in the table. How does the selection of the
four points influence the interpolated value?
Exercise 2.1.4 (Computational) Use the program INTERP to obtain five
equally spaced points in the interval (−5, 5) from y = exp(−x2 ) and then
2.1. POLYNOMIAL INTERPOLATION 17

determine a suitable interpolating polynomial (see example 2.1.3). You will

need to take the default prompt Select data option and then Data: from
keys remembering to clear the existing data before using the Data input
format: tabular option. For example, the point (−5, exp(−25)) can be
input as the tabular values X(1) = −5 and Y (1) =EXP(−25). When you
have constructed the interpolating polynomial through the points you have
selected, try adding other points and see how the results compare. What
happens as the number of points increases/decreases? Why is it better to
have an even number of points? Investigate what happens with non-equally
spaced points.
In general, given n+1 tabular points, (xi , yi), i = 0, . . . , n, we can proceed
as above and determine the coefficients of the interpolating polynomial Pn (x)
by solving a system of linear equations given by equation (2.1), or equivalently

Ac = y, (2.3)

where A is an n+1 by n+1 matrix and c and y are vectors of dimension n+1.
The columns and rows of A are numbered 0 to n and the components of A
are aij = xji , i = 0, . . . , n, j = 0, . . . , n. Similarly yi = y(xi ), i = 0, . . . , n, and
ci are the coefficients of the interpolating polynomial which is required. (See
equations (2.2) for a simple example with n = 4.) This system of equations
will have a unique solution provided the matrix A is non-singular, therefore,
we examine the possibility of it being singular or almost singular. In order to
do this we consider the determinant of A; this matrix is called a Vandermonde
matrix and it is possible to show that it has determinant
Y
(xi − xj ).
i=1,...,n, j<i

Exercise 2.1.5 (Analytical) Show that the determinant of the Vandermonde

matrix of order 3, given by

1 x0 x20
 

 1 x1 x21  ,
 

1 x2 x22

is (x1 − x0 )(x2 − x0 )(x2 − x1 ).

The importance of the above expression for the determinant of the Vander-
monde matrix is that if the tabulation points are close together then the
18 CHAPTER 2. INTERPOLATION

determinant may be small, i.e. the equations are nearly singular, and hence
the solution may be inaccurate. This type of inaccuracy is usually called
inherent ill-conditioning and is a fundamental part of the problem, i.e. in-
dependent of the method of solution. In order to demonstrate the problems
which may arise when determining interpolating polynomials in such circum-
stances let us consider the following example.
Example 2.1.4 (Computational)
Let us begin by finding the quartic polynomial which passes through the
points in table 2.3. Using the program INTERP produces the interpolating
polynomial shown in figure 2.6. In order to consider how sensitive this prob-
lem is let us perturb the y value of the third entry in table 2.3 from 1.05 to
0.95 and see how this affects the computed solution. The resulting quartic
polynomial is shown in figure 2.7. Compared with the original interpolation
quartic it is clear that this problem is inherently ill-conditioned.

Table 2.3:
x y
0.00 2.00
0.95 1.00
1.00 1.05
1.05 1.00
2.00 0.00

Example 2.1.4 gave an illustration of a problem which was inherently

ill-conditioned, i.e. a small change in the tabular values leads to a much
larger change in the interpolating polynomial. This was caused by several
of the tabular values being close together, resulting in a system of linear
equations which were nearly singular. We can also investigate the possibility
of errors being introduced in the interpolating polynomial by the method we
have selected to solve this problem. Such perturbations are called induced
errors. Ideally we would like small changes in the tabular values to produce
small changes in the interpolating polynomial and if this is the case then the
problem is said to be well conditioned. In order to demonstrate this idea
consider the following example.
2.1. POLYNOMIAL INTERPOLATION 19

0
0.5 1.0 1.5 2.0
-1

-2

-3

-4

-5

-6

Figure 2.6: The quartic interpolant for table 2.1.3

0
0.5 1.0 1.5 2.0

-2

Figure 2.7: The effect of perturbing entry 3 in table 2.1.3

20 CHAPTER 2. INTERPOLATION

Example 2.1.5 (Computational)

Applying the program INTERP to construct an interpolating polynomial
through the points in table 2.1 produces the polynomial

P4 (x) = −0.159796 + 3.164656x − 6.499406x2 + 8.274375x3 − 3.901563x4.

Now let us round the entries in table 2.1 to four decimal places and then con-
struct a polynomial through the resulting points. This gives the polynomial

−0.1597 + 3.1652x − 6.5044x2 + 8.2833x3 − 3.9063x4 .

The maximum relative change in any coefficient is less than 0.12% and so
this problem is relatively well conditioned.
Exercise 2.1.6 (Computational) Use the program INTERP to construct the
interpolating polynomial through the tabular values in the file SIN. Investi-
gate the sensitivity of the data provided by examining the effect of perturbing
one or more of the values given. Is this problem inherently well conditioned?
Is it susceptible to induced ill-conditioning?

2.2 Lagrangian Polynomials

In order to construct a polynomial of degree n through n + 1 distinct points
using equation (2.1) it is necessary to solve a system of linear equations. We
will now derive an alternative way of producing such a polynomial which does
not require the solution of a system of possibly ill-conditioned equations. As
there is only one polynomial of degree n which passes through n + 1 points
we would obtain the same polynomial as that obtained from equation (2.1)
if exact arithmetic were used. (See exercise 2.2.4.)
Let us begin by assuming that we have been given a set of data points,
(xi , yi ), i = 0, ..., n, and then examine the function

(x − x0 )(x − x1 )...(x − xi−1 )(x − xi+1 ) . . . (x − xn ).

Notice that there is no factor (x − xi ). This function can also be written as

the product
Πj6=i (x − xj ).
and is a polynomial of degree n which vanishes at each data point with the
exception x = xi . We now dene the Lagrangian polynomial functions, Li (x),
2.2. LAGRANGIAN POLYNOMIALS 21

i = 0, ..., n, as
Πj6=i(x − xj )
Li (x) =
Πj6=i(xi − xj )
We see that each of these polynomials has degree n and that Li (xk ) = 0if
k 6= ibut is equal to 1 when k = i. Now consider the function

i=n
X
P (x) = yi Li (x).(2.5)
i=0

Since Li (x) is a polynomial of degree n for each i, so is P (x) and by virtue

of the way in which each of the functions Li are dened P (x) will take the
values yi when x = xi , i = 0, ..., n. Therefore, P (x) and the interpolating
polynomial Pn (x) are identical since they are both polynomials of degree n
which agree at n + 1 points.
Example 2.2.1 (Analytical)
Let us consider the data points shown in table 2.1. The Lagrangian polyno-
mial L0 is given as follows:

(x − 0.3)(x − 0.5)(x − 0.7)(x − 0.9)

L0 (x) =
(0.1 − 0.3)(0.1 − 0.5)(0.1 − 0.7)(0.1 − 0.9)
= 26.04167x4 − 62.5x3 + 53.64583x2 − 19.375x + 2.460937.

We stress that in practice an explicit expression of such a polynomial would

not be determined; it is easier to evaluate it in the form of equation (2.5).
You should check that L0(x0 ) = 1, i.e. L0 (0.1) = 1, and that L0 (xi ) = 0,
i = 1, 2, 3, 4, i.e. L0 (0.3) = L0 (0.5) = L0 (0.7) = L0 (0.9) = 0. However,
because of the form of equation (2.5) it is beter to consider yi Li (x), for
example, y0 L0 (x) is plotted in figure 2.8.
Similarly, we can construct the polynomials L1 (x), L2 (x), L3 (x) and L4 (x).
The required interpolating quartic polynomial is then given by

P4 (x) = 0.09956L0 (x)+0.39646L1(x)+0.58813L2 (x)+0.77210L3 (x)+0.89608L4(x).

To evaluate P4 (0.51) we rstly nd L0 (0.51) = 0.004052 from equation (2.4) and

then, similarly, we have that L1 (0.51) = −0.03165, L2 (0.51) = 0.99688, L3(0.51) =
0.03498 and L4 (0.51) = −0.00426, hence P 4(0.51) = 0.59734 as before.
22 CHAPTER 2. INTERPOLATION

1.0

0.8

0.6

0.4

0.2

-0.5 0.5 1.0 1.5

-0.2

-0.4

Figure 2.8: The Lagrangian polynomial y0 L0 (x)

Exercise 2.2.1 (Computational) The program INTERP may be used to con-

struct a suitable interpolating polynomial through the points given in table
2.1. Load INTERP, cycle through the options available with the Se- lect
prompt until it reads method and then press RETURN. Cycle again until
the Method prompt is Lagrangian polynomials and press RETURN again.
Now use the default Graph option which is superimpose and a ashing cursor
will appear in the graphics window. Di?erent points can be selected using
the cursor keys and then pressing RETURN once again. For example, the
Lagrangian polynomial L0 (x) can be drawn by simply pressing RETURN
when the ashing cursor is positioned over the rst tabular point. However, in
order to make use of the form of equation (2.5) it is more conve- nient to use
the functions y0 L0 (x) and it is this function which is plotted in the graphics
window and displayed in the bottom window. In a similar way the other four
Lagrangian polynomials can be drawn.
Exercise 2.2.2 (Computational) Use the program INTERP to construct the
Lagrangian polynomials through the data in le DATA. Sum these poly- no-
mials as in equation (2.5) and show that the result is equivalent to the poly-
nomial given by direct construction of the interpolating polynomial.
Exercise 2.2.3 (Computational) Use the program INTERP to determine suit-
able third- and fourth-order polynomial interpolants for the data given in
table 2.1. (Hint: use the editor to delete points one at a time.)
Exercise 2.2.4 (Analytical) Show that there is a unique polynomial of degree
n which passes through the (n + 1) points (xi , yi ), i = 0, ..., n. (Hint: consider
2.3. DIVIDED DIFFERENCES 23

two polynomials pn (x) = a0 +a1 x+...+an xn and qn (x) = b0 +b1 x+...+bn xn ,

with ai 6= bi for at least one i, each of which interpolates the points given.
Now consider the non-trivial polynomial pn (x) − qn (x) at the points xi and
hence show that ai = bi , i = 0, ..., n.)
Exercise 2.2.5 (Analytical) Hermite Interpolation. Lagrangian interpolation
is based upon tabular values (xi , yi ), i = 0, ..., n; if in addition the gradients
y?i are supplied then higher order polynomials can be tted to the data. Show
that it is possible to t a polynomial of degree 2n + 1 in the form
i=n i=n
βi (x)yi′
X X
P2n+1 (x) = αi (x)yi +
i=0 i=0

where

αi (x) = [1 − 2Li (xi )(x − xi )]L2i (x),

βi (x) = (x − xi )L2i (x).

The result is called a Hermite interpolation polynomial.

2.3 Divided Differences

As another alternative to the construction of an interpolating polynomial
through the points (xi , yi), i = 0, ..., n, using Lagrangian polynomials we can
determine the interpolating polynomial of degree n as follows. If we dene
the function p0 (x) = y0 then, trivially, p0 (x0 ) = y0 . Now consider the linear
polynomial
p1 (x) = y0 + a1 (x − x0 ),
which satises p1 (x0 ) = y0 and then select a1 = (y1 − y0 )/(x1 − x0 ) so that

p1 (x1 ) = y0 + a1 (x1 − x0 ) = y1 .

Next dene
p2 (x) = a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ),
then since a0 = y0 and a1 = (y1 − y0 )/(x1 − x0 ) we have that p2 (x0 ) = y0 and
p2 (x1 ) = y1. Furthermore,

p2 (x) = p1 (x) + a2 (x − x0 )(x − x1 ).

24 CHAPTER 2. INTERPOLATION

We still have a2 at our disposal, but if

(p2 (x2 ) − p1 (x2 ))

a2 =
(x2 − x0 )(x2 − x1 )

then p2 (x2 ) = y2 , i.e. p2 (x) passes through (x0 , y0 ), (x1 , y1 ) and (x2 , y2 ). We
can repeat this process by dening

pj (x) = pj−1 (x) + aj (x − x0 )(x − x1 )(x − x2 ) . . . (x − xj−1 ), j = 1, ..., n,

and then selecting aj so that pj (xj ) = yj . The resulting polynomial, pn (x),

passes through (xi , yi), i = 0, ..., n, and is identical with that obtained by
using the Lagrangian polynomials.
Example 2.3.1 (Computational)
The interpolating polynomial through the rst three values in table 2.1 may
be constructed as follows. Using the above method gives

p0 (x) = y0 = 0.09956,
p1 (x) = p0 (x) + a1 (x − x0) = 0.09956 + a1 (x − 0.1).

(0.39646−0.09956)
Select a1 so that p1 (x1 ) = 0.39646, i.e. a1 = (0.3−0.1)
which gives

p1 (x) = 0.09956 + 1.4845(x − 0.1).

Next we dene p2 (x) by

p2 (x) = p1 (x) + a2 (x − 0.1)(x − 0.3)

and select a2 so that p2 (x2 ) = y2 , i.e. 2 (0.5) = 0.58813,

(p2 (x2 ) − p( x2 )) (0.58813 − 0.69336)

a2 = = = −1.315375.
(x2 − x0 )(x2 − x1 ) (0.5 − 0.1)(0.5 − 0.3)

This gives the polynomial

p2 (x) = 0.09956 + 1.4845(x − 0.1) − 1.315375(x − 0.1)(x − 0.3)

which passes through the first three entries in table 2.1.

2.3. DIVIDED DIFFERENCES 25

The process outline above can be automated in the following way. Given
the table of values (xi , yi), i = 0, ..., n, we dene the first divided differences
associated with these values by
(yi+1 − yi )
f [xi , xi+1 ] = .
(xi+1 − xi )
Next we compute second divided di?erences as
(f [xi+1 , xi+2 ] − f [xi , xi+1 ])
f [xi , xi+1 , xi+2 ] =
(xi+2 − xi )
and so on. When the (k − 1)th divided di?erences have been obtained,
i.e. f [xi , xi+1 , ..., xi+k−1 ] and f [xi+1 , xi+2 , ..., xi+k ] we dene the k th divided
di?erence as
(f [xi+1 , ..., xi+k ] − f [xi , ..., xi+k?1 ])
f [xi , xi+1 , ..., xi+k ] =
(xi+k − xi )
These differences can be shown schematically as in table 2.4. We can
then show that the polynomial pn (x), given by
pn (x) = y0 + f [x0 , x1 ](x − x0 ) + f [x0 , x1 , x2 ](x − x0 )(x − x1 )
= + . . . + f [x0 , x1 , . . . , xn ](x − x0 )(x − x1 ) . . . (x?xn?1),
passes through the points (xi , yi ), i = 0, ..., n, and is therefore identical to
the interpolating polynomial Pn (x). (See exercise 2.2.4.)
Exercise 2.3.1 (Analytical) Show that the above expression is the interpolat-
ing polynomial through the points (xi , yi), i = 0, ..., n.
Exercise 2.3.1 (Analytical) Show that f [x1 , x2 , . . . , xk ] is invariant with re-
spect to all permutations of the tabular points x1 , x2 , . . . , xk .
Example 2.3.2 (Calculator)
To demonstrate the use of divided differences we construct the interpolating
polynomial through the ve points in table 2.1. Firstly, we obtain the table
of divided differences given by table 2.5, as illustrated in table 2.4. The
required quartic polynomial is then given by
P4 (x) = 0.09956 + 1.48450(x − 0.1) − 1.31538(x − 0.1)(x − 0.3)
+ 2.031875(x − 0.1)(x − 0.3)(x − 0.5)
− 3.9015625(x − 0.1)(x − 0.3)(x − 0.5)(x − 0.7),
26 CHAPTER 2. INTERPOLATION

Table 2.4: Table of divided differences

x y First divided Second divided Third divided

difference difference difference
x0 y0
(y1 −y0 )
f [x0 , x1 ] = x1 −x0 )
f [x1 ,x2 ]−f [x0 ,x1 ]
x1 y1 f [x0 , x1 , x2 ] = x2 −x0
(y2 −y1 )
f [x1 , x2 ] = x2 −x1 )
f [x1 , x2 , x3 , x4 ]
f [x2 ,x3 ]−f [x1 ,x2 ]
x2 y2 f [x1 , x2 , x3 ] = x3 −x1
(y3 −y2 )
f [x2 , x3 ] = x3 −x2 )
x3 y3

Table 2.5: Table of divided differences

x y First divided Second divided Third divided Fourth divided

difference difference difference difference
0.1 0.09956
1.48450
0.3 0.39646 -1.31538
0.95835 2.031875
0.5 0.58813 -0.09625 -3.9015625
0.91985 -1.089375
0.7 0.77210 -0.74988
0.61990
0.9 0.89608
2.4. NEVILLE’S ALGORITHM 27

where the coe?cients are given by the rst entry in each column of the tableau
of divided di?erences. Expanding this expression gives the polynomial

P 4(x) = −0.159796 + 3.164656x − 6.499406x2 + 8.274375x3 − 3.901563x4,

which agrees with the interpolating polynomial through these points as com-
puted previously.
Exercise 2.3.3 (Analytical) Use the method given in section 2.1 to determine
an interpolating polynomial of degree 2 which passes through the points (0,
3), (1, 6) and (2, 11). Obtain the corresponding Lagrangian polynomials L0 ,
L1 and L2 and hence show that the polynomial

3L0 (x) + 6L1 (x) + 11L2 (x)

is identical with the interpolating polynomial which was derived using equa-
tion (2.1). Construct a table of divided di?erences for these data points and
hence construct a quadratic function which passes through the same points.
Conrm that all three methods give the same polynomial.
Exercise 2.3.4 (Programming) Write a simple program to read in a set of
tabular points and construct a table of divided differences. Test your pro-
gram on the data given in exercise 2.3.3. Extend the program to produce
polynomial interpolants as described in example 2.3.1.

2.4 Neville’s algorithm

Although the previous sections demonstrated how it was possible to construct
a suitable interpolating polynomial explicitly it is clear that a considerable
amount of effort is required in order to interpolate from the values given in
a table. Furthermore, neither the method outlined in section 2.1 nor the
Lagrangian method gives any indication of the order of polynomial required.
What we seek is some means of using a subset of the tabulated points, m
say, to construct a suitable interpolant and then by adding an additional
point construct a higher order polynomial. In this way if we can show that
the interpolants produced using subsets of m points and m + 1 points agree
then there is no point in investigating higher order polynomials. Neville’s
algorithm does precisely this.
28 CHAPTER 2. INTERPOLATION

Let us suppose we select a subset of m points from a table of values,

(x0 , y0), (x1 , y1), . . . , (xm−1 , ym−1 ). Next we construct the linear functions
(x − x0 )y1 − (x − x1 )y0
p01 (x) = ,
(x1 − x0 )
(x − x1 )y2 − (x − x2 )y1
p12 (x) = ,
(x2 − x1 )
..
.

(x − xm−2 )ym−1 − (x − xm−1 )ym−2

pm−2,m−1 (x) = .
(xm−1 − xm−2 )
Notice that pj,j+1(x) passes through the points (xj ,yj ) and (xj+1 ,yj+1), j =
0, . . . , m − 1. (These terms are sometimes called linear cross means.) Next
we construct the linear cross means between successive linear terms, e.g.
between p01 (x) and p12 (x), and then evaluate the quadratic function
(x − x0 )p12 (x) − (x − x2 )p01 (x)
.
(x2 − x0 )
This function passes through (x0 , y0 ), (x1 , y1) and (x2 , y2 ) and is denoted by
p02 (x). In this way we can recursively define a sequence of polynomials of
increasing order by
(x − xj )pj+1,k (x)−(x − xk )pj,k−1(x)
pjk (x) = , k > j, k = 1, .., m. (2.4)
xk − xj

This polynomial is of degree (k − j) and passes through the points (xi , yi )

for i = j, . . . , k, and is therefore equivalent to the Lagrangian polynomial of
the same degree (see exercise 2.2.4).
Exercise 2.4.1 (Analytical) Show that the quadratic polynomial p03 (x) satis-
fies the following: p03 (x0 ) = y0 , p03 (x1 ) = y1 , p03 (x2 ) = y2 .
In practice if we are given a set of data and require an estimate at a
non-tabular point x∗ then, rather than evaluating the general polynomial,
we only evaluate (2.4) at the point x∗ . Furthermore, it is usual to take the
tabular points beginning at the point closest to x∗ and taking subsequent
points in increasing distance from x∗ so that x∗ is near the middle of the
values used.
2.4. NEVILLE’S ALGORITHM 29

Example 2.4.1 (Computational)

Use Neville’s algorithm to find an estimate for y(0.51) in table 2.1 on page 9.
Load INTERP and cycle on the Select method prompt to select Neville’s
Algorithm. We shall use the option Direct so press RETURN again, and
then enter the value x = 0.51. The values in table 2.1 are set as default values
and should appear in a window on the right of the screen. We can now select
the entries from this table in order of closeness to the point x = 0.51. The
tabular value x = 0.5 is the closest one to x = 0.51, so that is the first one we
wish to select. To do so, use the up/down cursor keys to position the inverse
video block over the required value and then press RETURN. Another inverse
video block will now appear under the heading ‘x, y selected’, and you can
copy your chosen values into this just by pressing RETURN. The next tabular
values needed are those at x = 0.7, which should be selected and copied in a
similar way. At this point type Q to stop selecting values and the program will
then display a tableau based on the two values x = 0.5 and x = 0.7. Further
points can be added by typing Q and selecting the Select/edit data option.
Add further values until you obtain the tableau shown in table 2.6. Notice
that some editing of the selected values is allowed (should it be needed).
For example, you can delete an entry (the entries below then move up one
place), or copy one of the data points on top of another already selected
entry which then replaces that entry. (Only three columns of the table of
cross means are visible on the screen at any one time; other columns can be
viewed using the left/right cursor keys.) Notice that by taking the tabulation
points in such a way that the values |x∗ − xi | are in increasing order of size
the top entry in each column gives the interpolated value required. When
the entries in successive columns are sufficiently close we can terminate the
process. Additional values can be added to the bottom of the table in order
to improve the current interpolated value. In this example the second and
third entries agree to three decimal places and so it is sensible to quote the
result as y(0.51) ≈ 0.598. (In fact the value of p03 is much better than this;
from section 2.1, P4 (0.51) = 0.59734.)
Exercise 2.4.2 (Computational) Use the program INTERP with the op-
tion Neville’s algorithm to estimate y(0.51) by adding the final entry to
table 2.6 from table 2.1.
Unfortunately, Neville’s algorithm does not always work as well as shown
in example 2.4.1. If the tabulated data points are sufficiently smooth then
all is well but consider the following example.
30 CHAPTER 2. INTERPOLATION

Table 2.6: Neville’s algorithm applied to table 2.1.1

xi (0.51 − xi ) yi Linear Quadratic Cubic

0.5 +0.01 0.58813
p01 = 0.59733
0.7 −0.19 0.77210 p02 = 0.59751
p12 = 0.59367 p03 = 0.59795
0.3 +0.21 0.39646 p13 = 0.6149
p23 = 0.57133
0.9 −0.35 0.89608

Example 2.4.2 (Computational)

Table 2.7 illustrates the application of the program INTERP, with option
Neville’s algorithm, to estimate y(0.15), from the values given. The result
is poor because the tabulated values are obtained from the function y(x) =
1/x which changes very rapidly near x = 0. (The correct value of y(0.15) =
6.666666, illustrating the difficulties which may appear if the data points
vary greatly.)

Table 2.7: Results produced by Neville’s algorithm

x y Linear Quadratic Cubic Quartic

0.1 10.000000
7.500000
0.2 5.000000 7.083333
5.833333 6.927083
0.3 3.333333 6.145833 6.848958
4.583333 6.302083
0.4 2.500000 5.208333
3.750000
0.5 2.000000

Exercise 2.4.3 (Computational) Use the program INTERP to retrieve data

from file SIN and use Neville’s algorithm to estimate y( 31 π). The data in
2.4. NEVILLE’S ALGORITHM 31

this file is obtained from the function y = sin x. Insert additional points
and investigate how the accuracy of the interpolated point depends on the
spacing and number of points.

2.4.1 Inverse interpolation

Thus far we have concentrated upon the problem of interpolating non-tabular
values. In this section we will look briefly at the problem of inverse interpola-
tion which was mentioned on page 10. The simplest solution to this problem
is to reverse the roles played by the dependent and independent variables; for
example, we simply rewrite table 2.1 on page 9 as table 2.8. This technique
must be used with care because if there are two x values with the same y
value then we will not get a single-valued interpolating function. This is
equivalent to obtaining a singular matrix for equation (2.1). Furthermore,
we need the original tabular values to be monotonic otherwise the order of
the points will change when the roles of dependent and independent variables
are interchanged.

