MATHEMATICS OF COMPUTATION

Volume 66, Number 218, April 1997, Pages 699–717

S 0025-5718(97)00828-4

THE RESOLUTION OF THE GIBBS PHENOMENON

FOR SPHERICAL HARMONICS

ANNE GELB

Abstract. Spherical harmonics have been important tools for solving geo-
physical and astrophysical problems. Methods have been developed to effec-
tively implement spherical harmonic expansion approximations. However, the
Gibbs phenomenon was already observed by Weyl for spherical harmonic ex-
pansion approximations to functions with discontinuities, causing undesirable
oscillations over the entire sphere.
Recently, methods for removing the Gibbs phenomenon for one-dimensional
discontinuous functions have been successfully developed by Gottlieb and Shu.
They proved that the knowledge of the first N expansion coefficients (either
Fourier or Gegenbauer) of a piecewise analytic function f (x) is enough to
recover an exponentially convergent approximation to the point values of f (x)
in any subinterval in which the function is analytic.
Here we take a similar approach, proving that knowledge of the first N
spherical harmonic coefficients yield an exponentially convergent approxima-
tion to a spherical piecewise smooth function f (θ, φ) in any subinterval [θ1 , θ2 ],
φ ∈ [0, 2π], where the function is analytic. Thus we entirely overcome the
Gibbs phenomenon.

1. Introduction
Spherical coordinates arise in many problems studied by geophysicists and as-
trophysicists. Methods of representation of discrete functions on spheres have been
discussed in great detail, as there are many analytical and computational consider-
ations. One should refer to [10] and [11] for details. In this paper we consider
the spectral spherical harmonic representation. The spherical harmonics are a
two-dimensional basis set that automatically remedy the pole problem and offer
exponential convergence for analytic functions on the sphere. Also many numerical
algorithms in spherical harmonics already exist. Thus it seems natural to represent
a function in spherical coordinates with spherical harmonics.
The spherical harmonic expansion of a function f (θ, φ) defined on a sphere with
colatitude coordinate θ and longitude coordinate φ, 0 ≤ θ ≤ π, 0 ≤ φ ≤ 2π, is
defined as

Received by the editor October 26, 1995 and, in revised form, May 1, 1996.
1991 Mathematics Subject Classification. Primary 42A10, 41A10, 41A25.
Key words and phrases. Gibbs phenomenon, Gegenbauer polynomials, spherical harmonics,
exponential accuracy.
This work was supported in part by AFOSR grant F49620-95-1-0074, NSF grant DMS 95008
14, and DARPA/ONR AASERT grant N00014-93-1-0985.

c 1997 American Mathematical Society

699

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 700 ANNE GELB

Definition 1.1.
∞ X
X
(1.1) f (θ, φ) = aνq Yqν (θ, φ)
q=0 |ν|≤q

where the spherical harmonic Yqν (θ, φ) of degree q and order ν is

s
(2q + 1)(q − ν)! ν
Yqν (θ, φ) = Pq (cos θ)eiνφ
4π(q + ν)!

in terms of the associate Legendre functions Pqν (cos θ) [3].

The orthonormality of the spherical harmonics Yqν (θ, φ) over the sphere imply
that the coefficients aνq are given by
Z 2π Z π
(1.2) ν
aq = f (θ, φ)[Yqν (θ, φ)]∗ sin θ dθdφ,
0 0
where [Yqν (θ, φ)]∗ are the complex conjugates of Yqν (θ, φ).
The truncated spectral representation of f (θ, φ) is
X
N X
(1.3) gN (θ, φ) = aνq Yqν (θ, φ).
q=0 |ν|≤q

Lemma 1.1. If f (θ, φ) is infinitely differentiable, then gN (θ, φ) converges spec-

trally to f (θ, φ).
The proof of this lemma is presented in [10].
Now suppose f (θ, φ) is a discontinuous but piecewise smooth function. The
truncated series no longer converges exponentially to f . In fact, the Gibbs phe-
nomenon occurs at the point of discontinuity, causing oscillations over the entire
sphere (Figure 1) was initially observed by Herman Weyl [12].
In this paper it is shown that the first (N + 1)2 spherical harmonic coefficients
contain enough information to reconstruct a spectrally accurate approximation,
provided that f (θ, φ) is a piecewise analytic function.
Gottlieb and Shu [7] established exponential convergence and eliminated the
Gibbs phenomenon for one-dimensional piecewise analytic functions assuming
knowledge of the first N Fourier coefficients or the first N Gegenbauer coefficients
(in general). First the Gegenbauer coefficients based on the Gegenbauer polynomi-
1
als Cnλ (x), which are orthogonal in [−1, 1] with the weight function (1 − x2 )λ− 2 ,
were obtained from the known Fourier or Gegenbauer coefficients for λ ∼ N , and
then the Gegenbauer expansion was constructed. The convergence was shown to
be exponential in any subinterval for which the function was analytic. This paper
follows the same idea, only now in the two-dimensional spherical coordinates and
with knowledge of the first (N + 1)2 spherical harmonic coefficients. The procedure
consists of the same two steps as given in [7]:
1. Using the given first (N +1)2 spherical harmonic coefficients, aνq , based on
spherical harmonics Yqν (θ, φ), we recover, with exponential accuracy, the
first m ∼ N Gegenbauer expansion coefficients, based on a subinterval
[θ1 , θ2 ] ⊂ [0, π], φ ∈ [0, 2π], in which the function is presumably analytic.
This can be achieved for any L1 function, as long as λ, the order of the
Gegenbauer polynomial, is chosen to be proportional to N . The error

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 701

incurred at this stage is called the truncation error, and is investigated

in Section 3.
2. We then apply the previously established exponential convergence proof
[6] of the Gegenbauer expansion to analytic functions in [θ1 , θ2 ], φ ∈
[0, 2π], when the parameter λ is proportional to the number of terms
retained in the expansion. The error at this stage is labeled the regu-
larization error. Details can be found in [6]. The results are quoted in
Section 4.
Section 5 contains a numerical example to illustrate our results. In Section 2 we
establish some useful properties of the Gegenbauer polynomials and the associated
Legendre functions (Section 2 also appears in [7]).
Throughout this paper, A denotes a generic constant or at most a polynomial in
the growing parameters, as will be indicated in the text.

