JOURNAL OF COMPUTATIONAL PHYSICS 70, 63-88 (1987)

rthogonal Rational Functions on a

JOHN P. BOYD

Department of Atmospheric and Oceanic Science.

University of Michigan, Ann Arbor, Michigan 48109

Received April ! 1, 1986; revised September 12, 1986

By applying a mapping to the Chebyshev polynomials. we define a new spectral basis: the
“rational Chebyshev functions on the semi-infinite interval,” denoted by 7X,(y). Continuing
earlier work by the author and by Grosch and Orszag, we show tha? these rational functions
inherit most of the good numerical characteristics of the Chebyshev polynomials:
orthogonality, completeness, exponential or “infinite order” convergence, matrix sparsity for
equations with polynomial coefficients, and simplicity. Seven numerical examples illustrate
their versatility. The “Charney” stability problem of meteorology, for example, is solved to
show the feasibihty of applying spectral methods to a semi-infinite atmosphere. For functions
that are singular at both endpoints, such as K,(y), one may combine rational Chebyshev
functions with a preliminary mapping to obtain a single, exponentiaily convergent expansion
on y E [O, cc]. Finally, we successfullygeneralize the WKB method to obtain, for the Jo Bessel
function, an amplitude-phase approximation which is convergent rather than asymptotic and
is accurate not merely for large y but for all y, even the origin. 0 1987 Academic press, 1~.

1. INTR~DUCTI~~J

In an earlier paper, Boyd [ 1] combined an algebraic rnap~i~~ from [ - 1%1] to

[ - 03, cc ] with Chebyshev polynomials to provide a new set of basis functions for
solving differential equations on an infinite interval. In this follow-up, a similar set
of mapped Chebyshev polynomials are defined for the semi-infinite interval,
YE m al.
The earliest appearance of Chebyshev polynomials in tandem with an algebraic
change-of-coordinate was Grosch and Qrszag [2]. Boyd [3] used the metho
steepestdescent to provide guidelines for optimizing the mapping. Boyd 14-63
cussesalternatives to mappings and also the use of a change-of-coordinate to deal
functions with singularities on or near the expansion interval.
espite this prior work, there are still many issues left for the present article.
Programming mechanics, boundary conditions at both the origin and at infnnity,
functions that decay algebraically rather than exponentially as Y--P a3 and
numerical examples will be discussed in turn. The new basis functions, denoted by
EL,(y), are defined by

TL,( y) F T,(x) = cos(nt), (I.11

63
Copyright 0 1987 by Academc Press, Inc.
581/70/1-S Ail rights of reproduction in any form resewed.
64 JOHN P. BOYD

where L is a constant map parameter and the three coordinates are related by
y=L(l +x)/(1-x), x=(Y--L)l(Y+L) (1.2)
y = L cot2(t/2), t = 2 arccot( [ y/L] ‘j2). (1.3)
It may seem a little pretentious to introduce a new symbol for functions which
are merely the Chebyshev polynomials in disguise. However, the Chebyshev
polynomials in x are themselves the images of the terms of a Fourier cosine series
under the map x = cos(t). In both cases,the change-of-coordinates alters the shape
and properties of the basis set so profoundly that it is probably best to regard the
result of the mapping as a new species, deserving its own symbol.
Even so, the connection with the Fourier cosine functions is nonetheless very
important both for analyzing convergence, as done in the Appendix, and also for
writing practical computer programs, To avoid confusion as we leap from one coor-
dinate to another, we shall adopt the convention that y E [O, co] is the argument of
the TL,( y), x E [ - 1, 11 is the argument of the ordinary Chebyshev polynomials,
and t E [O, rc] (‘9” for “trigonometric”) is the argument of the cosines. We are free
to calculate in whichever of these three coordinates is most convenient.
We shall refer to the TL,(y) as the “rational Chebyshev functions on a semi-
infinite interval”; Table I lists the explicit form of the first nine functions for L = 1.
The graphs shown in Fig. 1 apply for any L if we replace y by (y/L); varying L
alters the width of the functions without changing their shape. Because the depen-
dence on L is so simple and also for notational simplicity, we shall often suppress
the map parameter by writing TL,( y) [instead of the more correct TL,( y; L)] and
by setting L = 1 in some figures and tables.
By merely changing the variable in the usual orthogonality integral for the
cosines, one can show that the rational Chebyshev functions are orthogonal:

j-k m=n=O

s
m mn(y;
0
L) mz(.Y; LW2/l:(y+ Lw2]) dy=
I0,
742,
mfn
m=n>O.

TABLE I

n W,(Y)

(y4-28y3+70y2-28y+l)/(yf1)4
(y5-45y4+210y3-210y2+45y- l)/(y+ 1)5
( y6 - 66y5 + 495y4 - 924y3 + 495y2 - 66y + I)/( y + 1)”
(y7-91p+ lOOly5- 3003y4+ 3003y3 - 1001y* + 9ly- l)/(y f 1)7
(y’- 120~’ + 1820y6-8008~~ + 12870y4-8008y3 + 1820y2- 120~ + l)/(y + 1)8
 RATIONAL CHEBYSHEV SERIES 65

0.0 1.0 20 30 4.0 5.0 6.0 7.0 8.0

FIG. 1. Graphs of TLl(y)[dot-dash], TL,(y)[dotted], TL,(y)[dashed], and TL,(y)[solid] on (a)

ye CO,Z] and (b) ye [O, S]. (The map parameter L= 1.)

owever, this orthogonality relationship has been listed principally to explain t

one never needs it! To apply Galerkin’s method, it is simpler to convert to
trigonometric argument t and then evaluate the needed integrals as in Boyd [l
In the next section, it will be shown that the rational Ghebyshev functions on the
semi-i&mite interval, like their cousins on the infinite interval described in
[ 11, give banded Galerkin matrices when the coefficients of the differential eq
are polynomials or rational functions of y. For general problems, however, the
pseudospectral method (alias “collocation” algorithm) is just as efficient and con-
siderably easier to program. In the next section, the mechanics of this and of
converting the problem from y to t will be describe
Section 3 explains that, again as true for the TB,1(y), the boundary condition at
infinity is usually “natural” in the sensethat it need not be explicitly im
basis functions. However, the Galerkin’s or pseudospectral matrix must be rnodi~e~
by adding a row to force the numerical solution to satisfy the boundary co~ditior~
at y = 0 unless the differential equation is singular at that point also-in which case
the boundary conditions may be “natural” at both endpoints.
Section 4 discussesfunctions which decay algebraically (as opposed to exponen-
tially) as y + CC.It is still possible to obtain exponentially fast convergence for such
S(y) provided that the function either does not oscillate or has the frequency of
oscillations tend to zero as y -+ co.
The fifth section is the heart of the paper: seven numerical examples that
illustrate each of the theoretical ideas in the article. Section 6 is a s~rnrn~~y and
prospectus. The details of the convergence theory, including the calcul . of the
contours of “equiconvergence” in the complex plane, have been bani to the
pendix because these concepts are only occasionally useful in ~~~~icati~~s. It is
66 JOHN P. BOYD

nonetheless satisfying that thanks to the connection between the orthogonal

rational Chebyshev functions and ordinary Fourier series expressed by (1.1), it is
possible to give a complete discussion of the theory of T&(y) expansions.

