error estimates for rayleigh?ritz approximations of eigenvalues and eigenfunctions of the mathieu...
TRANSCRIPT
DOI: 10.1007/s00365-002-0527-9
Constr. Approx. (2004) 20: 39–54
CONSTRUCTIVE
APPROXIMATION© 2003 Springer-Verlag New York Inc.
Error Estimates for Rayleigh–Ritz Approximations ofEigenvalues and Eigenfunctions of the Mathieu and
Spheroidal Wave Equation
Hans Volkmer
Abstract. Error estimates are derived for the computation of eigenvalues and eigen-
vectors of infinite tridiagonal matrices by the Rayleigh–Ritz method. The results are
applied to the Mathieu and spheroidal wave equation.
1. Introduction
It is well-known that self-adjoint operators associated with the Mathieu and spheroidal
wave equation can be represented by infinite symmetric tridiagonal matrices. Since pow-
erful software packages are widely available that compute eigenvalues of finite symmet-
ric tridiagonal matrices with almost arbitrary precision, it is natural to approximate the
eigenvalues of infinite symmetric tridiagonal matrices by eigenvalues of suitably cho-
sen finite submatrices. This method was proposed by Green and Michaelson [8] for the
Mathieu equation and by Hodge [9] for the spheroidal wave equation. Today the method
is widely employed by programs in software libraries [5], [16], [19]. Recently it was
used by Falloon [7] in his development of a software package for spheroidal wave func-
tions for Matematica. Of course, the method is just a special case of the Rayleigh–Ritz
method; see Weinstein and Stenger [17]. The Rayleigh–Ritz method is also mentioned
by the author in his chapter on spheroidal wave functions in the Digital Library of Mathe-
matical Functions (DLMF), the new edition of the Handbook of Mathematical Functions
by Abramowitz and Stegun [1].
The Rayleigh–Ritz method yields a decreasing sequence µr that approximates the
desired eigenvalue λ, and a sequence gr of elementary functions that approximates the
desired eigenfunction f (Mathieu or spheroidal wave function). The approximations µr
and gr are computed from an r × r submatrix of an infinite matrix. When writing the
chapter on spheroidal wave functions the author and the editors of the DLMF project
noticed several gaps in our knowledge about the convergence properties of the method.
For example, the editors asked whether the sequence gr converges uniformly to f as
Date received: January 16, 2002. Date revised: May 28, 2002. Date accepted: September 12, 2002. Commu-
nicated by Mourad Ismail. Online publication: February 21, 2003.
AMS classification: 33E10, 65L60, 65L70.
Key words and phrases: Rayleigh–Ritz method, Mathieu equation, Spheroidal wave equation, Tridiagonal
matrices, Error estimates.
39
40 H. Volkmer
r → ∞. More ambitiously, one may ask for error bounds for the maximum norm
‖gr − f ‖∞. Such bounds are needed when we have to decide how to truncate the infinite
matrix to a finite matrix for a prescribed precision. Another natural question is to ask for
the rate of convergence µr → λ and gr → f as r → ∞. This paper provides answers
to these questions, answers that the author was unable to find in the literature.
We know more about corresponding problems for the Mathieu equation but our knowl-
edge is also not satisfactory. There are lower bounds for λ derived from the well-known
method of intermediate problems due to Aronszaijn and Weinstein; see Weinstein and
Stenger [17, Chapter 5]. However, it seems impossible to estimate the rate of conver-
gence of µr → λ by this method. Of course, the method of intermediate problems is
applicable to a much larger class of eigenvalue problems than that treated in this paper.
Green and Michaelson [8] and the “Numerical Analysis” group at Delft [15] provide
error estimates for the computation of eigenvalues of the Mathieu equation but give no
rate of convergence and no error estimates for the approximations of eigenfunctions.
The estimates are different from those given in this paper. The derivation in [8], [15] is
based on the following simple lemma due to Wilkinson [18]:
Let A be an n × n symmetric matrix, µ ∈ R and x ∈ Rn , with Euclidian norm ‖x‖ = 1.
Let ε := ‖Ax − µx‖. Then there is an eigenvalue λ of A such that |λ − µ| ≤ ε.
Unfortunately, we face a technical problem when we try to apply this lemma. If we order
the eigenvalues of A is ascending order as λ1 ≤ λ2 ≤ · · · ≤ λn , then the lemma tells
us that there is i ∈ {1, . . . , n} such that |λi − µ| ≤ ε but the lemma does not tell us the
value of the subscript i . But knowledge of this value is required in the derivation of error
bounds for the Rayleigh–Ritz method. Papers [8] and [15] make a natural assumption
what i should be. In order to avoid this indexing problem we base the analysis on the
minimum–maximum principle for eigenvalues of self-adjoint operators. This principle
allows identification of the i th eigenvalue λi . Our error bounds are sharper than those in
[8] and [15] as indicated in Section 4.