Table 2.8: Interchanging variables in table 2.1.1

Yield, y Concentration of reagent, x

0.09956 0.10
0.39646 0.30
0.58813 0.50
0.77210 0.70
0.89608 0.90

Exercise 2.4.4 (Computational) Use interpolation applied to table 2.1 on

page 9 to determine x(0.4) using Neville’s algorithm and the option In-
verse. (Notice that applying inverse interpolation to this table is equivalent
to applying interpolation to table 2.8. The program INTERP retains the
original order of each (x, y) pair on the screen although their roles are re-
versed.) Check the result by using the option Construct interpolating
polyn. and evaluating it at the point you have determined. How would you
account for any discrepancy?
32 CHAPTER 2. INTERPOLATION

The procedure of interchanging the dependent and independent variable

can be very successful but must be used with care. It is clear that interpo-
lation works best when the xi values are not too close and the yi values do
not change too quickly. If this is the case then we can expect a reasonably
accurate interpolation, but if these conditions are not satisfied then inverse
interpolation may be difficult. However, in order to illustrate the usefulness
of inverse interpolation consider the following simple problem.

Table 2.9: Inverse interpolation applied to x2 − 3x + 1 = 0

x y
0.4 −0.040
0.38261
0.3 0.190 0.38189
0.38636 0.38195
0.5 −0.250 0.38261 0.38196
0.39130 0.38213 0.38196
0.2 0.440 0.38392 0.38202
0.36296 0.38258
0.1 0.710 0.37721
0.34483
0.0 1.000

Example 2.4.3 (Computational)

Use inverse interpolation to determine the solutions of the quadratic equation

x2 − 3x + 1 = 0.

Tabulating this function between x = 0 and x = 4 shows that the solutions

lie in the intervals (0, 0.5) and (2.5, 3). Now load the program INTERP
and enter values for f (x) = x2 − 3x + 1 for values of x between 0 and 0.5.
Using Neville’s algortihm and the option Inverse produces the results
in table 2.9. Notice that even though the roles of the dependent variable x
and the independent variable y have been interchanged they are displayed
in the same order on the screen. From this table the required solution
√ is at
x = 0.38196. (The solutions of the given quadratic are at x = 12 (3 ± 5), the
smallest of which is 0.381966011.)
2.5. SPLINE INTERPOLATION 33

Exercise 2.4.5 (Computational) Find the larger root of the quadratic equation
given in example 2.4.3 by tabulating the function f (x) in the interval (2.5, 3)
and then using inverse interpolation.
Exercise 2.4.6 (Computational) Load the data from file SIN and use inverse
interpolation to determine a value of x such that y = 0.5. The data in this file
are obtained from the function y = sin x; investigate the effect of additional
points and different orderings.

2.5 Spline interpolation

Example 2.1.2 showed how it was possible to fit a polynomial of degree 4
through five points, but the resulting function was quite oscillatory. We will
now investigate how it is possible to fit lower order interpolants which are
globally ‘smoother’. For example, through each pair of points (xi , yi ) and
(xi+1 , yi+1) there is a unique straight line, but such a function is unsatisfac-
tory since it will in general have kinks at each data point. (See figure 2.9.)
Mathematically speaking, the interpolating function will not be differentiable
at the points xi , i = 1, . . . , n − 1. As an alternative, we could fit a quadratic
function through (xi , yi ) and (xi+1 , yi+1 ). This problem does not have a
unique solution, but we can take advantage of this non-uniqueness to select
the quadratic curves whose derivatives match at each of the data points.
1.0

0.8

0.6

0.4

0.2

-0.2 0.2 0.4 0.6 0.8 1.0 1.2

Figure 2.9: Local linear interpolation

Let us suppose that we attempt to fit the quadratic function

Qi (x) = ai + bi x + ci x2 , i = 1, . . . , n, (2.5)
34 CHAPTER 2. INTERPOLATION

in the interval (xi−1 , xi ); then as there are n intervals we shall have 3n coeffi-
cients to find. First of all, we require that the resulting functions interpolate
the points (xi , yi ), i = 0, . . . , n, i.e.

Qi (xi−1 ) = yi−1 and Qi (xi ) = yi , i = 1, . . . , n, (2.6)

which gives 2n equations. (This condition will ensure that the resulting
functions are continuous at the internal points x1 , . . . , xn−1 .) Furthermore,
at each internal point, xi , we require that the derivatives of Qi and Qi+1
agree, i.e.
Q′i (xi ) = Q′i+1 (xi ), i = 1, . . . , n − 1. (2.7)
Therefore, we have 3n − 1 equations to determine the 3n coefficients ai ,
bi , ci , i = 1, . . . , n. It would appear that we can select any one coefficient
arbitrarily. Combining together the resulting quadratic functions gives a
quadratic spline.
1.5

1.0

0.5

-0.5 0.5 1.0 1.5

Figure 2.10: A quadratic spline through the data in table 2.1.1

The amount of effort in determining a quadratic spline can be reduced

by making use of the following procedure. Let us assume that in the interval
(xi−1 , xi ) the required spline component is Qi (x), as shown in figure 2.5.3.
We now assume that at xi−1 the slopes of both Qi−1 (x) and Qi (x) are di−1
and similarly at xi , Q′i (x) = Q′i+1 (x) = di . Since the derivative of a quadratic
function is linear, Q′i (x) is the linear cross mean of di and di−1 , i.e. we have
that
(x − xi−1 )di − (x − xi )di−1
Q′i (x) = ,
xi − xi−1
2.5. SPLINE INTERPOLATION 35

Q HxL d
d i i-1
i-1

x x
i-1 i

Figure 2.11: A quadratic spline component Qi (x) in (xi−1 , xi )

therefore, Q′i (xi ) = Q′i+1 (xi ) = di , which will ensure that the conditions (2.7)
are satisfied. Next we integrate this expression to get
(x − xi−1 )2 di − (x − xi )2 di−1
Qi (x) = + ci ,
2(xi − xi−1 )
where ci is a constant produced by the integration. Finally, we select the
constants ci to ensure that the combined spline is continuous across each
mesh node x = xi , i = 1, . . . , n − 1. (This ensures that equations (2.6) are
satisfied.) Notice that

Qi (xi ) = 12 (xi − xi−1 )di + ci = yi

and
Qi (xi−1 ) = 21 (xi−1 − xi )di−1 + ci = yi−1 .
We then eliminate ci between these equations to give
2(yi−1 − yi)
di = − di−1 . (2.8)
(xi−1 − xi )
This is a first-order recurrence relation and so all the required values, d1 , d2 , . . .
can be found once d0 is specified. However, in general we will not be able to
specify both d0 and dn . (See Harding and Quinney (1986).)
Example 2.5.1 (Computational)
Load INTERP and when the prompt shows Select method press RE-
TURN. Now change the option to Method: Quadratic spline and then
36 CHAPTER 2. INTERPOLATION

press RETURN again. Use the graph option to set up suitable plotting
ranges and then you will be requested to supply a value for the slope of the
spline at its left end, i.e. the value d0 . For the moment accept the default
value d0 = 1.4845 and then observe the corresponding quadratic spline being
plotted. (See figure 2.12.) Alternatively, we could take d0 = 0 which gives the
quadratic spline shown in figure 2.10. The value d0 = (y1 − y0 )/(x1 − x0 ) is
obtained by using a finite difference approximation for the slope at x0 which
explains why a better result is produced. (This is the default value which
the program adopts.)
Exercise 2.5.1 (Computational) Use the program INTERP to determine a
quadratic spline through the data points in table 2.1.1. Investigate the effect
of changing the value of d0 . (The quadratic spline for the default choice is
given by figure 2.12.)

1.5

1.0

0.5

-0.5 0.5 1.0 1.5

Figure 2.12: Quadratic spline for table 2.1.1 with d0 = 1.4845

From figure 2.12 we see that a quadratic spline and its derivative are
continuous at each node. However, the curvature is not smooth since
(d1 − d0 )
lim Q′′0 (x) =
x→x1 (x1 − x0 )
but
(d2 − d1 )
lim Q′′1 (x) =
x→x1 (x2 − x1 )
and these may not be the same. A further problem with quadratic splines
is that they tend to produce oscillatory functions as is also demonstrated
2.5. SPLINE INTERPOLATION 37

in figure 2.10. In order to ensure that both the first and second derivatives
are continuous we need to use a higher order local interpolant in each sub-
interval. One of the most popular is to use cubic functions in each sub-interval
and then attempt to select the coefficients so that both the slope and the
curvature are continuous across internal tabular points.

S HxL yi-1
i

yi S''i HxL M i-1

Mi

x x
i-1 i

Figure 2.13: Cubic polynomial through (xi , yi ) and (xi+1 , yi+1)

Let us suppose that we are given a table of data in the form of data points
(xi , yi), i = 0, . . . , n. In the interval (xi , xi+1 ) we use the cubic polynomial

Si (x) = ai + bi x + ci x2 + di x3 . (2.9)

In each sub-interval there are four coefficients to be determined, and since

there are n such intervals this will give a total of 4n unknown values. Clearly,

Si (xi ) = yi and Si (xi+1 ) = yi+1 , i = 0, . . . , n − 1,

which gives 2n conditions. Similarly, at each internal point we require

′
Si−1 (xi ) = Si′ (xi ), i = 1, . . . , n − 1,

and
′′
Si−1 (xi ) = Si′′ (xi ), i = 1, . . . , n − 1,
which gives a further 2n − 2 conditions (see figure 2.13). It would appear
that we can specify any two values arbitrarily. Once all 4n parameters are
determined we shall have a unique cubic in each subinterval. This collection
38 CHAPTER 2. INTERPOLATION

of functions is called a cubic spline. Even for a moderate number of tabu-

lation points this procedure will result in a large system of linear equations
but, fortunately, the total amount of effort can be considerably reduced in a
manner analogous to that used for quadratic splines.
Notice that in the interval (xi , xi+1 ) we have that

Si′′ (x) = 2ci + 6di x. (2.10)

Let us assume that the second derivative of Si at xi is given by some value

Mi , then

Si′′ (xi ) = 2ci + 6dixi = Mi ,

Si′′ (xi+1 ) = 2ci + 6dixi+1 = Mi+1 .

Solving these equations for ci and di gives

Mi+1 − Mi Mi xi+1 − Mi+1 xi
6di = , 2ci = .
xi+1 − xi xi+1 − xi
Therefore,
(x − xi )Mi+1 − (x − xi+1 )Mi
Si′′ (x) = .
xi+1 − xi
(This is just a linear cross mean.) Integrating this expression twice gives

Mi+1 (x − xi )3 − Mi (x − xi+1 )3
Si (x) = + αi x + βi , (2.11)
6hi
where hi = xi+1 − xi . Now we fit this expression to the data points (xi , yi )
and (xi+1 , yi+1 ), i.e.

Mi (xi − xi+1 )2 + 6αi xi + 6βi = 6yi,

Mi+1 (xi+1 − xi )2 + 6αi xi+1 + 6βi = 6yi+1 ,

from which we eliminate βi to produce

(yi+1 − yi ) 1
αi = − (xi+1 − xi )(Mi+1 − Mi ) (2.12)
(xi+1 − xi ) 6

and
βi = yi − αi xi − 61 Mi (xi+1 − xi )2 . (2.13)
2.5. SPLINE INTERPOLATION 39

However, Si′ (x) is required to be continuous across xi , therefore

′
Si−1 (xi ) = Si′ (xi ),

i.e.
Mi (xi − xi−1 ) + 2αi−1 = Mi (xi − xi+1 ) + 2αi
or
αi − αi−1 = 12 Mi (xi+1 − xi−1 ). (2.14)
Finally, we use equation (2.12) to eliminate αi and αi−1 from (2.14) to give
6(yi+1 − yi ) 6(yi − yi−1 )
hi−1 Mi−1 + 2(hi−1 + hi )Mi + hi Mi+1 = − . (2.15)
hi hi−1
This second-order recurrence relation holds at all internal points, there-
fore, we have a system of n − 1 linear equations in the n + 1 unknowns
M0 , M1 , . . . , Mn . In particular this system of equations has a special form
since in a matrix notation we obtain entries only on the main diagonal and
the diagonals directly above and below this diagonal. These equations are
therefore called tridiagonal and special methods have been developed to solve
them which take advantage of their structure. (For an introduction to second-
order recurrence relations and the solution of tridiagonal systems of linear
equations see Harding and Quinney (1986).) The general solution of this re-
currence relation will involve two arbitrary values which can be determined
if two additional conditions are supplied. There are two particularly popular
ways they can be specified: (i) free splines and (ii) clamped splines.

(i) Free splines (natural splines). The simplest way of imposing additional
conditions is to set both M0 and Mn equal to zero. (This is equivalent to
assuming that the interpolating function is approximately linear at x0 and
xn since Mi is the second derivative at xi .)

Example 2.5.2
Use the program INTERP to determine the natural spline through the data
points given in table 2.1. As there are five points there will be five second
derivative values to be determined. Let these values be M0 , M1 , M2 , M3 and
M4 , but because the spline is to be natural we have M0 = M4 = 0. Next we
write down the equation (2.15) at the internal points, i.e.
6 6
0.2M0 +0.8M1 +0.2M2 = (0.58813−0.39646)− (0.39646−0.09956)
0.2 0.2
40 CHAPTER 2. INTERPOLATION

6 6
0.2M1 +0.8M2 +0.2M3 = (0.77210−0.58813)− (0.58813−0.39646)
0.2 0.2
6 6
0.2M2 +0.8M3 +0.2M4 = (0.89608−0.77210)− (0.77210−0.58813)
0.2 0.2
or     
0.8 0.2 0 M1 −3.156899

 0.2 0.8 0.2 
 M 
2  =

 −0.231000 .


0 0.2 0.8 M3 −1.799699

These equations have the solution M1 = −4.306179, M2 = 1.440214 and
M3 = −2.609679. These values provide all that is necessary to construct the
required natural spline. The resulting function is shown in figure 2.14. If
we wish to evaluate the spline at a particular point it is first necessary to
decide in which sub-interval the required point lies. For example, to evaluate
this spline at x = 0.51 we need to evaluate the component of the spline
which is defined on the interval (0.5, 0.7) or (x2 , x3 ). The required function
is therefore

M3 (x − x2 )3 − M2 (x − x3 )3
S2 (x) = + α2 x + β2 ,
6(x3 − x2 )

where

(0.7721 − 0.58813) 1
α2 = − × 0.2(−2.660967 − 1.440214)
0.2 6
and
1
β2 = 0.58813 − 0.5α2 − × 1.440214(0.2)2
6
from (2.12) and (2.13). Evaluating these expressions gives the component
of the spline in the interval (0.5, 0.7) and hence S2 (0.51) = 0.59731 to five
decimal places.
Exercise 2.5.2 (Analytical) Determine an explicit expression for the natural
cubic spline through the data points of table 2.1. By direct evaluation show
that the resulting spline is continuous and has continuous first and second
derivatives at the internal points x1 ,...,xn−1 .
Exercise 2.5.3 (Computational) Use the program INTERP with the option
Method: Cubic spline - free to confirm the results of example 2.5.2.
2.5. SPLINE INTERPOLATION 41

1.5

1.0

0.5

-0.5 0.5 1.0 1.5

Figure 2.14: The free cubic spline for the points in table 2.1.1

(ii) Clamped splines. If it is possible to supply values for the slope of the
interpolating function at x0 and xn then we can use equation (2.11) to provide
the additional equations required. Let us suppose that we are given that
S0′ (x0 ) = d0 and Sn′ (xn ) = dn then differentiating (2.11) and substituting we
have
(y1 − y0 ) h0
d0 = − (M1 + 2M0 )
h0 6
or
6(y1 − y0 )
2h0 M0 + h0 M1 = − 6d0. (2.16)
h0
and
(yn − yn−1) hn−1
dn = + (Mn−1 + 2Mn )
hn−1 6
or
(yn − yn−1 )
hn−1 Mn−1 + 2hn−1 Mn = 6dn − 6 . (2.17)
hn−1
Next we combine equations (2.16) and (2.17) with the linear equations given
by (2.15) to obtain n + 1 linear equations in the unknown values Mi , i =
0, ..., n. Once these values are known the required cubic spline is uniquely de-
termined. The matrix system which is formed in this way is both tridiagonal
and diagonally dominant and can be solved in a relatively straightforward
way by Gaussian elimination without partial pivoting. (See Harding and
Quinney (1986).)
Example 2.5.3 (Computational)
42 CHAPTER 2. INTERPOLATION

Use the program INTERP to determine the clamped cubic spline which
passes through the data points in table 2.1 and has slope 1.4845 at x = 0.1
and 0.6199 at x = 0.9. (These values are default settings.) This clamped
cubic spline is shown in figure 2.15 and can be compared with the free spline
in figure 2.14.
1.5

1.0

0.5

-0.5 0.5 1.0 1.5

Figure 2.15: The clamped cubic spline in example 2.5.3

Exercise 2.5.4 (Computational) Use the program INTERP to determine the

clamped cubic spline which passes through the points given in table 2.1 and
has zero slope at both ends. Compare the result with the free spline and the
interpolating polynomial through the same points.
We can now compare the quadratic spline which passes through the points
in table 2.1, on page 9, which is shown in figure 2.12, the clamped cubic spline
shown in figure 2.15 and the corresponding free cubic spline which is shown
in figure 2.14. Clearly, the cubic spline is less prone to oscillations and we
can show that this is generally the case.
Theorem 2.1. Let C2 [x0 , xn ] be the set of all twice differentiable functions
which pass through the tabulation points, (xi , yi ), i = 0, . . . , n. The natural
cubic spline, S(x), which is a member of C2 [x0 , xn ], minimizes the average
curvature: s
1
Z xn 
(f ′′ (x))2 dx .
xn − x0 x0
Proof. Let S(x) be the required natural spline given by the solution of (2.15)
subject to M0 = Mn = 0 and let f (x) be any other twice differentiable
2.5. SPLINE INTERPOLATION 43

function over [x0 , xn ]. Now consider

Z xn Z xn
(f ′′ (x))2 dx = [S ′′ (x) + (f ′′ (x) − S ′′ (x))]2 dx
x0 x
Z x0 n Z xn
= S ′′ (x)2 dx + 2 S ′′ (x)[f ′′ (x) − S ′′ (x)]2 dx
x0 x0
Z xn
+ [f ′′ (x) − S ′′ (x)]2 dx
x0

and then look at the second of these terms, i.e.

Z xn
I= S ′′ [f ′′ − S ′′ ] dx.
x0

Integrating by parts gives

I = S ′′ (xn )[f ′ (xn ) − S ′ (xn )] − S ′′ (x0 )[f ′ (x0 ) − S ′ (x0 )]

Z xn
− [f ′ (x) − S ′ (x)]S ′′′ (x) dx.
x0

However, for a natural spline S ′′ (x0 ) = M0 = 0 and S ′′ (xn ) = Mn = 0

and therefore the first two terms vanish. Furthermore, in each sub-interval
(xi , xi+1 ) the cubic spline has constant third derivative which we shall call
Ki and so
i=n−1
X Z xi+1 i=n−1
[f ′ (x) − S ′ (x)]Ki dx = −
X
I =− Ki [f (xi ) − S(xi )] = 0.
i=0 xi i=0

From this we conclude that

Z xn Z xn
[f ′′ (x)]2 ≥ [S ′′ (x)]2 dx
x0 x0

as required.
In order to demonstrate the importance of this result consider the follow-
ing example.
Exercise 2.5.4 (Computational) Determine a polynomial approximation for
the data in table 2.10. (These data points are stored in the file SPLINE).
Now construct the natural cubic spline which passes through these points
and compare the results obtained.
44 CHAPTER 2. INTERPOLATION

Table 2.10: Data for exercise 2.5.4

x −1 −0.960 −0.860 −0.690 0.220 0.500 0.930
y −1 −0.151 0.594 0.986 0.895 0.500 −0.306

Exercise 2.5.5 (Computational) Use the program INTERP to determine a

quadratic spline through the data in the data file SPLINE which has zero
slope at its left-hand end, i.e. d0 = 0. Vary d0 and investigate the behaviour
of the resulting spline. Now fit a natural cubic spline through these data
points and compare the results. Which curve is the smoothest? Investigate
the behaviour of clamped splines for a variety of endpoint conditions. Is
it possible to improve upon the natural cubic spline? To see that using
a higher order interpolating function does not improve the behaviour, try
fitting a high order polynomial using the option Construct interpolating
polyn. (Hint: the quadratic spline is determined by a first order recurrence
relation for which an initial condition is supplied. Is it stable? Why is the
two-point boundary value problem formulation involved with cubic splines a
more stable process?)

2.6 Summary and conclusion

We have seen in this chapter how it is possible to produce a continuous func-
tion which passes through a given set of tabular values. The solution to this
problem is not unique. We can aim to fit a function which is composed of
a linear combination of simple powers of x, i.e. a polynomial, but this may
lead to an interpolating polynomial which is highly oscilllatory. Alterna-
tively, we can aim for a more local approximation and fit a spline which will
give a resulting function which is less likely to suffer from dramatic oscilla-
tions between tabular values. However, whichever method is used we cannot
say anything definite about non-tabular values. To illustrate this point see
figure 2.16. The interpolating function passes through all the tabular values
but is likely to be of little use in interpolation.
An underlying assumption with any interpolation problem is that the data
points have been determined from some unknown but continuous function.
If the underlying function is not continuous then no amount of continuous
functions can produce satisfactory interpolation near any singularity. Chap-
2.6. SUMMARY AND CONCLUSION 45

1.0

0.8

0.6

0.4

0.2

-3 -2 -1 1 2 3

Figure 2.16: An example of poor interpolation

ter 3 will consider this further, but for the moment we conclude this chapter
with the following general remarks.

1. Interpolation is usually more reliable than extrapolation; the latter

should be avoided if at all possible.

2. Except in all but the very simplest cases it is unwise to determine an

interpolating polynomial by considering expressions of the form a0 +
a1 x + a2 x2 + . . . + an xn . This problem is possibly ill-conditioned and
is likely to become more so as the number of tabular points increases.
However, the construction of a polynomial interpolant can be carried
out in a stable way using either Lagrangian polynomials or divided
differences. If the value of the interpolating polynomial is only required
at one or two points then it is better to use Neville’s algorithm.

3. Interpolation is a process whereby given a table of values (xi , y(xi)) and

a non-tabular value x∗ we attempt to find y(x∗ ). Inverse interpolation
is where we are given a value y and are asked to determine a value x
which would produce y.

4. The curve which has minimum curvature through a set of points and
yet has continuous first and second derivatives is a natural cubic spline.
Such a function is made up from a set of cubic functions, one in each
tabular interval, which are continuous and have continuous first and
second derivatives at each internal tabular point.
46 CHAPTER 2. INTERPOLATION

We have not considered the idea of rational interpolation, however, we

shall consider the rational approximation functions in Chapter 3. Given a
set of points and a rational function which passes through them it is clear
that rational interpolation can be carried out at non-tabular points.
Chapter 3

The Approximation of
Functions

‘I’ve got fourteen pots of honey left, or is it fifteen, as the case may be’,
said Pooh humbly.
‘Well, let’s call it sixteen’ said Rabbit.
With apologies to A A Milne.
In an age of the ever increasing use of pocket calculators and micro-computers,
the need to evaluate standard mathematical functions is also increasing.
Clearly, electronic devices cannot store all possible values of such functions;
the question therefore is how do such machines evaluate expressions such as
sin(1.4) or exp(−3.456)? Inherent in such a problem is not only the need
to carry out such an evaluation but also the need to know how accurate the
presented result is going to be. In this chapter we will examine this question
and try to provide some sensible answers.