2. MECHANICS: COORDINATE CONVERSION, GRID POINTS, AND BANDEDNESS

In the author’s experience, the simplest way to set up the Galerkin’s or

pseudospectral matrix for solving a boundary or eigenvalue problem is to exploit
the connection with a Fourier cosine series. Table II lists the formulas for transfor-
ming derivatives, which were computed using the algebraic manipulation language
REDUCE by iterating the elementary rule
d/dy = - (sin2( t/2) tan(t/2)} d/dt. (2.1)

TABLE II
Transformations of Derivatives for the Mapping y = L cot*(t/2) which Converts a Rational-Chebyshev
Series in Z,(y) into a Fourier Cosine Series in cos(nt). L Is a Constant, the “Map Parameter,” uY=
{ -sin3(f/2)/[L cos(2/2)]j u,, uYY= {sinS(t/2)/(2t2 cos3(t/2))(2cos(z/2) sin(f/2) u,,+ [3 - 2 sin*(t/2)]u,)

-
%Y
- 8 sin4 12 cos sin3 4 sin4
i-20 sin2 -18cossin -4 sin2
-1.5 [ xsin7(r/2)/j4L3 cos5(r/2)})

%
- 48 sin6 88 cos sin’ 48 sin6 - 8 cos sin5
+ 168sin4 -232~0s sin3 - 120 sin4 + 8 cos sin3
-210 sin* + 174 cos sin + 72 sin’
+105 [ x sin9/{8L4 cos7}]
%
-384 sin’ 800 cos sin’ 560 sin8 -160cossin’ -16sina
f1728 sin6 -2960~0s sin5 -2080 sin6 + 400 cos sin’ + 32 sin6
-3024 sin“ +39OOcos sin3 +2660 sin4 - 240 cos sin3 - 16 sin4
+2520 sin* -195Ocos sin -1140sinZ
-945 1:x sin”/{ 16L5 co?/]
%
3840 sin’0 -8768 cossin9 - 7200 sin” +272Ocos sin’ +480 sinlo - 32 cos5sin5
-21120 sins f41536cos sin’ $34800 sin* -10240~0s sin7 -1680sin’
+47520 sin6 - 76808 cos sins - 65040sin6 + 13160cos sin5 + 1920 sin6
-55440 sin4 +67440 cos sin3 + 56160sin4 -5640~0s sin3 - 720 sin4
+34650 sin2 - 25290cos sin - 18720sin*
-10395 [x(-l)sin13/{32L6cosi1}]

Note. The term [ x sinq/{2”LP cos’}] denotes that all entries in the table for the previous derivative
must be multiplied by this common factor. Note also that all sines and cosines in the table have
arguments of (t/2), not f.
 RATIONAL CHEBYSHEV SERIES 67

the main program may still use the physical coordinate y, however: Table IfI is a
FQRTRAN fragment that shows the trigonometric argument t can be entirely con-
fined to a single subroutine with just eight executable statements (wbicb eva~~~~es
the TE,( y) and its derivatives at a given y).
The pseudospectral grid in y is simply the image under the ma ping of an evenly
spaced Fourier grid
ti = 7c(2i+ 1)/(2N + 2), i = o,..., N (a.21
yi = L d(tJ2). (2.31

A differential equation of the form

Hzi=f, (2.4)
where H is a linear operator and u(y) and S(y) are functions, is ~seudospe~tra~~y
discretized to the matrix problem
Wa=f, (2.5)
where a is the column vector whose elements are the coefficients of the 1”&,(y)
series for u(y) and where

i/z (ff TL,llYl) I,v=yI~ i,j=O ,..., IV [pseudospectral] (2.6)

fief(Yi)2 i = cl,..., N. (2.7)

TABLE III
SUBROUTINE BASIS(A’, Y, L, TL, TLY, TtYY)
C A FORTRAN subroutine for computing the rational Chebyshev
c functions.
C INPUT: N (> = 0, the degree of the basis function)
c Y (on interval [0, Infinity])
c L (map parameter)
c OUTPUT: TL, TLY, TLYY are the,Nth Chebyshev functions
c and their first two derivatives.
c Tis the trigonometric argument (on interval [O, pi]).
REAL L
T= 2. * ACOT(SQR( Y/L))
TL = COS(FLOAT(N)*T)
C Derivatives of cos (Nt) with respect to t:
PT= -FLOAT(N) * SIN(FLOAT(N)*T)
PTi”= -FLOAT(N)*FLQAT(N) * TL
C Conversion of t-derivatives into x-derivatives
C=COS(T/Z.)
S = SIN( T/2.)
TLY = -PT*S*S*S/(L*C)
TLYY = (2.*C*S*PTT+ (3. - 2.*S*S)* PT) * (S**5/(2.*L*L*C**3))
RETURN
END
68 JOHN P. BOYD

Galerkin’s method replaces (2.6) and (2.7) by

H,r (TL,, H TL,), i, j=O ,‘.‘, N [Galerkin’s] (2.8)

fir (TL,,f), i = O,..., N, (2.9)
where the inner product is defined by [for arbitrary g(y), h(y)]

(g, h)3SZ g(LcoP[t/2])h(L cot2[t/2])& (2.10)

0

Since nonlinear solutions can be calculated by the Newton-Kantorovich method

(among others), solving a linear differential equation at each iteration [7, 81, we
discuss only linear problems in this article.
As explained in Gottlieb and Orszag [9], it is rarely worth the bother to replace
evaluation (pseudospectral) by integration (Galerkin’s) because the two methods
give almost equal accuracy for the same N, but collocation is simpler. There is one
exception: when the coefficients of the differential equation are polynomial or
rational in y, the coefficients will be rational functions of sin(t) and cos(t). After the
denominators have been multiplied out, the Galerkin matrix elements (2.8) and
(2.9) will be integrals of trigonometric basis functions multiplied by trigonometric
polynomials.
By using elementary trigonometric identities, one can show, as done in [ 11, that
if Pk(t) is a trigonometric polynomial of degree k, then letting c),,(t) denote either
sin(nt) or cos(nt) as appropriate,
nfk

Pkct) hdt) = c &dm(t). (2.11)

m=n-k

If k is the highest degree of the trigonometric polynomials that appear in the dif-
ferential equation (after clearing denominators), then (2.11)--and the orthogonality
of the sines and coefficients-imply that the Galerkin’s matrix will be banded in the
sensethat
H,=O if /i-j1 > k. (2.12)

This is an extremely important property because banded matrices can be solved in

a small fraction of the expense of inverting full matrices. Thus in spite of its greater
complexity, Galerkin’s method is still a key ingredient of many semi-implicit time-
stepping codes. Cain et al. [lo] seemto have been the first to appreciate this point;
they also mapped Fourier series to a semi-infinite or infinite domain, but did not
make the connection with Chebyshev polynomials that is central to [l] and [Z].
The present paper will not dwell on the issue of bandedness because the precise
bandwidth varies from problem to problem. Many older spectral papers went to
great lengths to apply recurrence relations and elaborate algebra to analytically
evaluate the Galerkin’s matrix. This is reasonable only when the end result is a
 RATIONAL CHEBYSHEVSERIES 69

sparse matrix and the programmer intends to exploit the bandedness. ~tberwisej
the pseudospectral algorithm is simpler to program and easier to adapt from one
differential equation to another. All the numerical examples given SnSection 5 were
solved pseudospectrally.