Blanch [6] treats error estimates for the computation of eigenvalues and eigenfunctions
of the Mathieu equation based on continued fractions. However, in [15, p. 41] it is
stated that this method in some cases will not produce the correct result. The method
requires the knowledge of a “sufficiently good approximation” for the desired eigenvalue
λ as an initial value for a recursive procedure. Such approximations are usually derived
from power series and asymptotic expansions of eigenvalues. Alhargan [2], [3] provides
an elaborate system of initial approximations whose suitability is tested by numerical
experimentation. Falloon [7] points out that, today, it is easier to find such approximations
using the Rayleigh–Ritz method. So even if we use continued fractions an error estimate
of the Rayleigh–Ritz approximation is useful. The main advantage of using the Rayleigh–
Ritz method to compute eigenvalues and eigenfunctions of the Mathieu and spheroidal
wave equation is that we can rely on sophisticated linear algebra software packages while
almost no additional programming is needed.
Shirts [13], [14] shows how to compute a Mathieu function of noninteger order using
finite submatrices of doubly infinite matrices. Errors are investigated by numerical ex-
perimentation. It would be of interest to derive rigorous error bounds along the lines of
this paper.
Error Estimates for Rayleigh–Ritz Approximations 41
2. The Rayleigh–Ritz Method for Self-Adjoint Operators
Let H be a Hilbert space with inner product (·, ·) and norm ‖ · ‖ and let A : H ⊃D(A) → H be a self-adjoint operator, bounded below with compact resolvent; see [10].
Then the spectrum of A agrees with the set of eigenvalues of A, it is bounded below,
and has no finite point of accumulation. Therefore, we may order the eigenvalues of A
in ascending order and counted according to multiplicity as
λ1 ≤ λ2 ≤ λ3 ≤ · · · .
The minimum–maximum principle [17, p. 12] states that, for all k ∈ N,
λk = minEk
max{(Ax, x) : x ∈ Ek, ‖x‖ = 1},(2.1)
where the minimum is taken over all linear subspaces Ek of D(A) with dimension k.
Let E be a linear subspace of D(A) of finite dimension r , let P : H → H be the
orthogonal projection of H onto E , and let Q := I − P . Let
µ1 ≤ µ2 ≤ · · · ≤ µr
denote the eigenvalues of the operator PA : E → E counted according to multiplicity.
Again we have the minimum–maximum principle
µk = minFk
max{(Ax, x) : x ∈ Fk, ‖x‖ = 1},(2.2)
where the minimum is taken over all linear subspaces Fk of E with dimension k.
Equations (2.1) and (2.2) imply the Poincare inequalities [17, p. 10],
λk ≤ µk, k = 1, . . . , r.(2.3)
We wish to estimate µk − λk . Let x1, x2, x3, . . . denote an orthonormal basis of H
such that Axj = λj x j . We introduce k × k Hermitian matrices L = (L ij), M = (Mij) by
L ij := (Qxi , x j ), Mij := ((PAP − A)xi , x j ), i, j = 1, . . . , k.
We assume that ‖L‖ < 1, where ‖L‖ denotes the spectral norm of L . Then the vectors
Px1, . . . , Pxk are linearly independent and, thus, Fk := span{Px1, . . . , Pxk} is a linear
subspace of E with dimension k. We use this space Fk in (2.2) in order to derive an upper
bound for µk . Let
0 �= w =k
∑
i=1
αi xi .
Since
(Aw, w) =k
∑
i=1
λi |αi |2 ≤ λk‖w‖2 = λk(‖Pw‖2 + ‖Qw‖2),
42 H. Volkmer
we obtain
(APw, Pw) − λk‖Pw‖2 ≤ ((PAP − A)w, w) + λk‖Qw‖2
≤ ‖M + λk L‖‖w‖2.
Now
‖w‖2 = ‖Pw‖2 + ‖Qw‖2 ≤ ‖Pw‖2 + ‖L‖‖w‖2
yields
(APw, Pw) − λk‖Pw‖2 ≤‖M + λk L‖
1 − ‖L‖‖Pw‖2.
This inequality combined with (2.2) implies the following lemma:
Lemma 1. If ‖L‖ < 1, then
µk − λk ≤‖M + λk L‖
1 − ‖L‖.
Let y1, . . . , yr be an orthonormal basis of E such that PAyi = µi yi and (xk, yk) ≥ 0.
We wish to estimate ‖yk − xk‖. Write
Pxk =r
∑
i=1
βi yi , βi = (xk, yi ).
Then we have
r∑
i=1
|βi |2(µi − λk)2 =
∥
∥
∥
∥
∥
r∑
i=1
βi (µi − λk)yi
∥
∥
∥
∥
∥
2
= ‖PAPxk − PAxk‖2.
Let
K := ‖PAPxk − PAxk‖, := rmaxk �=i=1
‖µi − λk |−1,
and assume that < ∞. Then
r∑
k �=i=1
|βi |2 ≤ 2 K 2.(2.4)
Since
1 − ‖Qxk‖2 = ‖Pxk‖2 =r
∑
i=1
|βi |2,
we obtain, from (2.4),
1 − ‖Qxk‖2 − 2 K 2 ≤ β2k .
