THE APPROXIMATION OF ALMOST TIME AND

BAND LIMITED FUNCTIONS BY THEIR

EXPANSION IN SOME ORTHOGONAL
POLYNOMIALS BASES
Philippe Jaming, Abderrazek Karoui, Susanna Spektor

To cite this version:

Philippe Jaming, Abderrazek Karoui, Susanna Spektor. THE APPROXIMATION OF ALMOST
TIME AND BAND LIMITED FUNCTIONS BY THEIR EXPANSION IN SOME ORTHOGONAL
POLYNOMIALS BASES. Journal of Approximation Theory, 2016, 212, pp.41-65. �hal-01103588�

HAL Id: hal-01103588

https://hal.archives-ouvertes.fr/hal-01103588
Submitted on 15 Jan 2015

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 THE APPROXIMATION OF ALMOST TIME AND BAND LIMITED
FUNCTIONS BY THEIR EXPANSION IN SOME ORTHOGONAL
POLYNOMIALS BASES

PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

Abstract. The aim of this paper is to investigate the quality of approximation of almost time
and almost band-limited functions by its expansion in three classical orthogonal polynomials bases:
the Hermite, Legendre and Chebyshev bases. As a corollary, this allows us to obtain the quality of
approximation in the L2 −Sobolev space by these orthogonal polynomials bases. Also, we obtain
the rate of the Legendre series expansion of the prolate spheroidal wave functions. Some numerical
examples are given to illustrate the different results of this work.

1. Introduction
Time-limited functions and band-limited functions play a fundamental role in signal and image
processing. The time-limiting assumption is natural as a signal can only be measured over a finite
duration. The band-limiting assumption is natural as well due to channel capacity limitations. It is
also essential to apply sampling theory. Unfortunately, the simplest form of the uncertainty principle
tells us that a signal can not be simultaneously time and band limited. A natural assumption is
thus that a signal is almost time- and almost band-limited in the following sense:
Definition. Let T, Ω > 0 and εT , εΩ > 0. A function f ∈ L2 (R) is said to be
• εT -almost time limited to [−T, T ] if
Z
2
|f (t)|2 dt ≤ ε2T kf kL2 (R) ;
|t|>T

• εΩ -almost band limited to [−Ω, Ω] if

Z
2
|fb(ω)|2 dω ≤ ε2Ω kf kL2 (R) .
|ω|>Ω

Here and throughout this paper the Fourier transform is normalized so that, for f ∈ L1 (R),
Z
1
fb(ω) := F[f ](ω) := √ f (t)e−itω dt.
2π R
Of course, given f ∈ L2 (R), for every εT , εΩ > 0 there exist T, Ω > 0 such that f is εT -almost
time limited to [−T, T ] and εΩ -almost time limited to [−Ω, Ω]. The point here is that we consider
T, Ω, εT , εΩ as fixed parameters. A typical example we have in mind is that f ∈ H s (R) and is
time-limited to [−T, T ]. Such an hypothesis is common in tomography, see e.g. [14], where it is
required in the proof of the convergence of the filtered back-projection algorithm for approximate
inversion of the Radon transform. But, if f ∈ H s (R) with s > 0, that is if
Z
2
kf kH s (R) := (1 + |ω|)2s |fb(ω)|2 dω < +∞,
R

Date: January 15, 2015.

1991 Mathematics Subject Classification. 41A10;42C15,65T99.
Key words and phrases. Almost time and band limited functions; Hermite functions; Legendre Polynomials, Cheby-
shev polynomials, prolate spheroidal wave functions.
1
2 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

then
(1 + |ω|)2s b
Z Z
|fb(ω)|2 dω ≤ |f (ω)|2 dω
|ω|>Ω |ω|>Ω (1 + |Ω|)2s
2
kf kH s (R)
≤ .
(1 + |Ω|)2s
1 kf kH s
Thus f is s
-almost band limited to [−Ω, Ω].
(1 + |Ω|) kf kL2 (R)
An alternative to the back projection algorithms in tomography are the Algebraic Reconstruction
Techniques (that is variants of Kaczmarz algorithm, see [14]). For those algorithms to work well it
is crucial to have a good representing system (basis, frame...) of the functions that one wants to
reconstruct.
Thanks to the seminal work of Landau, Pollak and Slepian, the optimal orthogonal system for
representing almost time and band limited functions is known. The system in questions consists
of the so called prolate spheroidal wave functions, ψkT , and has many valuable properties (see [16,
10, 11, 17, 18]). Among the most striking properties they have is that, if a function is almost time
limited to [−T, T ] and almost band limited to [−Ω, Ω] then it is well approximated by its projection
on the first 4ΩT terms of the basis:
X
(1.1) f' f, ψkT ψkT .
0≤k<4ΩT