2. Preliminaries
2.1. Gegenbauer polynomials. In this section we present some properties of the
Gegenbauer polynomials and the associated Legendre functions which are necessary
to prove the results in Sections 3 and 4. One should see Bateman [2] for details.
Definition 2.1. The Gegenbauer polynomial Cnλ (x), for λ ≥ 0, is defined by
1 dn h 1
i
(2.1) (1 − x2 )λ− 2 Cnλ (x) = G(λ, n) n (1 − x2 )n+λ− 2 ,
dx
where G(λ, n) is given by
(−1)n Γ(λ + 12 )Γ(n + 2λ)
(2.2) G(λ, n) = .
2n n!Γ(2λ)Γ(n + λ + 12 )
Formula (2.1) is also called the Rodrigues’ formula [1, page 175].
Under this definition, for λ > 0,
Γ(n + 2λ)
(2.3) Cnλ (1) =
n!Γ(2λ)
and
(2.4) |Cnλ (x)| ≤ Cnλ (1), −1 ≤ x ≤ 1.
The Gegenbauer polynomials are orthogonal under the weight function
1
(1 − x2 )λ− 2 , thus
Z 1
1
(2.5) (1 − x2 )λ− 2 Ckλ (x)Cnλ (x)dx = δk,n hλn ,
−1

where, for λ > 0,

1 Γ(λ + 12 )
(2.6) hλn = π 2 Cnλ (1) .
Γ(λ)(n + λ)
The approximation of the Gegenbauer polynomials for large n and λ is dependent
upon the well-known Stirling’s formula for Γ(x) given by
(2π) 2 xx+ 2 e−x ≤ Γ(x + 1) ≤ (2π) 2 xx+ 2 e−x e 12x ,
1 1 1 1 1
(2.7) x ≥ 1.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 702 ANNE GELB

Lemma 2.1. There exists a constant A independent of λ and n such that

1 1
λ2 λ2
(2.8) A−1 Cnλ (1) ≤ hλn ≤ A C λ (1).
(n + λ) (n + λ) n
The proof follows from (2.6) and Stirling’s formula (2.7).
The following lemma to be used later is easily obtained from the Rodrigues
formula (2.1).
Lemma 2.2. For any λ ≥ 1
d h 1
i G(λ, n) 3
(2.9) (1 − x2 )λ− 2 Cnλ (x) = (1 − x2 )λ− 2 Cn+1
λ−1
(x).
dx G(λ − 1, n + 1)
The proof follows from taking the derivative on both sides of the Rodrigues
formula (2.1), and then using it again on the right-hand side.
The following formula [1, page 176] will also be needed:
1 d  λ 
(2.10) Cnλ (x) = ( C (x) − Cn−1
λ
(x) ),
2(n + λ) dx n+1
which is true for all λ ≥ 0.
The associated Legendre functions are defined as
Definition 2.2.
dm Pl (x) m
(2.11) Plm (x) = (1 − x2 ) 2
, 0 ≤ m ≤ l.
dxm
Employing Rodrigues’ formula for Pl (x), the corresponding Rodrigues formula
for Plm (x) is
m
(1 − x2 ) 2 dl+m 2
Plm (x) = (x − 1)l , 0 ≤ m ≤ l.
2l l! dxl+m
The equation (2.11) allows rapid development of many properties of the Plm ,
particularly the recurrence relations [8, page 1005],
for l
(l − m + 1)Pl+1
m
− (2l + 1)xPlm + (l + m)Pl−1
m
= 0,
and for m
p p
1 − x2 Plm+1 + 2mxPlm + (l + m)(l − m + 1) 1 − x2 Plm−1 = 0.
These equations lead to other useful relations as well, notably those used in the
truncation error proofs in Section 3.

3. Truncation error in a subinterval

Assume that f (θ, φ) is an L1 function defined for θ ∈ [0, π] and φ ∈ [0, 2π], and
that f (θ, φ) is periodic in φ and analytic for θ in a subinterval [θ1 , θ2 ] ⊂ [0, π].
Also assume that the spherical harmonic partial sum of f (θ, φ) is known
X
N X
ν
(3.1) gN (θ, φ) = aνq Yqν (θ, φ).
q=0 |ν|≤q

The spherical harmonics Yqν (θ, φ) are defined by

s
(2q + 1)(q − ν)! ν
(3.2) Yqν (θ, φ) = Pq (cos θ)eiνφ ,
4π(q + ν)!

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 703

where Pqν (cos θ) are the associated Legendre functions (2.11).