Hence
0 ≤ 1 − βk ≤ 1 − β2k ≤ ‖Qxk‖2 + 2 K 2
Error Estimates for Rayleigh–Ritz Approximations 43
and
‖Pxk − yk‖2 =r
∑
k �=i=1
|βi |2 + |βk − 1|2
imply the following lemma:
Lemma 2. We have
‖Pxk − yk‖2 ≤ 2 K 2 + (‖Qxk‖2 + 2 K 2)2
and
‖xk − yk‖2 ≤ ‖Qxk‖2 + 2 K 2 + (‖Qxk‖2 + 2 K 2)2.
3. The Rayleigh–Ritz Method for Infinite Tridiagonal Matrices
Let H = l2(N) with natural orthonormal basis en, n ∈ N. Let cn, dn, n ∈ N, be two real
sequences such that cn is bounded, cn �= 0 for all n, and dn converges to infinity. Let
S : H ⊃ D(S) → H be the linear operator defined by Sen = dnen with domain
D(S) =
{
∞∑
n=1
unen ∈ H :
∞∑
n=1
|dnun|2 < ∞
}
.
Let T : H → H be the linear operator defined by Ten = cn−1en−1 + cnen+1, where
c0 := 0. Then S is self-adjoint, bounded below with compact resolvent, while T is
bounded and self-adjoint. It follows from [10, p. 291] that A := S + T is self-adjoint,
bounded below with compact resolvent. Formally, A is given by the infinite symmetric
tridiagonal matrix
A =
d1 c1 0 0 0 . . .
c1 d2 c2 0 0 . . .
0 c2 d3 c3 0 . . .
......
......
...
.(3.1)
We denote the eigenvalues and eigenvectors of A as in Section 2 and we set xi,n :=(xi , en). Let Pr denote the orthogonal projection of H onto Er := spanr
j=1 ej and let
Qr := I − Pr . Let
µ1,r < µ2,r < · · · < µr,r
denote the eigenvalues of the r × r tridiagonal matrix Ar consisting of the first r rows
and columns of (3.1). It should be mentioned that other choices of principal submatrices
of A are possible which sometimes may lead to better error estimates.
Let y1,r , . . . , yr,r be an orthonormal basis of Er such that Ar yj,r = µj,r yj,r and
(yj,r , x j ) ≥ 0. In order to estimate µk,r −λk and ‖yk,r −xk‖ we assume given computable
quantities �i , �i,k, i = 1, . . . , k, such that
λi ≤ �i , λk − λi ≤ �i,k .
44 H. Volkmer
In applications we will use �i = µi,s with s ≥ i which is possible by the Poincare
inequality (2.3). For an estimate of �i,k it will be sufficient to note that
λ1 ≥∞infj=1
(dj − |cj | − |cj−1|).
Lemma 3. The matrix elements of L , M , and the quantity K (corresponding to P = Pr )
admit the estimates
|L ij| ≤ ‖Qr xi‖‖Qr x j‖,
|Mij + λk L ij| ≤ |cr ||xi,r+1||x j,r | + �i,k‖Qr xi‖‖Qr x j‖,
K ≤ |cr ||xk,r+1|.
Proof. The estimate of L ij = (Qr xi , Qr x j ) follows from the Cauchy–Schwarz in-
equality. Since
APr xi = −cr xi,r+1er + cr xi,r er+1 + λi Pr xi ,
we obtain
Pr APr xi − Axi = −cr xi,r+1er − λi Qr xi .
Thus
|Mij + λk L ij| ≤ |cr ||xi,r+1||x j,r | + (λk − λi )|L ij|
≤ |cr ||xi,r+1||x j,r | + �i,k‖Qr xi‖‖Qr x j‖.
The estimate for K follows from
Pr APr xk − Pr Axk = −cr xk,r+1er .
Let i ∈ N. Choose ℓi ≥ 2 so large that
dn > |cn−1| + |cn| + �i for n ≥ ℓi .(3.2)
For n ≥ ℓi we define f(0)
i,n := 1 and then, recursively,
f(q+1)
i,n :=|cn−1|
dn − �i − |cn| f(q)
i,n+1
.
By induction on q we see that
1 ≥ f(q)
i,n > f(q+1)
i,n > 0.
The following lemma estimates the eigenvector xi of A; see [12, p. 91] for a similar
estimate:
Lemma 4. For n ≥ ℓi and q ≥ 0, we have
|xi,n| ≤ f(q)
i,n |xi,n−1|.(3.3)
Error Estimates for Rayleigh–Ritz Approximations 45
Proof. Since the sequence xi,n, n ∈ N, is in l2(N), there are arbitrarily large n ≥ ℓi
such that |xi,n+1| ≤ |xi,n|. Then
cn−1xi,n−1 + (dn − λi )xi,n + cn xi,n+1 = 0(3.4)
implies that
|xi,n| ≤|cn−1|
dn − �i − |cn||xi,n−1| ≤ |xi,n−1|.