3.1 Taylor series approximation

It is well known that provided a function is sufficiently smooth near a par-
ticular point x = a then it is possible to write
(x − a)2 ′′
f (x) = f (a) + (x − a)f ′ (a) + f (a) + ...
2!
where the infinite series converges for all x which satisfy |x − a| ≤ R and R
is called the radius of covergence (see Appendix A). This is called a Taylor’s

47
48 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

series expansion of the function f about the point x = a. An alternative

form is
(x − a)2 ′′
f (x) = f (a) + (x − a)f ′ (a) + f (a) + . . .
2!
(x − a)n (n)
... + f (a) + Rn , (3.1)
n!
where the remainder term Rn is given by

(x − a)(n+1) (n+1)
Rn = f (θ), (3.2)
(n + 1)!

and θ lies between x and a. Therefore, if we knew the value of f (a), f ′ (a),
f ′′ (a), . . . , f (n) (a), we could use equation (3.1) to estimate f (x). Further-
more, the maximum possible error of such an approximation is given by
max|Rn | for all values of θ between x and a.
Example 3.1.1
We can use a Taylor series with a = 0, which is sometimes called a Maclaurin
series, to estimate exp(0.1). We can then consider questions such as the
following. If five terms of this series are used what is the maximum possible
error? How many terms are required to evaluate exp(0.5) correct to six
decimal places? The Taylor series expansion of exp(x) about x = 0 is given
by
x2 x3 x4 xn
exp(x) = 1 + x + + + + ...+ + ....
2! 3! 4! n!
Taking the first five terms gives

0.12 0.13 0.14

exp(0.1) ≈ 1 + 0.1 + + + ,
2 6 24
therefore, exp(0.1) ≈ 1.105170831 which can be compared with the correct
value of exp(0.1) = 1.10517092 . . . . The remainder term R4 associated with
the above Taylor series approximation may be used to estimate the error
involved with this approximate value. From equation (3.2) with x = 0 and
a = 0.1 we have that
 
 0.15 exp(θ)  exp(0.1)
|R4 | ≤ max  ≤ 0.15 ≤ 9.2 × 10−8 ,
 
120 120

0≤x≤0.1  
3.1. TAYLOR SERIES APPROXIMATION 49

compared with the actual error of 9.0 × 10−8 . Suppose we now wanted to
find exp(0.5) using the same quartic polynomial expression then

0.52 0.53 0.54

exp(0.5) ≈ 1 + 0.5 + + + = 1.6484374
2 6 24
but this compares less favourably with exp(0.5) = 1.6487213 . . .. For this
value of x, max |R4 | = 4.29 × 10−4 . Clearly, the range of values for which we
can use this polynomial approximation is limited, but one way around this
problem would appear to be to include further terms. Alternatively, we can
use the remainder term, Rn , to see just how many terms are required. For
example, in order to ensure that the approximation has a maximum error,
for all |x| ≤ 12 , which is less than 1 part in 106 we need to select an n such
that
|Rn | ≤ 0.5 × 10−7 , for all |x| ≤ 21 .
Therefore, from (3.2) we need to find an integer n such that
 
 0.5n+1 exp(x) 
max1  ≤ 0.5 × 10−7
 
(n + 1)! 

|x|≤
2

which can be achieved if n ≥ 9. This result implies that if we require an

accuracy of six decimal places each time we evaluate the exponential function
for |x| < 21 then a polynomial of degree at least 9 is required. Clearly, for
larger values of x more and more terms would be needed.
Example 3.1.2 (Computational)
We shall demonstrate a Taylor series approximation for the function exp(x),
|x| ≤ 1, using the program APPROX. (The program has a default func-
tion setting for this problem.) Load the program APPROX and when the
prompt Function is shown accept the default value as set which will produce
a new prompt Method. Cycle through the options available by repeatedly
pressing the SPACE BAR and when the option shows Taylor Series press
RETURN. Again cycle through the options available, to familiarize your-
self with the program, and when it shows Option: calculate coeffs press
RETURN again. Now set the value of degree to 3, which will ask for a
cubic approximation, and finally put a = 0. After a short pause the prompt
will change to Plot: function; press RETURN to accept this. The prompt
should now be Graph data: superimpose; however, clear the screen by
changing this option to Graph data: clear – same axes and then press
50 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

RETURN. The function will now be plotted in the graphics window and
then the program will return to the option Plot: function. Now change
the Plot option to approximation by pressing the SPACE BAR and press
RETURN to show the required cubic Taylor series approximation for exp(x)
on this interval. It is also possible to plot the error between the function and
approximation by changing the plotting option, Plot, to show Plot: error.
The result is shown in figure 3.1. The associated error function is plotted on
the same x axis but the corresponding y axis is scaled by a suitable factor
which is displayed on the screen. In order to find this approximation it is
necessary to determine suitable values for the coefficients in equation (3.1)
and with the option calculate coeffs this is done by using finite difference
approximations. (See Appendix B.) Exercise 3.1.1 demonstrates how it is
possible for the user to supply these coefficients manually.

2.5

2.0

1.5

1.0

-1.0 -0.5 0.5 1.0

Figure 3.1: Taylor series expansion for exp(x) about x = 0

Exercise 3.1.1 (Computational) Load the program APPROX and the func-
tion exp(x), |x| ≤ 1. Select the Method: Taylor Series and the Option:
calculate coeffs and then plot the function, its cubic Taylor series approxi-
mation and the associated error as described in example 3.1.2. Exit from this
part of the program by taking the option Plot: quit and follow the instruc-
tions given. This will return the program to the prompt Option: calculate
coeffs. Again cycle through the options available at this point which in-
clude the possibility of storing the current coefficients on a suitable disc file,
reading another set of coefficients which you have previously stored, or set
coefficients manually. The degree of the current approximation should be
3 and the point about which the expansion is determined should be a = 0;
3.1. TAYLOR SERIES APPROXIMATION 51

if not reset these values. The correct values for the first four coefficients in
the Taylor series of exp(x) about x = 0 are 1, 1, 21 and 61 . Accept the option
to set coefficients manually and then input these values; notice that the
current values, which are determined by finite differences, are shown on the
screen and are very close to the correct values. Compare the approximations
which are determined.
Exercise 3.1.2 (Computational) Use the program APPROX to determine a
cubic Taylor series expansion about x = 0 for exp(x), |x| ≤ 2. Compare the
results with a corresponding expansion about the points x = 1 and x = −1.
Exercise 3.1.3 (Analytical) How many terms are required to evaluate the
function log(1 + x), |x| < 21 , correct to eight decimal places?
As the evaluation of polynomials plays a fundamental role in this type of
approximation it is appropriate to consider the most efficient way of carrying
out this task. We shall illustrate a technique called nested multiplication.
This method permits the evaluation of a polynomial expression without the
storage of intermediate results and therefore it can be carried out on a hand
calculator which does not have a memory.
Exercise 3.1.4 (Analytical) Polynomial evaluation by nested multiplication.
Let pn (x) = an xn + an−1 xn−1 + . . . + a1 x + a0 . Determine the number of
multiplications and additions required to evaluate this polynomial directly
for a fixed value of x. Show that if we define a sequence uj by

un = an , uj−1 = xuj + aj−1 , j = n, n − 1, . . . , 1, (3.3)

then pn (x) = u0 . This is called nested multiplication. Determine the num-

ber of operations required to evaluate the polynomial pn using this method.
Which method is the most efficient? Notice that by using nested multiplica-
tion pn (x) can be calculated without any intermediate memory.
For a fixed value of x the technique of nested multiplication introduced in
exercise 3.1.4 is equivalent to a first-order recurrence relation. However, the
recurrence relation (3.3) is stable only if |x| ≤ 1 (see Harding and Quinney
(1986)). To demonstrate the importance of this comment let us suppose that
we wish to approximate sin(x), for a value of x in the interval [10, 11], by
using a Taylor series expansion of degree 3 about the point x = 10. This
gives

p3 (x) = sin(10)+(x−10) cos(10)− 21 (x−10)2 sin(10)− 16 (x−10)3 cos(10). (3.4)

52 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

Expanding out this expression gives the cubic

p3 (x) = a0 + a1 x + a2 x2 + a3 x3 for x ∈ [10, 11], (3.5)

where a0 = −104.798, a1 = 35.6743, a2 = −3.92335 and a3 = 0.139845,

correct to six significant figures. Now suppose we wish to find p3 (10.5). As
a3 is accurate to six decimal places, when it is be multiplied by a factor
(10.5)3 the resulting value of p3 (10.5) may only have two or possibly three
decimal places correct. Clearly, the evaluation of the polynomial (3.5) may
be prone to induced errors and should be avoided. Instead we leave it in the
form (3.4), i.e.

p3 (x) = −0.544021 − 0.839072(x − 10) + 0.272011(x − 10)2

+ 0.139845(x − 10)3

which ensures that all the terms raised to powers are less than 1; hence we
should have a stable recurrence relation and reasonable conditioning. This
introduces a very important point, for clearly this problem will always arise
if we attempt to approximate any function outside the interval either [0, 1]
or [−1, 1]. However, we can always scale any interval [a, b] to either of these
standard intervals by a suitable linear scaling procedure.
Exercise 3.1.5 (Calculator) By examining the possible errors in the coeffi-
cients of the polynomial (3.5) determine upper and lower error functions for
this expression.
Exercise 3.1.6 (Analytical) Scale the interval [2, 10] to the standard intervals
[0, 1] and [−1, 1].

3.2 Polynomial approximation

We have seen how it is possible to obtain a polynomial approximation for a
given function by truncating a corresponding Taylor series expansion. Now
we will consider an alternative approach using interpolating polynomials. Let
us suppose that we are given a function f (x), x in (a, b), and attempt to find
a polynomial function of degree n, pn (x), which is close to f in some sense.
Example 3.2.1 (Computational/Calculator)
Let us determine a polynomial approximation of degree 3 for the function
f (x) = exp(x), for values of x in [−1, 1]. We begin by considering the Taylor
3.2. POLYNOMIAL APPROXIMATION 53

series expansion of exp(x) in this interval. This is given by

exp(x) = 1 + x + 12 x2 + 61 x3 + R3 (x),

where the remainder term satisfies |R3 (x)| ≤ max|x|≤1 |f iv |/4! ≤ 0.113. This
approximation is shown in figure 3.2 and the error function, R3 (x), in fig-
ure 3.3. As an alternative method we can make the function and approximat-
ing polynomial agree at a given set of points, i.e. we select the coefficients a
nth degree polynomial so that pn (xi ) = f (xi ), i = 0, ..., n+ 1. Thus the prob-
lem becomes one of constructing an interpolating polynomial through n + 1
points. Firstly, we need to decide where we are going to put the tabulation
points xi , i = 0, 1, 2, 3, at which the polynomial and function agree. For the
moment, let us put them at equally spaced points in [−1, 1], i.e. −1, − 13 , 13
and 1, and assume that the required cubic polynomial is

p3 (x) = a0 + a1 x + a2 x2 + a3 x3 .

Then at x = −1 we require that

p3 (−1) = a0 − a1 + a2 − a3 = exp(−1).

We obtain a similar equation at the remaining three points which gives a

system of four linear equations in the unknown coefficients a0 , a1 , a2 and a3
(see section 2.1). These equations can be solved using Gaussian elimination
to produce the cubic approximation

p3 (x) = 0.995196 + 0.999049x + 0.547885x2 + 0.176152x3 . (3.6)

This function is also plotted in figure 3.2 and its error in figure 3.3. Notice
that the maximum error associated with the polynomial (3.6) is much less
than the maximum error associated with the corresponding Taylor expansion
of the same order.
We have already seen how it is possible to estimate the error associated
with a Taylor series approximation of a given function. Now we will con-
struct an error function for polynomial interpolation. Firstly, we recall that
f (x) and the interpolant pn (x) agree at the n + 1 points xi , i = 0, 1, . . . , n,
therefore,

en (x) = f (x) − pn (x) = K(x)(x − x0 )(x − x1 ) . . . (x − xn ),

54 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

2.5

2.0

x
1.5 e

1.0

0.5

-1.0 -0.5 0.5 1.0

Figure 3.2: Taylor series approximation and p3 (x) for exp(x)

0.025

0.020 Error in cubic Taylor series

0.015

0.010

0.005

-1.0 -0.5 0.5 1.0

-0.005 Error in interpolating cubic

-0.010

Figure 3.3: Error functions associated with figure 3.2.1

3.2. POLYNOMIAL APPROXIMATION 55

where K(x) is to be determined. For any given value of x let us define the
function φ by
φ(t) = f (t) − pn (t) − K(x)w(t),
where w(t) = (t−x0 )(t−x1 ) . . . (t−xn ), and then notice that φ vanishes at the
points n + 2 points x, x0 , . . . ,xn . By Rolle’s theorem φ′ must vanish between
any two points at which φ vanishes and so φ′ has n + 1 zeros. Similarly, φ′′
has n zeros and so on until we find that the (n + 1)th derivative of φ, φ(n+1) ,
has exactly one zero; let this zero be θ, then
0 = φ(n+1) (θ) = f (n+1) (θ) − pn(n+1) (θ) − Kw (n+1) (θ).
However, p(n+1)
n is identically zero and w (n+1) = (n + 1)!, therefore,
f (n+1) (θ)
K= ,
(n + 1)!
or
f (n+1) (θ)
f (x) − pn (x) = (x − x0 )(x − x1 ) . . . (x − xn ) . (3.7)
(n + 1)!
(Recall that θ will depend upon the choice of x.) Notice the similarity be-
tween the form of this error function and that associated with the Taylor
series expansion given by equation (3.2).
Example 3.2.2 (Calculator/Computer)
We shall determine an upper bound on the error associated with the cubic
approximation for exp(x) which was derived in example 3.2.1, i.e.
p3 (x) = 0.995196 + 0.999049x + 0.547885x2 + 0.176152x3 .
From equation (3.7) with n = 3 we have that
|f (x) − p3 (x)| ≤ |(x + 1)(x + 13 )(x − 31 )(x − 1)|M/4!,
where M = sup |f iv |, x ∈ [−1, 1], and f (x) = exp(x). (We define the supre-
mum as follows: if M ≥ |f iv | and ∀ L ≥ |f iv |, M ≤ L then M is the
supremum.) We now examine the first four terms and note that for values √of
1
x between −1 and 1 this takes a maximum value of 0.19753 at x = ± 3 5.
(See figure 3.4.) Therefore, since |f iv (x)| = |ex | ≤ e, for x ∈ [−1, 1], the error
function e3 (x) satisfies |e3 (x)| ≤ 0.0224, which is the required error bound. In
fact this is a severe overestimate since the maximum error is actually 0.00481
but even a crude bound is better than none. (Recall that the corresponding
bound on the cubic Taylor series approximation was |R3 (x)| ≤ 0.113.)
56 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

0.4

0.3

0.2

0.1

-1.0 -0.5 0.5 1.0

-0.1

Figure 3.4: y = |(x + 1)(x + 31 )(x − 31 )(x − 1)|, |x| ≤ 1

Exercise 3.2.1 (Analytical) By determining upper and lower bounds on the

fourth derivative of exp(x) construct upper and lower bounds for the error
function e3 (x) given in example 3.2.1.
Exercise 3.2.2 (Computational) Use the program APPROX to determine
a polynomial approximation of order 2 for exp(x), |x| ≤ 1, by interpolation
at the points x = −1, − 31 and 1. Load the program APPROX, accept the
option Function as set and then select Method: Interpolation. Now
select Option: edit/inspect points and enter the three points on the curve
at x = −1, − 31 and 1. (When you have selected this option you will be given
the opportunity to change the plotting ranges if you wish to do so.) Select
the editing option, cycle through the edit options until it shows Add Point
and then press RETURN. The top window will now contain information
regarding how to input a point and, in the graphics window, there should be
a flashing cursor at the point (−1, 0). For example, if you press the COPY
key the cursor will move to the point (−1, exp(−1)). If you press RETURN
this value will be accepted as an interpolation point and the number of points,
N, which is shown in the bottom window will be increased by 1. Finally, the
edit option will be returned to Add Point. Now input two additional points
at x = − 13 , y = exp(− 31 ) and x = 1, y = exp(1). The x values can either
be input by pressing ‘X’ or by using the cursor which is controlled in the
usual way. (Each key press will move the cursor by an amount ‘dX’ and this
variable is in turn controlled by the ‘>’ and ‘<’ keys.) When you are satisfied
with the points which have been input quit the editor and select Option:
calculate coeffs then press RETURN. The computer will now determine
3.2. POLYNOMIAL APPROXIMATION 57

the polynomial through these points, display the resulting polynomial in the
bottom window and then display the prompt Plot: function. By changing
the prompt you can plot the approximation or the error between the function
and the computed polynomial. Use the editor to change the interpolation
points and investigate how the approximation and its error change. Now
return to the editor and add another two points and then determine quartic
polynomial approximations for exp(x), x ∈ [−1, 1].

Exercise 3.2.4 (Computational) Use the program APPROX with the option
Method: Interpolation to determine the cubic polynomial approximations
for the function exp(x), x in [−1, 1], by interpolation at −1, −0.5, 0.5 and
1. Investigate how the error between the interpolating polynomial and the
function changes as the tabular points change. Investigate how the error
changes as the number of tabular points increases. Find the polynomial
of smallest degree which approximates this function to an accuracy of at
worst 0.001 throughout this interval for interpolation points which are equally
spaced. Investigate how the error function behaves for different interpolation
points.

We have assumed that it is always possible to construct a suitable poly-

nomial approximation, but is this always the case? Fortunately, provided the
function f satisfies certain minimal requirements then there is always such a
polynomial. We state, without proof, the following theorem.

Theorem 3.2.1. Weierstrass’s approximation theorem. If f is a continuous

function on the finite interval [a, b], then for any ǫ > 0 there exists an nth
degree polynomial pn (x), for some n, such that

|f (x) − pn (x)| < ǫ for all x ∈ [a, b].

In other words, if we are given a continuous function f , defined on the interval

[a, b], then there is always a polynomial function which will approximate f
to within any prescribed error ǫ. Unfortunately, what Weierstrass’s theorem
does not tell us is how large n should be, nor how to find the associated
polynomial. Furthermore, this theorem assumes that we can carry out all
arithmetic exactly, which is not the case either.
58 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

3.3 Minimax polynomials

In the last section we saw that there are many polynomial approximations
of a given degree for any given function. For example,

exp(x) ≈ 1 + x + 21 x2 + 61 x3 , (3.8)

and
exp(x) ≈ 0.99520 + 0.99905x + 0.54788x2 + 0.17615x3 , (3.9)
are both cubic approximations. The first is a truncated Taylor series, the
second is an interpolating polynomial which agrees with exp(x) at the points
−1, − 31 , 13 and 1. We can compare these functions as shown in figure 3.2
but it is more informative to look at their respective error functions shown
in figure 3.3. If we were to obtain another interpolating cubic approximation
by adjusting the tabular points we could plot the associated error function
in the same way and compare it. In general we need to find a polynomial,
pn (x), which minimizes the maximum value of the error function, i.e. for any
other polynomial of degree n, qn (x) say, the minimax error ǫn satisfies

ǫn = max |f (x) − pn (x)| ≤ max |f (x) − qn (x)| for all x ∈ [a, b].

The resulting polynomial is called the minimax polynomial of degree n. By

adjusting the tabular points at which a polynomial qn (x) agrees with f (x) we
can look for the minimax polynomial, but this is likely to be a long process.
Instead we can use a very powerful result due to Chebyschev.
Theorem 3.3.1. Chebyshev’s equi-oscillation theorem. Let f (x) be a con-
tinuous function defined on the interval [a, b] and pn (x) be a polynomial
approximation, of degree n, for f on [a, b] with associated error function
en (x) = f (x) − pn (x). If there exists a set of at least n + 2 points, a ≤ x0 <
x1 < . . . < xn+1 ≤ b, where the error function takes maximum values such
that
en (xi ) = (−1)i ǫn , i = 0, 1, . . . , n + 1,
then pn (x) is the minimax approximation. This approximation is the only
one of degree ≤ n with this property.
Notice that the equi-oscillation theorem only requires alternating minima
and maxima in the error function; they need not be turning points as we shall
see in the next example. This theorem is extremely important for it enables
3.3. MINIMAX POLYNOMIALS 59

us to decide whether or not a particular polynomial is minimax amongst

all polynomials of a given degree. For example, the error associated with
the third-order approximation (3.9), which is shown in figure 3.3, has five
alternating minima and maxima at the points P , Q, R, S and T , but since
they are not all of the same magnitude this polynomial is not minimax.

2.5

2.0

x
1.5 e

1.0

0.5

-1.0 -0.5 0.5 1.0

Figure 3.5: Linear minimax approximation of exp(x), |x| ≤ 1

Example 3.3.1 (Analytical)

In order to demonstrate the minimax equi-oscillation property we will con-
struct a minimax linear approximation for exp(x) on the interval [−1, 1]. Let
us write the linear minimax approximation for exp(x) as p1 (x) = ax+ b, then
the error function e1 (x) is

e1 (x) = exp(x) − ax − b.

As the degree of the polynomial approximation is 1 we need to determine

three points x0 , x1 and x2 such that e1 (x) takes maximum absolute values
which are of the same size but alternating sign, i.e.

e1 (x0 ) = E, e1 (x1 ) = −E, and e1 (x2 ) = E.

At least one of x0 , x1 and x2 must be an internal point of the interval [−1, 1]

at which there will be a turning point, i.e. de1 /dx = 0, which gives

exp(x) − a = 0.
60 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

This equation has only one solution, therefore only one of x0 , x1 and x2
is a turning point. Let this be x1 . Accordingly, e1 (x) will have maximum
absolute value at both x0 and x2 neither of which will be turning points.
Therefore,

exp(−1) + a − b = E
exp(x1 ) − ax1 − b = −E
exp(1) − a − b = E

which gives three equations in the four unknowns; the unknown point x1 , the
coefficients a and b, and the error E. However, at x1 the error function has
a turning point and so
de1
(x1 ) = exp(x1 ) − a = 0,
dx
which gives a fourth equation. The solution of these equations is:

(exp(1) − exp(−1))
a = = 1.17520119,
2
x1 = loge (a) = 0.161439361,
(exp(x1 ) + exp(1) − ax1 − a)
b = = 1.26427905
2
E = exp(1) − a − b = 0.278801586.

The resulting polynomial approximation is shown in figure 3.5 together with

its error function in figure 3.6. Recall that we are seeking a polynomial of
degree 1 and notice that the error function in figure 3.6 has three alternating
minima and maxima at P , Q and R, all of which are equal to E in magnitude.
Therefore, this function is the linear minimax approximation for exp(x), for
x in [−1, 1] and E is the minimax error ǫ1 . In order to investigate this
idea further consider an alternative linear approximation for exp(x) given by
q1 (x) = 1 + x, the error function for which is also shown in figure 3.6. It is
clear that the maximum error of q1 is greater than that of p1 . The Chebyshev
equi-oscillation theorem states that for any other linear approximation the
associated error has a maximum value greater than the error associated with
the minimax linear approximation p1 (x). For linear minimax approximations
it is possible to determine the correct expression algebraically, but in general
this is not the case.
3.3. MINIMAX POLYNOMIALS 61

0.2

0.1

-1.0 -0.5 0.5 1.0

-0.1

-0.2

Figure 3.6: Linear minimax error function for exp(x), |x| ≤ 1

Example 3.3.2 (Analytical/Computational)

We shall attempt to construct a cubic minimax approximation for exp(x)
for x in [−1, 1]. Firstly, let us examine the problems of determining the
minimax approximation analytically. We need to find a set of five points,
(n = 3 since we seek a cubic approximation), x0 , x1 , x2 , x3 and x4 , where
the error function
e3 (x) = exp(x) − a3 x3 − a2 x2 − a1 x − a0
has alternating minima and maxima all of which are of equal magnitude.
Differentiating this function shows that de3 /dx has no more than three zeros
in [−1, 1] and so x0 = −1 and x4 = 1 at which the error function will have
maximum absolute value. (To see this, sketch the curve y = exp(x) and any
other cubic.) Therefore, we have to solve the equations
exp(−1) + a3 − a2 + a1 − a0 = E
exp(x1 ) − a3 x31 − a2 x21 − a1 x1 − a0 = −E
exp(x2 ) − a3 x32 − a2 x22 − a1 x2 − a0 = E (3.10)
exp(x3 ) − a3 x33 − a2 x23 − a1 x3 − a0 = −E
exp(1) − a3 − a2 − a1 − a0 = E
plus the conditions for turning points at x1 , x2 and x3 , i.e.
exp(x1 ) − 3a3 x21 − 2a2 x1 − a1 = 0
exp(x2 ) − 3a3 x22 − 2a2 x2 − a1 = 0
exp(x3 ) − 3a3 x23 − 2a2 x3 − a1 = 0
62 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

which gives a system of eight non-linear equations in the unknown values x1 ,

x2 , x3 , the error E, and the coefficients a0 , a1 , a2 and a3 . In general it would
be extremely difficult to solve these equations.
Exercise 3.3.1 (Analytical) Construct a cubic minimax approximation for the
function cos(x) for values of x in [−1, 1]. (Hint: use symmetry to reduce the
size of the problem.)
Example 3.3.2 demonstrates the difficulty in constructing minimax ap-
proximations of order n > 1. We will now consider an alternative iterative
approach. Consider the cubic approximation for exp(x) given by (3.6), i.e.

exp(x) ≈ 0.99520 + 0.99905x + 0.54788x2 + 0.17615x3 ,

which is shown in figure 3.2. The associated error, as shown in figure 3.3,
has the correct number of minima and maxima but they are of different
magnitude. The internal stationary points of the error function occur at
approximately −0.7, 0 and 0.8, therefore let us solve the equations (3.10)
with x1 = −0.7, x2 = 0 and x3 = 0.8. This produces the linear equations
    
1 −1 1 −1 1 a0 exp(−1)
2 3
1 −0.7 0.7 −0.7 −1   a exp(−0.7)
    
 1
   
   

 1 0 0 0   a2
1   
 = 
 exp(0) 

2 3

 1 0.8 0.8 0.8 −1   a3
 


 exp(0.8) 

1 1 1 1 1 E exp(1)

and in turn the solution of these equations gives the cubic

q3 (x) = 0.99465 + 0.99643x + 0.54308x2 + 0.17877x3. (3.11)

The error e3 (x) = exp(x) − q3 (x) is shown in figure 3.7, together with the
computed value for E, and has turning points at −0.687312, 0.045814 and
0.727528 in addition to the maxima at −1 and +1. (These extremum points
are denoted by P , Q, R, S and T ). As we seek a minimax approximation
of degree 3 the error should have 3+2 alternating extrema of the same mag-
nitude. From figure 3.7 we can see that the error function does have five
extrema which have alternating sign but they are of different magnitude.
If the extrema were of the same magnitude then E would be the minimax
error ǫ3 for exp(x). We could now substitute these five points into (3.10)
to determine another approximation and hence iteratively refine our cubic
3.3. MINIMAX POLYNOMIALS 63

approximation. (Notice that only the x co-ordinate is required to deter-

mine the next approximation.) This is the basis of the Remes algorithm
for determining minimax polynomial approximations which can be described
schematically as in figure 3.8.
0.006

0.004

0.002

-1.0 -0.5 0.5 1.0

-0.002

-0.004

Figure 3.7: Error curve associated with the cubic (3.3.4)

Not available

Figure 3.8: The Remes algorithm

Example 3.3.3 (Computational)

We demonstrate the Remes algorithm by considering a cubic minimax ap-
proximation for exp(x), −1 ≤ x ≤ 1. Load the program APPROX, for
which this is the default function. When the prompt shows Function: as
set press RETURN and select the option Method: Minimax/Remes.
As we seek the minimax approximation of order 3, we will need to supply
five points at which the required approximation for exp(x) has alternating
minima and maxima. For the moment let us use the values x = −1, −0.7,
0, 0.8 and 1 which can be input using the prompt Option: edit/inspect
points. When you are satisfied that suitable points have been entered quit
this option and change the prompt to Option: calculate coeffs. (It is also
possible to supply a first approximation for the minimax polynomial using
the option Taylor Series). Now press RETURN and the computer will de-
termine the solution of the equations in step 2 of the Remes flowchart, shown
in figure 3.8, and display the prompt Plot: function. If the function has
64 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

not been plotted prior to this point then pressing RETURN will display it
in the graphics window, otherwise, change the prompt to Plot: approxi-
mation and compare the computed approximation with the function. It is
likely that the function and the computed approximation are very close; the
error between them can be examined by using the option Plot: error. For
this example you should obtain figure 3.7 together with a dotted line which
denotes the value for E in equations (3.10). The coefficients of the computed
approximation can be examined by using prompt Option: tabulate, which
in this case gives the cubic (3.11). Notice that the error associated with this
approximation has maximum values at the points P , Q, R, S and T . We can
select these points to be used in the next stage of the Remes algorithm by us-
ing the prompt Option: edit/inspect points and then return to Option:
calculate coeffs to determine the next approximation. (Only the x coordi-
nates are required, therefore, it is unnecessary to enter the y coordinates.)
This process can then be repeated until convergence is obtained.
Exercise 3.3.2 (Computational) Use the program APPROX to carry out
the next iteration of the Remes algorithm using the minima/maxima of the
error function associated with the cubic approximation (3.11).
Exercise 3.3.3 (Computational) Use the program APPROX with the op-
tion Remes/minimax to determine the quartic minimax approximation
for exp(x), |x| ≤ 1.
Exercise 3.3.4 (Computational) Use the program APPROX with the op-
tion Remes/minimax to determine the quartic minimax approximation
for log(1 + x), |x| ≤ 21 . Try different starting polynomial approximations to
examine the speed at which this algorithm converges to the required minimax
approximation.
At each stage of the Remes algorithm it is to be expected that the maxi-
mum value of the associated error function will be reduced but it will certainly
be an upper bound on the minimax error ǫn for polynomial approximations
of order n. In fact we can also produce a lower bound for the minimax error
ǫn .
Theorem 3.3.2. (de la Vallee Poussin).
Given that qn is an nth degree polynomial approximation for a function
f (x), x ∈ [a, b], and xi , i = 0, . . . , n+ 1, are any n+ 2 points in [a, b] such that
a ≤ x0 < x1 < . . . < xn+1 ≤ b where the error function en (x) = f (x) − qn (x)
3.3. MINIMAX POLYNOMIALS 65

alternates in sign at successive points, then

min
x
|en (xi )| ≤ ǫn ≤ max |en (x)|,
i x∈[a,b]

where ǫn is the minimax error.

