Appendix H

Properties of Chebyshev polynomials

The basic properties are now fairly well known [103, 209]; however, for the sake
of completeness they are described here along with one or two less well-known
results.

H.1 Definitions and orthogonality relations

The definition of the Chebyshev polynomial of order n is

Tn (x) = cos(n cos 1 (x)); jxj  1

Tn (x) = cosh(n cosh 1 (x)); jxj  1: (H.1)

It is not immediately obvious that this is a polynomial. That it is follows

from applications of De Moivre’s theorem. For example

T3 (x) = cos(3 cos 1 (x)) = 4 cos3 (cos 1 (x)) 3 cos(cos 1 (x)) = 4x3 3x:
(H.2)
The Chebyshev polynomials are orthogonal on the interval [ 1; 1] with
weighting factor w(x) = (1 x 2 ) 2 which means that
1

Z 1 
dx w(x)Tn (x)Tm (x) = (1 + Æn0 )Ænm (H.3)
1 2
where Ænm is the Kronecker delta.
The proof of this presents no problems: first the substitution y = cos 1 (x)
is made; second, making use of the definition (H.1) changes the integral (H.3) to
Z 
dy cos(my) cos(ny) (H.4)
0
and this integral forms the basis of much of Fourier analysis. In fact, Chebyshev
expansion is entirely equivalent to the more usual Fourier sine and cosine

601
602 Properties of Chebyshev polynomials

expansions. Returning to the integral, one has

if m 6= n
8
Z  < 0;
dy cos(my) cos(ny) = ; if m = n = 0
: ;
(H.5)
0 if m = n 6= 0.
2
With the help of the orthogonality relation (H.3) it is possible to expand any
given function in terms of a series of Chebyshev polynomials, i.e.
m
X
f (x) = ai Ti (x): (H.6)
i=0
Multiplying through by w(x)T j (x) and using the relation (H.3) gives for the
coefficients Z 1
ai = Xi dx w(x)Ti (x)f (x) (H.7)
1
where Xi = 1= if i 6= 0 and Xi = 2= if i = 0.
The extension to a double series is fairly straightforward. If an expansion is
needed of the form
XmX n
f (x; y) = Cij Ti (x)Tj (y) (H.8)
i=0 j =0
then Z +1 Z +1
Cij = Xi Xj dx dy w(x)w(y)Ti (x)Tj (y)f (x; y): (H.9)
1 1
The orthogonality relations can also be used to show that the Chebyshev
expansion of order n is unique. If
m
X m
X
f (x) = ai Ti (x) = bi Ti (x) (H.10)
i=0 i=0
then multiplying by w(x)T j (x) and using the relation (H.3) gives a i = bi .

H.2 Recurrence relations and Clenshaw’s algorithm

Like all orthogonal polynomials, the Chebyshev polynomials satisfy a number of
recursion relations. Probably the most useful is

Tn+1 (x) = 2xTn (x) Tn 1 (x): (H.11)

The proof is elementary. If y = cos 1 (x) then

Tn+1 (x) = cos((n + 1)y) = cos(ny) cos(y) sin(ny) sin(y)

Tn 1(x) = cos((n + 1)y) = cos(ny) cos(y) + sin(ny) sin(y) (H.12)
 Recurrence relations and Clenshaw’s algorithm 603

and adding gives

Tn+1 (x) + Tn 1 (x) = 2 cos(ny) cos(y) = 2xTn (x) (H.13)

as required.
It is clear that if the recurrence begins with T 0 (x) = 1 and T1 (x) = x,
equation (H.11) will yield values of T n (x) for any n. This is the preferred means
of evaluating T n (x) numerically as it avoids the computation of polynomials.
In order to evaluate how good a Chebyshev approximation is, one compares
the true function to the approximation over a testing set. This means that one is
potentially faced with mamy summations of the form (H.6). Although current
computers are arguably powerful enough to allow a brute force approach, there
is in fact a much more economical means of computing (H.6) than evaluating the
polynomials and summing the series. The method uses Clenshaw’s recurrence
formula. In fact this can be used for any polynomial which uses a recursion
relation although the version here is specific to the Chebyshev series. The general
result is given in [209].
First define a sequence by

yn+2 = yn+1 = 0; yi = 2xyi 1 yi + ai : (H.14)

Then

f (x) = [yn 2xyn+1 + yn+2 ]Tn(x) +


+ [yi 2xyi+1 + yi+2 ]Ti (x)
+


+ [a0 y2 + y2 ]T0(x) (H.15)

after adding and subtracting y 2 T0 (x). In the middle of this summation one has

+


+ [yi+1 2xyi+2 + yi+3 ]Ti+1 (x) + [yi 2xyi+1 + yi+2 ]Ti (x)
+ [yi 1 2xyi + yi+1 ]Ti 1 (x) (H.16)

so the coefficient of y i+1 is

Tn+1 (x) 2xTn (x) + Tn 1 (x) (H.17)

which vanishes by virtue of the recurrence relation (H.11). Similarly all the
coefficients vanish down to y 2 . All that remains is the end of the summation
which is found to be
f (x) = a0 + xy1 y2 : (H.18)
Therefore to evaluate f (x) for each x, one simply passes downwards through the
recurrence (H.14) to obtain y 1 and y2 and then evaluates the linear expression
(H.18). Unfortunately there is no obvious analogue of Clenshaw’s result for two-
dimensional expansions of the form (H.8). This means that in evaluating a double
series, one can only use the recurrence if the function f (x; y ) splits into single-
variable functions, i.e. f (x; y ) = g (x) + h(y ). Of all the examples considered in
604 Properties of Chebyshev polynomials

chapter 7, only the Van der Pol oscillator fails to satisfy this condition, although
it would be unlikely to hold in practice.
Clenshaw’s algorithm can also be used algebraically in order to turn
Chebyshev expansions into ordinary polynomials. However, one should be aware
that this is not always a good idea [209].