The coefficients aνq in (3.1) are given by

Z 2π Z π
(3.3) ν
aq = f (θ, φ)[Yqν (θ, φ)]∗ sin θ dθdφ,
0 0
where [Yqν (θ, φ)]∗ are the complex conjugates of Yqν (θ, φ).
ν
It has been shown that gN (θ, φ) will not converge fast to f (θ, φ) if there exist
any discontinuities [12].
The coefficients of f (θ, φ) must satisfy
Assumption 3.1. |aνq | ≤ A independent of q.
Note that this is true for f (θ, φ) ∈ L1 .
The goal is to recover f (θ, φ) for θ in subinterval [θ1 , θ2 ] ⊂ [0, π], and φ ∈ [0, 2π]
for a fixed φ. The approximation expansion is written in terms of cos θ because of
the associated Legendre functions Pqν (cos θ). Since φ is fixed, the function f will be
considered only as a function of θ and will be denoted by f (θ). Hence we make the
transformation from θ in [0, π] to ξ in [−1, 1] and apply the following definitions.
Definition 3.1. The local variable ξ = cos θ is defined by
x(ξ) = ξ + δ,
β−α
 = ,
2
β+α
δ = ,
2
(3.4) −1 ≤ α ≤ β ≤ 1,
where α = cos θ1 , β = cos θ2 , and θ2 < θ1 .
Definition 3.2. The Gegenbauer partial sum is defined by
X
m
(3.5) λ
fm (x) = fˆλ (l)Clλ (x),
l=0

where the first m coefficients fˆλ (l) based upon the Gegenbauer polynomials Clλ (x)
1
with weight function (1 − x2 )λ− 2 for any constant λ ≥ 0 are defined by
Z 1
1 1
(3.6) fˆλ (l) = λ (1 − x2 )λ− 2 Clλ (x)f (x)dx, 0 ≤ l ≤ m,
hl −1
and the Gegenbauer polynomials Clλ (x) are defined by (2.1).

The Gegenbauer expansion of f (x) for the subinterval [α, β] is

∞
X
(3.7) f (ξ + δ) = fˆλ (l)Clλ (ξ), −1 ≤ ξ ≤ 1,
l=0

where the coefficients fˆλ (l) are

Z 1
1 1
(3.8) fˆλ (l) = λ (1 − ξ 2 )λ− 2 Clλ (ξ)f (ξ + δ)dξ.
hl −1

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 704 ANNE GELB

Of course, fˆλ (l) is not known, but rather an approximation to fˆλ (l), denoted ĝλ (l),
based on the transformation in equation (3.4) of the spherical harmonic partial
ν
sum, gN (ξ + δ). These approximate coefficients ĝλ (l) are defined as
Z 1
1 1
(3.9) ĝλ (l) = λ (1 − ξ 2 )λ− 2 Clλ (ξ)gN
ν
(ξ + δ)dξ.
hl −1
The truncation error describes how well the coefficients ĝλ (l) approximate fˆλ (l).
Definition 3.3. The truncation error is defined by
Xm
(3.10) T E(λ, m, N, , ν) = max | (fˆλ (l) − ĝλ (l))Clλ (ξ)|,
−1≤ξ≤1
l=0

where fˆλ (l) and ĝλ (l) are defined in equations (3.8) and (3.9).
Lemma 3.1. The truncation error can be estimated by
s
X∞ X Clλ (1) X ν (2q + 1)(q − ν)!
m
T E(λ, m, N, , ν) ≤ | |aq |
hλl 4π(q + ν)!
q=N +1 l=0 |ν|≤N
Z 1
1
(3.11) × (1 − ξ 2 )λ− 2 Clλ (ξ)Pqν (ξ + δ)dξ|.
−1

Proof. Substituting equation (3.8) and equation (3.9) into equation (3.10) yields
(3.12)
Z 1
1 1
fˆλ (l) − ĝλ (l) = λ (1 − ξ 2 )λ− 2 Clλ (ξ)(f (ξ + δ) − gN
ν
(ξ + δ))dξ.
hl −1

Applying the estimate (2.4) on Clλ (ξ), and the equations (3.1) and (3.2), the esti-
mate (3.11) is obtained.
It must be shown now that the estimate (3.11) is small. For simplicity of nota-
tions denote
Z 1
1
(3.13) λ,l
Sq,ν = (1 − ξ 2 )λ− 2 Clλ (ξ)Pqν (ξ + δ)dξ
−1
and
λ,l
λ,l
Sq,ν
(3.14) Iq,ν = ,
G(λ, l)
where G(λ, l) is defined in (2.2).
λ,l
Sq,ν must be bounded effectively so that the truncation error is small. This is
λ,l
done by finding a recurrence relation for Iq,ν ∼ Sq,ν
λ,l
, and then an upper bound for
|ν| ≤ N .
λ,l
Lemma 3.2. Iq,ν satisfies the following recurrence relation
 
2 (ν + 1) (q + ν) λ−1,l+1 (q − ν + 1) λ−1,l+1
λ,l
Iq,ν = I − I
(2q + 1) (ν + 2) (q − ν) q−1,ν (q + ν + 1) q+1,ν
ν
(3.15) + I λ,l .
(q − ν)(q + ν + 1)(ν + 2) q,ν+2
See proof of Lemma 3.2 in the appendix.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 705

λ,l
It is now possible to bound Sq,ν of (3.13) by applying Lemma 3.2.
λ,l
Lemma 3.3. For Sq,ν in (3.13) where 0 ≤ ν ≤ N ,

(3.16)
s
(q + ν)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
λ,l
Sq,ν ≤ max(ν ρ , 1) ,
(q − ν + 1)! λ Γ(q) |G(0, l + λ)|

where G(λ, l) is defined in (2.2) and subject to the conditions

4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
0 ≤ ν ≤ N < q,
λ − 1 > κ > 2,
ρ > 3.