Notice that the upper bound in theorem 3.3.2 is over the whole interval,
but that the lower bound is only at the points x0 , x1 , . . . , xn+1 . For example,
the cubic approximation (3.11) produces the bounds

0.0053526 ≤ ǫn ≤ 0.0059120.

(The lower bound is taken over the points −1, −0.7, 0, 0.8 and 1.0 and occurs
at x = −0.7. The upper bound is over the whole of the interval and occurs at
x ≈ 0.7275.) If these bounds are sufficiently close to each other then there is
little point in carrying out another iteration of the Remes algorithm. What
is more, the lower bound tells us that if we wish to have a minimax error
less than 0.0053526 then we must construct a polynomial approximation of
higher order.
The Remes algorithm is based upon the selection of exactly n + 2 al-
ternating minima/maxima from a given error function, but it is possible to
obtain an error curve for a polynomial approximation of degree n which has
more than n + 2 such points. (Although pn (x) cannot have more than n − 1
turning points f (x) − pn (x) could have any number.) In this case we sim-
ply select n + 2 values where the error function has alternating minima and
maxima, the remainder are discarded, the only proviso is that we must take
the largest minima/maxima amongst the points we select. We will not give
details of how to find the relevant minima/maxima but in general this is
usually accomplished by a linear search along the error curve, not by differ-
entiating. However, the location of suitable points by hand is a relatively
simple procedure as the following example demonstrates.
Example 3.3.4 (Analytical)
Let us suppose that we have constructed a quadratic approximation for some
function and the error curve looks like figure 3.9. The minimax approxima-
tion should have four alternating minima and maxima, this curve has five.
We need to select four of the values shown in order to proceed with the
Remes algorithm; in this case we take x1 , x2 , x3 and x4 as these are alter-
nating minima and maxima and the maximum error is at x4 . The point x0
will no longer be used.
66 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

x x
1 3
x x x
0 2 4

Figure 3.9: A possible quadratic approximation error curve

As the order of the polynomial approximation increases the time required

by the Remes algorithm increases rapidly and the closer the initial approxi-
mation is to the minimax the better. Furthermore, as the number of internal
points increases equation (3.10) may become ill conditioned. The reason for
this is clear. As n increases then the matrix corresponding to equations (3.10)
becomes very close to the Vandermonde matrix discussed in Chapter 2 which
is known to be ill conditioned as the tabular points get closer together. A
solution to this problem will be discussed in a later section.

Exercise 3.3.5 (Computational) Use the program APPROX to construct a

minimax polynomial approximation for exp(x) which is of minimal order but
has error no worse that 0.001 for −1 ≤ x ≤ 1.

Exercise 3.3.6 (Computational) Use the points −1, 0, 1 to construct a linear

minimax approximation for x3 over the interval [−1, 1]. Explain why the
Remes algorithm fails in this case and suggest how it is possible to overcome
this difficulty. (Hint: look at the quadratic minimax approximation.)

3.4 Chebyshev approximation

All the approximations we have discussed so far have been simple power
series approximations, i.e. of the form
i=n
ai xi + an+1 xn+1 + . . . ,
X
f (x) = (3.12)
i=0
3.4. CHEBYSHEV APPROXIMATION 67

in the hope that, provided the series converges as n → ∞, taking a finite

number of terms will provide a suitable approximation. If we wish to ob-
tain an approximation close to the minimax by truncating equation (3.12)
after n terms then the remainder must satisfy the Chebyshev equi-oscillation
property. Clearly, using this simple power series cannot produce this phe-
nomenon, therefore, we will now look at alternative expansions. We begin
by looking at Chebyshev polynomials.
Let us consider a sequence of functions which are defined, for all values
of x ∈ [−1, 1], as follows:
T0 (x) = 1, T1 (x) = x,
Tn+1 (x) = 2xTn (x) − Tn−1 (x). (3.13)
From the recurrence relation (3.13) we can generate
T2 (x) = 2xT1 (x) − T0 (x) = 2x2 − 1,
T3 (x) = 2xT2 (x) − T1 (x) = 4x3 − 3x,
and so on. These functions have some remarkable properties. For example,
notice that Tn (x) is a polynomial of degree n and that for n odd Tn is an
odd function whilst for n even Tn is even. The functions Tn (x) are called
Chebyshev polynomials. The first four are shown in figure 3.10.
1.0 T0

T1
0.5

-1.0 -0.5 0.5 1.0

T2
-0.5

T3
-1.0

Figure 3.10: The Chebyshev polynomials Ti (x), i = 0, 1, 2, 3

Exercise 3.4.1 (Analytical) Show that Tn (x), x ∈ [−1, 1], is a polynomial of

degree n and that it is odd when n is an odd integer and even when n is an
even integer. (Hint: use induction.)
68 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

We can also define Chebyshev polynomials in another way, i.e.

Tn (x) = cos(n cos−1 (x)). (3.14)
To see this note that
Tn+1 (x) = cos((n + 1) cos−1 (x))
= cos(cos−1 (x)) cos(n cos−1 (x))
− sin(cos−1 (x)) sin(n cos−1 (x))
and
Tn−1 (x) = cos((n − 1) cos−1 (x))
= cos(cos−1 (x)) cos(n cos−1 (x))
+ sin(cos−1 (x)) sin(n cos−1 (x))
Adding together these equations gives
Tn+1 (x) + Tn−1 (x) = 2 cos(cos−1 (x)) cos(n cos−1 (x)) = 2xTn (x), (3.15)
which is equivalent to (3.13). The definition (3.14) now enables us to draw
another conclusion about the polynomial Tn (x): it has exactly n + 1 alter-
nating minima and maxima, with equal magnitude, in the interval [−1, 1].
The importance of this comment can be demonstrated as follows.
Let us assume that instead of (3.12) we can write
i=n
X
f (x) = ci Ti (x) + cn+1 Tn+1 + cn+2 Tn+2 + . . . . (3.16)
i=0

If this is the case then the dominant term in the error when this series is
truncated after n + 1 terms will be cn+1 Tn+1 and as Tn+1 has exactly n + 2
equal and opposite minima and maxima we should obtain an approximation
close to the minimax! We must now justify the expansion (3.16) of any
function in terms of Chebyshev polynomials. First of all, if the series (3.12)
converges for x ∈ [−1, 1] then so will any re-arrangement of its terms, but this
is just what the Chebyshev expansion is. Therefore, if (3.12) converges then
so will (3.16). The next stage is to find an expression for the coefficients
cj , j = 0, 1, . . .. This turns out to be quite straightforward because the
Chebyshev polynomials possess an orthogonality property, i.e.

π, i = j = 0
1 Ti (x)Tj (x)
Z 

q dx =  π2 , i = j > 0 .
−1 2
(1 − x ) 
0, i=6 j
3.4. CHEBYSHEV APPROXIMATION 69

To see this notice that the substitution x = cos θ reduces

Z 1
Ti (x)Tj (x)
q dx
−1 (1 − x2 )
to the standard trigonometric integral

Z π 
 π, i = j = 0
π
cos iθ cos jθ dθ = , i=j>0 .
0  2
0, i=6 j


Therefore, in order to determine the coefficients cj in equation (3.16) we mul-

tiply both sides by Tj (x)/(1 − x2 )1/2 and integrate over the interval (−1, 1),
i.e.
1 1 f (x) 1 π
Z Z
c0 = q dx = f (cos θ) dθ, (3.17)
π −1 (1 − x2 ) π 0
2 1 Tj (x)f (x) 2 π
Z Z
cj = q dx = f (cos(θ)) cos jθ dθ. (3.18)
π −1 (1 − x2 ) π 0

Alternatively, we can redefine the Chebyshev series (3.16) as

i=n i=∞
1 X X′
f (x) = c0 + ci Ti (x) + cn+1 Tn+1 + cn+2 Tn+2 + . . . = ci Ti (x), (3.19)
2 i=1 i=0

where the symbol denotes that the first coefficient is 21 c0 . All the coefficients
′

are then given by (3.18).

Example 3.4.1 (Analytical)
Determine the Chebyshev coefficients for the function 3x(1 − x2 )1/2 . Notice
that this is an odd function, i.e. f (x) = −f (−x), and so all the even coeffi-
cients will vanish; we need only determine the odd coefficients. Therefore,
q
2
Z 1 3x (1 − x2 )T1 (x)
c1 = q dx
π −1 (1 − x2 )
4 1 2 4
Z
= 3x dx = ,
π 0 q
π
2 1 3x (1 − x2 ) T3 (x)
Z
c3 = q dx
π −1 (1 − x2 )
4Z1 12
= 3x(4x3 − 3x) dx = − ,
π 0 5π
70 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

and so on. The function given by c1 T1 (x) + c3 T3 (x) is compared with f (x) =
3x(1 − x2 )1/2 in figure 3.11 and the associated error is shown in figure 3.12.
Notice that the error function has six alternating minima/maxima, but of
unequal size. Although we have approximated f by a cubic, the six min-
ima/maxima cause no problems because the coefficient c4 is zero and there-
fore the quartic approximation for f is given by a cubic.

1.5

1.0

0.5

-1.0 -0.5 0.5 1.0

-0.5

f HxL -1.0

-1.5

Figure 3.11: Cubic Chebyshev approximation for 3x(1 − x2 )1/2

0.4

0.2

-1.0 -0.5 0.5 1.0

-0.2

-0.4

Figure 3.12: The error associated with figure 3.4.2

Example 3.4.2 (Analytical)

In order to illustrate approximation using Chebyshev polynomials we shall
compare a linear Chebyshev approximation for the function ex , |x| ≤ 1,
3.4. CHEBYSHEV APPROXIMATION 71

with the corresponding linear minimax approximation p1 (x) = 1.2464279 +

1.175201x, which was determined in example 3.3.1. The coefficients for the
required approximation are given, from equation (3.18), as c0 = 2.53213212
and c1 = 1.130318, which gives the Chebyshev approximation

exp(x) ≈ 1.266066 + 1.130318T1(x) = 1.266066 + 1.130318x

which is close to the minimax approximation of degree 1. (Notice that the

first term is 12 c0 .)
Exercise 3.4.2 (Computational) Use the program APPROX with the op-
tion Method: Chebyshev to find approximations for the function exp(x),
x ∈ [−1, 1], of degree 1,2 and 3. Compare the computed approximations
with the minimax approximations derived in the last section. (The resulting
Chebyshev approximations are shown in figure 3.13.)

2.5

2.0

1.5

Linear 1.0

Quadratic 0.5

-1.0 -0.5 0.5 1.0

Figure 3.13: Chebyshev approximations for exp(x)

Given the coefficients of a Chebyshev approximation we can compare it

with the minimax polynomial. For example, the minimax cubic approxima-
tion for exp(x), x in [−1, 1], is given by

p3 (x) = 0.9945858 + 0.9956685x + 0.5429732x2 + 0.1795327x3

which has minimax error ǫ3 = 0.0055216. The corresponding Chebyshev

approximation is

P3 (x) = 1.266066T0 + 1.130318T1 + 0.271495T2 + 0.044337T3

72 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

which gives the equivalent polynomial

P3 (x) = 0.994573 + 0.997307x + 0.542986x2 + 0.177348x3,
and this has maximum error 0.006065. At this point we emphasize that there
is no need to convert a Chebyshev series in this way; in fact to do so may lead
to significant errors. The reason for this is the size of the coefficients in the
corresponding series; the smaller the coefficients the better. Notice that for
this example the coefficients of the higher powers of the Chebyshev series are
much smaller than those of the minimax approximation. It is better to eval-
uate a given Chebyshev series by making use of the recurrence relation (3.13)
in order to minimize induced errors. In general given a Chebyshev series
i=n
X ′
f (x) = ci Ti (x)
i=0

it can be evaluated as follows. For a given value of x define un (x) = cn

un−1(x) = cn−1 + 2xun ,
uj (x) = cj + 2xuj+1 − uj+2 , j = n − 2, . . . , 1, (3.20)
then
i=n
′
u0 (x) = 12 c0 + xu1 (x) − u2 (x) =
X
cj Tj (x) = f (x).
i=0
Notice that (3.20) is a second-order recurrence relation, for any fixed value
of x, and is relatively stable provided |x| ≤ 1 which is precisely the range of
values in which we are interested. (This technique for evaluating Chebyshev
series can be compared with nested multiplication for power series.) In order
to show this we rearrange the equations (3.20) to give
cn = un (x),
cn−1 = un−1 (x) − 2xun (x),
cj = uj (x) − 2xuj+1 (x) + uj+2(x), j = n − 2, . . . , 1,
1
c =
2 0
u0 (x) − xu1 (x) + u2 (x).
Now recall that
f (x) = cn Tn + . . . + c1 T1 + 21 c0
= un (x)(Tn (x) − 2xTn−1 (x) + Tn−2 (x)) +
un−1(x)(Tn−1 (x) − 2xTn−2 (x) + Tn−3 (x)) + . . . +
u1 (T1 (x) − xT0 (x)) + u0 (x)T0 (x).
3.4. CHEBYSHEV APPROXIMATION 73

Finally, we use the recurrence relation (3.13), in which case all the terms
except the last on the right-hand side vanish; this last term is u0 since T0 = 1.
This technique would not be used to produce a explicit expression for a
Chebyshev series as a power series but to evaluate the Chebyshev series at a
particular point.
Example 3.4.3 (Calculator)
Evaluate the Chebyshev series

1.266066T0 + 1.130318T1 + 0.271495T2 + 0.044337T3

at the point x = 21 . We set u3 = 0.044337 and then produce

u2 = 0.271493 + 2 × 21 × u3 = 0.271493 + 0.044337 = 0.315832

u1 = c1 + 2 × 21 × u2 − u3 = 1.130318 + 0.315832 − 0.044337
= 1.401813
u0 = 21 c0 + 12 u1 − u2 = 1.6511405.

(Recall that 21 c0 = 1.266066 from equation (3.18).)

In general the determination of the Chebyshev coefficients is not as sim-
ple as demonstrated in example 3.4.1 and it may be necessary to resort to
numerical integration in order to evaluate them. (This is what the option
Chebyshev in the program APPROX does.) However, the Chebyshev
expansion (3.16) has one other bonus. Clearly, given any other set of poly-
nomials we could expand any suitable function in terms of a similar expan-
sion in which case we would still need to determine the relevant coefficients.
Amongst all1 such expansions it can be shown that the coefficients in the
Chebyshev series decrease more rapidly than any other expansion. Further-
more, given an expansion in terms of the first n such polynomials the error
in truncating the series after n terms, i.e.

cn Tn + cn+1 Tn+1 + . . .

is closely bounded above by the single term |cn Tn |, provided the coefficients
tend to zero sufficiently rapidly. (See Fox and Parker (1972).) However,
|Tn (x)| ≤ 1 and so we can bound the truncation error by the single term |cn |.
1
Strictly speaking amongst all expansions in terms of ultra-spherical polynomials for
which the associated weight function is w(x) = (1 − x)−p , p = 21 gives Chebyshev polyno-
mials, p = 0 gives Legendre polynomials.
74 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

Therefore, if we require an approximation which has an overall error bound

ǫ we determine successive Chebyshev coefficients until we have |cn | ≤ ǫ and
this will give the required polynomial.
Example 3.4.4 (Computational)
The cubic Chebyshev approximation for exp(x), x ∈ [−1, 1] is given by

P3 (x) = 1.266066T0 + 1.130318T1 + 0.271493T2 + 0.044337T3

where the next coefficient is given by c4 = 0.005474. The corresponding cubic

minimax is given by

p3 (x) = 0.9945858 + 0.9956685x + 0.5429732x2 + 0.1795327x3,

which has minimax error ǫ3 = 0.0055216. From this we can see that c4 is a
reasonable estimate for ǫ3 .

Table 3.1: Minimax and Chebyshev maximum absolute errors

Taylor series Minimax Chebyshev Coefficient
Order error error error neglected
3 0.04744 0.005529 0.006063 0.005474
exp(x) 4 0.00656 0.000547 0.000587 0.000522
x ∈ [−1, 1] 5 0.00292 0.000045 0.000064 0.000046

2 0.28737 0.029830 0.033308 -0.029437

2
loge (1 + x ) 4 0.15340 0.003424 0.003887 0.003198
x ∈ [−1, 1] 6 0.07556 0.000444 0.000689 -0.000432

sin(x) 3 0.00897 0.000499 0.000503 0.000521

x ∈ [−1, 1] 5 0.00009 0.000003 0.000003 -0.000003

In many cases Chebyshev approximations will be very close to the mini-

max polynomial. To illustrate this we list in table 3.1 various functions and
compare the minimax error, corresponding Chebyshev maximum error and
also the size of the first coefficient ignored in the Chebyshev approximation2 .
2
The values given are those calculated by the software. If exact coefficients are supplied
then different entries may be obtained.
3.4. CHEBYSHEV APPROXIMATION 75

If the Chebyshev approximation is not sufficiently close to the minimax ap-

proximation then we simply take it as a first approximation in the Remes
algorithm. Alternatively, we can start the Remes algorithm by constructing
a polynomial approximation using the points where the maximum errors oc-
cur at the extrema of the required Chebyshev polynomial. For example, if we
require a polynomial approximation of degree n, then we know that the error
function must have n + 2 alternating minima/maxima; we take as our first
iπ
approximation for the extrema the points xi = cos (n+1) , i = 0, 1, . . . , n+1,
which are the n + 2 extrema of Tn+1 (x). This choice of points should produce
an approximation which is close to the minimax.
Exercise 3.4.3 (Analytical) As an alternative to the Chebyshev polynomials,
given on page 67, we can define the Modified Chebyshev polynomials Tn∗ (x),
for 0 ≤ x ≤ 1, according to
Tn∗ (x) = Tn (2x − 1), 0 ≤ x ≤ 1.
Derive the first five Modified Chebyshev polynomials and sketch the corre-
sponding curves. Show that
∗
Tn+1 (x) = 2(2x − 1)Tn∗ (x) − Tn−1
∗
(x).
The best quadratic Chebyshev approximation for x3 + x2 over the range
[−1, 1] is x2 + 43 x which has a maximum error 0.25, i.e.
x3 + x2 = x2 + 43 x + 41 T3 (x).
Show that the best Chebyshev quadratic approximation for this function over
the range [0, 1] is given by
x3 + x2 = 1
32
(10T0∗ + 15T1∗ + 6T2∗ ) + x2 + 1 ∗
T .
32 3
1
Therefore, the best quadratic approximation is 32 (1 − 18x + 80x2 ) which has
1
a maximum error of 32 , one eighth that of the approximation on the interval
[−1, 1].
Exercise 3.4.4 (Analytical). Runge’s problem revisited. In example 2.1.3 and
exercise 2.1.4 the construction of a polynomial approximation for exp(x) was
considered. It was demonstrated that if the interpolation points are equally
spaced then as the number of points increases the interpolating polynomial,
pn (x), oscillates wildly. This would seem to imply that
max |en (x)| = max | exp(−x2 ) − pn (x)| → ∞ as n → ∞.
x0 ≤x≤xn x0 ≤x≤xn
76 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

However, recall that in general the error between an interpolating polynomial

pn (x) and a function, f (x), for which it is an approximation satisfies
Mn+1 i=nY
|f (x) − pn (x)| ≤ (x − xi ),
(n + 1)! i=0

where Mn = maxx0 ≤x≤xn |f (n) (x)| and xi , i = 0, . . . , n, are the interpolation

points selected. Show that by selecting
i=n
(x − xi ) = 2−n Tn+1 (x)
Y
(3.21)
i=0

the error function en (x) satisfies

Mn+1
|en (x)| ≤ .
2n (n + 1)!
Therefore, provided Mn+1 is finite for all n, en (x) → 0 as n → ∞. The
choice (3.21) looks rather strange but simply requires that the zeros of both
sides of this equation are identical, i.e. we need to select the interpolation
points to be the zeros of Tn+1 which are
!
(2i + 1)π
xi = cos , i = 0, 1, . . . , n. (3.22)
2n + 2

3.4.1 Polynomial economization

Let us suppose that we are given a polynomial of degree n and ask is it
possible to construct a polynomial of degree ≤ n which is minimax? This
apparently difficult problem is in fact very easy to solve using Chebyshev
polynomials. In order to see this let
Qn (x) = a0 + a1 x + . . . + an−1 xn−1 + an xn ,
and then note that Tn (x) = 2n−1xn + Pn−1 (x), where Pn−1 is a polynomial
of degree n − 1. Therefore,
xn = 2(1−n) (Tn (x) − Pn−1 (x))
which gives
Qn (x) = a0 +a1 x+. . .+an−1 xn−1 +an 2(1−n) (Tn (x) − Pn−1 (x))
= a0 +a1 x+. . .+an−1 xn−1 −an 2(1−n) Pn−1 (x)+an 2(1−n) Tn (x).
3.4. CHEBYSHEV APPROXIMATION 77

Therefore, the (n − 1)th degree polynomial

a0 + a1 x + . . . + an−1 xn−1 − an 2(1−n) Pn−1 (x)

differs from Qn by a term an 2(1−n) Tn (x) which has exactly n + 1 equal and
opposite minima/maxima of size an 2(1−n) and is therefore the minimax ap-
proximation of this degree.
Example 3.4.5 (Analytical)
Construct a quadratic minimax approximation for the polynomial

Q3 (x) = 1 + x + x2 + x3 .

The Chebyshev polynomial T3 (x) = 4x3 − 3x, therefore,

x3 = 41 (T3 (x) + 3x)

and

Q3 = 1 + x + x2 + x3 = 1 + x + x2 + 41 (T3 (x) + 3x)

= 1 + 74 x + x2 + 41 T3 (x).