3. BOUNDARY CONDITIONS:
IMPLICIT VERSUSEXPLICIT, NATURAL VERSUSESSENTIAL

When applying Chebyshev polynomials to solve a differential equation on

[ - 1, I], it is usually necessary to modify the matrix defined above by reser
two rows to impose the boundary conditions. The only e ption is when the
ferential equation is singular at one or both endpoints. e might suppose
such singularities would cause enormous difficulties, but they often make the
problem simpler: the desired solution is the only linearly independent solution
which is analytic at the singularity. One can therefore apply (2.6) witbo~t
modification; the Chebyshev series will converge to the bounded solution without
imposition of explicit constraints.
In the language of finite element methods, boundary con
explicitly imposed on the numerical solution are called “essential.”
~~c~~~~~~~~e~ sum of basis functions will converge to the correct
behavior-although not exactly satisfying the boundary conditions for any finite
N-the boundary conditions are said to be “natural.”
It is possible to construct differential equations which are analytic at jollity, but
the tlsual case is that all solutions of the differential equation except the desired one
blow up as y -+ co. The boundary condition at infinity is then “natural.”
At y = 0, it may be necessary to reserve one row of the pseudospectral matrix to
impose an “essential” condition such as u(0) = 1. e shah show how this is done in
pction 5 in the context of the two numerical examples for which it is necessary (the
solution is K,(y) and the “Charney” problem of meteorologic. On the
the differential equation may be singular at t e origin as well as at
infinity; this is true for the Laguerre and Whittaker ~~ge~pr~b~emsof Section 5.
normal cases are “natural-natural” or ““essential-natural” for the boun-
itions at the origin and at infinity, res
When both boundary conditions are natural, tion.
The only constraints on the boundary behavior are the i osedl
by the differential equation.

4. SOLUTIONSWHICH DECAY ALGEBRAICALLY WITH

y OR ASYMPTOTETO A CONSTANT

Laguerre functicns, which are an alternative to the rational Chebysbev f~~~t~o~~;~

also have series coefficients that decay exponentially with n if S(y), the solution
70 JOHN P. BOYD

being expanded, decays exponentially with y as y -+ co. However, the Laguerre

functions are the product of exp( -y/2) with the nth Laguerre polynomial, ~0 a]1
these basis functions decay exponentially with y for large y. Iff(y) does not decay
exponentially, too-if instead, f(y) decreases as some inverse power of Y, or
perhaps asymptotes to a constant-then the mismatch in asymptotic behavior
between the function and the Laguerre basis will cause the Laguerre coefficients to
decreasealgebraically with n.
Because the i%,(y) and its derivatives all decrease as inverse powers of y (or
tend to a constant), they retain the property of exponentially rapid convergence
even for such slowly decaying f(y), and are therefore the method of choice for this
situation. In Section 5, we shall discuss a couple of examples where we exploit this
strength of the rational Chebyshev functions to compute very accurate
approximations for the ground state eigenvalue of the quantum anharmonic
oscillator and for the Bessel function J,(y).
There is, alas, a restriction. If f(y) oscillates, then the frequency of these
oscillations must tend to 0 as y + og unless f(y) decreasesexponentially fast with y.
For example,
f(v) = exp( 9) MY) (4.1)
has an infinite number of sign changes on [0, 001. To obtain an approximation
with a uniformly small absolute error E, however, it is only necessary to resolve the
oscillations on y E [0, -log(s)] because /f( y)l <E for all larger y. The T&(y) series
for (4.1) converges exponentially fast with n.
Unfortunately,
f(y)=(l+Y)1’2Jo(Y) -a(y) NY -n/4 + 4CYl), Y-+co (4.2)
is a counterexample, and its rational Chebyshev series is useless for all practical
purposes. The problem is that the amplitude of the oscillations in the Bessel
function does not decay. If the spectral series cannot resolve a region of sinusoidal
variation, even one for very, very large y, then this will give an 0( 1) absolute error.
A uniform approximation on y E CO,00] through a direct expansion is impossible.
Another way of (i) reaching the same conclusion and (ii) testing a function f(y)
for convergence is to apply ordinary Fourier convergence theory to f(L cot2(t/2)).
(The Fourier coefficients of this function of t are identical with the TL, coefficients
of f(y).) For (4.2), for example, f(L cot2(t/2)) N [constant] cos(4L/t* + constant)
as t -+ 0. Since cos(4L/t2) lacks bounded derivatives of any order at t = 0, it follows
that the Fourier (and TL,) coefficients of the Besselfunction cannot decreaseeven
as fast as 0(1/n).
Fortunately, however, there are, at least sometimes, alternatives: for (4.2) the
amplitude a(y) and the derivative of the phase d,(y) both asymptote to constants,
and these two functions do have rapidly convergent Z,(y) series. An explanation
of how to extract two expansions from one f(y) will be given in Section 5-but it
requires a special trick. A single expansion in terms of any standard basis set will
fail for (4.2).
 RATIONAL CHEBYSHEV SERIES 71

5. NUMERICAL EXAMPLES

In this section, we discuss seven problems, which subdivide into a simple boun-
dary value problem, three eigenproblems, and three expansions of known functions.
If numerical analysis is an art and not merely a science, it is hoped that this art
show will illustrate the little tricks that are important in a~~li~atio~s.
Example One: K1 Bessel Boundary Value Problem

y2u, + yu, - (1 f JJ”)u = 0, (51)

41)=K1(1) (5.2)

with the exact solution

where K,(y) is the usual “imaginary” Besselfunction. This boundary value problem
requires two slight modifications to the standard procedures explained earlier.
First, if the problem is posed on the interval YE: [yo, 00 ] instead of [O, m], then
we merely generalize (1.3) to

y = yo + L cot2(t/2), bj,if)
where in the present casey. = I. Since this is merely a translation by a constant, the
derivative formulas given in Table II are unchanged.
The second modification is that when the basis includes { TLo,..., TL,), the
collocation grid is changed to

yi= YO + L COt2(tj/2), i= (5.5a)

ti= n(2i-k 1)/(2N), i = o,..., N - 1,

which is the same form as (2.2) except that there are only N points on the interior
of [O, n]. N rows of the matrices H and f are given by (2.6) and (2.7), but the last
row is altered to