See proof of Lemma 3.3 in the appendix.

Note that the upper bound on κ and the lower bound on ρ are required to
λ,l
estimate Sq,ν (see proof of Lemma 3.3 and Lemma 3.4). The lower bound on κ
ensures convergence of the truncation error (Theorem 3.1). These parameters have
not been optimized.
λ,l
An estimate is also needed for Sq,ν , where −N ≤ ν ≤ 0.
λ,l
Lemma 3.4. For Sq,ν in (3.13), where −N ≤ ν ≤ 0,

(3.17)
s
(q + ν − 1)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
λ,l
Sq,ν ≤ max(|ν| , 1) ρ
,
(q − ν)! λ Γ(q) |G(0, l + λ)|

where G(λ, l) is defined in (2.2) and subject to the conditions

4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
0 ≥ ν ≥ −N,
λ − 1 > κ > 2,
ρ > 3,
|N | < q.

See proof of Lemma 3.4 in the appendix.

λ,l
Combining Lemma 3.3 and Lemma 3.4 yields the following estimate for Sq,|ν| .

λ,l
Lemma 3.5. For Sq,|ν| in (3.13) we have the following estimate

(3.18)
s
λ,l (q + ν)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
Sq,|ν| ≤ max(|ν| , 1) ρ
,
(q − ν)! λ Γ(q) |G(0, l + λ)|

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 706 ANNE GELB

where G(λ, l) is defined in (2.2) and subject to the conditions

4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
|ν| ≤ N < q,
λ − 1 > κ > 2,
ρ > 3.
λ,l
Finally it is possible to estimate |Sq,|ν| | in terms of the truncation error param-
eters λ, m, , and ν.
Lemma 3.6.
s
λ,l (m + 2λ)m+2λ 1 (q + ν)!
(3.19) Sq,|ν| ≤A ,
(2λ)λ mm q κ (q − ν)!
1
where A grows at most as (m + λ) 2 N ρ . The estimate is subject to the conditions
4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
|ν| ≤ N < q,
λ − 1 > κ > 2,
ρ > 3.
Proof. Starting from estimate (3.18) and using the definition of G(λ, l) in (2.2), the
following estimate is obtained:
s
λ,l Γ(q − λ) |G(λ, l)| (q + ν)! ρ
Sq,|ν| ≤ |ν| (q − l)λ−κ
λ Γ(q) |G(0, l + λ)| (q − ν)!
Γ(q − λ) Γ(λ + 12 )Γ(l + 2λ) 2l+λ (l + λ)Γ(l + λ + 12 )
≤
λ Γ(q) 2l l!Γ(2λ)Γ(l + λ + 12 ) Γ(l + λ)
s
(q + ν)! ρ
× |ν| (q − l)λ−κ
(q − ν)!
Γ(q − λ) Γ(λ + 12 )Γ(l + 2λ) 2l+λ (l + λ)Γ(l + λ + 12 )
≤
λ Γ(q) 2l l!Γ(2λ)Γ(l + λ + 12 ) Γ(l + λ)
s
(q + ν)! ρ
× |ν| (q − l)λ−κ
(q − ν)!
s
Γ(q − λ) Γ(λ)Γ(l + 2λ)2λ (q + ν)! ρ
≤ |ν| (q − l)λ−κ
λ Γ(q) l!Γ(2λ) (q − ν)!
s
Γ(q − λ) Γ(λ)Γ(m + 2λ)2λ (q + ν)! ρ λ−κ
≤ |ν| (q)
λ Γ(q) m!Γ(2λ) (q − ν)!
s
(q − λ)q−λ e−(q−λ) λλ e−λ (m + 2λ)m+2λ e−(m+2λ) 2λ (q + ν)! 1
≤ A
λ q q e−q mm e−m (2λ)2λ e−2λ (q − ν)! q κ
s
(m + 2λ)m+2λ (q + ν)! 1
≤ A .
(2λ)λ mm (q − ν)! q κ

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 707

Equation (2.2) is used in the second inequality, the monotonicity with respect
to l in the fourth inequality, and Stirling’s formula (2.7) in the fifth inequality.

We are now ready for the main theorem of this section.

Theorem 3.1. Let the truncation error be defined in (3.10). Let λ = αN and
m = βN with 0 < α, β < 1, λ − 1 > κ ≥ 2, and ρ > 3. Then for |ν| ≤ N
 N
(β + 2α)β+2α
(3.20) T E(αN, βN, N, ) ≤ A ,
2α αα β β
2
where A grows at most as N 1+ρ . In particular, if α = β < 27 , then

(3.21) T E(αN, αN, N, ) ≤ ApN ,

where
 α
27α
(3.22) p= < 1.
2
Proof. The theorem follows from the definition of the truncation error (3.11)
s
∞
X X Clλ (1) X ν (2q + 1)(q − ν)!
m
T E(λ, m, N, , ν) ≤ | |aq |
hλl 4π(q + ν)!
q=N +1 l=0 |ν|≤N
Z 1
1
× (1 − ξ 2 )λ− 2 Clλ (ξ)Pqν (ξ + δ)dξ|.
−1

λ,l
Recalling the bound on Sq,ν in (3.19)
s
λ,l (m + 2λ)m+2λ 1 (q + ν)!
Sq,|ν| ≤A ,
(2λ)λ mm q κ (q − ν)!

the assumption (3.1)

|aνq | ≤ A,

and the estimate of hλn in (2.8)

1 1
λ2 λ2
A−1 Cnλ (1) ≤ hλn ≤ A C λ (1),
(n + λ) (n + λ) n
the truncation error estimate is obtained
∞
r
X X (2q + 1) (m + 2λ)m+2λ 1
T E(λ, m, N, , ν) ≤ A .
4π (2λ)λ mm q κ
q=N +1 |ν|≤N

Substituting in the values λ = αN and m = βN ,

 N
(β + 2α)β+2α
T E(αN, βN, N, ) ≤ A ,
2α αα β β
where A grows at most as N 1+ρ .