For more details, see [10]. This is a remarkable fact as this is exactly the heuristics given by
Shannon’s sampling formula (note that to make this heuristics clearer, the functions are usually
almost time-limited to [−T /2, T /2] and this result is then known as the 2ΩT -Theorem, see [10]).
However, there is a major difficulty with prolate spheroidal wave functions that has attracted a lot
of interest recently, namely the difficulty to compute them as there is no inductive nor closed form
formula (see e.g. [2, 3, 4, 13, 21]). One approach is to explicitly compute the coefficients of the prolate
spheroidal wave functions in terms of a basis of orthogonal polynomials like the Legendre polynomials
or the Hermite functions basis. The question that then arises is that of directly approximating almost
time and band limited functions by the (truncation of) their expansion in the Hermite, Legendre
and Chebyshev bases. This is the question we address here.
An other motivation for this work comes from the work of the first author [8] on uncertainty
principles for orthonormal bases. There, it is shown that an orthonormal basis (ek ) of L2 (R) can not
have uniform time-frequency localization. Several ways of measuring localization were considered,
and for most of them, the Hermite functions provided the optimal behavior. However, in one case,
the proof relied on (1.1): this shows that the set of functions that are εT -time limited to [−T, T ]
and εΩ -band limited to [−Ω, Ω] is almost of dimension 4ΩT . In particular, this set can not contain
more than a fixed number of elements of an orthonormal sequence. As this proof shows, the optimal
basis here consists of prolate spheroidal wave functions. As the Hermite basis is optimal for many
uncertainty principles, it is thus natural to ask how far it is from optimal in this case.
Let us now be more precise and describe the main results of the paper. In Section 2, we first
give a brief description of the asymptotic approximation of the Hermite functions in terms of the
sine and cosine functions. Then, we use the asymptotic behaviour of the Hermite function and
give an error analysis of the uniform approximation of the Hermite function projection kernel
Xn
kn (x, y) = hk (x)hk (y) by an appropriate Sinc kernel. Here, hk denotes the k−th L2 -normalized
k=0
Hermite function. Then, based on the previous asymptotic approximation of the Hermite kernel, we
give the quality of almost time- and band-limited functions by Hermite functions. In Section 3, we
use the explicit formula for the finite Fourier transform of the Legendre polynomials in terms of the
Bessel function and give the convergence rate of the Legendre series expansion of a c−band-limited
 ALMOST TIME AND BAND LIMITED FUNCTIONS 3

function. Then, we extend this result to the case of almost time- and band-limited function. In
Section 4, we show the results obtained for the Legendre polynomials to the case of Chebyshev
polynomials. Section 5 is divided into two parts. In the first part, we first give an application of the
results of Section 3 related to the convergence rate of the Legendre series expansion of the prolate
spheroidal wave functions (PSWFs). Note that for a given bandwidth c > 0, and an integer n ≥ 0,
the n−th PSWF, denoted by ψn,c is a c−band-limited function, given as the n−th eigenfunction of a
sin c(x − y)
compact integral operator Qc , defined on L2 ([−1, 1]) with the sinc kernel Kc (x, y) = . In
π(x − y)
the second part of Section 5, we give various numerical examples that illustrate the different results
of this work.

2. Approximation of almost band limited functions by Hermite functions basis.

In this section, we study the quality of approximation of band limited and almost band limited
functions by the Hermite and scaled Hermite functions. For this purpose, we first need to review
the asymptotic uniform approximation of the Hermite functions by the sine and cosine functions.
This is the subject of the following paragraph.