Therefore, the function 1 + 47 x + x2 is the required quadratic minimax ap-

proximation for Q3 (x) since it will have precisely four equal and opposite
minima/maxima of size 41 . No other quadratic polynomial can have a smaller
error.
Exercise 3.4.5 (Analytical) Construct a linear minimax approximation for
the quadratic function
Q2 (x) = 1 + x + 12 x2
which is the Taylor series expansion of exp(x) about x = 0. Determine the
minimax approximation error involved.

3.4.2 Chebyshev minimax approximation

In section 3.3 the Remes algorithm was used to construct the minimax ap-
proximation for a given function. This algorithm is relatively easy to imple-
ment as was demonstrated in exercise 3.3.2 but as the order of the approxi-
mating polynomial increases convergence may be slow and it is possible that
the process becomes ill conditioned. In the previous section it was shown
78 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

how to construct a Chebyshev approximation which is almost as good as

the minimax and so it would seem sensible to combine the Remes algorithm
together with Chebyshev polynomials; we will briefly investigate this further
in this section.
Let f (x) be a continuous function for which pn (x) is the minimax polyno-
mial of degree n. The error function en (x) = f (x) − pn (x) has exactly n + 2
alternating minima and maxima of equal magnitude. However, as pn (x) is a
polynomial of degree n we can express it in the form

pn (x)+ = 21 c0 T0 + c1 T1 (x) + c2 T2 (x) + . . . + cn Tn (x).

This form of expressing the minimax polynomial has several advantages over
an expansion in simple powers of x, not the least of which is that the coef-
ficients will be numerically smaller which means that the evaluation of this
expression will be less prone to rounding errors. Any other linear combination
of Chebyshev polynomials will have a larger maximum error bound than that
given by the minimax error ǫn = max |en (x)|. Let us assume that we have
obtained an approximation for the minimax polynomial in terms of a Cheby-
shev series and that we have determined (n+2) points at which the associated
error function has alternating minima and maxima, x0 ≤ x1 ≤ . . . ≤ xn+1 .
Now we determine a better polynomial approximation, qn (x), as

qn (x) = 21 b0 + b1 T1 (x) + . . . + bn Tn (x)

by solving the (n + 2) linear equations

f (xi ) − qn (xi ) = (−1)i+1 E, i = 0, . . . , n + 1, (3.23)

to determine the coefficients b0 , b1 , . . . , bn . The similarity between the equa-

tions (3.23) and (3.10) is striking but the former equations have the advantage
of being far better conditioned as n increases. In order to demonstrate this
procedure consider example 3.4.6.

Example 3.4.6 (Computational)

Let us construct the cubic Chebyshev minimax approximation for the func-
tion exp(x), −1 ≤ x ≤ 1. We begin by specifying five points as extrema for
the error function

e3 (x) = exp(x) − c0 T0 − c1 T1 − c2 T2 − c3 T3 .
3.4. CHEBYSHEV APPROXIMATION 79

0.004

0.002

-1.0 -0.5 0.5 1.0

-0.002

-0.004

Figure 3.14: The error exp(x) − Q3 (x). See equation (3.4.12)

We use the
 extrema
 of the Chebyshev polynomial T4 (x) which are given by
1
xi = cos 4 iπ , i = 0, 1, 2, 3, 4. (See page 75.) This gives the points x0 = −1,
√ √
x1 = − 12 2, x2 = 0, x3 = 12 2 and x4 = 1. (Compare these values with
those used in example 3.3.3). We now solve the five linear equations
exp(xi ) − c0 T0 (xi ) − c1 T1 (xi ) − c2 T2 (xi ) − c3 T3 (xi ) = (−1)i E,
i = 0, 1, 2, 3, 4.
These equations can be rewritten as
    
1 −1
√ 1 −1 √ 1 c0 exp(−1)
√
1 12 2 0 − 12 2 −1   c1   exp(− 12 2)
    
 
    

 1 0√ −1 √0   c2  = 
1     exp(0)
√


1 1 1

 1 −2 2 0 2
2 −1



 c 3



 exp( 2
2)


1 1 1 1 1 E exp(1)
and produce the cubic Chebyshev polynomial
Q3 (x) = 1.266066 + 1.130321T1 + 0.271540T2 + 0.044880T3 (3.24)
which has the error function shown in figure 3.14. This function has maxi-
mum absolute value 0.00583 but the corresponding value of E is 0.005474.
We see that this cubic is not minimax but we can repeat this process to
improve it. We can also compare this cubic with the first four terms of the
Chebyshev series for exp(x) given by
exp(x) ≈ 1.266066 + 1.130318T1 + 0.271495T2 + 0.044337T3
which has a maximum error of 0.00607.
80 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

Exercise 3.4.6 (Computational) Write out the equations in the next step of
the Remes algorithm for determining the cubic Chebyshev minimax approx-
imation for exp(x). Solve the corresponding linear equations and investigate
whether the resulting cubic is minimax.

3.5 Orthogonal functions

The Chebyshev polynomials introduced in section 3.4 are extremely useful
when an approximation to a given function is required. This is because they
possess many useful properties, one of the most important of these being
orthogonality. By this we mean that for a set of polynomial functions, φn (x),
−1 ≤ x ≤ 1, n = 0, 1, . . . , there exists a weight function w(x) > 0 such that
Z 1
w(x)φn (x)φm (x) dx = Mn δmn ,
−1

where Mn is a constant and δmn is 1 if n = m and 0 otherwise. We could

take any other set of linearly independent polynomials, φn (x), x ∈ [−1, 1],
and consider expansions similar to (3.12) or (3.16). (The condition of linear
independence is needed to ensure that no one member of the set can be
expressed in terms of a linear combination of the rest and would, therefore,
be redundant.) Ignoring the problems of convergence of infinite series, let us
for the moment write
f (x) = c0 φ0 + c1 φ1 + c2 φ2 + . . . .
However, if the functions φ0 , φ1 , . . . are orthogonal we can determine the
coefficients cn as
1
Z 1
cn = w(x)f (x)φn (x) dx.
Mn −1
Example 3.5.1
The following are sets of orthogonal polynomials
(i) Chebyshev polynomials. Tn (x), x ∈ [−1, 1], w(x) = (1 − x2 )−1/2 .
T0 (x) = 1, T1 (x) = x,
Tn+1 (x) = 2xTn (x) − Tn−1 (x).
(ii) Legendre polynomials. Pn (x), x ∈ [−1, 1], w(x) = 1.
P0 (x) = 1, P1 (x) = x,
3.6. FOURIER SERIES EXPANSIONS 81

(n + 1)Pn+1 (x) + nPn−1 (x) = (2n + 1)xPn (x).

Further details of these polynomial functions and their properties can be
found in Fox and Parker (1972) or Churchill (1963).
The idea of orthogonal polynomials can be generalized further to sets of
orthogonal functions.
Example 3.5.2
The following are sets of orthogonal functions
q
(i) Exponential functions exp(inx), i = (−1), w(x) = 1.
(ii) Fourier series.
(
sin(nπx)
φn (x) = , w(x) = 1.
cos(nπx)

Exercise 3.5.1 (Analytical) Show that the sets of functions given in examples
3.5.1 and 3.5.2 are orthogonal. In each case, assuming that there exists an
expansion of the form,
n=∞
X
f (x) = cn φn (x), (3.25)
n=0

determine an expression for the coefficients cn , n = 0, 1, . . .

Whenever we express a function in terms of an infinite series such as (3.25)
we must always consider whether the series representation is valid, i.e. does
the series converge and if so does it converge to the function we require. If
the function is continuous then there are few problems, however, even when
the function is discontinuous it is sometimes possible to determine a suitable
expansion. We will consider this problem in the next section.

3.6 Fourier series expansions

The idea of orthogonal functions was briefly introduced in the last section,
we can demonstrate how useful this notion can be. Consider the following
expansion in terms of the orthogonal functions cos(nπx), i.e.
1 1
f (x) = cos(2πx) + cos(6πx) + 25
9
cos(10πx) + . . .
1
+ cos(2(2n − 1)πx) + . . . . (3.26)
(2n − 1)2
82 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

For the moment let us ignore the problem of whether this infinite series
converges and try instead to plot it. Of course we cannot add in all the
terms but for the moment let us see what happens if we include the first
four, five or perhaps six terms. Now we have a finite expression which we
can plot. The result is shown in figure 3.15.

1.0

0.5

-1.0 -0.5 0.5 1.0

-0.5

-1.0

Figure 3.15: The cosine series (3.6.1)

At first sight it appears that as we include more and more terms the sum
of this series of continuous functions appears to be a function which is not
differentiable at various points. Conversely, we seem to be able to write the
sawtooth waveform as a sum of eminently smooth functions. We can now
ask whether this is always the case. However, before we answer this question
let us take a closer look at the graph drawn in figure 3.15, in particular near
the peaks. By plotting in more detail we see that the graph in fact does
not change direction sharply but does so in a smooth way (see figure 3.16).
However as we include more and more terms the magnification required to see
this smooth change becomes greater and greater and to all practical purposes
the sawtooth wave is generated.
Let us now turn to a general function f (x), |x| ≤ 1, and examine the
conditions under which it is possible to write this function as an expansion
of the form
a0 n=∞
X
f (x) = + (an cos(nπx) + bn sin(nπx)) , |x| ≤ 1, (3.27)
2 n=1

where the ′ denotes that the first coefficient is given by 12 a0 . The expan-
sion (3.27) is called the Fourier series expansion of f (x). If we are given a
3.6. FOURIER SERIES EXPANSIONS 83

-0.95

-1.00

-1.05

-1.10

-1.15

-1.20

0.40 0.45 0.50 0.55 0.60

Figure 3.16: Enlargement of figure 3.6.1 near x = 0.5, y = −1.23

function f (x) for a ≤ x ≤ b then we can always map the function to the
interval |x| ≤ 1 by a suitable linear transformation. Alternatively, if the
function is defined on an infinite interval then since the sine and cosine func-
tions are periodic, i.e. sin(x + 2π) = sin(x) and cos(x + 2π) = cos(x), we
shall require that the function f (x) is also periodic. Therefore, without loss
of generality we shall only consider functions which are defined for |x| ≤ 1.
Next we consider the notion of sectional continuity.
Definition 3.6.1. Let f (x) be a function which is continuous at all points
of a finite interval a ≤ x ≤ b except possibly at a set of points a ≤ x1 <
. . . < xp ≤ b. If f has a finite limit as x approaches the ends of each interval
(xi , xi+1 ), i = 1, . . . , p − 1, from its interior then f is said to be sectionally
continuous.
Definition 3.6.2. A function is said to have a right- (left-) hand derivative at
a point if
f (x + h) − f (x)
lim
h→0 h
exists, where h tends to zero through negative (positive) values.

Example 3.6.1
The functions shown in figure 3.17 are simple examples of sectionally contin-
uous functions.

Theorem 3.6.1. Let f be a sectionally continuous function on the interval

84 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

1.5

1.0

0.5

-1.5 -1.0 -0.5 0.5 1.0 1.5

-0.5

-1.0

-1.5

Figure 3.17: Sectionally continuous functions

(−1, 1) then the Fourier series

n=∞
1
X
a
2 0
+ (an cos(nπx) + bn sin(nπx)),
n=1

where
Z 1
an = f (x) cos(nπx) dx, n = 0, 1, 2, . . .
−1
Z 1
bn = f (x) sin(nπx) dx, n = 1, 2, . . .
−1

converges to the value

1
2
[f (x + 0) + f (x − 0)]
at every point where f has a right- and left-hand derivative. (By f (x ± 0) we
shall mean the limit of f at x through values which approach x from above
or below.)
Example 3.6.2 (Computational)
Load the program APPROX and select the Method: Fourier. We will
now investigate the Fourier series which correspond to the function
(
1, −1 < x < 0,
f (x) =
0, 0 < x < 1.

The function editor can be used to input this function piece by piece. To
return to the function editor press ESCAPE and then when the prompt
3.6. FOURIER SERIES EXPANSIONS 85

shows Action: Continue press RETURN which should produce the option
Function: as set. (Notice that the current function and range is shown in
the bottom window.) By pressing any key except RETURN cycle through
the options available at this time until it reads Function: inspect/edit and
then press RETURN again. This will display another prompt which you will
need to alter to Change f(x) before pressing RETURN. A flashing cursor
will appear in the bottom window and we can type in the value 1 which
now becomes the current function. The next stage is to split the range using
the prompt Split section, recall that we need to split the range at x = 0.
The bottom window will now show that we have the function f (x) = 1
for negative values of x as the first of two sections. Next we shall amend
the second section using the prompt Next section and then cycling until
the prompt reads Change f(x), in which case we can set the function to
f (x) = 0. Finally, we can plot the current function using the option Plot
function. (There are several other options available at this level which
can be used to move between the various sections of the function, edit or
delete sections, or add other sections if required.) At this point the required
function should be plotted; notice in particular the form of the curve near the
discontinuous point x = 0. This is caused by using only a finite number of
points when plotting. Now quit the function editor and return to the prompt
Method: Fourier and then press RETURN again. Again at this point
there are various options available which include supplying the coefficients
manually, reading them from file, storing the current coefficients in a file
and calculating the coefficients corresponding to the current function. This
last option requires the evaluation of equation (3.27) and in general this is
carried out numerically. Therefore, accept the option calculate coeffs and
then set the upper limit on the number of terms to four. After a short pause,
whilst the coefficients are being calculated, the option will return to Option:
calculate coeffs. Again cycle through the possibilities available at this point
and plot the function, approximation and error. Increase the number of terms
included in the series and compare the approximations which are determined.
Notice that each approximation takes the value 21 at the origin which agrees
with the value given by theorem 3.6.1. How do you account for the behaviour
of the Fourier series at ±1?

Exercise 3.6.1 (Analytical) Determine analytically the Fourier coefficients for

the function given in example 3.6.2 and compare them with those determined
by the program APPROX.
86 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS
q
Exercise 3.6.2 (Analytical) Show that the function f (x) = |x|, |x| < 1 does
not have a one-sided derivative at the point x = 0, but that this function
does have a convergent Fourier series on the interval |x| < 1 which is valid at
x = 0. This demonstrates that the existence of one-sided derivatives is not
a necessary condition for the existence and convergence of Fourier series.
Exercise 3.6.3 (Computational/Analytical) Use APPROX to determine the
Fourier series expansion of exp(x), |x| ≤ 1. Compare the result with cor-
responding polynomial approximations. Investigate the behaviour of the
Fourier series which is obtained for values |x| > 1.
Exercise 3.6.4 (Computational) Determine the Fourier series associated with
the function | sin(πx)|, |x| ≤ 1.
Exercise 3.6.5 (Computational) Load the program APPROX and then input
the file RAMP. Plot the function which is specified by this file and then
determine its Fourier series. Compare the coefficients obtained with the exact
analytical results. Compare the function and its Fourier approximation as
the number of harmonics increases.

3.7 Rational approximation

In previous sections we have seen how it is possible to determine approx-
imations for continuous functions in terms of polynomial expansions and
sectionally continuous periodic functions in terms of Fourier series. How-
ever, if we are given a function which has an apparent singularity neither of
these methods can give a satisfactory approximation. Consider the following
example.
Example 3.7.1 (Computational)
Load the program APPROX and determine a cubic approximation for the
function f (x) = (x2 − 3x + 2)−1 , |x| ≤ 1.5. Using the interpolation points
−1.5, −1, 0, 1.5 produces the cubic shown in figure 3.18. The poor ac-
curacy is due to the behaviour of f near x = 1. This demonstrates that
attempting to approximate a function which has a singularity by continuous
functions cannot lead to a satisfactory result. Let us reconsider the function
f (x) = (x2 − 3x + 2)−1 and set F (x) = (x − 1)f (x). If we now determine a
polynomial approximation for F (x) then we obtain a very satisfactory min-
imax approximation. This suggests that the correct way of approximating
3.7. RATIONAL APPROXIMATION 87

the function f (x) is to seek approximations of the form

Pn (x)
f (x) ≈ ,
Qm (x)
where Pn (x) and Qm (x) are polynomials of degree n and m respectively.
Clearly, we could multiply Pn and Qm by any factor and so by convention we
assume that the coefficient of xm in Qm is 1. The resulting approximation is
denoted by Rnm (x).

2 f HxL

-1.5 -1.0 -0.5 0.5 1.0 1.5

Cubic approximation
-2

-4

-6

Figure 3.18: Cubic approximation for (x2 − 3x + 2)−1 , |x| ≤ 1.5

Example 3.7.2 (Analytical)

In order to illustrate rational approximation let us determine R1,1 (x) for the
function exp(x), |x| ≤ 1. Let the rational approximation R1,1 be
a0 + a1 x
R1,1 (x) =
b0 + x
then we need to determine values for a0 , a1 and b0 . As with polynomial
interpolation we can only make R1,1 and exp(x) agree at a finite set of points
and since we have three coefficients to select we ensure that
exp(xi ) = R1,1 (xi ), i = 1, 2, 3.
For example, let us use the tabulation points x1 = −1, x2 = 0 and x3 = 1;
then
at x = −1 : (b0 − 1) exp(−1) = a0 − a1 ,
at x = 0 : b0 = a0 ,
at x = 1 : (b0 + 1) exp(1) = a0 + a1 .
88 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

Solving these equations produces a1 = −1, a0 = b0 = −2.1639534. This gives

the following rational approximation for exp(x):
−2.1639534 − x
R1,1 (x) =
−2.1639534 + x
which is shown in figure 3.19. Notice that this function is singular at x =
−2.1639534 but this is outside the region where the approximation is re-
quired. The error function corresponding to this approximation, e1,1 (x) =
exp(x) − R1,1 (x), is shown in figure 3.20 and has an error bound of 0.0575.
This error bound can be compared with the linear minimax error bound of
0.2788 and cubic minimax error bound of 0.0552.
3.0

2.5

2.0

1.5
Rational approximation

1.0

0.5

-1.0 -0.5 0.0 0.5 1.0

Figure 3.19: Rational approximation R1,1 (x), for exp(x), |x| ≤ 1

In general if we seek a rational approximation, Rnm (x), for a given func-

tion f (x) then we can match the function and the approximation at n+m+1
points since this is the number of coefficients to be determined, i.e. we select
points xi , i = 0, . . . , (n + m), and then solve
Qm (xi )f (xi ) = Pn (xi ), i = 0, . . . , (n + m).
Since the coefficient of xm in Qm is unity we can rewrite these equations as
j=n j=m−1
aj xji − bj xji f (xi ) = xm
X X
i f (xi ), i = 0, . . . , (n + m). (3.28)
j=0 j=0

This gives a system of (n + m + 1) linear equations in the unknown coeffi-

cients which can be solved by Guassian elimination with partial pivoting, if
necessary.
3.7. RATIONAL APPROXIMATION 89

0.01

-1.0 -0.5 0.5 1.0

-0.01

-0.02

-0.03

-0.04

-0.05

Figure 3.20: The error R1,1 (x) − exp(x), |x| ≤ 1

Exercise 3.7.1 (Computational) Use the program APPROX with the option
Method: Rational to determine the R2,2 (x) approximation for exp(x),
|x| ≤ 1, at equally spaced points. Compare the results obtained by using
other points. Also compare the results with R3,1 and R1,3 approximations
for this function.
Exercise 3.7.2 (Computational) Use the program APPROX to construct
rational approximations, Rmn (x), for exp(x), |x| ≤ 2, for which m and n are
less than 4. Compare them with the cubic minimax approximation.
As with polynomial approximation of functions based on interpolation
at a set of points, rational approximation can suffer from induced ill condi-
tioning. This can be overcome to some extent by using rational Chebyshev
approximations and further details are given by Fox and Parker (1972). How-
ever, there are other problems which can occur.
(i) The equations (3.28) may turn out to be singular.
Example 3.7.3 (Analytical)
Let us consider a rational approximation, R1,1 (x), for the cubic function
f (x) = x3 − 3x2 + 4x,
which is determined by interpolation at x = 0, x = 1 and x = 2. The
equations (3.28) become
    
1 0 0 a0 0
 1 1 −2   a1  =  2 
    

1 2 −4 b0 8
90 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

and these equations are inconsistent.

(ii) In example 3.7.1 we saw that the rational approximation R1,1 for
exp(x), |x| ≤ 1, determined using interpolation at −1,0 and 1 is
−2.1639534 − x
R1,1 (x) =
−2.1639534 + x
which has a singularity at x = 2.1639534. Fortunately, this singularity is
outside the region where the approximation is required, but this may not
always be the case.
Example 3.7.4 (Analytical)
Consider the function

f (x) = −3 + 20x − 8x2 , 0 ≤ x ≤ 2.

Using the tabulation points 0, 1 and 2 produces the rational approximation

1.5 + 3x
R1,1 =
−0.5 + x
1
which has a singularity at x = 2
where the original function is continuous.
(iii) As Pn and Qm are polynomials they may have a common factor which
can be cancelled out.
Example 3.7.5 (Analytical)
Consider the cubic

f (x) = 0.5x3 − 4x2 + 10.5x − 6

for which
f (1) = 1, f (2) = 3, f (3) = 3, f (4) = 4.
The rational function R2,1 determined from equation (3.28) is

x2 − 2x x(x − 2)
R2,1 (x) = =
x−2 x−2
= x, if x 6= 2.

In order that the resulting function is continuous at x = 2 we need to define

R2,1 (2) = 2 which is inconsistent with the data provided.
3.7. RATIONAL APPROXIMATION 91

We have seen that it is posible that rational approximation may introduce

singularities, but this possibility can be advantageous as the following simple
exercise demonstrates.
Exercise 3.7.3 (Computational) The function f (x) = exp(x)/(1 − 2 cos(x))
has a singularity at x = 13 π. Construct a rational approximation for f (x) by
considering the function (x − 31 π)f (x).

3.7.1 Rational minimax approximation

Clearly, given a function f (x) there will be many rational approximations
Rnm . We can now ask is it possible to find one rational approximation
amongst this class which minimises the maximum error. It is possible to ob-
tain results analogous to the Chebyshev equi-oscillation theorem and also de
la Vallee Poussin but we refer the interested reader to Ralston (1965). Simi-
larly, there is a procedure which corresponds to the Remes algorithm which
determines the rational minimax approximation iteratively and is usually
called Maehly’s algorithm. This requires the solution of non-linear equations
at each step. Unfortunately, due to the non-linearity of Maehly’s algorithm
there may not be a unique solution, but, it is possible to show that there is
only one solution which does not introduce additional singularities into the
required rational approximation. The determination of this unique rational
approximation is very difficult and beyond the scope of this text.
Execise 3.7.4 (Computational) Use the program APPROX to contruct a
rational approximation for exp(x), |x| ≤ 1, by interpolation at five equally
spaced points. Examine the error between the computed approximation and
the function and adjust the interpolation points to try to reduce the max-
imum error. Recompute a new rational approximation based on the new
interpolation points and examine the error function which is produced.

3.7.2 Padé approximation

Just as it is possible to consider a truncated Taylor series of a continuous
function as an approximation, the equivalent expression in terms of a rational
expansion is called a Padé approximation. To illustrate this consider the
following example.
92 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

Example 3.7.6
We shall construct a Padé approximation for exp(x) which is of the form
R1,1 , i.e.
a0 + a1 x
exp(x) ≈ ,
b0 + x
where the coefficients a0 , a1 and b0 are selected so that the function and
the approximation have as many derivatives as possible equal when x = 0.
Firstly, we re-arrange this expression to give
(b0 + x) exp(x) = a0 + a1 x (3.29)
then when x = 0, i.e.
b0 exp(0) = b0 = a0 . (3.30)
Next we differentiate equation (3.29) to obtain
(b0 + x) exp(x) + exp(x) = a1
and then set x = 0 to produce
b0 + 1 = a1 . (3.31)
As we have three parameters at our disposal we can repeat this procedure
once more to give
(b0 + x) exp(x) + exp(x) + exp(x) = 0
at x = 0 which gives b0 = −2. Substituting back into equation (3.31) gives
a1 = −1 and into equation (3.30) to get a0 = −2. Therefore, we have the
Padé approximation
−2 − x
exp(x) ≈ .
−2 + x
This approximation can be compared with the rational approximation ob-
tained by interpolation at x = −1, x = 0 and x = 1 in example 3.7.2 which
was
−2.1639534 − x
R1,1 (x) = .
−2.1639534 + x
The errors associated with these two rational approximations is shown in
figure 3.21.
Exercise 3.7.5 (Analytical and Computational) Compare the rational approx-
imation R1,3 , R2,2 and R3,1 determined by interpolation at equally spaced
points in (−1, 1) with the corresponding Padé approximation.
3.8. APPLICATIONS AND CONCLUSIONS 93

0.06

0.04

0.02

-1.0 -0.5 0.5 1.0

-0.02

-0.04

-0.06

-0.08

Figure 3.21: Padé and rational, R1,1 , errors for ex

3.8 Applications and conclusions

We now return to the problem posed at the beginning of this chapter; how
does a computer evaluate sin(1.4) or exp(−3.456)? Different manufacturers
have different methods but we shall illustrate the basic principles which are
used. In general no approximation will be valid for all values of the argument
and it is usual to define a principal range and then reduce the evaluation for a
particular argument to this range. For example, we could use the periodicity
of the trigonometric function to produce a principal range of (0, 2π), in fact
we can use other smaller principal ranges as we now demonstrate. One other
point to notice is the heavy emphasis placed on binary arithmetic which is
the basic form of computer arithmetic.
Example 3.8.1 (Analytical)
As a first example we shall illustrate how the function sin(x), x in radians, is
evaluated. (The actual implementation will differ from machine to machine
but the basic idea is common to all machines.)