This demonstrates that |xi,n| ≤ |xi,n−1| for all n ≥ ℓi . Hence (3.3) is true for q = 0. Now
induction on q using (3.4) proves (3.3) for all q .
We pick q > 0 and set
fi,n := f(q)
i,n .(3.5)
The following error estimates improve when we increase q although the choice q = 1
or q = 2 will usually be sufficient.
Lemma 5. For i ∈ N and n ≥ ℓi , we have
|xi,n| ≤ ai,n :=n
∏
j=ℓi
fi, j ,(3.6)
‖Qn xi‖ ≤ bi,n := (1 − f 2i,n+2)
−1/2ai,n+1.(3.7)
Proof. Since |xi,n| ≤ 1 for all n, we obtain (3.6) from (3.3). Moreover,
‖Qn xi‖ ≤ |xi,n+1|(1 + f 2i,n+2 + f 2
i,n+2 f 2i,n+3 + · · ·)1/2.
Since 0 < fi, j < 1 for all j ≥ ℓi , we can estimate the sum under the root by a geometric
series to obtain the desired bound (3.7).
In order to estimate µk,r − λk we now proceed as follows. For the choice of ℓi , i =1, . . . , k, we have to verify (3.2) which comprises infinitely many conditions. Therefore,
we assume given a computable sequence Dn → ∞ and a computable bounded sequence
Cn such that
Dn ≤ Dn+1, Cn ≥ Cn+1, dn ≥ Dn, |cn| ≤ Cn for n ∈ N.(3.8)
Then (3.2) is satisfied if
Dℓ > Cℓ + Cℓ−1 + �i , ℓ = ℓi .
Next, for r ≥ ℓk , we estimate the entries of the k × k matrices L and M + λk L by
applying Lemmas 3 and 5. Then we bound the spectral norms of these matrices by their
Schur norms. Finally, we apply Lemma 1 to estimate µk,r − λk .
46 H. Volkmer
This way we obtain the following theorem:
Theorem 1. Using the quantities ai,n and bi,n from Lemma 5 we have, for r ≥ ℓk ,
‖L‖ ≤k
∑
i=1
b2i,r ,
‖M + λk L‖ ≤ |cr |
(
k∑
i=1
a2i,r
k∑
j=1
a2j,r+1
)1/2
+
(
k−1∑
i=1
�2i,kb2
i,r
k∑
j=1
b2j,r
)1/2
.
If ‖L‖ < 1, then
µk,r − λk ≤‖M + λk L‖
1 − ‖L‖.
In order to estimate ‖yk,r − xk‖, we find an upper bound for K according to Lemmas 3
and 5. We derive an upper bound for
= rmaxk �=i=1
|µi,r − λk |−1
by using the estimate for λk obtained above. Then we apply Lemma 2.
Theorem 2. If r ≥ ℓk , we have
‖yk,r − xk‖2 ≤ b2k,r + 2c2
r a2k,r+1 + (b2
k,r + 2c2r a2
k,r+1)2.
Several obvious simplifications in these estimates lead to somewhat cruder error
bounds. For example, we may use that ai,n and bi,n are increasing in i .
We note two results on the rate of convergence that follow from these estimates.
Theorem 3. Let k ∈ N. For every ε > 0, we have
µk,r − λk = O(εr ) as r → ∞
and
‖yk,r − xk‖ = O(εr ) as r → ∞.
The O-constants may depend on k but are independent of r .
Theorem 4. Assume that dn > 0 for all n (which we may achieve by replacing dn by
dn + τ for sufficiently large τ ), and that
∞∑
n=1
1
dn
< ∞.
Let k ∈ N. Then we have
µk,r − λk = O
(
c21 · · · c2
r
d21 · · · d2
r dr+1
)
as r → ∞,
Error Estimates for Rayleigh–Ritz Approximations 47
and
‖yk,r − xk‖ = O
(
|c1 · · · cr+1|d1 · · · dr+1
)
as r → ∞.
The O-constants may depend on k but are independent of r .
It should be noted that explicit expressions for the O-constants in Theorems 3 and 4
can be derived from the estimates in Lemmas 1, 2, 3, and 5.
4. Application to the Mathieu Equation
We consider the problem of finding eigenvalues λ for which the Mathieu equation
w′′ + (λ − 2h2 cos(2t))w = 0(4.1)
has nontrivial odd solutions of period π . Here h2 is a given real parameter. Other self-
adjoint eigenvalue problems for the Mathieu equation can be treated similarly. The
well-known representation of the differential operator
w �→ −w′′ + 2h2 cos(2t)w
with respect to the orthonormal basis
2√
πsin(2nt), n ∈ N,
in L2(0, π/2), is of the form (3.1) where
cn = h2, dn = 4n2, n ∈ N.