H,,,,=(-l)j, j = O,...,M, f, = u(yo) (5.6)

to impose the boundary condition. Note that we have used

TL,( yo) = cos(nn) = ( - 1)” (5.7)
in the left half of (5.6).
Boyd [3] has argued that one of the virtues of using mapped ~bebysh~~
polynomials as the basis is that the efficiency is insensitive to the map parameter L.
Figure 2, which compares three solutions computed for three values of E that
72 JOHN P. BOY73

0.0
0.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Y
FIG. 2. A comparison of numerical solutions with N= 6 (six interior collocation points plus the
boundary condition) for Example One, whose exact solution is K,(y). Exact solution [solid], L=O.25
[dotted], L= 1.25 [solid], and L = 6.25 [dashed] are shown. Note that the exact solution and the
approximation with map parameter = 1.25 are graphically indistinguishable, so both are shown as the
solid curve.

by a factor of 25, certainly confirms his assertion. The numerical solution for
L = 1.25 is indistinguishable from the exact solution to within the thickness of the
curve, but the approximations for L five times smaller and five times larger also
have small absolute errors: no worse than 0.01 for the former and 0.015 for
L = 6.25.
No comparison is made with Boyd’s formulas [3] for predicting the optimum L
becausethese are asymptotic formulas for large N-and we only need N= 6 for this
problem.
Example Two: Whittaker’s Equation Eigenproblem
The differential equation

u,+ [-l/4+ l/y+&$l u=o, ye co,001, (5.8)

where d is the eigenvalue is a special case of Whittaker’s equation. The exact
solution is u G exp( -y/2) for A = - 1 plus
u z e -“.5yyL~(y) (5.9a)
/Zrn, It>0 is an integer, (5.9b)

where Lf, is the associated Laguerre polynomial of first order and degree n. Because
the differential equation is singular at both endpoints, an unmodified TL,(y) basis,
as is appropriate for “natural” boundary conditions at both y = 0 and y = co, gave
exceptionally accurate results.
One special difficulty with eigenvalue problems (whether on an unbounded inter-
val or not) is that N basis functions will never give enough resolution to even
 RATIONAL CHEBYSHEV SERIES 73

crudely represent all of the first N eigenmodes. It follows that of t

of the pseudospectral matrix, only the lowest few will be go approximations t
those of the differential equation while the larger matrix eig dues are numerica
artifacts that must be discarded. The only effective way to determine which eigen-
values are good and which are inappropriate is to repeat the calculation with dif-
ferent N and accept only those eigenvalues that are approximately the same in both
computations.
arenthetically, we note that for this example, some of these spurious eigenvalues
complex even though all eigenvalues of the differential equation are real.
Galerkin’s method, which gives a symmetric matrix H if applied to a ~e~f-a~~~~~~
operator, is better in this respect becauseall its numerical eigenvalues are real. Even
w&h Gale&in’s, however, the large eigenvalues of the matrix are stih hopelessly
wrong, so the fact that some pseudospectral eigenvalues are complex is irrelevant
unless the eigenproblem is solved using an iterative method that requires real eigeu-
values.
The discretized matrix problem is a “generalized” eigen~r~b~em in which t
eige lue ii is multiplied by a full matrix rather than the identity matrix.
the algorithm from the EISPACK library will compute all the eigenv
eigenfunctions-real or complex-without requiring any input except the two
square matrices themselves.
Bn Figs. 3 and 4, we have arbitrarily defined a “good” eigenvalue as one which is
within 0.05 of the corresponding exact eigenvalue and then plotted the number of
“‘good” eigenvalues as a function of the map parameter E for N = 10 and ,$J= 40,
respectively. The graphs show several things: (i) as for the first example, ~GCIJ~~CJTis
rather insensitive to L. Indeed, for N= 40 one can vary L by three orders of
magnitude and still obtain the first two eigenvalues correctly! (ii) The optimum

FIG. 3. The number of “good” eigenvalues for the Whittaker eigenvalue equation for iv= 10 as a
function of map parameter L. A “good” eigenvalue is arbitrarily defined to be one which is within 0.05 of
the exact eigenvalues.
74 JOHN P. BOYD

" i 2 4 8 16 32 64 128256515
L (Map parameter)
FIG. 4. Same as Fig. 2 but for N=40.

eigenvalue increases roughly linearly with N as predicted by the steepest descent

theory of Boyd [3] for solutions which are entire functions. (iii) Boyd’s prediction
that
Loptimum=1.414N
is an overestimate-but just barely; the accuracy is so insensitive to L in the
neighborhood of the optimum L that it is difficult to judge. (iv) The fact that the
differential equation is singular at both endpoints and u(y) is singular at infinity
seemsquite irrelevant; the prediction (5.10) depends only on the knowledge that the
solutions are without singularities for any finite y.
Boyd [ 1l] offered the rule-of-thumb that in a pseudospectral eigenvalue
calculation, the number of “good” eigenvalues is roughly N/2. Figure 5 tests this

” IO 20 30 40 50 60 70
N

FIG. 5. The number of “good” eigenvalues as a function of N with L = 0.8N for the Whittaker eigen-
value problem.
 RATIONAL CHEBYSHEV SERIES 75

conjecture by plotting the number of “good” eigenvalues against N where L =

The graph clearly shows that the number of accurate eigenvalues does rise linearly
with A’, but is equal to 0.3N to 0.4N (with the larger ratio applying for larger IV),
which is a little smaller than Boyd’s rule-of-thumb. Clearly, however, the penalty
extracted by the unboundedness of the interval and the singularities of the differen-
tial equation at both endpoints is quite modest.

Example Thee: Laguerre Eigenproblem

he differential equation
yu,+ (y-t 1) u,+/lu=o (5.11)

e exact eigensolutions
u= 1, a=0 (5.123)
u = e-‘L,( y), A E n + 1, n 3 0, an integer, (5.12bj
where the L,,(y) are the usual Laguerre polynomials. This problem is clearly a close
cousin of the previous example, and the boundary conditions are “natural” at bot
endpoints. However, excluding the trivial eigenvaiue (A = 0), numerical solutions o
(5.11) were extremely disappointing: with L = 4 an N= 30, only one eigenvalue
was even close, and the approximation was the complex conjugate pair

aapprox= 0.9998 It 0.064i zz 1. (5.13)

The eigenfunction is also complex, but is a good approximation (within 1 %) of
exp( - y). Still, this is astonishingly poor accuracy for such a simple problem with
so large an N, and varying the map parameter did not produce much improvement.
The problem is that the second linearly independent solution of (5.11) has the
asymptotic form

u- l/y”, y -+ co. (5.14)