1. If |x| ≤ 10−8 then return sin(x) ≈ x.

2. Reduce the argument by periodicity so that |x| ≤ 21 π and set t = 2x/π;

then
n=∞
An Tn∗ (t2 ),
X
sin(x) = t
n=0
94 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

n=∞
Bn Tn∗ (t2 ),
X
cos(x) =
n=0

where Tn∗ is the nth modified Chebyshev polynomial and the coefficients
An and Bn are given in table 3.2.

Table 3.2: Modified Chebyshev coefficients for sin(x) and cos(x)

n An Bn
0 1.276278962 0.472001216
1 −0.285261569 −0.499403258
2 0.009118016 0.027992080
3 −0.000136587 −0.000596695
4 0.000001185 0.000006704
5 −0.000000007 −0.000000047

Accuracy. For |x| ≤ 12 π the absolute error in the evaluation of sin(x) does
not exceed 0.25 × 10−8 and for cos(x) it does not exceed 0.42 × 10−7 .
Example 3.8.2 (Analytical)
Evaluate sin(1.4). The required argument lies within (− 12 π, 21 π) therefore we
set t = (2 × 1.4)/π ≈ 0.891267681. Next we use the recurrence relations

T0∗ (t2 ) = 1, T1∗ (t2 ) = 2t2 − 1,

∗
Tn+1 (t2 ) = 2(2t2 − 1)Tn∗ (t2 ) − Tn−1
∗
(t2 ).

Then
T0∗ (t2 ) = 1, T1∗ (t2 ) = 0.588716160,
T2∗ (t2 ) = −0.306826567, T3∗ (t2 ) = −0.949983676,
T4∗ (t2 ) = −0.811714915, T5∗ (t2 ) = −0.005755699.
from which we obtain sin(1.4) ≈ 0.985449729 which has absolute error 0.78×
10−9 . Similarly, cos(1.4) ≈ 0.169967143 which has absolute error 0.36×10−10 .
Exercise 3.8.1 (Calculator) Evaluate sin(1.0) using the algorithm described
in example 3.8.1.
Example 3.8.3 (Analytical)
The evaluation of exp(x).
3.8. APPLICATIONS AND CONCLUSIONS 95

1. If x < −180 return a value 0.

2. Divide x by loge (2) and write

 
1
x = 4a − b − 16
c loge (2) − d,

where a, b and c are integers with 0 ≤ b ≤ 3, 0 ≤ c ≤ 15 and 0 ≤ d <

1
16
loge (2), then

exp(x) = 16a 2−b 2−c/16 exp(−d).

3. Compute exp(−d) using a minimax approximation polynomial approx-

imation of degree 6.

4. Multiply exp(−d) by 2−c/16 then halve the result b times.

5. Multiply by 16a by adding the hexadecimal exponent a to the exponent

of the previous operation to give exp(x).

Accuracy. If −1 < x < 1 then the maximum relative error in exp(x) is less
than 0.21 × 10−15 , otherwise it is less than 0.43 × 10−15 .
Example 3.8.4 (Analytical)
The natural logarithm is evaluated using a rational approximation.

1. If x = 1 return loge (x) = 0.

2. Determine an integer n such√ that 2n−1 ≤√x ≤ 2n and then put r = 2−n x
( 21 ≤ r < 1). Set u = (r − 12 2)/(r + 12 2) and find
1+u
 
v = loge
1−u
20790 − 21545.27u2 + 4223.9187u4
!
≈ u
10395 − 14237.635u2 + 4778.8377u4 − 230.41913u6
then  
1
loge (x) ≈ n − 2
loge 2 + v.

Accuracy. The maximum absolute error in loge (x) is less than 1.02 × 10−10 .
Exercise 3.8.2 (Calculator/Computational) Use the algorithms given in ex-
amples 3.8.3 and 3.8.4 to evaluate exp(3.456) and loge (3.456).
96 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS

Not all standard functions use polynomial approximations. For example,

the evaluation of the square root function uses an iterative approach.

Example 3.8.5 (Analytical)

√
The evaluation of x.

1. If x = 0 then return 0.

1
2. Let x = 162p−q f , where p is an integer, q = 0 or 1 and 16
< f < 1 then

√ q
x = 16p 4−q f.

√
3. Take as a first approximation for f

(0.14605 + 1.8342f )
y0 = 16p 4−q .
(0.98026 + f )

4. Apply the Newton–Raphson iteration

!
1 x
yn+1 = 2
yn +
yn

twice. The final iteration is taken as y2 = 12 (y1 − x

y1
) + x
y1
in order to
minimize computational errors.

Accuracy. The maximum relative error is 0.48 × 10−6 .

Example 3.8.6 (Calculator)
√
Let us determine 3. We can write 3 = 162−1 0.1875, therefore p = 1,
q = 1 and f = 0.1875. Using these values we obtain y0 = 1.67829862,
y1 = 1.73291159, y2 = 1.73205102 and y3 = 1.73205081. This last value is
correct to eight decimal places.

Exercise 3.8.3
√ (Calculator) Use the algorithm outlined in example 3.8.5 to
determine 1.76. Square the result and hence estimate the error. Compare
the error with the bound which is given.
3.8. APPLICATIONS AND CONCLUSIONS 97

In this chapter we have seen how it is possible to construct both polyno-

mial and rational approximations for a given function. It is difficult to say
which type of approximation will in general be the best; this will depend on
the accuracy required and the frequency with which the approximation is go-
ing to be used. It is arguable that in most cases a Chebyshev approximation
will be sufficient for most purposes. The determination of a minimax approx-
imation requires considerable effort, but once determined only the coefficients
in the relevant polynomial need to be stored and evaluation is particularly
easy.
98 CHAPTER 3. THE APPROXIMATION OF FUNCTIONS
Chapter 4

Least Squares Approximation

Of course the first thing to do was to make a grand survey of the country she
was going to travel through. “It’s something like learning geography”, thought
Alice, as she stood on tiptoe in hopes of being able to see a little farther.
Lewis Carroll.

4.1 Introduction
In Chapter 2 the problem of fitting a function through a set of data points was
considered. It was shown that as the number of points increases then poly-
nomial approximation becomes more oscillatory. In order to overcome this
problem the concept of spline interpolation was introduced whereby lower
order polynomials were employed. In this chapter we shall consider an alter-
native approach; to do this consider the following example.
Example 4.1.1 (Analytical)
Newton’s law of cooling states that the rate at which a body loses heat is
proportional to the difference between the temperature of the body and the
surrounding ambient temperature T0 . Let T be the temperature at time t;
then we can write
dT
= −K(T − T0 ),
dt
where K is some constant, which can be integrated to give

T (t) = T0 + A exp(−Kt). (4.1)

99
100 CHAPTER 4. LEAST SQUARES APPROXIMATION

In an experiment the ambient temperature was found to be 10 ◦C, and the

data in table 4.1 were observed; how can we find the most suitable value for
the constant K? The values of T − 10 are plotted against t in figure 4.1, as
are the curves T0 + A exp(−Kt) for an assortment of A and K. The problem
of minimizing the error between the data points and any one member of this
family of curves is quite difficult.

Table 4.1: Data obtained for example 4.1.1

t (min) T ◦ C loge (T − 10)
1 65 4.007333
2 48 3.637586
3 30 2.995732

100

80

60 A=95, k=0.4

A=95, k=0.6 A=95, k=0.5

40

A=90, k=0.5

20

0 1 2 3

Figure 4.1: Values from table 4.1.1 and 10 + A exp(−Kt)

As an alternative we can rewrite equation (4.1) as

loge (T − T0 ) = B − Kt, B = loge A,
and plot the new variable y = loge (T − T0 ) against t. We then have a linear
relationship between y and t, and the slope of this line is related to the value
K we require. For a particular choice of B and K it is unlikely that the
selected line will pass through all the data points and so we write the error
at a point ti as
ei = y(ti ) − yi, (4.2)
4.1. INTRODUCTION 101

0 1 2 3

Figure 4.2: Plot of t against loge (T − 10) for table 4.1.1

where y(ti) is the value determined from the particular choice of the param-
eters and yi is the observed value at ti . In order to obtain the ‘best possible’
fit we could attempt to select the parameters B and K to minimize the sum
of squares, S, of the errors. In this case

S = (B−K −4.007333)2 +(B−2K −3.637586)2+(B−3K −2.995732)2. (4.3)

This function has a turning point where ∂S/∂B = ∂S/∂K = 0 which gives
two linear equations for K and B. Solving these equations produces the least
squares approximation for the given data as

loge (T − T0 ) = 4.558485 − 0.5058005t, (4.4)

or equivalently

T (t) = 10 + 95.43878 exp(−0.5058005t). (4.5)

The corresponding sum of squares, S, given by equation (4.3) is then 1.234 ×

10−2 . Notice that (4.4) is a linear least squares fit for the function loge (T −T0 )
but that (4.5) is not necessarily a non-linear least squares fit for T . We shall
discuss linear least squares in some detail; non-linear least squares is far more
difficult. A detailed discussion is given by Stoer and Burlisch (1980).
102 CHAPTER 4. LEAST SQUARES APPROXIMATION

4.2 Linear least squares

In the last section we used a simple demonstration which fitted a straight
line by minimizing the sum of squares of the associated errors. We will now
extend this approach to fitting a linear function to a general set of values
and then generalize the method to fit a linear combination of functions.
Let us suppose that we have been given a set of tabular values (xj , yj ),
j = 1, . . . , m, and that we wish to find a linear function,

y = c0 + c1 x, (4.6)

which minimizes some overall measure of the difference between the tabular
values yj and the corresponding y(xj ) as provided by the approximating
function. There is no unique way of providing a measure of the discrepancy
between yj and y(xj ). For example, we could use the sum of the magnitudes
of the errors to compare different values of c0 and c1 , i.e.
j=m
X j=m
X
|y(xj ) − yj | = |c0 + c1 xj − yj |. (4.7)
j=0 j=0

It is possible to determine values of c0 and c1 which minimize this functional

but it requires linear programming techniques beyond the scope of this text.
(If such an approximation is found it is called an l 1 approximation.) However,
in practice it is easier to minimize the sum of squares of the errors, i.e.
j=m
(c0 + c1 xj − yj )2 .
X
S(c0 , c1 ) = (4.8)
j=0

This is called the least squares fit or an l2 approximation. We could select

values for c0 and c1 at random and compute the value of S(c0 , c1 ) but a much
simpler way is to use the calculus.
Exercise 4.2.1 (Computational) Load the program LSQ and change the
prompt to Select data. Now change the current option to Data: from
keys and clear the existing data. Select the prompt Data input format:
tabular and input the entries in the first and third columns of table 4.1.
(Recall that loge (65 − 10) can be input as LN(55).) The data points can be
plotted on suitable axes by using the option Data: plot&/or edit which
should produce a screen display like figure 4.2. (Recall that these data can be
stored in a file using the option Data: store on file but you will need a disc
4.2. LINEAR LEAST SQUARES 103

of your own which is not write protected.) Exit from this set of options by
using the prompt Data: quit and when the option shows Method: Linear
press RETURN. Select the prompt Option: interactive fitting and a de-
fault line will appear on the screen together with the current sum of squares
of deviations (SSD) as given by equation (4.8). The position and slope of the
current line can be changed by using the cursor keys.
1. ↑ ↓ shift the current line with respect to the axis y = 0.

2. ← → rotate the current line about the middle of the screen.

3. < > increase or decrease the increment by which the line is changed.
As the line is moved the sum of squares of deviations (SSD) is changed
and displayed in the top window. It is also possible to examine the current
values of the coefficients in equation (4.8), c0 and c1 , by selecting the prompt
Option: edit/inspect coeff. Try to adjust the line to minimize the sum
of squares of deviations.
Exercise 4.2.1 illustrates some of the problems in finding a least squares
approximation by trial and error; therefore, we will now consider a more di-
rect approach. The functional S(c0 , c1 ), given by equation (4.8), has turning
points whenever
∂S ∂S
= = 0, (4.9)
∂c1 ∂c0
therefore, necessary conditions for a minimum are
j=m
X
xj (c0 + c1 xj − yj ) = 0,
j=0
j=m
X
(c0 + c1 xj − yj ) = 0.
j=0

Finally, these equations can be rewritten in the form

j=m j=m j=m
x2j =
X X X
c0 xj + c1 xj yj ,
j=0 j=0 j=0
(4.10)
j=m
X j=m
X
(m + 1) c0 + c1 xj = yj .
j=0 j=0
104 CHAPTER 4. LEAST SQUARES APPROXIMATION

For the example in section 4.1 these equations reduce to the linear equations
! ! !
6 14 c0 20.269701
= .
3 6 c1 10.640651

The solution of these equations is c0 = 4.558485 and c1 = −0.5058005 as

in section 4.1. These coefficients produce the fitted linear function y =
4.558485 − 0.5058005t.
Exercise 4.2.2 (Computational) Load the program LSQ and select the prompts
Method: Linear and Option: calculate coeffs. The program will now
determine the least squares linear fit by using equation (4.11) and then re-
turn to this prompt. It is possible to display the fitted function by selecting
the prompt Option: plot and examine the computed coefficients using the
prompt Option: edit/inspect coeffs. Compare the computed least squares
fit with the results of exercise 4.2.1.
Exercise 4.2.3 (Computational) Repeat exercise 4.2.2 but investigate the sen-
sitivity of the computed line with respect to small preturbations of the data
points. Is this problem well conditioned?
We could generalize the above argument to consider functions of the form

y(x) = c0 + c1 x + c2 x2 + . . . + cn xn

but these can in turn be considered as a special case of

i=n
X
y(x) = c0 φ0 + c1 φ1 + . . . + cn φn = ci φi (x), (4.11)
i=0

where φi (x), i = 0, . . . , n, are any suitable basis functions. It is clear that

n ≤ m but if n = m the problem reduces to constructing an interpolating
polynomial through the data points. As before we attempt to determine
coefficients ci , i = 0, . . . , n, in order to minimize the functional
j=m
(yj − y(xj ))2 .
X
S(c0 , c1 , . . . , cn ) = (4.12)
j=0

At a turning point we shall have

∂S
= 0, i = 0, . . . , n, (4.13)
∂ci
4.2. LINEAR LEAST SQUARES 105

which produces a system of n+1 linear equations in the unknown coefficients

ci , i = 0, . . . , n. The equations (4.12) may be rewritten as
j=m i=n
!2
X X
S(c0 , . . . , cn ) = yj − ci φi (xj ) . (4.14)
j=0 i=0

Therefore, differentiating with respect to ck produces

j=m
X i=n
X
!
φk (xj ) yj − ci φi (xj ) = 0, k = 0, . . . , n. (4.15)
j=0 i=0

which can in turn be re-arranged to give

 
i=n
X j=m
X j=m
X
 φk (xj )φi (xj ) ci = φk (xj )yj , k = 0, . . . , n. (4.16)
i=0 j=0 j=0

Equivalently, we may rewrite these equations as

Ac = b,
where the components of the matrix A are given by
j=m
X
aki = φk (xj )φi (xj )
j=0

j=m
and those of the vector b by bk = j=0 φk (xj )yj . Notice that A is indepen-
P

dent of the values yj , j = 0, . . . , m. The system of linear equations (4.16) are

called normal equations and can be solved by Gaussian elimination without
the need to partially pivot since the matrix of coefficients is symmetric and
positive definite.
Exercise 4.2.4 (Computational) Load the program LSQ for which the default
data points are given in table 2.1. When the prompt shows Select method
press RETURN and then use the option Method: Linear to determine
a linear least squares approximation for these data points using the option
interactive fitting and then plot it. Determine the least squares fit by
using the option calculate coeffs and compare the computed results. Now
select the option Method: Polynomial and fit polynomials of degree 2, 3
and 4. (Why is it not possible to fit higher order polynomials to these data
points?) Modify the data points using the option edit/inspect points and
observe the way the fitted function changes.
106 CHAPTER 4. LEAST SQUARES APPROXIMATION

1.5
Linear

Quadratic
1.0

0.5

-0.5 0.5 1.0 1.5

Quartic

Figure 4.3: Linear and polynomial fitting for table 2.1.1

Exercise 4.2.5 (Computational) Repeat exercise 4.2.4 but fitting

(i) Chebyshev polynomials of degree 2, 3 and 4.
(ii) Fourier series of order 1 and 2.
(iii) The functions 1, exp(x), exp(2x), . . ..
For (iii) what is the maximum number of such terms which can be used?
Exercise 4.2.6 (Computational) Repeat exercise 4.2.4 but try specifying any
suitable set of basis functions of your choice. (Recall that they must be
linearly independent.)
Exercise 4.2.7 (Analytical) Show that the functional, S, given in equation (4.12)
may be written in the matrix form
S(c0 , . . . , cn ) = (Nc − y)T (Nc − y),
where N is a rectangular matrix for which Nij = φi (xj ), y and c are vectors
whose components are yj , j = 0, 1, . . . , m and ci , i = 0, . . . , n, respectively.
By differentiating this form directly show that the normal equations become
NT Nc = NT y,
and deduce that these equations are identical to the equations (4.16). Fur-
thermore, show that the matrix N is symmetric and positive definite.
The matrix A in the normal equations (4.16) will in general be full but
if the basis functions φi (x), i = 0, . . . , n, satisfy
j=m
X
φr (xj )φs (xj ) = 0, if r 6= s (4.17)
j=0
4.3. CONTINUOUS LEAST SQUARES APPROXIMATION 107

then all the off-diagonal entries of A vanish. If this is the case then the
functions φi (x), i = 0, . . . , n, are said to be orthogonal with respect to the
points xj , j = 0, . . . , m. Under this condition the normal equations reduce
to a diagonal system and the coefficients of (4.11) are given by
Pj=m
φi (xj )yj
ci = Pj=0
j=m 2
, i = 0, . . . , n. (4.18)
j=0 φi (xj )

Given any set of linearly independent functions it is possible to generate an-

other set which satisfies (4.17) but we shall consider an alternative approach
in a later section.
Exercise 4.2.8 (Computational) Load the program LSQ and input the data
file SPLINE. Compare linear, quadratic and cubic least squares fits for these
data. Explain why the fitted function is so poor. Repeat this exercise for the
datafile SIN.

4.3 Continuous least squares approximation

In the last section we discussed the problem of fitting a function of the form

y(x) = c0 φ0 (x) + . . . + cn φn (x), (4.19)

to a set of discrete points. This idea can be generalised to consider the

approximation of a given function f (x), |x| ≤ 1, by expressions of this form.
In order to do this we consider the functional
i=n
!2
Z 1 X
S(c0 , . . . , cn ) = f (x) − ci φi (x) dx (4.20)
−1 i=0

as a measure of the error between the function and its approximation. It is

also possible to consider a weighted measure of the discrepancy by defining

i=n
!2
Z 1 X
S(c0 , . . . , cn ) = w(x) f (x) − ci φi (x) dx, (4.21)
−1 i=0

where w(x) is a suitable weight function. We have assumed, without loss of

generality, that the function f (x) is defined on the interval [−1, 1]; if not then
108 CHAPTER 4. LEAST SQUARES APPROXIMATION

we can always transform it to be so. As before we differentiate S(c0 , . . . , cn )

with respect to ck to produce the corresponding normal equations, i.e.
i=n
X Z 1  Z 1
w(x)φk (x)φi (x) dx ci = w(x)φk (x)f (x) dx, k = 0, . . . , n.
i=0 −1 −1
(4.22)
This produces a system of linear equations in the form Ac = b, where
Z 1
aki = w(x)φk (x)φi (x) dx
−1

and Z 1
bk = w(x)φk (x)f (x) dx.
−1

However, if the basis function φi (x), i = 0, . . . , n, are orthogonal with respect

to the weight function w(x), i.e.
Z 1
w(x)φi (x)φj (x) dx = 0, i 6= j,
−1

then the normal equations are particularly simple and the coefficients ci ,
i = 0, . . . , n, are given by
R1
−1 w(x)φi (x)f (x) dx
ci = R1
2
. (4.23)
−1 w(x)φi (x) dx

Example 4.3.1 (Analytical)

The Chebyshev polynomials form an orthogonal basis. (See section 3.4 and
example 3.5.1.) Hence, we can construct the Chebyshev least squares ap-
proximation for a function f (x), x ∈ [−1, 1] as follows. We consider an
approximation in the first (n + 1) Chebyshev polynomials given by
i=n
X
f (x) ≈ ci Ti (x).
i=0

Then from equation (4.23) the least squares coefficients are given by

1Z1 f (x) 2 Z 1 Ti (x)f (x)

c0 = q dx, ci = q dx.
π −1 (1 − x2 ) π −1 (1 − x2 )
4.4. DISCRETE LEAST SQUARES APPROXIMATION 109

Exercise 4.3.1 (Analytical) Show that for the coefficients given in example
4.3.1 the functional S(c0 , c1 , . . . , cn ) given in equation (4.21), has minimum
value
1 f 2 (x)
Z
q dx − 12 π(2c20 + c21 + . . . + c2n ).
−1 (1 − x2 )

Exercise 4.3.2 (Analytical) Show that the Fourier series approximation

a0 k=n
X
f (x) ≈ + (ak cos kx + bk sin kx)
2 k=1

with the Fourier coefficients defined by

1 π 1 π
Z Z
ak = f (x) cos kx dx, bk = f (x) sin kx dx,
π −π π −π

is a least squares approximation for f (x), with unit weight, in |x| ≤ π. (The
question of the convergence of the Fourier series as n → ∞ will not be
discussed here. For a full discussion see Churchill (1963).)

4.4 Discrete least squares approximation

In the last section we saw how it was possible to construct a continuous
least squares approximation for a given function, however, to do so it was
necessary to evaluate the coefficients by performing numerous integrations.
We will now return to consider how it is possible to construct a discrete
variant which avoids this problem.
In section 4.2 we saw that given a set of points (xj , yj ), j = 0, . . . , m, we
could construct a linear least squares fit, i=n
i=0 ci φi (x), by solving the normal
P

equations given by
 
i=n
X j=m
X j=m
X
 φk (xj )φi (xj ) ci = φk (xj )yj , k = 0, . . . , n, (4.24)
i=0 j=0 j=0

but if the basis functions are orthogonal with respect to the tabular points,
i.e.
j=n
X
φk (xj )φi (xj ) = 0, i 6= k, (4.25)
j=0
110 CHAPTER 4. LEAST SQUARES APPROXIMATION

then the coefficients in the least squares approximation are given by the
simple form
Pj=m
φi(xj )yj
ci = Pj=0
j=m 2
, i = 0, . . . , n. (4.26)
j=0 φi (xj )

Therefore, given any function f (x) we can construct the least squares ap-
proximation to it if (m + 1) points, (xj , f (xj )), j = 0, . . . , m, are known such
that equation (4.25) is satisfied. Conversely, if we have m + 1 points which
satisfy this condition then the interpolating polynomial constructed in this
way will be the least squares approximation.
In order to illustrate this approach consider the Chebyshev approximation
given by
i=n
X ′
f (x) ≈ ai Ti (x), −1 ≤ x ≤ 1, (4.27)
i=0
′
where the denotes the fact that the first coefficient is halved. Then if the
tabulation points xj , j = 0, . . . , m, are selected to be the zeros of Tm+1 , i.e.
!
(2j + 1) π
xj = cos , j = 0, . . . , m, (4.28)
(m + 1) m

it is possible to show1 that


j=m 
 0, r 6= s,
′′
1
X
Tr (xj )Ts (xj ) = 2
m, r = s 6= 0, (4.29)
m, r = s = 0,

j=0 

′′
where the denotes that both the first and last terms in the summation, for
any r or s, are to be halved. In this case the coefficients in equation (4.27)
are given by
2 j=m
X ′′
ai = f (xj )Ti (xj ). (4.30)
m j=0
This is the discrete Chebyshev least squares approximation. If n = m, i.e.
the number of tabulation points is equal to the number of coefficients, then
′ ′′
the in equation (4.27) is replaced by for an exact fit. Notice that the
tabulation points xj are not equally spaced.
1
See Cheney (1966).
4.4. DISCRETE LEAST SQUARES APPROXIMATION 111

Example 4.4.1 (Analytical)

Let us consider the function exp(x), x ∈ [−1, 1]. The Chebyshev series which
represents this function is given by

exp(x) = 1.266066T0 + 1.130318T1 + 0.271495T2 + 0.044337T3

+ 0.005474T4 + 0.000543T5 + . . . .

The Chebyshev discrete least squares approximation of degree 2 which is

constructed by using least squares at the zeros of T4 is given by

1.266066T0 + 1.130315T1 + 0.271450T2.