The eigenvalue λk of the corresponding operator A agrees with the Mathieu eigenvalue
b2k(h2) in the notation of Arscott [4, p. 53].
By Theorem 4, we obtain, for fixed k and h2,
µk,r − λk = O
(
h4r
42r+1r !2(r + 1)!2
)
as r → ∞.
The O-constant can be computed from Theorem 1 for given k and h2. For practical
work a lookup-table of such O-constants for various ranges of k and h2 could be created.
For example, if k = 1 and −1 ≤ h2 ≤ 1, using �1 = µ1,1 = 4, ℓ1 = 2, and q = 1 in
(3.5), we obtain
a1,n ≤ b1,n−1 ≤ 2.425
(
|h2|4
)n−11
n!2
which leads to
µ1,r − λ1 ≤94h4r
42r+1r !2(r + 1)!2for r ≥ 3.(4.2)
48 H. Volkmer
Table 1. Estimates from Theorem 1 for distance between Mathieu eigenvalue
b2(10) and approximation µ1,r .
r µ1,r µ1,r − b2(10) Error estimate
5 −2.3821550258499152723614319 0.32 · 10−5 1.16 · 10−5
6 −2.3821582248055205665354767 0.11 · 10−7 0.40 · 10−7
7 −2.3821582359350987602677340 0.22 · 10−10 0.79 · 10−10
8 −2.3821582359569297267420560 0.26 · 10−13 0.94 · 10−13
9 −2.3821582359569556955120616 0.20 · 10−16 0.71 · 10−16
10 −2.3821582359569557153243100 0.10 · 10−19 0.36 · 10−19
11 −2.3821582359569557153344461 0.36 · 10−23 1.30 · 10−23
This estimate is quite sharp. For example, if r = 10, h2 = 1, then
λ1 = 3.91702477299847118670341688529379991002933424614053 . . . ,
µ1,10 = 3.91702477299847118670341688529379991002989666177165 . . . ,
so the left-hand side of inequality (4.2) is approximately 0.56·10−39 whereas the estimate
on the right-hand side gives 1.01 · 10−39. These and the following computations were
carried out in high precision arithmetic with Maple.
Table 1 shows error estimates in the case k = 1, h2 = 10 for the computation of the
Mathieu eigenvalue
b2(10) = −2.3821582359569557153344497960307538858413 . . . .
We use �1 = µ1,5 and q = 2 in (3.5).
Table 2 shows error estimates in the case k = 6, h2 = 10 for the computation of the
Mathieu eigenvalue
b12(10) = 144.3502080084893260019912658518645935310531 . . . ,
using �i = µi,i+4 for i = 1, . . . , 6 and q = 2 in (3.5).
Surprisingly, the quality of the error bounds, when compared with the true error,
improves if k is increased. As was to be expected, the speed of convergence becomes
Table 2. Estimates from Theorem 1 for distance between Mathieu eigenvalue
b12(10) and approximation µ6,r .
r µ6,r µ6,r − b12(10) Error estimate
7 144.3816507440327643267900414 0.31 · 10−1 0.40 · 10−1
8 144.3503676402783310493153412 0.16 · 10−3 0.19 · 10−3
9 144.3502083582352186876740700 0.35 · 10−6 0.40 · 10−6
10 144.3502080088932724039297084 0.40 · 10−9 0.46 · 10−9
11 144.3502080084896019655539877 0.28 · 10−12 0.31 · 10−12
12 144.3502080084893261223614652 0.12 · 10−15 0.14 · 10−15
13 144.3502080084893260020266843 0.35 · 10−19 0.40 · 10−19
Error Estimates for Rayleigh–Ritz Approximations 49
Table 3. Estimates from Theorem 1 for distance between Mathieu eigenvalue b2(100) and µ1,r .
r µ1,r µ1,r − b2(100) Error estimate
7 −141.26840751579981679504 0.12 · 10−1 0.82 · 10−1
8 −141.27933852281769510426 0.72 · 10−3 5.06 · 10−3
9 −141.28002555042468333569 0.31 · 10−4 2.20 · 10−4
10 −141.28005582197545941739 0.99 · 10−6 6.94 · 10−6
11 −141.28005678552116753108 0.23 · 10−7 1.62 · 10−7
12 −141.28005680821069021504 0.41 · 10−9 2.88 · 10−9
13 −141.28005680861389319213 0.56 · 10−11 3.91 · 10−11
worse when h2 is increased. Table 3 shows error bounds in the case k = 1, h2 = 100 for
the computation of the Mathieu eigenvalue
b2(100) = −141.2800568086194528253587605820418627831331 . . . ,
using �1 = µ1.5 and q = 2 in (3.5).