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 708 ANNE GELB

4. Regularization error
For the one-dimensional case, as shown in [7], the second part of the error, called
the regularization error, is caused by the finite Gegenbauer expansion based on a
subinterval [α, β] ⊂ [−1, 1] approximation to a function f (x) which is assumed
analytic in this sub-interval. Instead of f (x) on a subinterval [α, β] ⊂ [−1, 1] as
in [7], we now have f (θ, φ) on a subinterval [θ1 , θ2 ] ⊂ [0, π], φ ∈ [0, 2π]. The
regularization error will be the same result as shown in [7]. We will thus just quote
the result (substituting f (θ, φ) for f (x)).
Assume that f (θ, φ) is an analytic function on [θ1 , θ2 ] ⊂ [0, π], φ ∈ [0, 2π], satis-
fying:
Assumption 4.1. There exist constants τ ≥ 1 and C(τ ) such that, for every k ≥ 0,
dk f k!
(4.1) max k
(θ, φ) ≤ C(τ ) k , 0 ≤ φ ≤ 2π.
θ1 ≤θ≤θ2 dx τ
This is a standard assumption for analytic functions. τ is the distance from
[θ1 , θ2 ] to the nearest singularity of f (θ, φ) in the complex plane (see for example
[9]).
As in the case of the truncation error, the approximation expansion is in terms
of cos θ. Recall the transformation that was made previously from θ in [0, π] to the
local variable ξ in [−1, 1]:
x(ξ) = ξ + δ,
β−α
 = ,
2
β+α
δ = m,
2
−1 ≤ α ≤ β ≤ 1,
where α = cos θ1 , β = cos θ2 , and θ2 < θ1 .
Consider the Gegenbauer partial sum of the first m terms for the function
f (ξ + δ):
X
m
(4.2) λ
fm (ξ) = fˆλ (l)Clλ (ξ),
l=0

with ξ,  and δ defined above, and the Gegenbauer coefficients based on [α, β]
defined by
Z 1
1 1
(4.3) fˆλ (l) = λ (1 − x2 )λ− 2 Clλ (ξ)f (ξ + δ)dξ.
hl −1
The regularization error in the maximum norm is defined by:
X
m
(4.4) RE(λ, m, ) = max f (ξ + δ) − fˆλ (l)Clλ (ξ) .
−1≤ξ≤1
l=0
The following estimate of the regularization error, when λ ∼ m, is obtained:
Theorem 4.1. Assume λ = γm where γ is a positive constant. If f (θ, φ) is ana-
lytic in [θ1 , θ2 ] ⊂ [0, π], φ ∈ [0, 2π], and satisfies Assumption 4.1, then the regular-
ization error defined in (4.4) can be bounded by
(4.5) RE(γm, m, ) ≤ Aq m ,

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 709

where q is given by
(1 + 2γ)1+2γ
(4.6) q= ,
τ 21+2γ γ γ (1+ γ)1+γ
which is always less than 1. In particular, if γ = 1 and m = βN where β is a
positive constant, then
(4.7) RE(βN, βN, ) ≤ Aq N ,
with
 β
27
(4.8) q= .
32τ
Combining the estimates for truncation error and regularization error it is pos-
sible to obtain the main theorem of this chapter, showing the complete removal of
the Gibbs phenomenon for a discontinuous function on a sphere.
Theorem 4.2. Consider a L1 function f (θ, φ) analytic in [θ1 , θ2 ] ⊂ [0, π], φ ∈
[0, 2π], and satisfying Assumption 4.1. Assume that the first N spherical harmonic
coefficients
Z 2π Z π
aνq = f (θ, φ)[Yqν (θ, φ)]∗ sin θ dθdφ,
0 0

for |ν| ≤ N ≤ q, are known. Let ĝλ (l), defined in (3.9) for 0 ≤ l ≤ m be the
Gegenbauer expansion coefficients of the Gegenbauer partial sum, defined in (3.5),
based on the subinterval [θ1 , θ2 ], φ ∈ [0, 2π]. Then for λ = m = βN with β < 27
2
,
X
m


(4.9) max f (ξ + δ) − ĝλ (l)Clλ (ξ) ≤ A qTN + qR
N
,
−1≤ξ≤1
l=0

where
 β  β
27β 27
qT = < 1, qR = < 1,
2 32τ
and A grows at most as N 1+ρ , where ρ > 3.
Proof. Just combine the results of Theorem 3.1 and Theorem 4.1.