The maximum difference between these functions is 5.045 × 10−2 and the
sum of squares of deviations is 3.835 × 10−3 .
We can compare the coefficients of a Chebyshev series and those of the
discrete Chebyshev least squares as follows. Recall that the coeffcients of the
Chebyshev series, ci , are given by
2Zπ
ci = f (cos θ) cos iθ dθ
π 0
and if we apply the composite trapezium rule to this integral with m + 1
points we find

2 k=m
X ′′ kπ
ci ≈ f (cos θk ) cos iθk , θk = ,
m k=0 m
2 k=m
X ′′
= f (xk )Ti (xk ) = ai ,
m k=0

from (4.30), i.e. the Chebyshev series of f (x) is the Fourier series of f (cos(x)).
Exercise 4.4.1 (Computational) Load the program LSQ and change the
prompt to Select function. When the current option shows Function:
disabled change it to enabled and press RETURN. Now change the option
to Function: inspect/edit and change the function to ex , |x| ≤ 1. Quit
this section of the program and change the prompt to Select data. Use
the option Data: standard grid to select a uniform grid of four points and
then quit. The program should now return to the prompt Method: Linear;
change the option to Polynomial and fit a polynomial of degree 2. Com-
pare the result with a Chebyshev least squares fit of the same degree. What
112 CHAPTER 4. LEAST SQUARES APPROXIMATION

do you deduce? Increase the number of points and repeat the comparison.
Return to the option Select data and use the zeros of T4 (x) as data points,
construct a least squares fit for ex , |x| ≤ 1, based on these points and then
compare it with your results for equally spaced points.
Exercise 4.4.2 (Computational) Construct an interpolating polynomial ap-
proximation for ex , |x| ≤ 1, which passes through six uniformly spaced points.
(Use INTERP). Compare the result with least squares polynomial fits for
polynomials of degree 3, 4 and 5. Repeat this exercise using the zeros of
T6 (x). What do you conclude?
Exercise 4.4.3 (Computational) Construct the minimax cubic polynomial ap-
proximation for ex , |x| ≤ 1, using the Remes algorithm in APPROX. Com-
pare it with least squares Chebyshev approximations of order 3,4 and 5.
Estimate the relative efficiency of determining these approximations.
Chapter 5

Integration

‘Quid pro Quadrature?’

Although many functions are known to be integrable, that is to say they

have an indefinite integrand, there may be no expression for the integrand in
terms of elementary functions. Perhaps the simplest example of this is the
integral Z 1
exp(−x2 ) dx.
0
In this chapter we will see how it is possible to determine an estimate for
such integrals. To do this we first find an approximation for the integrand,
integrate the approximation analytically in order to estimate the integral and
then estimate the overall error. This process is called quadrature.

5.1 Simple quadrature

Let f (x) be a continuous function; then we may use its Taylor series expan-
sion about x = a to approximate it, i.e.
(x − a)2 ′′ (x − a)n (n)
f (x) = f (a)+(x−a)f ′ (a)+ f (a)+. . .+ f (a)+. . . . (5.1)
2 n!
Therefore,
#b
(x − a)2
"
Z b
f (x) dx = f (a) [x − a]ba ′
+ f (a) + ...
a 2 a
= (b − a)f (a) + 21 (b − a)2 f ′ (a) + . . . .

113
114 CHAPTER 5. INTEGRATION

By adding in sufficient terms, and providing (b − a) is sufficiently small, we

might hope to estimate the integral. If we were to replace the expansion (5.1)
by a Taylor series with remainder we might even be able to estimate the error
but as we have seen the Taylor series expansion is not the best approximation
available. As an alternative, for any set of points xj , j = 0, . . . , m, we can
find the Lagrangian polynomial approximation
j=n
X
pn (x) = Lj (x)f (xj ), (5.2)
j=0

which has error

f (n+1) (θ) j=n

Y
en (x) = f (x) − pn (x) = (x − xj ), (5.3)
(n + 1)! j=0

where θ depends on x and lies between x0 and xn . Therefore,

Z b Z b
f (x) dx = (pn (x) + en (x))dx (5.4)
a a
Z b j=n
X Z b
= Lj (x)f (xj ) dx + en (x) dx (5.5)
a j=0 a

j=n
X
= αj f (xj ) + En , (5.6)
j=0

where
Z b
αj = Lj (x) dx, (5.7)
a
1
Z b j=n
(x − xj ) f (n+1) (θ) dx.
Y
En = (5.8)
(n + 1)! a j=0

The approach now is to select the points xj in order to minimize the error En .
Notice that the quadrature formula (5.6) involves (2n + 2) parameters; the
interpolation points xj and the coefficients αj , j = 0, . . . , n, but that these
values are not independent. For any arbitrary choice of interpolation points,
usually called the abscissa, the coefficients, which are called the weights,
are determined from equation (5.7). As a result we should be able to make
the quadrature formula exact for polynomials of degree n since n + 1 points
5.1. SIMPLE QUADRATURE 115

determine a polynomial of degree n exactly. This is confirmed by the fact

that en is proportional to f (n+1) which vanishes for all polynomials of degree
less than n + 1. Quadrature formulae of this type are called Newton–Cotes
methods and will be considered in section 5.2. Alternatvely, we can generalize
this approach and begin with the formula

Z b j=n
X
f (x) dx = αj f (xj ) + En (5.9)
a j=0

and then select values for all the xj ’s and αj ’s to make this a method of order
higher than n + 1. Such formulae are called Gaussian quadrature methods
and they will be considered in section 5.3.

Example 5.1.1 (Computational)

The program QUAD is intended as a demonstration program for simple
quadrature formulae, whereas INTEGR is intended for practical numerical
quadrature. In order to illustrate how the choice and number of interpolation
points influences the accuracy of numerical quadrature load the program
QUAD for which the default problem is
Z 1
x5 dx.
0

Accept the prompts Function: as set and Select data by pressing RE-
TURN. The prompt will now change to Option: integrate. Notice that in
the graphical window three equally spaced points are placed on the default
integrand x5 . This is confirmed in the bottom window where N = 3 is dis-
played. Furthermore, as three points are available it is possible to construct
an interpolating polynomial of degree 2 which is confirmed by n = 2 in the
top window. Now press RETURN. The bottom screen will display the esti-
mate for the integral which has been determined as the Current Estimate
together with an estimate of the quadrature error as determined by approxi-
mating equation (5.8). It is possible to examine the interpolating polynomial
which has been used to approximate the integrand by changing the prompt
to Option: plot, pressing RETURN and then accepting the option Plot:
approximation. In addition, the difference between the integrand and the
approximating polynomial can be highlighted using the option Plot: shade
and the error can be displayed using the option Plot: error.
116 CHAPTER 5. INTEGRATION

Exercise 5.1.1 (Computational) Load the program QUAD with the default
problem
Z 1
x5 dx
0

and by changing the prompt Select quadrature to Select data investigate

how the number of points in a uniform mesh influences the accuracy of the
computed estimate for the integral. Why are six points sufficient to give an
exact answer?
Example 5.1.2 (Computational)
Load the program QUAD to compute an approximation for
Z 1
exp(−x2 ) dx
−1

using a uniform grid of five points1 . (Hint: Use the option Select function
to change the integrand.) Return to the option Select data and then select
Data: plot &/or edit and change the interpolation points to be at x =
0, ±0.5, ±1. Quit the data editor and select the prompt Option: integrate.
An estimate for the value of the integral will now be determined based on
these points together with an estimate of the quadrature error. Notice that
the previous estimate remains visible in order that you can compare it.
Exercise 5.1.2 (Computational) Repeat example 5.1.2 for other sets of five
interpolation points and try to determine the best possible choice, i.e. mini-
mize the computed error estimate.

5.2 Newton–Cotes methods

Let us now return to consider Newton–Cotes formulae which are based on
replacing the integrand by an interpolating polynomial. We begin by con-
sidering the simplest possible case where we replace the integrand, f (x), by
a linear approximation p1 (x), a ≤ x ≤ b, i.e.

b−x x−a
f (x) ≈ p1 (x) = f (a) + f (b), (5.10)
b−a b−a
1
Some computers treat the expression EXP(−X 2 ) as EXP((−X)2 ) whereas we require
EXP(−(X)2 ).
5.2. NEWTON–COTES METHODS 117

which is the Lagrange polynomial through (a, f (a)) and (b, f (b)). Therefore,
Z b Z b
f (x) dx = p1 (x) dx + E1
a a
f (a) Z b f (b) Z b
= (b − x) dx + (x − a) dx + E1
(b − a) a (b − a) a
= 12 (b − a)[f (a) + f (b)] + E1 .

This is the simple trapezium rule. We may use equation (5.3) to determine
the error, E1 , in this approximation as
b b f ′′ (θ(x)) f ′′ (θ) b
Z Z Z
e1 (x) dx = (x − a)(x − b) dx = (x − a)(x − b) dx
a a 2! 2! a
(5.11)
where the mean value theorem for integration has been used to simplify the
integral. Integrating the last term gives

(b − a)3 ′′
E1 = − f (θ), (5.12)
12
where a ≤ θ ≤ b. The accuracy of the trapezium rule is proportional to
the cube of the interval width, provided the integrand has a bounded second
derivative, and if we require higher accuracy then it is necessary to include
further interpolation points.
Exercise 5.2.1 (Analytical) Show that by sub-dividing the interval (a, b) into
two sub-intervals such that x0 = a, x1 = 21 (a + b), x2 = b and then replacing
the integrand by a Lagrangian polynomial approximation of degree 2 we
obtain
Z b Z x2
f (x) dx = p2 (x)dx + E2
a x0
1
= 6
(b − a) [f (x0 ) + 4f (x1 ) + f (x2 )] + E2 ,

which is Simpson’s rule. Show that

h5 iv
E2 = − f (θ), a ≤ θ ≤ b.
90

where h = 21 (b − a).
118 CHAPTER 5. INTEGRATION

The trapezium rule will integrate any polynomial of degree ≤ 1 exactly since
the error is proportional to f ′′ , whereas Simpson’s rule will integrate any
cubic without error. In general a quadrature formula is said to have degree
of precision p if it will integrate all polynomials of degree ≤ p without error.
The trapezium rule has degree of precision 1, Simpson’s rule has degree of
precision 3. By introducing further points we can increase the degree of
precision further.
Example 5.2.1
The Newton–Cotes closed formulae for the integral ab f (x) dx are given by
R

the following. Divide the interval (a, b) into n sub-intervals such that a = x0 ,
b = xn , xj = a + jh, j = 0, . . . , n, h = (b − a)/n then xx0n f (x) dx is given
R

in table 5.1. Notice that for n even the degree of precision is n + 1, but for
n odd it is n and this is generally the case (see Isaacson and Keller (1966)).
Accordingly, even formulae are usually preferred.

Table 5.1: Closed Newton–Cotes formulae

1
n=1 2
h[f (x0 )
+ f (x1 )]
1 3 ii
2 point The trapezium rule: p = 1 : E1 = − 12 h f (θ)

1
n=2 3
h[f (x0 )
+ 4f (x1 ) + f (x2 )]
3 point Simpson’s 13 rd rule: p = 3: E2 = − 90
1 5 iv
h f (θ)

3
n=3 8
h[f (x0 )
+ 3f (x1 ) + 3f (x2 ) + f (x3 )]
4 point Simpson’s 38 th rule: p = 3 : E3 = − 80 3 5 iv
h f (θ)

2
n=4 45
h[7f (x0 )
+ 32f (x1 ) + 12f (x2 ) + 32f (x3 ) + 7f (x4 )]
8
5 point Milne’s rule: p = 5 : E4 = − 945 h7 f vi (θ)

The formulae given in example 5.2.1 are termed closed since they use
the end points of the interval of integration; however, it is useful to define
formulae which do not have this property.
Example 5.2.2
The Newton–Cotes open formulae are given as follows. We divide the interval
(a, b) into n + 2 intervals so that a = x−1 and b = xn+1 , i.e. xj = a + (j + 1)h,
5.3. GAUSS QUADRATURE FORMULAE 119

Table 5.2: Open Newton–Cotes formulae

n=0 2hf (x0 )

1 point The midpoint rule: p = 1 : E0 = 31 h3 f ′′ (θ)

3
n=1 2
h[f (x0 ) + f (x1 )]
2 point p = 1 : E1 = 43 h3 f ′′ (θ)

4
n=2 3
h[2f (x0 ) − f (x1 ) + 2f (x2 )]
14 5 iv
3 point p = 3 : E2 = 45 h f (θ)

jR = 0, . . . , n, with h = (b − a)/(n + 2); then open quadrature methods for

b
a f (x) dx are given in table 5.2.
Example 5.2.3
Consider the integral 01 x−1/2 dx. It is not possible to apply the closed
R

Newton–Cotes formulae to this integral even though the integral is well de-
fined because the integrand is undefined at x = 0. However, there is no such
difficulty in applying open formulae which do not require the integrand at
x = 0.
As the number of points increases we obtain formulae with higher preci-
sion but unfortunately the resulting methods become unsuitable. For exam-
ple, the nine-point closed Newton–Cotes formula is given by
Z x9
4h
f (x) dx ≈ 14175 [989f (x0 ) + 5888f (x1 ) − 928f (x2 ) + 10496f (x3 )
x0
−4540f (x4 ) + 10496f (x5 ) − 928f (x6 ) + 5888f (x7 ) + 989f (x8 )].
This formula has large coefficients which alternate in sign and this can lead
to large cancellation errors unless special care is taken.

5.3 Gauss quadrature formulae

We shall now attempt to find quadrature formulae of the form
Z 1 j=n
X
f (x) dx = αj f (xj ) + En , (5.13)
−1 j=0
120 CHAPTER 5. INTEGRATION

which have degree of precision > n + 1. Such formulae are not restricted
to the interval (−1, 1), since given any other interval we can simply map it
to this range. (The reason for using this particular range will become clear
later.) We will now select the (2n + 2) parameters, i.e. the weights, αj , and
the abscissae xj , j = 0, . . . , n, to make the formula as accurate as possible.
Example 5.3.1 (Analytical)
For n = 1 equation (5.13) becomes
Z 1
f (x) dx = α0 f (x0 ) + α1 f (x1 ) + E1 . (5.14)
−1

As we are looking for a method which has degree of precision > 2 the error
E1 must vanish if f (x) = 1. Therefore,
Z 1
1 dx = 2 = α0 + α1 . (5.15)
−1

Similarly, if f (x) = x then E1 = 0 provided

Z 1
x dx = 0 = α0 x0 + α1 x1 . (5.16)
−1

We can repeat this for f (x) = x2 to obtain

2
3
= α0 x20 + α1 x21 (5.17)
and for f (x) = x3 to obtain
0 = α0 x30 + α1 x31 . (5.18)
If all the equations (5.15)–(5.18) are satisfied then the resulting quadrature
formula will be exact for all possible cubics, i.e. it will have degree of precision
3. Notice that the solution of these equations would be identical if we were
to integrate from 1 to −1 and so α0 = α1 and x0 = −x1 and these conditions
enable us to determine
1
α0 = α1 = 1, x0 = −x1 = √ .
3
This gives the Gauss quadrature formula
! !
1 1 1
Z
f (x) dx = f − √ + f √ + E1 ,
−1 3 3
where E1 is proportional to f iv (θ), −1 ≤ θ ≤ 1, since this method has degree
of precision 3.
5.3. GAUSS QUADRATURE FORMULAE 121

Example 5.3.1 demonstrates the method of undetermined coefficients and

clearly this approach can be generalized to the Gaussian quadrature formula
given by equation (5.13). In equation (5.13) there are (2n + 2) parameters,
(n + 1) abscissae xj , j = 0, . . . , n, and (n + 1) weights αj , j = 0, . . . , n.
Therefore, we should be able to select the abscissae and weights to give a
degree of precision of at least 2n+1. It is also clear that to use this method for
the derivation of Gaussian quadrature formulae will lead to lengthy algebraic
manipulation if n > 1. Fortunately, there is an alternative approach.
Recall that a set of polynomials, φj (x), j = 0, . . ., −1 ≤ x ≤ 1, are said
to be orthogonal with respect to a weight function w(x) if
Z 1
w(x)φj (x)φk (x) dx = 0, j 6= k.
−1

We will now show that if we use the (n + 1) zeros of φn+1 (x) as the abscissa
in equation (5.13) then we can obtain degree of precision 2n + 1. To do this
it is necessary to show that φn (x) has exactly n real zeros in the interval
(−1, 1). Firstly, φ0 is a constant C 6= 0, therefore,
Z 1 Z 1
0= w(x)φn (x)φ0 (x) dx = C w(x)φn (x) dx
−1 −1

and since w(x) ≥ 0 this implies that φn (x) has at least one change of sign
in (−1, 1) and hence has at least one zero. Recall that φn is a polynomial
of degree n and by the Fundamental Theorem of Algebra has at most n real
zeros. Let us assume that the number of real zeros of φn in (−1, 1) is j < n,
and that these zeros are at zi , i = 1, . . . , j, where −1 < z1 < z2 < . . . <
zj < 1. With no loss in generality we may assume that φn > 0 in (−1, z1 ),
φn < 0 in (z1 , z2 ), and so on. Now define P (x) = (−1)j i=j i=1 (x − zi ); then
Q

P (x)φn (x) > 0, and so

Z 1
w(x)P (x)φn dx > 0. (5.19)
−1

However, P is a polynomial of degree j < n and so it can be expressed as a

linear combination
i=j
X
P (x) = ci φ i ,
i=0
therefore,
Z 1 i=j
X Z 1
w(x)P (x)φn dx = ci w(x)φi φn dx = 0
−1 i=0 −1
122 CHAPTER 5. INTEGRATION

which contradicts (5.19). Hence all n zeros of φn (x) lie within (−1, 1).
We have already seen examples of orthogonal polynomials in Chapter 3.
In particular the Legendre polynomials are given by P0 (x) = 1, P1 (x) = x,
and

(n + 1)Pn+1 (x) = (2n + 1)xPn (x) − nPn−1 (x), −1 ≤ x ≤ 1 (5.20)

which have weight w(x) = 1. From above we see that Pn+1 (x) has n + 1 real
zeros in the interval (−1, 1), and we will now show that if we select these
zeros as the abscissa we obtain a quadrature formula of the form (5.13) which
has degree of precision 2n + 1. To show this let P (x) be any polynomial of
degree ≤ 2n + 1, then there are polynomials Q(x) and R(x) of degree ≤ n
such that
P (x) = Q(x)Pn+1 (x) + R(x).
Furthermore, we can express Q(x) as a linear combination of Legendre poly-
nomials of degree ≤ n, i.e.
i=n
X
Q(x) = di Pi (x)
i=0

hence
Z 1 i=n
X Z 1
Q(x)Pn+1 (x) dx = di Pi (x)Pn+1 (x) dx = 0
−1 i=0 −1

and so
Z 1 Z 1 Z 1 Z 1
P (x) dx = Q(x)Pn (x) dx + R(x) dx = R(x) dx. (5.21)
−1 −1 −1 −1

Given any set of abscissae we can select weights αj such that

Z 1 j=n
X
R(x) dx = αj R(xj ) (5.22)
−1 j=0

where the integration is exact since R(x) is a polynomial of degree less than
n. Finally, notice that if the interpolation points are selected to be the zeros
of Pn+1 (x) then

P (xj ) = Q(xj )Pn+1 (xj ) + R(xj ) = 0 + R(xj ) = R(xj )

5.3. GAUSS QUADRATURE FORMULAE 123

and so
Z 1 j=n
X
P (x) dx = αj P (xj )
−1 j=0

with no error, i.e. the quadrature method has degree of precision 2n + 1. It

is possible to show2 that the error of such a formula is given by

h2n+3 ((n + 1)!)2

En = f (2n+2) (θ) (5.23)
(2n + 3)(2(n + 1)!)3

where h is the interval width. For the interval (−1, 1) we have h = 2.

Example 5.3.2 (Computational)

In order to illustrate the effectiveness of using the zeros of the Legendre
polynomials load the program QUAD and then estimate the integral
Z 1
x5 dx
0

using three uniformly spaced points. (This is the default problem for the
program QUAD.) Press ESCAPE and change the prompt to Select data
and then press return. Now use a standard grid of three points which are
the zeros of the required Legendre polynomial. This produces an estimate
for the integral of 0.1666669 compared with the exact value of 16 . Recall that
in example 5.1.1 it takes six equally spaced points to evaluate this integral
exactly; here we have used only three. Notice also that the function which
is used to approximate the integrand differs from x5 by as much as 0.207
and yet the estimate for the integral is very good. The reason for this is
given by equation (5.23). Using three points we have n = 2 and so the error
term will involve f vi (θ) which is identically zero for this integrand. (Notice
that the program automatically deals with the problem of mapping the zeros
of Legendre polynomials, usually defined on the interval ±1, to a suitable
range.)
Example 5.3.2 (Analytical)
Let us find the Gauss–Legendre quadrature method based on two points, i.e.
n = 1. The second Legendre polynomial is

P2 (x) = 12 (3x2 − 1)
2
See Froberg (1987).
124 CHAPTER 5. INTEGRATION
q q
which has zeros at x0 = − 13 and x1 = 1
3
; these points are used as the
abscissa. The weight α0 is given by
1 x − x1
Z
α0 = dx
−1 x0 − x1
q
1
Z 1 (x − 3
)
= q dx
1
−1 −2 3
 s 1
1 x2 1
= − q  −x
2 1 2 3
3 −1
= 1

Similarly α1 = 1. This gives the quadrature formula

! !
1 1 1
Z
f (x) ≈ f − √ + f √
−1 3 3
which was obtained in example 5.3.1. Finally, the error is given from equa-
1
tion (5.23) as 135 f iv (θ), |θ| ≤ 1. We see that this method which is obtained
by approximating the integrand at two points has degree of precision 3. Re-
call that using two equally spaced points the maximum degree of precision
is 1.
Example 5.3.3
Values for the weight and absicca of Gauss–Legendre quadrature formulae
are given in table 5.3.
Example 5.3.4 (Calculator/Computational)
Let us estimate the integral
Z 1
exp(−x2 ) dx
0

using Gauss–Legendre quadrature. The range of integration is (0, 1) which

can be transformed to (−1, 1) using the substitution t = 2x − 1 but notice
that Z 1 Z 1
exp(−x2 ) dx = 12 exp(−x2 ) dx.
0 −1

Applying the Gaussian quadrature formula with n = 1, i.e. two points,

gives the result 0.7165315 whilst the exact value is 0.74682413. . . . This is
5.3. GAUSS QUADRATURE FORMULAE 125

Table 5.3: Gauss–Legendre quadrature weights and abscissa

n xj αj
1 ±0.5773502692 1.0000000000

2 0 0.8888888889
±0.7745966692 0.5555555556

3 ±0.3399810436 0.6521451549
±0.8611363116 0.3478548451

4 0 0.5688888889
±0.5384693101 0.4786286705
±0.9061798459 0.2369268851

rather disappointing but the reason for the low accuracy is clear if we sketch
the integrand. From figure 5.1 the integrand has a relatively large peak at
x = 0 which is ignored by Gaussian formulae which use an even number
of points, i.e. n is odd. However, if we use the Gaussian formula with
n = 4, i.e. five points we obtain the estimate 0.746832, which is correct
to four decimal places. Using the program QUAD applied to this integral
directly, i.e. without changing the range of integration, and using the zeros
of the second Legendre polynomial as abscissa produces an estimate for the
integral of 0.7465947. Using the zeros of P4 produces an estimate of 0.7468245
whereas using four equally spaced points gives 0.7469923.

Exercise 5.3.1 (Computational) Load the program QUAD and estimate the
integral
Z 1
exp(−x2 ) dx
0

using two, three, and four equally spaced points. In each case examine the
interpolating polynomial which is constructed and compare it with the inte-
grand. Notice that the program computes an error bound between the inte-
grand and the interpolating polynomial and also an estimate of the quadra-
ture formula. Repeat this exercise using the zeros of the corresponding Leg-
endre polynomials.
126 CHAPTER 5. INTEGRATION

1.0

0.8

0.6

0.4

0.2

-2 -1 0 1 2

Figure 5.1: The integrand exp(−x2 ), −1 ≤ x ≤ 1

The Gauss–Legendre formulae were deduced from orthogonal polynomials

which are defined only on the interval (−1, 1) and integrals over other ranges
must first be reduced to this interval. However, it is frequently the case that
other sets of orthogonal polynomials can be used to evaluate integrals defined
on other intervals.
Example 5.3.5
Alternative weight functions.
(i) Laguerre polynomials. (See table 5.3.2.) 0∞ w(x)f (x) dx
R

w(x) = e−x : 0 ≤ x ≤ ∞, Ln (x) = ex (dn /dxn )(e−x xn ).