In [15, p. 45] it is mentioned that the computation of the Mathieu eigenvalue b2(1000)
with double precision requires a 515 × 515 matrix based on the error estimates given in
[8], compared with a 57 × 57 matrix based on the error estimates derived in [15]. Since
b2(1000) = −1811.5224151493534312803605191615791051136948 . . . ,
double precision requires µ1,r − b2(1000) < 10−13. Theorem 1 with �1 = µ1,10 and
q = 4 in (3.5) gives the estimate µ1,26 − b2(1000) < 0.59 · 10−13, so a 26 × 26 matrix is
sufficient. In fact, a 26 × 26 matrix is the smallest matrix that is needed to approximate
b2(1000) within an error less than 10−13. For 57×57 and 515×515 matrices, we obtain
µ1,57 − b2(1000) < 0.32 · 10−71 and µ1,515 − b2(1000) < 0.12 · 10−2274.
The Mathieu eigenfunction se2k(t) = se2k(t, h2)belonging to the eigenvalue b2k(h2) =
λk is given by
se2k(t) =∞
∑
n=1
xk,n sin(2nt),
where xk = (xk,1, xk,2, . . .) is the eigenvector of A belonging to the eigenvalue λk and∑∞
n=1 x2k,n = 1; see [4, p. 55]. We use the eigenvector yk,r = (u1, . . . , ur , 0, 0, . . .)
of Pr APr belonging to the eigenvalue µk,r to approximate se2k by the trigonometric
polynomial
pk,r (t) =r
∑
n=1
un sin(2nt).
The L2(0, π/2)-norm
‖se2k − pk,r‖ =√
π
2‖xk − yk,r‖
can be estimated as indicated in Section 3. In particular, by Theorem 4,
‖se2k − pk,r‖ = O
(
h2r+2
4r+1(r + 1)!2
)
as r → ∞.(4.3)
50 H. Volkmer
Table 4. Estimates from Theorem 2
for L2(0, π/2) distance between Math-
ieu function se12(·, 10) and approximation
p6,r .
r ‖se12 − p6,r ‖ Error estimate
7 0.14 · 10−1 0.19 · 10−1
8 0.74 · 10−3 1.12 · 10−3
9 0.29 · 10−4 0.46 · 10−4
10 0.86 · 10−6 1.39 · 10−6
11 0.20 · 10−7 0.33 · 10−7
12 0.37 · 10−9 0.63 · 10−9
13 0.58 · 10−11 1.00 · 10−11
Table 4 gives L2-error estimates in the case k = 6, h2 = 10 for the computation of
the Mathieu function se12. We choose q = 1 in (3.5).
The following theorem estimates the maximum norm of se2k − pk,r :
Theorem 5. For every t ∈ R, we have
|se2k(t) − pk,r (t)| ≤ ‖Pr xk − yk,r‖
(
r∑
n=1
sin2(2nt)
)1/2
+∞
∑
n=r+1
|xk,n|| sin(2nt)|,
where the quantities on the right-hand side of the inequality can be estimated according
to Lemmas 2, 3, and 5.
Proof. We write
se2k(t) − pk,r (t) =r
∑
n=1
(xk,n − un) sin(2nt) +∞
∑
n=r+1
xk,n sin(2nt)
and estimate the first sum using the Cauchy–Schwarz inequality.
Table 5 shows error estimates in the case k = 6, h2 = 10 for the computation of
se12(0.2, 10) = 0.8143833501330276753773030610857527020833 . . . .
5. Application to the Spheroidal Wave Equation
We consider the problem of finding eigenvalues λ for which the spheroidal wave equation
((1 − t2)w′)′ +(
λ + γ 2(1 − t2) −m2
1 − t2
)
w = 0(5.1)
has nontrivial bounded even solutions in (−1, 1). Here γ 2 is a given real parameter and
m = 0, 1, 2, 3, . . . . The eigenvalue problem for bounded odd solutions can be treated
similarly.
Error Estimates for Rayleigh–Ritz Approximations 51
Table 5. Estimates from Theorem 5 for distance between se12(0.2, 10) and approximation p6,r (0.2).
r p6,r (0.2) se12(0.2, 10) − p6,r (0.2) Error estimate
7 0.81649100562612785984 −0.21 · 10−2 1.28 · 10−2
8 0.81400537451258646982 0.38 · 10−3 1.30 · 10−3
9 0.81440948923732886079 −0.26 · 10−4 0.71 · 10−4
10 0.81438236755585789352 0.98 · 10−6 2.53 · 10−6
11 0.81438337423093520623 −0.24 · 10−7 0.63 · 10−7
12 0.81438334972786824766 0.41 · 10−9 1.16 · 10−9
13 0.81438335013757800301 −0.46 · 10−11 1.56 · 10−11
In L2(0, 1) we work with the orthonormal basis
�mn (t) =
Pmn (t)
sn,m
, sn,m =
√
1
(2n + 1)
(n + m)!
(n − m)!, n = m, m + 2, m + 4, . . . ,
where Pmn denotes the associated Legendre function. The well-known representation of
the differential operator
w �→ −((1 − t2)w′)′ +(
m2
1 − t2− γ 2(1 − t2)
)
w
with respect to this orthonormal basis has the form (3.1) where
dj = (m + 2 j − 2)(m + 2 j − 1) − 2γ 2bj ,
bj =(m + 2 j − 2)(m + 2 j − 1) − 1 + m2
(2m + 4 j − 5)(2m + 4 j − 1),
cj = γ 2
√(2 j − 1)2 j
√2m + 2 j − 1
√2m + 2 j
(2m + 4 j − 1)√
2m + 4 j − 3√
2m + 4 j + 1.