5. Numerical results
In this section we show numerical results for one example. Suppose we are given
the function

2 + sin 4 , 0 ≤ θ ≤ 2 ,
11φ
cos 3θ π
(5.1) f (θ, φ) =
sin φ, π
2 < θ ≤ π.
Note the discontinuity at θ = π2 . We use N = 20, 40, and 80 latitudinal points
θ ⊂ [0, π], and 36 longitudinal points φ ⊂ [0, 2π].
Notice in Figure 1 that the Gibbs phenomenon is prevalent at both boundaries
as well as in the middle. The spherical harmonic expansion of f (θ, φ) produce the
L1 errors displayed in Table 1. Although the errors away from the discontinuity
appear to slowly converge, oscillations can be seen all over the sphere.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 710 ANNE GELB

Figure 1. Spherical harmonic approximation for 80 latitudinal

points. (1a) is the contour plot, (1b) is the cut at φ = π2

Table 1. L1 errors for the spherical harmonic expansion of the

function f (θ, φ)

π π
N average error error at 4 error at 2 error at 0
20 .16 6.3 × 10−2 .75 .96
−2 −2
40 9.7 × 10 1.3 × 10 .73 .96
80 6.1 × 10−2 1.9 × 10−2 .72 .96

Fixing φ = φf ∈ [0, 2π], the Gegenbauer approximation is obtained in the fol-

lowing way:
ν
1. The spherical harmonic expansion approximation gN (θ, φf ) on the subinterval
[θ1 , θ2 ] is computed from the given spherical harmonic coefficients aνq .
2. The first m ∼ N Gegenbauer coefficients for each fixed value φf
Z 1
1 1
ĝλ (l) = λ (1 − ξ 2 )λ− 2 Clλ (ξ)gN
ν
(ξ + δ, φf )dξ
hl −1
are approximated for λ ∼ N .
3. Finally the Gegenbauer expansion approximation
X
m
λ,
gm (θ, φf ) = ĝλ (l)Clλ (ξ)
l=0
is formed for the first m Gegenbauer polynomials.
The results shown in Figure 2 show that the Gegenbauer polynomial harmonic
expansion approximation simultaneously eliminates the oscillations at both bound-
aries (for each subinterval).
The L1 errors of the Gegenbauer expansion where N is the number of latitu-
dinal points, λ is the order of the Gegenbauer polynomial, and m is the number
of Gegenbauer polynomials is seen in Figure 3 and in Table 2. (The number of
longitudinal points remains constant.) The errors listed in both of the tables are
for the interval [0, π2 ].

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 711

Figure 2. (2a) Gegenbauer approximation for 40 latitudinal

points. (2b) Gegenbauer approximation for 80 latitudinal points

Figure 3. Errors in log scale for the Gegenbauer expansion ap-

proximation of f (θ, φ). The solutions for each subinterval [0, π2 ],
[ π2 , π] are found separately for φ = π2

Table 2. Errors for the Gegenbauer approximation of the function

f (θ, φ) using 36 longitudinal points

π π
N λ m average error error at 4 error at 2 error at 0
20 1 2 1.9 × 10−2 1.1 × 10−3 8.6 × 10−2 .39
40 7 3 6.3 × 10−4 7.4 × 10−6 3.3 × 10−3 6.1 × 10−4
80 8 5 5.2 × 10−5 1.2 × 10−7 2.4 × 10−4 3.4 × 10−6

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 712 ANNE GELB

6. Notes
This paper did not address some important considerations, which will be ad-
dressed in future papers:
1. The parameters λ and m have not been optimized.
2. The approximation gets worse for large latitudinal N . The error seems
to occur in the spherical harmonic approximation, where there is a con-
siderable amount of computation taking place. We used the subroutines
from the NCAR library to compute the spherical harmonic partial sum.
3. The errors are an order of magnitude worse at the discontinuities than
anywhere else. This is of course to be expected, but there are possible
ways to improve upon this error that have not been addressed.
4. We are assuming exact knowledge of the points of discontinuity. When
this information is not available, methods for locating discontinuities
must be employed.
The author would like to acknowledge David Gottlieb for his enthusiastic support
and consistently helpful advice.

Appendix
λ,l
Lemma 3.2. Iq,ν
satisfies the following recurrence relation:
 
2(ν + 1) (q + ν) λ−1,l+1 (q − ν + 1) λ−1,l+1
λ,l
Iq,ν = I − I
(ν + 2)(2q + 1) (q − ν) q−1,ν (q + ν + 1) q+1,ν
ν
+ I λ,l
(q − ν)(q + ν + 1)(ν + 2) q,ν+2
Proof. Pqν (t), for t ∈ [−1, 1], is defined in [8] by
ν ν+ 1
(A.1) Pqν (t) = H(ν)(1 − t2 ) 2 Cq−ν2 (t),
where
(−1)ν (2ν)!
(A.2) H(ν) = .
2ν ν!
Also the following relations for the associated Legendre functions are given in [8]
(A.3)
 
νt ν ν(q + ν + 1)(q − ν) ν
√ Pqν+1 (t) = − Pqν+2 (t) + Pq (t) .
1 − t2 2(ν + 1) 2(ν + 1)

(A.4)
p 1  
1 − t2 Pqν+1 (t) = (q − ν)(q − ν + 1)Pq+1
ν
(t) − (q + ν + 1)(q + ν)Pq−1
ν
(t) .
2q + 1
For clarification purposes, let x = x(ξ) = ξ + δ, and also
a = (q − ν),
b = (q + ν),
c = (q − ν + 1),
d = (q + ν + 1).