R∞
(ii) Hermite polynomials. (See table 5.3.3.) −∞ w(x)f (x) dx

w(x) = exp(−x2 ): −∞ ≤ x ≤ ∞.
Hn (x) = (−1)n exp(x2 )(dn /dxn )(exp(−x2 )).
5.3. GAUSS QUADRATURE FORMULAE 127

Table 5.4: Gauss–Laguerre weights and abscissa

n xj αj
1 0.585786 0.853553
3.414213 0.146447

2 0.415775 0.711093
2.294280 0.278518
6.289945 0.010389

3 0.322547 0.603154
1.745746 0.357419
4.536620 0.038888
9.395071 0.000539

4 0.263560 0.521756
1.413403 0.398667
3.596426 0.075942
7.085810 0.003612
12.640801 0.000023

Table 5.5: Gauss–Hermite weights and abscissa

n xj αj
1 ±0.707107 0.886227

2 0 1.181636
±1.224745 0.295409

3 ±0.524648 0.804914
±1.650680 0.081313

4 0 0.945309
±0.958572 0.393619
±2.0201829 0.019953
128 CHAPTER 5. INTEGRATION

5.4 Composite quadrature

We have seen that the error associated with simple quadrature is proportional
to a power of the interval width. It would seem sensible, therefore, to make
the range of integration as small as possible. To do this we write
Z b Z a+h Z a+2h Z b
f (x) dx = f (x) dx + f (x) dx + . . . + f (x) dx (5.24)
a a a+h b−h

where h = (b − a)/N. We then replace the contribution made in each sub-

interval by the corresponding quadrature formula. The result is called a
composite quadrature formula.
Example 5.4.1 (Composite trapezium rule)
The simple trapezium rule applied to an interval (xi , xi+1 ) is given by
Z xi+1
f (x) dx = 12 h[f (xi ) + f (xi+1 )] + E1i ,
xi

1 3 ′′
where h = (xi − xi+1 ) and E1i = − 12 h f (θi ), xi ≤ θi ≤ xi+1 . If we set
xi = a + ih, with h = (b − a)/N, then equation (5.24) becomes
Z b 1
f (x) dx = h[f (a) + f (a + h)] + E10
a 2
1
+ h[f (a + h) + f (a + 2h)] + E11 +
2
..
.
1
+ h[f (b − h) + f (b)] + E1N −1
2
1
= h[f (a)+2f (a+h)+2f (a+2h)+. . .+2f (b−h)+f (b)]
2
+ECT ,

where the composite trapezium error, ECT , is given by

i=N
X−1 h3 ′′ Nh3 ′′
ECT = − f (θi ) = − f (θ).
i=0 12 12

(The mean value theorem for integrals has been used to replace θi , i =
0, . . . , N − 1, by θ with a ≤ θ ≤ b. See Appendix A.) However, Nh = (b − a)
and so
1 2
ECT = − 12 h (b − a)f ′′ (θ). (5.25)
5.4. COMPOSITE QUADRATURE 129

Notice that the error of the composite trapezium rule is O(h2 ) compared with
O(h3 ) of the simple trapezium rule, but that the h in the composite rule can
be made small by introducing more sub-intervals whereas with the latter it
is the full interval of integration. By systematically increasing N we should
be able to reduce the quadrature error.
Example 5.4.2 (Computational)
Consider the simple trapezium rule for the following integral:
Z 1
exp(x2 ) dx = 1
2
[f (0) + f (1)] + E1
0
= 12 [1 + 2.7182818] + E1
= 1.8591409 + E1 .

We can bound the error, E1 , involved in this procedure by using equa-

tion (5.25), i.e.
 
1 ′′
|E1 | = − 12 f (θ) ≤ 1
6 exp(1) = 21 e1 = 1.35914
 
12

since |f ′′ (x)| = |(4x2 + 2) exp(x2 )| ≤ 6. Therefore,

Z 1
exp(x2 ) dx = 1.8591409 ± 1.35914.
0

We can improve upon this estimate for 01 exp(x2 ) dx by using the composite
R

trapezium rule. For example, loading the program INTEGR, for which
the default problem is this inegral, we obtain the results given in table 5.6.
Notice that the program gives an estimated error bound which is determined
by estimating an upper bound for |f ′′ | using finite differences3 . In all cases
this appears to be an overestimate of the quadrature error which is also given
in table 5.6. The actual error in determining this integral is given in the final
column of this table.

Notice that the errors in table 5.6 decrease by a factor of four whenever the
step size is halved, thus confirming that the error in the composite trapezium
rule is proportional to h2 . Clearly, as h tends to zero the value produced by
the quadrature formula appears to converge and by systematically reducing
h we should obtain an accurate result. Alternatively, we can estimate the
3
See Appendix B.
130 CHAPTER 5. INTEGRATION

R1
Table 5.6: Composite trapezium rule for 0 exp(x2 ) dx

N Estimate Estimated error bound Error bound Error

2 1.5715832 0.0958526 0.3397852 0.1089310
4 1.4906789 0.0410165 0.0894463 0.0280267
8 1.4697123 0.0143884 0.0212366 0.0070601
16 1.4644203 0.0043409 0.0053091 0.0017681

number of intervals required to achieve a given accuracy. For example, let

us suppose that we require a final estimate which is correct to six decimal
places, i.e. |ECT | ≤ 0.5 × 10−6 . To do this we select h such that
1 2
|ECT | ≤ 12
h max |f ′′| ≤ 0.5 × 10−6
0≤x≤1

or
1
h2 ≤ × 10−6
e
h ≤ 0.6065 × 10−3 .

From this we conclude that N = 1/h ≥ 1649 intervals. Using this value for
N gives
Z 1
exp(x2 ) dx = 1.462651912 ± 0.00000017.
0

By increasing N further we might expect to be able to increase the accuracy

further but we are limited by the accuracy of the machine being used. To see
this recall that any estimate for an integral which is computed by quadrature
formulae of this form is determined by the addition of N terms and if each
term is evaluated to a precision ǫ then the total has a maximum possible error
Nǫ; as N becomes larger the accumulated error will dominate the quadrature
error.
Exercise 5.4.1 (Composite Simpson’s rule). Show that by replacing an inte-
gral by
Z b Z a+2h Z a+4h Z b
f (x) dx = f (x) dx + f (x) dx + . . . + f (x) dx
a a a+2h b−2h
5.4. COMPOSITE QUADRATURE 131

where h = (b − a)/2N, and then applying Simpson’s rule we obtain

Z b
1
f (x) dx = 3 h[f (a) + 4f (a + h) + 2f (a + 2h) + 4f (a + 3h)+
a
. . . + 2f (b − 2h) + 4f (b − h) + f (b)] + ECS . (5.26)

1
The error of this composite Simpson’s rule is ECS = − 180 h4 (b − a)f iv (θ),
a ≤ θ ≤ b.
Exercise 5.4.2 (Computational) Load the program INTEGR and estimate
the integral 01 exp(−x2 ) dx using the option Trapezium. Use the error term
R

ECT to estimate the number of intervals required to determine this integral

correct to five decimal places. Repeat this exercise with the composite Simp-
son rule using the prompt Simpson’s 1/3rd.
It is clear that we can develop composite integration formulae for any
of the higher order Newton–Cotes or Gauss methods. Notice that for hand
calculation when using the Newton–Cotes formulae and doubling the number
of points from N to 2N we can re-use previous calculations. For example
with the composite trapezium rule with N intervals we have
Z b
f (x) dx ≈ T (h) = 12 h[f (a) + 2f (a + h) + . . . + 2f (b − h) + f (b)]
a

and for 2N intervals

   
1 1
T 2
h = 4
h[f (a) + 2f a + 21 h + 2f (a + h)
 
+2f a + 32 h + . . .
 
+2f (b − h) + 2f b − 21 h + f (b)]
1
= 2
T (h) +
h      i
1
2
h f a + 21 h + f a + 32 h + . . . + f b − 21 h .

Therefore, to find T ( 12 h) we need only N more function evaluations than

when finding T (h). This is generally the case with Newton–Cotes formulae.
The Gaussian quadrature formula with 2N points has no abscissa in common
with the N-point formula. The composite trapezium method has another
advantage; the weights at each abscissa remain the same as the number of
intervals is doubled but the same is not true with higher order methods.
132 CHAPTER 5. INTEGRATION

Example 5.4.3 (Analytical)

Consider Simpson’s rule applied with two intervals to produce
Z b
f (x) dx ≈ 31 h[f (a) + 4f (a + h) + f (b)],
a

where h = 21 (b − a). If we double the number of sub-intervals then

Z b 1

1

3
   
f (x) dx ≈ h f (a)+4f a+ h +2f (a+h)+4f a+ h +f (b)
a 6 2 2
and we see that the coefficient of f (a + h) has changed from 4 to 2.
Exercise
R1
5.4.3 (Computational) Load the program INTEGR and evaluate
2
0 exp(−x ) dx using the options Gaussian 2 point and Gaussian 4 point.
Notice that at each stage you will be asked to specify the number of panels,
where the total number of points used will be equal to the number of panels
times the number of points in the formula. Compare the results obtained
with those in exercise 5.4.2.
Exercise 5.4.4 (Computational) Use the program INTEGR to estimate the
integral
1
Z 1
√ dx.
0 x
Why do the trapezium rule and Simpson’s rule fail? Compare the results pro-
duced by the open Newton–Cotes three-point rule and the Gauss–Legendre
methods. Which is the most efficient?

5.5 Romberg integration

Whenever we approximate an integral by a quadrature formula it is essential
that some estimate of the quadrature error is made. We have seen that using
the composite trapezium rule it is possible to show that
Z b
1 2
I= f (x) dx = T (h) − 12
h (b − a)f ′′ (θ), a ≤ x ≤ b. (5.27)
a

where
T (h) = 21 h[f (a) + 2f (a + h) + . . . + f (b)].
If we halve h then we have
 
I =T 1
2
h − ( h)2 (b
1 1
12 2
− a)f ′′ (θ′ ). (5.28)
5.5. ROMBERG INTEGRATION 133

Now multiply equation (5.28) by four and subtract (5.27) from the result.
We obtain
1 2
3I = 4T ( 12 h) − T (h) − 12
h (b − a)[f ′′ (θ′ ) − f ′′ (θ)]. (5.29)
If θ′ ≈ θ then
         
1 1 1 1 1
3
4T 2
h − T (h) = T 2
h + 3
T 2
h − T (h)

should be a better approximation for I than either T (h) or T ( h2 ). This process

is called Richardson’s method.
Example 5.5.1 (Computational)
Consider the integral 01 exp(x2 ) dx. Using the program INTEGR we obtain
R

values for the trapezium rule in table 5.7. Let us now apply Richardson’s
method to the results in the second column. For example, applying equa-
tion (5.29) with h = 0.25 we have
h i
T ( 21 h) + T ( 12 h) − T (h)
1
3
1.469712276 − 1.490678861
= 1.46971276 + = 1.462723415.
3
Likewise, we can apply this technique to the remaining entries in the second
column of table 5.7. Notice that we can achieve 6 correct decimal places
using only 32 points instead of the 1649 predicted in example 5.3.1.

R1
Table 5.7: Richardson’s method applied to 0 exp(x2 ) dx

N h Trapezium Error Richardson Error

4 0.25 1.490678861 2.8 × 10−2
8 0.125 1.469712276 7.1 × 10−3 1.462723415 7.2 × 10−5
16 0.0625 1.464420310 1.8 × 10−3 1.462656321 4.6 × 10−6
32 0.03125 1.463094102 4.4 × 10−4 1.462652033 2.8 × 10−7
64 0.015625 1.462762347 1.1 × 10−4 1.462651762 1.8 × 10−8

Exercise 5.5.1 (Analytical) If S(h) denotes the result of applying Simpson’s

rule with step h to estimate an integral I = ab f (x) dx then
R

h4
I − S(h) = (b − a) f iv (θ), a ≤ θ ≤ b. (5.30)
180
134 CHAPTER 5. INTEGRATION

Show that h i
16S( 21 h) − S(h) S( 21 h) − S(h)
= S( 21 h) + (5.31)
15 15
gives a better approximation for I.
Richardson’s method can be applied in a similar way to any other quadra-
ture rule but it can be developed to produce a systematic extrapolation pro-
cedure. Let us assume that I(h) is a result produced by some quadrature
rule when applied to an integral I and that the error associated with this
approximation is given by

I = I(h) + Ahp + O(hp+1 ) (5.32)

where A is some constant independent of h. Similarly,

I = I( 12 h) + A( 21 h)p + O(( 21 h)p+1 ). (5.33)

Now we multiply (5.33) by 2p and then subtract (5.32) from the result. This
produces

(2p − 1)I = 2p I( 12 h) − I(h) + O(hp+1 ) (5.34)

2p I( 21 h) − I(h)
I = + O(hp+1 ). (5.35)
2p − 1

Therefore, if I(h) and I( 21 h) are of precision p − 1 then

2p I( 21 h) − I(h) 1 I( 12 h) − I(h)
= I( 2
h) + (5.36)
2p − 1 2p − 1
has degree of precision ≥ p. This process can be repeated and is called
Romberg extrapolation.
Example 5.5.2 (Analytical)
It can be shown that if T (h) is the result of applying the trapezium rule to
estimate an integral I = ab f (x) dx then
R

I = T (h) + Ah2 + Bh4 + Ch6 + . . . (5.37)

where A, B, C, . . . , are constants independent of h. These constants depend

only upon f in the sense that A is related to f ′′ , B to f iv and so on. Therefore,
5.5. ROMBERG INTEGRATION 135

R1
Table 5.8: Romberg extrapolation applied to 0 exp(x2 ) dx

N h T (h) T (h, 12 h) T (h, 12 h, 14 h)

4 0.25 1.490678861
8 0.125 1.469712276 1.462723415
16 0.0625 1.464420310 1.462656321 1.462651849
32 0.03125 1.463094102 1.462652033 1.462651747

provided f has a bounded second derivative we can eliminate the term Ah2
to obtain equation (5.44) with p = 2, i.e.
h i
T (h, 12 h) = T ( 21 h) + 1
3
T ( 12 h) − T (h) ,

which has error proportional to h4 . From this we have that

I = T (h, 21 h) + B ′ h4 + C ′ h6 + . . .

where B ′ is independent of h but does depend on the fourth derivative of f . If

f has a bounded fourth derivative we can apply (5.44) to eliminate the term
′ 4
B
R1
h and so on. As an example let us use the program INTEGR to evaluate
2
0 exp(x ) dx. This particular integral is set up as default, therefore, select
the option Romberg and start with h = 0.5 to produce the Romberg tableau
shown in table 5.9. Notice that we look down the columns for convergence
because the terms across the rows must necessarily approach one another by
the way in which they are defined.
Although Romberg extrapolation can produce a dramatic increase in the
accuracy of a quadrature rule it must be used with care.
Example 5.5.3 (Analytical)
R √
Consider the integral 01 x log(x) dx. The integrand is shown in figure 5.2
and is clearly defined at x = 0 but this function does not have a bounded
second derivative throughout the interval and so the expression (5.45) is not
valid. This integral can be carried Rout by making a change of variable, x = t2 ,
which transforms the integral to 01 4t2 log(t) dt, the integrand of which has
bounded derivatives of all orders.
136 CHAPTER 5. INTEGRATION

0.2 0.4 0.6 0.8 1.0

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6

-0.7

√
Figure 5.2: The integrand x log(x)

Romberg extrapolation is based on the existence of the expression (5.45),

but even if such an expression is not valid it is sometimes possible to find an
expression for the error. For example, for integrals of the form
Z 1
I= x1/2 log(x)g(x) dx,
0

where g(x) is sufficiently smooth, the error in the trapezium rule, T (h), is
3 3 5 5
I − T (h) = Ah 2 log(h) + Bh 2 + Ch2 + Dh 2 log(h) + Eh 2 + . . . .

Given a sequence of values h1 , h2 , h3 , . . . the terms in this series can be elim-

inated systematically. (See Fox and Mayers (1965).)
Exercise 5.5.2 (Computational) Use the option Romberg in the program
INTEGR to evaluate 01 x7 dx. Why do three extrapolations produce an
R

exact answer?
Exercise 5.5.3 (Analytical) If T (h) and S(h) are results produced by the
trapezium and Simpson’s method show that
h i
S( 21 h) = T ( 12 h) + 1
3
T ( 12 h) − T (h) .

Furthermore, show that the first extrapolation of Simpson’s rule, and there-
fore, the second extrapolation of the trapezium rule, produces the closed
Newton–Cotes rule for four points given in table 5.1. (The resulting method
is called Simpson’s 83 th rule.)
5.6. ADAPTIVE INTEGRATION 137

Exercise 5.5.4 (Computational) Use the program INTEGR with option

Trapezium to estimate the integral 01 exp(x2 ) dx. Repeat this exercise for
R

the prompts Simpson’s 1/3rd and Simpson’s 3/8th. Confirm that these
values are the first, second and third columns of the corresponding Romberg
table.

5.6 Adaptive integration

With all quadrature formulae the choice of the step size, h, is critical. Too
large a value and the result will be poor, too small a value and the amount of
work required will be large and the result may be prone to rounding errors.
In this section we will now outline an procedure which can automatically
adjust the step size. To see the importance of this approach consider the
following example.

10

2 4 6 8 10

Figure 5.3: The integrand [(x − 5)2 + 0.1]−1

Example 5.6.1 (Analytical)

Let us estimate the integral
10 1
Z
dx.
0 (x − 5)2 + 0.1
The integrand is shown in figure 5.3. If we were to apply Simpson’s rule with
a uniform step size h to produce an estimate S(h) then the error involved
is proportional to h4 f iv . Throughout most of the interval f iv will be very
138 CHAPTER 5. INTEGRATION

small; only in the neighbourhood of x = 5 will it be significant. However, to

ensure that we achieve an overall precision ǫ we need to select h such that
 
 h4 f iv 
≤ǫ
 
 180 
 

i.e. !1/4
180ǫ
h≤
|f iv |max
in which case we will be forced to use a small value for h even though the
fourth derivative is very small except near x = 5. In order to overcome this
difficulty we will now permit the sub-division of the range of integration and
use non-uniformly spaced abscissa. Let Iab (h) be an approximation for the
integral I = ab f (x) dx which has degree of precision p, i.e.
R

I − Iab (h) = Ahp+1 f (p+1) (θ). (5.38)

If we require a final accuracy ǫ then we should select h so that

!1/p
ǫ
h≤
|Af (p+1) |max

but in most cases |f (p+1) |max is unknown. In order to overcome this we sub-
divide the range of integration (a, b) into (a, c) and (c, b), where c = 21 (a + b),
then
 p+1
Iac − Iac ( 21 h) = A 1
2
h f (p+1) (θ1 )
 p+1
Icb − Ibc ( 12 h) = A 1
2
h f (p+1) (θ2 ),

where a ≤ θ1 ≤ c and c ≤ θ2 ≤ b. Therefore, there will exist a value θ′ in

(a, b) such that

hp+1 (p+1) ′
I − Iac ( 21 h) − Icb ( 21 h) = Af (θ ).
2p
If we subtract equation (5.38) from this then

Iab (h) − Iac ( 21 h) − Icb ( 21 h) ≈ Ahp+1 f (p+1) (θ)(1 − 2−p )

5.7. THE CLENSHAW–CURTIS METHOD 139

provided θ ≈ θ′ . This last equation enables us to estimate f (p+1) (θ), or more

precisely
 
b c 1 b 1
Ia (h) − (Ia ( 2
h) + Ic ( 2
h))
Ahp+1 f (p+1) (θ) = .
(1 − 2−p )
Hence, if
|Iab (h) − Iac ( 21 h) − Icb ( 21 h)| ≤ (2p − 1)ǫ
then
|I − (Iac ( 12 h) + Icb ( 21 h))| ≤ ǫ
as required. If not we reconsider each sub-interval independently and use a
criterion of 12 ǫ in each sub-interval.
Example 5.6.2 (Computational)
If we use the program INTEGR with the option Simpson 1/3rd to eval-
uate the integral
Z 10
1
dx
0 (x − 5)2 + 0.1
then 512 intervals are required to ensure that two successive estimates agree
to eight decimal places. Using the option Gaussian 4 point Rule 128 panels
of four points are needed to achieve the same accuracy and so there is little to
be gained in using this method. However, using the adaptive procedure with
a relative error tolerance of 10−8 only 161 function evaluations are required.

5.7 The Clenshaw–Curtis method

We have seen in the previous sections how it is possible to estimate an in-
tegral by approximating the integrand by a suitable polynomial, integrate
the approximation and then estimate the error. In order to obtain the best
polynomial approximation, of a given order, it would seem obvious to use
the minimax approximation. However to do this we need to expend more
effort in finding the minimax approximation than we would use with a poorer
approximation of higher order. As a compromise we could use an approxi-
mation which is almost minimax but which can be found a great deal easily,
for example a Chebyshev approximation. This observation is the basis of the
Clenshaw–Curtis method.
140 CHAPTER 5. INTEGRATION

Let us assume that we have a Chebyshev approximation for a function

f (x) which is given by
f (x) = 21 c0 T0 + c1 T1 (x) + c2 T2 (x) + . . . , (5.39)
then the indefinite integral of f (x) is given by
Z Z Z
f (x) dx = 12 c0 T0 dx + c1 T1 (x) dx + . . . . (5.40)

However,
Z Z
T0 dx = 1 dx = x + C0 = T1 (x) + C0 ,
Z Z
T1 dx = x dx = 21 x2 + C1 = 14 T2 (x) + C1′ ,

where C0 , C1 and C1′ are constants of integration, and

1 Tn+1 Tn−1
Z  
Tn (x) dx = − , n ≥ 2. (5.41)
2 n+1 n−1
This last result can be obtained as follows:
Z Z
Tn (x) dx = cos(n cos −1 x) dx
Z
= − cos(nθ) sin θ dθ, (x = cos θ)
Z
= − 12 [sin(n + 1)θ − sin(n − 1)θ] dθ
1 Tn+1 Tn−1
 
= − .
2 n+1 n−1
We now recall that T2n is an even function and that T2n+1 is an odd
function and then use these results to show that
[T2n+1 ]1−1 = 2, and [T2n ]1−1 = 0.
Substituting these expressions into equation (5.40) we determine
1 c0 1 1−3 1 3−5
Z      
f (x) dx = 2 + c2 + c4 + ...
−1 2 2 1×3 2 3 × 5
c0 c2 c4

= 2 − − − ...
2 1×3 3×5
∞
X 1 + (−1)r
= c0 − cr
r=2 r2 − 1
5.7. THE CLENSHAW–CURTIS METHOD 141

and if we knew the Chebyshev coefficients of the function f (x) we could find
the required integral. Notice that the above series will converge much faster
than the Chebyshev series for the integrand because of the term r 2 − 1 in the
denominator. Unfortunately, the determination of the Chebyshev coefficients
is time consuming and in general will involve integrae, if

|Iab (h) − Iac ( 21 h) − Icb ( 21 h)| ≤ (2p − 1)ǫ

then
|I − (Iac ( 12 h) + Icb ( 21 h))| ≤ ǫ
as required. If not we reconsider each sub-interval independently and use a
criterion of 12 ǫ in each sub-interval.
Example 5.6.2 (Computational)
If we use the program INTEGR with the option Simpson 1/3rd to eval-
uate the integral
1
Z 10
dx
0 (x − 5)2 + 0.1
then 512 intervals are required to ensure that two successive estimates agree
to eight decimal places. Using the option Gaussian 4 point Rule 128 panels
of four points are needed to achieve the same accuracy and so there is little to
be gained in using this method. However, using the adaptive procedure with
a relative error tolerance of 10−8 only 161 function evaluations are required.
Now we multiply (5.33) by 2p and then subtract (5.32) from the result.
This produces

(2p − 1)I = 2p I( 21 h) − I(h) + O(hp+1 ) (5.42)

2p I( 12 h) − I(h)
I = + O(hp+1 ). (5.43)
2p − 1

Therefore, if I(h) and I( 21 h) are of precision p − 1 then

2p I( 12 h) − I(h) 1 I( 21 h) − I(h)
= I( 2
h) + (5.44)
2p − 1 2p − 1

has degree of precision ≥ p. This process can be repeated and is called

Romberg extrapolation.
142 CHAPTER 5. INTEGRATION

R1
Table 5.9: Romberg extrapolation applied to 0 exp(x2 ) dx

N h T (h) T (h, 12 h) T (h, 21 h, 14 h)

4 0.25 1.490678861
8 0.125 1.469712276 1.462723415
16 0.0625 1.464420310 1.462656321 1.462651849
32 0.03125 1.463094102 1.462652033 1.462651747

Example 5.5.2 (Analytical)

It can be shown that if TR(h) is the result of applying the trapezium rule to
estimate an integral I = ab f (x) dx then

I = T (h) + Ah2 + Bh4 + Ch6 + . . . (5.45)

where A, B, C, . . . , are constants independent of h. These constants depend

only upon f in the sense that A is related to f ′′ , B to f iv and so on. Therefore,
provided f has a bounded second derivative we can eliminate the term Ah2
to obtain equation (5.44) with p = 2, i.e.
h i
T (h, 12 h) = T ( 21 h) + 1
3
T ( 12 h) − T (h) ,

which has error proportionarect to four decimal places using the method
of Clenshaw and Curtis? If the trapezium rule were applied directly to this
integral how many intervals are required to determine a result which is correct
to four decimal places? Which method is the most efficient and why?