The eigenvalue λk of the corresponding operator A agrees with the spheroidal eigenvalue
λmm+2(k−1)(γ
2) in the notation of [4].
Note that the sequence bj converges to 14
as j → ∞. The sequence bj is monotonically
decreasing for m ≥ 1, and monotonically increasing for m = 0 except for the first term
b1 = 13. Therefore, we can choose a sequence Dj satisfying (3.8) as
Dj =
(2 j − 2)(2 j − 1) − 12γ 2, if m = 0, γ 2 > 0, j ≥ 2,
dj , if m ≥ 1, γ 2 > 0,
dj , if m = 0, γ 2 < 0, j ≥ 2,
(m + 2 j − 2)(m + 2 j − 1) − 12γ 2, if m ≥ 1, γ 2 < 0.
The sequence cj converges to 14γ 2. The sequence |cj | is monotonically increasing for
m ≥ 1 and monotonically decreasing for m = 0. Hence, we choose the sequence C j
satisfying (3.8) as
C j =
{
|cj | if m = 0,
14|γ 2| if m ≥ 1.
52 H. Volkmer
Table 6. Estimates from Theorem 1 for distance between spheroidal eigenvalue
λ02(10) and approximation µ2,r .
r λ02(10) µ2,r − λ0
2(10) Error estimate
3 1.7962473200113952199510987 0.58 · 10−2 1.25 · 10−2
4 1.7904107081973358415378761 0.16 · 10−4 0.36 · 10−4
5 1.7903944464734022475175084 0.15 · 10−7 0.35 · 10−7
6 1.7903944312365846389547950 0.62 · 10−11 1.48 · 10−11
7 1.7903944312303699054175945 0.13 · 10−14 0.31 · 10−14
8 1.7903944312303686201621693 0.15 · 10−18 0.37 · 10−18
9 1.7903944312303686200127650 0.11 · 10−22 0.26 · 10−22
By Theorem 4, we obtain
µk,r − λk = O
(
γ 4r
42r+1(m + 2r − 1)!2(m + 2r + 1)!2
)
as r → ∞.
We use that cj = (γ 2/4)(1 + O( j−2)) for m = 0 to prove this statement for m = 0.
Table 6 shows error estimates in the case m = 0, k = 2, γ 2 = 10 for the computation
of the spheroidal eigenvalue
λ02(10) = 1.7903944312303686200127544999612315906840 . . . .
For the error estimates in this section we choose �1 = µ1,6, �2 = µ2,7, and q = 1
in (3.5).
The spheroidal wave eigenfunction Psmn (t) = Psm
n (t, γ 2), n = m+2(k−1), belonging
to the eigenvalue λk , is given by
Psmn (t) = sn,m
∞∑
j=1
xk, j�mm+2( j−1)(t),
where xk = (xk,1, xk,2, . . .) is the eigenvector of A belonging to λk and∑∞
n=1 x2k,n = 1;
see [4]. We use the eigenvector yk,r = (u1, . . . , ur , 0, 0, . . .) of Pr APr belonging to the
eigenvalue µk,r to approximate Psmn by the quasi-polynomial
pk,r (t) = sn,m
r∑
j=1
u j�mm+2( j−1)(t).
The L2(0, 1)-norm
‖Psmn − pk,r‖ = sn,m
(
∞∑
j=1
(xk, j − u j )2
)1/2
can be estimated as indicated in Section 3. In particular, by Theorem 4,
‖Psmn − pk,r‖ = O
(
|γ 2|r+1
4r+1(m + 2r + 1)!2
)
.
Table 7 gives L2-error estimates in the case m = 0, k = 2, γ 2 = 10 for the computation
of the spheroidal wave function Ps02(·, 10).
Error Estimates for Rayleigh–Ritz Approximations 53
Table 7. Estimates from Theorem 2 for
L2(0, 1) distance between spheroidal wave
function Ps02(·, 10) and approximation p2,r .
r ‖Ps02(·, 10) − p2,r ‖ Error estimate
3 0.59 · 10−2 0.85 · 10−2
4 0.22 · 10−3 0.34 · 10−3
5 0.54 · 10−5 0.84 · 10−5
6 0.91 · 10−7 1.44 · 10−7
7 0.11 · 10−8 0.18 · 10−8
8 0.11 · 10−10 0.17 · 10−10
9 0.79 · 10−13 1.29 · 10−13
The following theorem estimates the maximum norm of Psmn − pk,r :
Theorem 6. For every t ∈ [−1, 1], k ∈ N, n = m + 2(k − 1), we have
|Psmn (t) − pk,r (t)| ≤ sn,m‖Pr xk − yk,r‖
(
r∑
j=1
(�mm+2( j−1)(t))
2
)1/2
+ sn,m
∞∑
j=r+1
|xk, j ||�mm+2( j−1)(t)|,
where the quantities on the right-hand side of the inequality can be estimated as indicated
in Sections 2 and 3 and
|�mn (t)| ≤
{√(2n + 1)/2 if m = 1, 2, . . . ,
√2n + 1 if m = 0.