H.3 Chebyshev coefficients for a class of simple functions

In chapter 7, the Chebyshev expansion for the restoring force f (y; y_ ) is estimated
for a number of simple systems. In order to form an opinion of the accuracy of
these estimates, one needs to know the exact values of the coefficients. A function
sufficiently general to include the examples of chapter 7 is

f (x; y) = ax3 + bx2 + cx + dy + ey2 + fx2 y: (H.19)

The x and y are subjected to a linear transformation

x !  (x) = x = x 2
1
y ! _(y) = y = y 2 (H.20)
1
where

1 = 12 (xmax xmin ); 2 = 12 (xmax + xmin )

1 = 12 (ymax ymin); 2 = 12 (ymax + ymin): (H.21)

The form of f in the (x; y) coordinate system is given by

f (x; y) = f (x; y) = f ( 1 (x); _ 1 (y)) = f ( 1 x + 2 ; 1 y + 2 ): (H.22)

A little algebra produces the result

2 2
f (x; y) = ax3 + bx + cx + dy + ey2 + fx y + gxy + h (H.23)

where

a = a 31
b = 3a 21 2 + b 21 + f 1 2
c = 3a 1 22 + 2b 1 2 + 2f 1 2 2
d = d 1 + 2e 1 2 + f 22 1
e = e 12
f = f 21 1
g = 2f 1 2 2
h = a 2 + b 2 + c 2 + d 2 + e 22 + f 22 2 :
3 2 (H.24)
 Least-squares analysis and Chebyshev series 605

One can now expand this function as a double Chebyshev series of the form
XmX n
f (x; y) = Cij Ti (x)Tj (y) (H.25)
i=0 j =0
either by using the orthogonality relation (H.9) or by direct substitution. The exact
coefficients for f (x; y ) are found to be
C00 = h + 12 (b + e)
C01 = d + 12 f
C02 = 12 e
C10 = 21 a + c
C11 = g
C12 = 0
C20 = 12 b
C21 = 12 f
C22 = 0
C30 = 12 a: (H.26)

H.4 Least-squares analysis and Chebyshev series

It has already been noted in chapter 7 that Chebyshev polynomials are remarkably
good approximating polynomials. In fact, fitting a Chebyshev series to data is
entirely equivalent to fitting a LS model. With a little extra effort one can show
that this is the case for any orthogonal polynomials as follows [88]:
Let f i (x); i = 1; : : : ; 1g be a set of polynomials orthonormal on the
interval [a; b] with weighting function w(x), i.e.
Z b
dx w(x) i (x) j (x) = Æij : (H.27)
a
(The Chebyshev polynomials used in this work are not orthonormal. However, the
set 0 (x) =  2 and i (x) = (2= ) 2 are.) Suppose one wishes to approximate
1 1

a function f (x) by a summation of the form

Xn
f^(x) = ci i (x): (H.28)
i=0
A least-squared error functional can be defined by
Z b
In [ci ] = dx w(x)jf (x) f^(x)j2
a
Z b n
X
= dx w(x)jf (x) ci i (x)j2 (H.29)
a i=0
606 Properties of Chebyshev polynomials

and expanding this expression gives

Z b n Z b
X
In [ci ] = dx w(x)f (x)2 + 2 ci dx w(x)f (x) i (x)
a i=0 a
n X
X n Z b
+ ci cj dx w(x) i (x) j (x): (H.30)
i=0 j =0 a
Now, the Fourier coefficients for a i for an expansion are defined by
Z b
ai = dx w(x)f (x) i (x) (H.31)
a
so using this and the orthogonality relation (H.27) gives
Z b n
X n
X
In [ci ] = dx w(x)f (x)2 2 ai ci + c2i (H.32)
a i=0 i=0
and finally completing the square gives
Z b n
X n
X
In [ci ] = dx w(x)f (x)2 a2i + (ci ai )2 : (H.33)
a i=0 i=0
Now, the first two terms of this expression are fixed by the function f (x) and
the Fourier coefficients, so minimizing the error functional by varying c i is simply
a matter of minimizing the last term. This is only zero if a i = ci . This shows
clearly that using a Fourier expansion of orthogonal functions is an LS procedure.
The only point which needs clearing up is that the usual LS error functional is
Z b
In [ci ] = dx jf (x) f^(x)j2 (H.34)
a
without the weighting function. In fact, for the Chebyshev expansion changing
the variables from x to y = cos(x) changes (H.28) to
Z 
In [ci ] = dy jf (cos 1 (y)) f^(cos 1 (y))j2 (H.35)
0
which is the usual functional over a different interval.