Thus, the form of the differential equation forces both solutions to vanish as y -+ co.
When no boundary constraints are applied, i.e., when the boundary condition at
infinity is treated as natural, the numerical solution is still able to dist~~~~~s
between the two solutions and converge to the one which is decaying ex~on~~tia~~y
rather than algebraically with y-but just barely. Imposing the explicit constraint
U(KI ) = 0 will be quite useless since both solutions satisfy this condition.
The difficulty can be cured by introducing the new unknown.
w(y) E eY’“u(y) g5.15)
which satisfies the differential equation

yw, + w, + [(a - %/2) - (1/4)/y] W= 0. (5.16)

76 JOHN P. BOYD

Now, the offending solution blows up exponentially instead of a slow, inverse

power decay. When (5.16) is pseudospectrally solved using N= 40, the result is 15
eigenvalues to within 3 %-and all 41 matrix eigenvalues are real whereas the
“bad” version of the same problem, (5.11), yields only three “good” eigenvalues for
N = 40 even if one is willing to ignore the small imaginary parts.
The moral of the story is quite clear: it is important to analyze the asymptotic
behavior of the differential equation and its solutions.
Example Four: “Charney” Eigenproblem
The physical background of the “Charney” problem is described in Pedlosky
[ 121. Here, it will suffice to note that Charney’s work is typical of one-dimensional
hydrodynamic stability theory, motivated in his case by the need to explain the
growth of the travelling storm systems (“synoptic cyclones,” alias “baroclinic
instability”) that are the “Highs” and “Lows” on a daily weather map. The differen-
tial equation is
(y-c)u,,+ [r-y/4+c/4] u=O, (5.17)

where y is height above the ground, r is a constant parameter (which depends on

the nondimensional longtitudinal wavenumber of the unstable modes) and c, the
eigenvalue, is the phase speed of the waves. For the special case known as the
“Boussinesq approximation,” the boundary condition is
cu,+u=o at y =O. (5.18)

Most of the principles that apply to the standard eigenproblems above must be
discarded. The only physically relevant mode is the unstable one, that is, the mode
with a complex phase speed c. The matrix has real eigenvalues, too, but [12]
explains why these should be ignored.
The unstable modes have branch points (“critical levels”) for complex y which
move closer and closer to the real interval y E [0, W] when r tends to either 0 or 1.
To track the instability in these limits, one must use the special mapping tricks
explained in Boyd [S]. Fortunately, the most interesting modes are those that grow
most rapidly since they will leave their more slowly amplifying brethren behind and
dominate the weather. For strongly unstables waves, we can obtain good accuracy
with small N and no special tricks.
The boundary condition at infinity [the top of the atmosphere] is “natural,” but
(5.18) must be explicitly imposed. Curiously, the fact that the eigenvalue appears in
the boundary condition actually simplifies life because the result of applying a
pseudospectral discretization to both the differential equation and boundary
condition can still be written in the standard form
Aa = cBa, (5.19)

where A and B are square matrices such that A- cB = H, where H is defined by

(2.6) except for the last row, where (5.18) is imposed. If the boundary condition did
 RATIONAL CHEBYSHEV SERIES 77

not involve c, then the square matrix B would have a row of zeros, so before apply-
ing the QZ algorithm, a preprocessing step to reduce the problem to standard form
would be necessary [18].
The derivative in the boundary condition, however, does require special treat-
ment because Table II gives

d%AYYdY = n { sin’(t/2)/cos( t/2) > srn(nt), (5.20)

where t = 2 arccot( [y/L] ‘1’). Unfortunately, both sin(nt) and cos(1/2) vanish at
t = 71,which corresponds to y = 0. Thus, to follow the strategy of transforming from
y to the trigonometric argument t, we must apply Wopital’s rule [or equivalently,
Taylor expand the vanishing functions and then take the ratio] to obtain
dTL,/dy lyzo = -2n2 cos(nn)/L. (521)

It then follows that the elements of the last rows of A an are given by
A, = TL,(O) = cos( jz) and B, = -dTL,/dy(O) as given by (5.21).
For Y= 0.5 (near the r of maximum growth rate), the exact phase speed for ihe
one unstable mode is

c = 0.901016+ i0.544987. (5.22)

Figure 6 shows the errors in the real and imaginary parts of c for L ranging from
L= l/l4 to t = 16 for N= 10. Despite the small number of basis functions, the
branch point in u(y) for complex y, and the tremendous variations in the map
parameter, the errors are all rather modest. For L between 0.25 and 4, the reiaiiue
error in both the real and imaginary parts is never more than 2 O/o.
Boyd [3] showed that when a solution has singularities for complex y, this will
make the optimum L much smaller. Figure 6 clearly supports this. For the hit-

000
2 4 8 16

FIG. 5. The errors in the real part (left half of each pair) and imaginary part (right half) of the eigcn-
value c for the “Charney” problem for I = 0.5 with N = 10, plotted as a function of map parameter L,.
78 JOHN P. BOYD

taker eigenproblem, the optimum LoPtimum(N)=8 for N= 10 versus only

Loptimum(lO)= 1 here even though both solutions decay as exp( -y/2). The dif-
ference is simply that the Laguerre functions are entire functions whereas the
Charney waves have branch points.
Example Five: “Global” Expansion of Smallest Eigenvalue, Quantum Anharmonic
Oscillator
For this and the next two examples, the standard way to approximate the
function is by using (i) a Taylor series for small values of the argument and (ii) an
asymptotic expansion for Jyj $1. Because the three examples have quite different
behavior for small and large y, it is impossible to extend the asymptotic series
(which is divergent!) to small y or vice versa. Nonetheless, by using mapping, we
shall obtain rapidly convergent approximations that are “global,” that is to say, are
a single expansion accurate over the whoZeinterval y E [0, co].
Boyd [ 133 gives the physical background of the anharmonic oscillator. Here, our
goal is to approximate the dependence of the smallest (“ground state”) eigenvalue
on the coupling constant A. Boyd [I33 calculated two separate Chebyshev series
approximations, one on I, E [0,0.2] and the other on [0.2, co], by solving the
problem at a set of discrete ;1 chosen to be the Chebyshev interpolation points on
each interval. To express the results in a simpler and more transparent form, these
Chebyshev approximations were converted to ordinary polynomials to obtain
E&I) z 0.50077+ 0.664202- 0.792801*, ;1E [O, 0.23
(5.23)
E,(,?) z A”‘(O.66839 + 1.41059/2-2’3- 0.005977i-4’3), /ZE[.2, CD].
Since the error in these two approximations is only 1 part in 700 and 1 part in
1500, respectively, it is clear that they are an efficient method for representing the
energy of the ground state of this quantum oscillator. ’ Nonetheless, it is still appeal-
ing to find a single expression that works for all A.
The form of (5.23) determines the form of our “global” expansion. First, since the
asymptotic expansion for large /z involves fractional powers, we define
y f A’i3. (5.24)

Second, since EO(lz)is unbounded as Iz-+ co, we apply the TL,(y) expansion to

G(Y) = Eo(v3)/U + ~1, (5.25)

where the denominator causes EJ y) to asymptote to a constant as y + co. Figure 7

is a bar graph that dramatically shows the rapid convergence of the series.