Proof. We argue as in the proof of Theorem 5. The bound for �mn (t) follows from the
addition formula for Legendre functions; see [4, p. 259] or [11]. The latter reference also
contains some sharper bounds.
Table 8 shows error estimates in the case m = 0, k = 2, γ 2 = 10 for the computation of
Ps02(0.5, 10) = 0.1073214125922448426909203465692704761396 . . . .
Table 8. Estimates from Theorem 6 for distance between Ps02(0.5, 10) and approximation p2,r (0.5).
r p2,r (0.5) Ps02(0.5, 10) − p2,r (0.5) Error estimate
3 0.09952067097265232504 0.78 · 10−2 1.26 · 10−2
4 0.10723677849603458567 0.84 · 10−4 2.62 · 10−4
5 0.10732612678700087639 −0.47 · 10−5 1.11 · 10−5
6 0.10732152152220085124 −0.11 · 10−6 0.24 · 10−6
7 0.10732141295182414931 −0.36 · 10−9 1.61 · 10−9
8 0.10732141258305621825 0.92 · 10−11 2.47 · 10−11
9 0.10732141259215136846 0.93 · 10−13 2.35 · 10−13
54 H. Volkmer
Acknowledgment. The author thanks the referees for pointing out additional refer-
ences, suggesting inclusion of more tables, and other useful hints.
References
1. M. ABRAMOWITZ, I. STEGUN (eds.) (1964): Handbook of Mathematical Functions. National Bureau of
Standards Applied Mathematics Series, No. 55. U.S. Government Printing Office, Washington, DC.
2. F. A. ALHARGAN (1996): A complete method for the computations of Mathieu characteristic numbers
of integer orders. SIAM Rev., 38:239–255.
3. F. A. ALHARGAN (2000): Algorithms for the computation of all Mathieu functions of integer orders.
ACM Trans. Math. Software, 26:390–407.
4. F. ARSCOTT (1964): Periodic Differential Equations. New York: Macmillan.
5. L. BAKER (1992): C Mathematical Function Handbook. New York: McGraw-Hill. (Includes diskettes.)
6. G. BLANCH (1966): Numerical aspects of Mathieu eigenvalues. Rend. Circ. Mat. Palermo, 15:51–97.
7. P. FALLOON (2001): Theory and Computation of Spheroidal Harmonics with General Arguments. Master
of Science thesis. University of Western Australia.
8. D. J. GREEN, S. MICHAELSON (1965): Series solution of certain Sturm–Liouville eigenvalue problems.
Comput. J., 7:322–336.
9. D. B. HODGE (1970): Eigenvalues and eigenfunctions of the spheroidal wave equation. J. Math. Phys.,
11:2308–2312.
10. T. KATO (1980): Perturbation Theory for Linear Operators. New York: Springer-Verlag.
11. G. LOHOFER (1998): Inequalities for the associated Legendre functions. J. Approx. Theory, 95:178–193.
12. J. MEIXNER, F. W. SCHAFKE (1954): Mathieusche Funktionen and Spharoidfunktionen. Berlin: Springer-
Verlag.
13. R. B. SHIRTS (1993): The computation of eigenvalues of Mathieu’s differential equation for noninteger
order. ACM Trans. Math. Software, 19:377–390.
14. R. B. SHIRTS (1993): Algorithm 721 MTIEU1 and MTIEU2: Two subroutines to compute eigenvalues
and solutions to Mathieu’s differential equation for noninteger and integer order. ACM Trans. Math.
Software, 19:391–406.
15. THE GROUP “NUMERICAL ANALYSIS” AT DELFT UNIVERSITY OF TECHNOLOGY (1973): On the compu-
tation of Mathieu functions. J. Engrg. Math., 7:39–61.
16. W. J. THOMPSON (1997): Atlas for Computing Mathematical Functions: An Illustrated Guide for Practi-
tioners, with Programs in C and Mathematica. New York: Wiley. (Includes CD-ROM. Fortran 90 edition
exists also.)
17. A. WEINSTEIN, W. STENGER (1972): Methods of Intermediate Problems for Eigenvalues. New York:
Academic Press.
18. J. H. WILKINSON (1961): Rigorous error bounds for computed eigenvalues. Comput. J., 4:230–241.
19. S. ZHANG, J. JIN (1996): Computation of Special Functions. New York: Wiley. (Includes diskette with
Fortran programs.)
H. Volkmer
Department of Mathematical Sciences
University of Wisconsin-Milwaukee
P. O. Box 413
Milwaukee, WI 53201
USA