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 713

The definition of Pqν (t) in equation (A.1) is substituted into equation (3.13) to
obtain
Z 1
1 ν ν+ 1
λ,l
Sq,ν = H(ν) (1 − ξ 2 )λ− 2 Clλ (ξ)(1 − x2 ) 2 Cq−ν2 (x)dξ.
−1

Then applying the Rodrigues formula (2.1) and integrating by parts yields
Z 1
1 dq−ν
(1 − ξ 2 )λ− 2 Clλ (ξ)(1 − x2 )− 2
1 ν
= G(ν + , q − ν)H(ν) (1 − x2 )q dξ
2 −1 dxq−ν
Z
G(ν + 12 , q − ν)H(ν) 1
d h i
(1 − ξ 2 )λ− 2 Clλ (ξ)(1 − x2 )− 2
1 ν
= −
q−ν −1 dξ
dq−ν−1
× (1 − x2 )q dξ.
dxq−ν−1

The equation (2.9) is used and Rodrigues’ formula (2.1) is applied again to obtain
Z 1
H(ν)(2ν + 1) G(λ, l) 3 ν+ 32 ν
= (1 − ξ 2 )λ− 2 Cl+1
λ−1
(ξ)Cq−ν−1 (x)(1 − x2 ) 2 +1
ad G(λ − 1, l + 1) −1
Z
H(ν)(2ν + 1) 1 1 ν+ 32 ν
+ νx(1 − ξ 2 )λ− 2 Clλ (ξ)Cq−ν−1 (x)(1 − x2 ) 2 dξ.
ad −1

Applying the definition of G(λ, l) in (2.2) and the definition of Pqν (t) in (A.1) yields
  Z 1 p
1 G(λ, l) 3
= − (1 − ξ 2 )λ− 2 Cl+1
λ−1
(ξ)Pqν+1 (x) 1 − x2 dξ
ac G(λ − 1, l + 1) −1
Z 1
1 νx
+ (1 − ξ 2 )λ− 2 Clλ (ξ)Pqν+1 (x) √ dξ.
−1 1 − x2

Using equations (A.3) and (A.4) gives

(A.5)
Z 1  
G(λ, l) 3 b ν c ν
= (1 − ξ 2 )λ− 2 Cl+1
λ−1
(ξ) Pq−1 (x) − Pq+1 (x) dξ
(2q + 1)G(λ − 1, l + 1) −1 a d
(A.6)
Z 1  
ν 2 λ− 12 λ 1 ν+2
+ (1 − ξ ) Cl (ξ) P (x) − Pq (x) dξ.
ν
2(ν + 1) −1 ad q

Finally substituting equation (3.13) into equation (A.5) yields

 
1 (q + ν) λ−1,l+1 (q − ν + 1) λ−1,l+1
λ,l
Iq,ν = Iq−1,ν − Iq+1,ν
(2q + 1) (q − ν) (q + ν + 1)
ν λ,l ν
+ I + I λ,l .
2(ν + 1)(q − ν)(q + ν + 1) q,ν+2 2(ν + 1) q,ν

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 714 ANNE GELB

λ,l
Lemma 3.3. For Sq,ν in ( 3.13) where 0 ≤ ν ≤ N we have:
s
(q + ν)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
λ,l
Sq,ν ≤ max(ν ρ , 1) ·
(q − ν + 1)! λ Γ(q) |G(0, l + λ)|

where G(λ, l) is defined in (2.2) and subject to the conditions:

4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
0 ≤ ν ≤ N < q,
λ − 1 > κ > 2,
ρ > 3.

Proof. (by induction) For ν = 0

Z 1
1
λ,l
Sq,ν = (1 − ξ 2 )λ− 2 Clλ (ξ)Pq (ξ + δ)dξ
−1
Z 1
1 1
= (1 − ξ 2 )λ− 2 Clλ (ξ)Cq2 (ξ + δ)dξ.
−1

Upon defining
Z 1
1
λ,l
Fq,µ = (1 − ξ 2 )λ− 2 Clλ (ξ)Cqµ (ξ + δ)dξ,
−1

the equality holds

λ,l
Sq,ν=0 = Fq,λ,l1 ,
2

λ,l
Fq,µ is the integral evaluated in the Legendre case, [7]. Using Lemma 3.7 of [7]

Γ(q − λ) |G(λ, l)|

Fqλ,l ≤ A ,
λ Γ(q) |G(0, l + λ)|

where A grows at most as q 2µ−1 and G(λ, l) is defined in (2.2). Therefore for ν = 0
s
λ,l q! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
Sq,0 ≤ .
(q + 1)! λ Γ(q) |G(0, l + λ)|

λ,l λ,l
Applying the recurrence relation for Iq,ν in (3.15) to Sq,ν yields

λ,l (ν + 2)(q − ν)(q + ν + 1) λ,l

Sq,ν+2 = Sq,ν
ν
G(λ, l) 2(ν + 1)
+
G(λ − 1, l − 1) ν(2q + 1)
h i
λ−1,l+1 λ−1,l+1
× (q − ν)(q − ν + 1)Sq+1,ν − (q + ν)(q + ν + 1)Sq−1,ν .