1 Parenthetically, we note a couple of minor errors in [13]: the first entry in column one of Table II
should be 1.11059... instead of 1.1059..., and also the claim that 1-” E [0,0.731 J should be replaced by
CO,2.9241. All other expressions in [13] are correct.
 RATIONAL CHEBYSHEV SERIES

7025
-u
020
015/j
0. IO
0.05
0.00 I
012345
J
n
FIG. 7. The absolute values of the coefficients of the rational Chebyshev expansion of the ground
state of the quantum quartic oscillator. The argument of the T&,(y) is y = /I’j3, where 1 is the “coupling
constant” of the oscillator and the function expanded is the ground state eigenvalue divided by (I + i.“j)
so that the TL, series is applied to an fly) which is bounded as y + m3.

Table IV lists the numerical value of the coefficients. Recalling that

I ELI d 1, y e [0, cc 1 for all M, (5.25)
we see that the error in truncating the series after a given number of terms is boun-
ded by the sum of the absolute values of all the neglected coefficients. Thus, the

TABLE IV

n ati

0 0.491194
1 0.093215
2 0.091423
3 -0.011715
4 -0.004460
5 0.00285 1
6 - 0.000334
7 -0.000360
8 0.000206
9 -0.000014
10 - 0.000037
11 0.000020
12 -0.00ooo1
13 - o.OcOOo4
14 0.000002

Note. This lists the TL,(y) coef-

ficients for the ground state eigenvalue of
the quartic oscillator of quantum
mechanics.

581/70/l-6
80 JOHN P. BOYD

three-term approximation has a relative error of no worse than 1 in 25 over the

whole semi-infinite interval-poorer than either part of (5.23), but still remarkably
good for such a low order approximation-and keeping the terms up to and
including TL,(y) gives an absolute error of less than 0.004.
Example Six: “Global” Expansion of K,( Y)
The Bessel function K,(y) has a simple pole at y = 0, so we will compute
approximations to

f(v) -.YK,(.Y). (5.27)

Unfortunately, K,(y) also has a logarithm at the origin, and this cannot be mul-
tiplied out since the coefficient of the logarithm is an infinite series [that for Z,(y)].
Fortunately, Stenger [14] has observed that it is possible to force the error to
decrease exponentially with N by using a mapping from y E [0, cc] to
YE [-co, co] of exponential character, that is to say, one which clusters the
pseudospectral interpolation points around y = 0 so that the density increases
exponentially with N. Boyd [6] notes that once the mapping has been applied, one
can use any of a variety of basis sets on YE [I-co, co] including the rational
Chebyshev functions of Boyd [ 11. To approximate (5.27), we therefore need to
combine ideas from all three of these previous works: [ 14,6, 11.
Since the K,(y) decays as exp( -y) as y -+ co, we need to transform y only for
small y, so we choose a map such that Y(v) is proportional to y as y -+ co. A simple
choice from [ 141 is
y = sinh - ‘(e “), YE[-co,m]. (5.28)
The rational Chebyshev functions TB,( Y) are in turn the images of cos(nt) under
the mapping Y = L cot(r). Thus, the coefficients of the required global expansion
are those of the Fourier cosine series of
f(t) = {sinh-‘[exp(L cot(t))] > K,{sinh-‘[exp(L cot(t))] > (5.29)
and the series approximation can be written

f(y) = C a, cos(n cot -‘[(l/L) log(sinh[Iyl)l>. (5.30)

A well-known theorem asserts that a polynomial approximation to a function

which is singular at one endpoint-even if f(v) is bounded at that point-cannot
converge faster than algebraically as the number N of terms in the series is
increased. We are able to bypass this restriction becauseour double mappings have
replaced polynomials (or cosines) by the more complicated transcendentals shown
in (5.30).
Indeed, the convergence is almost too good! Boyd Cl, 31 explains that when f(y)
is singular at an endpoint, the convergence must be “subgeometric” in the sense
 RATIONAL CHEBYSHEV SERIES 81

that the decay, although exponential with n, is slower than that of exp(-pn) for
any positive pi Figure 8 graphs the logarithm of the error bound (obtaine
adding up the absolute values of all the neglected coefficients) for (5.30). For a
series whose terms decrease like those of a geometric series (the usual case for a
Chebyshev expansion on x E [ - 1, l]), the graph would asymptote to a straight
line. When the series has only “subgeometric” convergence a defined in [4], the
graph should flatten for large n so that the slope approaches
Nonetheless, Fig. 8 approximates a straight line. If one has good eyesight, one
can detect that the slope is slowly decreasing as n increases--but for n = 24, B
approximation is already accurate to better than five decimal places and t
deviation from a straight line is still small. The concept of ““geometric” versus
“‘subgeometric” convergence is an asymptotic notion. Figure 8 shows that sometimes
the distinction between the two is detectable only for such large 12as to have no
practical significance. [Table VIII of Boyd [ 131 makes the same point.] It also
implies that the mapping treatment of a singularity as weak as y2 log(y) is
unnecessary unless one wants more than five decimal places.
This series differs from the other six examples in that we do not actually use the
2X,(y). However, the desired expansion interval is semi-infinite, and this example
does illustrate the way in which different mappings can be combined to deal with
the simultaneous problems of (i) a logarithm at an endpoint and (ii) an u~bou~de
interval.
Example Seven: Amplitude and Phase Expansions of Bessel Fwxtion JO(y)
The Besselfunction J,(y) can be approximated by a power series in y for small y
and by a divergent asymptotic series of the form
4d~) - wbyV2 {[II - 91~28~2) + ow)j COS(~ - z/4j
+ [-1/(8y)+Q(yd3)1 sin(y-n/4)). (5.31)

0 4 8 16 20 24
n
FIG. 8. A plot of the logarithm (base 10) of the error bound E, (obtained by summing the absolute
values of ail coefkients with n>N) versus N for the function yK,(y) as computed using the arcsinh-
exponential mapping with L = 1.
82 JOHN P. BOYD

Since no finite sum of basis functions can possibly hope to uniformly approximate
an irz@zite number of oscillations on y E [O, co], it is obvious that for this example,
the rational Chebyshev series and all other direct spectral approximations will fail
miserably.
The remedy is to define a modified f(v) with a multiplicative factor that
eliminates the leading square root in (5.13) and then compute two separate expan-
sions for the functions P(y) and Q(y) in

f(Y) = (1 +YP2 Jo(Y) (5.32)

= P(y) cos(y - n/4) + Q(y) sin( y - z/4). (5.33)
The form of the asymptotic expansion suggests what numerical experiments con-
firm: P(y) and Q(y) have rapidly convergent TL, expansions whose coefficients are
listed in Table V. With L = 2, the absolute error with a combined total of seventeen
coefficients is uniformly less than 2 x 10e7 on y E [0, co].
Nonetheless, there are some serious and unresolved issues hidden behind this glit-
ter of success.First, (5.33) represents one function in terms of two: how do we
choose P(y) and Q(y), out of the infinite range of possibilities, to maximize
numerical efficiency?
Krylov and Bogoliubov, as reviewed by Nayfeh [ 151, successfully solved this
uniqueness problem for perturbation theory by demanding that

u = 4.4 ~WKY~) (5.34a)

uy =. -a(y) sin($[y]). (5.34b)
Their motive is that in the asymptotic regime (y&l for our example), the
amplitude a and derivative of the phase 4 vary slowly with y. They therefore
imposed (5.34b), which would be trivially true if a and 4, were constant. (Note that

TABLE V

n P” Qn

0 1.10499 0.112125
1 -0.422833 -0.158486
2 0.157614 0.060126
3 -0.055648 -0.016579
4 0.017623 0.003121
5 - 0.004784 - 0.000308
6 0.001132
7 -0.000258
8 0.000060
9 -0.000013
10 o.oOOoO3
 RATIONAL CHEBYSHEV SERIES 83

an expression like (5.34a) can always be converted into (5.33) through elementary
trigonometry, so we shall refer to approximations in either form as “am~l~tu
phase” expansions.) This pair of equations is equivalent to
u(y) = (CL2
+ [u Y]y, (5.35)

which gives us a unique, explicit expression for the amplitude. We can expand this
as a IL!&(Y) series through either a matrix multiplication or a Fast Fourier Trans-
form after evaluating a(Y) at the usual interpolation points (2.3).
The fatal flaw is that if we define E to be their ~e~t~rbat~o~ parameter, the
Krylov-Bogoliubov amplitude function has an O(E) part which oscillates for all real
Y. For our present f(Y), one can show from (5.31) that

a’(v) - 1 -t l/Y - M4Y2)

+ (cos[2(y - 7x/4)] - 4 sin[2(Y- 7-c/4)1)/@Y’). (5.36)
Numerical experiments (not presented) show that this oscillation destroys the con-
vergence of the EL,(Y) series-the coefficients level off and do not become small
even for large n. Worse still, the Krylov-Bogoliubov procedure is only asyrn~t~t~~,
so it is impossible to refine it.
We therefore tried the simple-minded alternative of expanding P(Y) an (v) i
(5.33) as

flv) = f P,Kz(Y),
?I=0

and then imposing the collocation conditions

~YY,) GOS(Y~- ~‘4) + Q(Y~) sin(yi- n/4) =f(yi),

i=Q,..., N+M+ 1, (5.38)

where
yj = L cot2(tJ2),
tj=n(2i+1)/(2M+2N+4), i=O,..., N+M+l (5.39)

are the usual rational Chebyshev interpolation points. This gives a simple set of
linear equations to determine the unknown coeffcients {Pn> and {Q,>. (Note that
an approximation in the form u = a(y) cos(4[y]) depends nonlinearly on the
(6, so (5.38) is preferable.) Taking M= N = 8 gave a maximum absolute err
y E [0, 02] of only 0.00002.
Unfortunately, increasing M and N gave no improvement even in 16-
precision because the set of linear equations is ill-conditioned. The wo~~~~,
however, is that the interpolation method works at all. Although we can rigorously
justify the pseudospectral method for a given non-oscillatory f(y) [by usin
84 JOHN P. BOYD

change-of-coordinates from y to t and then invoking the theory of trigonometric

interpolation], there is no theory for a double expansion such as we attempt here.
The interpolation (5.37) is a shot in the dark-albeit a successful one,
Some experimentation showed that taking M > N reduced the ill-conditioning
becausethe factors of E,(y) sin(y - 7t/4) closely resemble T&(Y) cos(y - 7~/4)for
large n, which is why Table V lists 11 coefficients for the multiplier of cos(y - 7~/4)
and only 6 for that of sin(y - n/4). Additionally, we improved the stability of the
calculation by using the interpolation points (5.3) to define a discrete inner product
(an approximation to the integral on [0, co] with the weight given in (1.4)) and
then applied a Gram-Schmidt orthogonalization to define new basis functions.
After applying this least-squares procedure using 30 quadrature points and 21 basis
functions, we converted the coefficients back into the form of (5.37). Figure 9 shows
the amplitude and phase functions when (5.33) is converted into f(y)=a(y)
COS(tKYl)~
When we replace interpolation by a least-squares procedure, we are guaranteed
that adding more terms will reduce the residual at least at the quadrature points.
However, there is still no theory to guarantee that minimizing errors at any finite
set of points will give small errors over the continuous interval y E [0, co]. In
addition, M and N are still limited because of as yet poorly understood round-off
errors.
In short, this amplitude-phase approximation would make a pure mathematician
recoil in horror. Nonetheless, it gives seven decimal places of accuracy. Clearly, it is
possible for creative use of interpolation to give approximations on unbounded
intervals that mimic asymptotic expansions and the WKB method, and yet have (i)
a much wider range of usefulness and (ii) are free of the curse of asymptoticity,
which limits the accuracy of WKB formulas for fixed y even when y is large.

--._ ------
0.01 “-r-+++ l--Lb-__
0.0 I.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
Y
FIG. 9. The amplitude (solid line) and phase (dashed) in the amplitude-phase approximation to
f(r) = (1 +JJP* Jo(v).
 RATIONAL CHEBYSHEV SERIES 85

6. SUMMARY
y making a change-of-coordinate, we have shown that Chebyshev poIy~omiaIs
(or eq~~va~e~tly,a Fourier cosine series) can be appl to a semi-infinite interval as
tirst suggested by Grosch and Orszag [2]. By explo g the mapping
nection between Fourier cosine series and the new basis functions deli
rational Chebyshev functions Z!&(y), we are able to fill in the gaps of
In the Appendix, we display the “equiconvergence” contours in the complex
plane. In the body of the text, we discuss the issues of natural versus essential boun-
dary co tions, banded Galerkin matrices, and explain how the 2&(y)-in con-
trast to guesre functions-can be applied to bounded functions with asymptotic
expansions in inverse powers of y as 4’ 4 03.
The seven numerical examples illustrate the possibilities. The Cba~~ey
merical weather prediction, which usesmixed spectral-elite
ical harmonics in the horizontal, fourth order differences i
can instead employ pure pseudospectral codes by using the TL., in h
success of the uniform amplitude-phase expansion for Jo(y) shows
possible to replace the WKB method by an approximation which is ~~~~~~~~~~
than asymptotic and which can be applied for a2.ly, not just when y is large.
retical work is needed to put such “‘double expansions” on a firm
tion, we have not discussed condition numbers, the eigenvalue spec-
ditioning” methods for pseudospectral matrix problems. In one
-conditioned iterative procedures are unnecessary (most of the
calculations in article took less than five seconds on an
dimensions, the best choice of pre-conditioner is still a resea
ordinary Chebyshev expansions [16J.
Similarly, we have avoided the important ic of time-stepping algorithms com-
bined with EL,(y) series for the spatial coo nates. As emphasized by Tre~e~~e~l
and Trummer [I191, the eigenvalue spectr of the first d&vat& operator is
poorly understood even for ordinary Chebyshev polynomials. At
know how to predict whether a pseudospectral representatio of a di~eKe~tia1
operator will have complex eigenvalues (which could cause time-step
~~StabiIi~y)even for a self-adjoint equation; the examples above show only that th,e
spectrum is sometimes purely real-and sometimes not.
Clearly, orthogonal rational functions offer many topics for future researc
etheless, the examples and analysis given here show that these mapped Chebyshev
polynomials are an extremely effective tool for solving difhcult problems on a semi-
infinite interval.