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 715

Thus
s
λ,l ν +2 ρ (q + ν)! (q − l)λ−κ
Sq,ν+2 ≤ (ν )(q − ν)(q + ν + 1)
ν (q − ν + 1)! λ
Γ(q − λ) |G(λ, l)|
×
Γ(q) |G(0, l + λ)|
1 (ν + 1)(ν ρ ) |G(λ, l)|
+
λ ν |G(0, l + λ)|
" s
(q − ν)(q − ν + 1) Γ(q + 2 − λ) (q + 1 + ν)!
× (q − l)λ−κ−1
q Γ(q + 1) (q + 2 − ν)!
s #
(q + ν)(q + ν + 1) Γ(q − λ) (q − 1 + ν)!
+ (q − l − 2) λ−κ−1
,
q Γ(q − 1) (q − ν)!
and therefore
" s #
λ,l (q + ν + 2)! (q − l)(λ−κ)
Γ(q − λ) |G(λ, l)|
Sq,ν+2 ≤ (ν + 2)ρ
(q − ν − 1)! λ Γ(q) |G(0, l + λ)|
" ρ−1
ν (q − ν)(q + ν + 1)
× p
ν +2 (q + ν + 2)(q + ν + 1)(q − ν + 1)(q − ν)
(ν + 1)ν ρ−1
+
(ν + 2)ρ−1
(q − ν)(q − ν + 1)(q − λ + 1)(q − λ)
× p
(q + ν + 2)(q − ν + 2)(q − ν + 1)(q − ν)(q − l)q 2
!#
(q + ν)(q + ν + 1)(q − 1)
+ p .
(q + ν + 2)(q + ν + 1)(q + ν)(q − ν)(q − l − 2)q
Hence
s
λ,l (q + ν + 2)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
Sq,ν+2 ≤ (ν + 2) ρ
.
(q − ν − 1)! λ Γ(q) |G(0, l + λ)|
λ,l
Lemma 3.4. For Sq,ν in (3.13) where −N ≤ ν ≤ 0 we have:
s
(q + ν − 1)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
λ,l
Sq,ν ≤ max(|ν|ρ , 1) ·
(q − ν)! λ Γ(q) |G(0, l + λ)|
where G(λ, l) is defined in (2.2) and subject to the conditions:
4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
λ − 1 > κ > 2,
0 ≥ ν ≥ −N,
ρ > 3,
|N | < q.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 716 ANNE GELB

Proof. Starting with the recurrence relation

λ,l (ν − 2)
Sq,ν−2 = S λ,l
ν(q − ν + 2)(q + ν − 1) q,ν
G(λ, l) 2(ν − 1) 1 h i
λ−1,l+1 λ−1,l+1
+ Sq+1,ν −Sq−1,ν .
G(λ − 1, l − 1) ν(2q + 1) (q − ν + 2)(q + ν − 1)
Choosing η = −ν, so that 0 ≤ η ≤ N , the following recurrence relation is obtained
(A.7)
λ,l (η + 2)
Sq,−(η+2) = S λ,l
η(q + η + 2)(q − η − 1) q,−η
G(λ, l) 2(η + 1) 1
+
G(λ − 1, l − 1) η(2q + 1) (q + η + 2)(q − η − 1)
h i
λ−1,l+1 λ−1,l+1
× Sq+1,−η − Sq−1,−η .
It must be shown that
s
λ,l (q − η − 1)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
Sq,−η ≤ max(|η|ρ , 1) ,
(q + η)! λ Γ(q) |G(0, l + λ)|
where G(λ, l) is defined in (2.2) and with the same parameters
4 ≤ λ ≤ q,
0 ≤ l ≤ m ≤ N,
0 ≤ η ≤ N < q,
λ − 1 > κ > 2,
ρ > 3.
For η = 0 the same first inductive step as in Lemma 3.3 is used.
Substituting into the recurrence relation (A.7) above, the inequality is directly
obtained
s
λ,l (q − η − 3)! (q − l)(λ−κ) Γ(q − λ) |G(λ, l)|
Sq,−(η+2) ≤ (η + 2)ρ .
(q + η + 2)! λ Γ(q) |G(0, l + λ)|

References
1. M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions, Dover, 1970.
2. H. Bateman, Higher Transcendental Functions, v2, McGraw-Hill, 1953. MR 84h:33001b
3. E. Butkov, Mathematical Physics, Addison-Wesley Publishing Company, 1968.
4. D. Gottlieb and S. Orszag, Numerical Analysis of Spectral Methods: Theory and Applications,
SIAM-CBMS, Philadelphia, 1977. MR 58:24983
5. D. Gottlieb, C.-W. Shu, A. Solomonoff and H. Vandeven, On The Gibbs Phenomenon I: recov-
ering exponential accuracy from the Fourier partial sum of a non-periodic analytic function,
J. Comput. Appl. Math. 43 (1992), 81–92. MR 94h:42006
6. D. Gottlieb and C.-W. Shu, On The Gibbs Phenomenon III: recovering exponential accuracy
in a sub-interval from the spectral partial sum of a piecewise analytic function, SIAM J.
Numer. Anal., 33:1 (1996), 280–290. CMP 96:09
7. D. Gottlieb and C.-W. Shu, On The Gibbs Phenomenon IV: recovering exponential accuracy
in a sub-interval from a Gegenbauer partial sum of a piecewise analytic function, Math.
Comp. 64:211 (1995), 1081–1095. CMP 95:11
8. I. Gradshteyn and I. Ryzhik, Tables of Integrals, Series, and Products, Academic Press, 1980.
MR 81g:33001
9. F. John, Partial Differential Equations, Springer-Verlag, 1982. MR 87g:35002

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 RESOLUTION OF THE GIBBS PHENOMENON FOR SPHERICAL HARMONICS 717

10. S. Orszag, Fourier Series on Spheres, Mon. Wea. Rev. 102 (1978), 56–75.
11. P. N. Swarztrauber, On the Spectral Approximation of Discrete Scalar and Vector Functions
on a Sphere, Siam J. Numer. Anal., 16:6 (1979), 934–949. MR 81c:65011
12. H. Weyl, Die Gibbssche Erscheinung in der Theorie der Kugelfunktionen, Gesammelte
Abhandlungen, Springer-Verlag, 1968, 305–320.

Division of Applied Mathematics, Brown University, Providence, Rhode Island 02912

E-mail address: [email protected]

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use