APPENDIX: CONTOURS OF EQUICONVERGENCE

The domain of convergence of a Fourier cosine series is a strip parallel to the real
t-axis,that is, the series converges within the region
Itil dP Fourier domain of convergence (A,?)
86 JOHNP.BOYD

where p is the largest possible positive constant such that the interior of the strip
(but not its boundary) contains no singularities of the function f(t). Although we
are usually interested in evaluating a function only for real t, (A.l) is of con-
siderable practical importance because the asymptotic rate at which the Fourier
coefticients decrease with n is directly related to the width of the strip of the
convergence via
a, - I: l(VW, n-+co, (A.21
where the empty [ ] denote functions that vary more slowly than S-H and where
161= ep. (A.3)
Thus, the wider the strip of convergence, the faster the rate at which the Fourier
coefficients decrease.This motivates the following.

DEFINITION. The “contours of equiconvergence” are the curves in the complex

plane where the asymptotic rate of convergence is a constant, that is, are the
contours of constant 6 (or p).
For Fourier series, the equiconvergence contours are straight lines parallel to the
real t-axis. The “equiconvergence” contours for the X,(y) series are the images of
these straight lines under the mapping t + y. Therefore, if f(y) has a singularity at
the point y = yr f iyi, then the width of the strip of convergence in the t-plane is
p=Im{2arccot[([y,i-iy,]/L)1’2]} (A.41
= -~/~~~~~cs~+~i~+~~21/[Is,2+~~i-~~21} (A.51

via the logarithmic representation of the arccot function where

r = (y/L)“‘. (‘4.6)

Knowledge of p(y,, yi) determines the asymptotic behavior of the TL, series via
(A.2) and (A.3). When the location of the singularities of f(y) are known in
advance, either becausef(y) itself is known or because it satisfies a linear differen-
tial equation with coefficients whose singularities are known, then one can evaluate
(A.4) for each branch point or pole, and the smallest p so obtained determines the
region of convergence. Figure 10 shows the contours of p(y) in the complex
y-plane.
One can directly trace these contours by defining
srsinh(p) (A.7)
and then solving the equation (with L = 1 for simplicity)
y?[ -s4] + 2y?[2 - s4 - 2s2(.? + 1) yr - s4yf]
+s”(y,+ 1)” { -4y,-?(y,+ 1)Z) =o (A.81
 RATIONAL CHEBYSHEV SERIES

-2 0 2
Re(y)

FIG. 10. The equiconvergence contours in the upper half of the complex y-plane for the rationai
Chebyshev functions on the semi-infinite interval. The map parameter L = I; only the upper half-plane is
shown becausethe curves in the lower half-plane are the reflection with respect to the real y-axis of those
for Im(y) > 0.

for yf as a function of y, and p. Equation (A.7) is derived by decomposing t

trigonometric functions in (1.3) into their real and imaginary parts, solving for
cos(t,) and sin(t,) and then applying the identity cos(t,j2 + sin(l,j2 = 1 to eliminate
the real part of t. Although (A.7) can be solved as a quadratic in yf, the equation is
nonetheless quartic, which implies that the equiconvergence contours for the
T&,(y) are more complicated than the ellipses that play the same role for the
ordinary Chebyshev polynomials, 7’,(x).
The reason that this analysis has been banished to this appendix is that unf~r-
tunately it applies only to functions that are analyk at infinity, such as certain
rational functions and other simple forms. If f(y) has poles, branch points, or
essential singularities at infinity, then the TL,(y) series will converge only on the
real interval YE LO,co]. Equation (A.2) does not describe the asymptotic form of
the coefficients and the contours in Fig. 10 become irrelevant.
Fortunately, if f(y) has all its derivatives bounded at infinity, the series coef-
ficients will still decrease exponentidy with n, but the (asymptotic!) rate of con-
vergence will be “subgeometric” in the sense defined in yd 141. Characterizing
this requires the more subtle steepest descent analysis o oy$ [3,4] rather than
the equiconvergence contours described here.

ACKNOWLEDGMENT

This work was supported by the NSF under Grant OCE8509923

88 JOHN P. BOYD

REFERENCES

1. J. P. BOYD, J. Comput. Phys. 69, 112-142 (1987).

2. C. E. GROSCH AND S. A. ORSZAG, J. Comp. Phys. 25, 273 (1977).
3. J. P. BOYD, J. Comput. Phys. 45, 43 (1982).
4. J. P. BOYD, J. Comput. Phys. 54, 382 (1984).
5. J. P. BOYD, J. Comput. Phys. 51, 454 (1985).
6. J. P. BOYD, J. Comput. Phys. 64, 266 (1986).
I. H. J. NORTON, Comput. J. I, 76 (1964).
8. J. P. BOYD, Physicn D 21, 227 (1986).
9. D. GOTTLIEB AND S. A. ORSZAG, “Numerical Analysis of Spectral Methods: Theory and
Applications” (SIAM Philadelphia, 1977).
10. A. B. CAIN, J. H. FERZIGER, AND W. C. REYNOLDS, J. Comput. Phys. 56, 272 (1984).
11. J. P. BOYD, Monthly Weather Rev. 106, 1192 (1978).
12. J. PEDLOSKY, “Geophysical Fluid Dynamics” (Springer-Verlag, New York, 1979).
13. J. P. BOYD, J. Math. Phys. 19, 1445 (1978).
14. F. STENGER,SIAM Rev. 23, 165 (1981).
15. A. H. NAYFEH, “Perturbation Methods” (Wiley, New York, 1973).
16. C. CANUTO AND A. QUARTERONI, J. Comput. Phys. 60, 315 (1985).
17. J. P. BOYD, J. Sci. Comput. 1, 183 (1986).
18. J. GARV AND R. HELGASON, J. Comput. Phys. 5, 169 (1970).
19. L. N. TREFETHENAND M. R. TRUMMER, SIAM J. Numer. Anal. (1986), in press.