2.1. Approximating Hermite functions with the WKB method. Let Hn be the n-th Hermite
polynomial, that is
2 d
n 2
Hn (x) = ex e−x .
dxn
Define the Hermite functions as
2 1
hn (x) = αn Hn (x)e−x /2 where αn = √ .
π 1/4 2n n!
As is well known:
(i) (hn )n≥0 is an orthonormal basis of L2 (R).
(ii) hn is even if n is even and odd if n is odd, in particular h02p (0) = 0 and h2p+1 (0) = 0. Further
s √ s
(−1)p (2p − 1)!! 0 (−1)p 4p + 2 (2p − 1)!!
h2p (0) = 1/4 and h2p+1 (0) = .
π (2p)!! π 1/4 (2p)!!
(iii) hn satisfies the differential equation h00n (x) + (2n + 1 − x2 )hn (x) = 0.
We will now follow the WKB method to obtain an approximation of hn . In order to√simplify
notation, we will fix n and drop all supscripts during the computation. Let h = hn , λ = 2n + 1,
and define for |x| < λ
Z x
p 1
p(x) = λ2 − x2 , ϕ(x) = p(t) dt and ψ± (x) = p exp ±iϕ(x).
0 p(x)
Note that ψ± have been chosen to have
0 0
ψ+ (x)ψ− (x) − ψ− (x)ψ+ (x) = −2i
and 0 2
p0 p0 2λ2 + 3x2
 
1 1
y 00 + (p2 − q)y = 0 where q = − =− .
2 p 4 p 4p(x)4
Note that h00 (x) + p(x)h(x) = 0 so that
(h0 ψ± − ψ±
0
h)0 = h00 ψ± − ψ±
00
h = −qhψ± .
Let us now define Z x
Q± (x) = q(t)h(t)ψ± (t) dt.
0
4 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

Integrating the previous differential equation between 0 and x, we obtain the system
 0 0
h (x)ψ+ (x) − h(x)ψ+ (x) = h0 (0)ψ+ (0) − h(0)ψ+0
(0) − Q+ (x)
0 0 0 0 .
h (x)ψ− (x) − h(x)ψ− (x) = h (0)ψ− (0) − h(0)ψ− (0) − Q− (x)
It remains to solve this system for h to obtain the principal term of h:
√
Theorem 2.1. Let n ≥ 0, λ = 2n + 1. Then, for |x| ≤ λ,
√ cos ϕn (x) h0n (0) sin ϕn (x)
(2.2) hn (x) = λhn (0) 2 + √ + En (x)
(λ − x2 )1/4 λ (λ2 − x2 )1/4
where
Z xp  5/2
2 2
5 λ
(2.3) ϕn (x) = λ − t dt and |En (x)| ≤ .
0 4 λ 2 − x2
Further, if |x|, |y| ≤ T ≤ λ2 ,
√
ϕn (x) = 2n + 1x − en (x),
where
T3 T2
(2.4) |en (x)| ≤ and |en (x) − en (y)| ≤ |x − y|,
3λ λ
while
2 7
(2.5) |En (x)| ≤ and |En (x) − En (y)| ≤ |x − y|.
λ3 λ5/2
Remark. One may explicitly compute ϕ:
2n + 1 x xp
ϕn (x) = arcsin √ + 2n + 1 − x2 .
2 2n + 1 2
√
Also, ϕn has a geometric interpretation: it this the area of the intersection
√ of a disc of radius 2n + 1
centered at 0 with the strip [0, x] × R+ . In particular, when x → 2n + 1, ϕn (x) ∼ π4 (2n + 1).
The result is not entirely new (e.g. [5, 6, 9, 12, 15]), except for the Lipschitz bounds of E.
Therefore we will only sketch the proof of this theorem in Appendix A. √
Using standard asymptotic of h2p (0) and of h02p+1 (0) and the fact that λ2 − x2 ' λ when
λ → ∞, one may further simplify this result to the following:
Corollary 2.2. Let T ≥ 2 and let n ≥ 2T 2 . Then, for |x| ≤ T , we obtain that
– if n is even, n = 2p
(−1)p
(2.6) h2p (x) = √ 1/4 cos ϕ2p (x) + Ẽ2p (x);
πp
– if n is odd, n = 2p + 1
(−1)p
(2.7) h2p+1 (x) = √ 1/4 sin ϕ2p+1 (x) + Ẽ2p+1 (x),
πp
where, for |x|, |y| ≤ T ,
3T 2 T2
(2.8) |Ẽn (x)| ≤ and |Ẽn (x) − Ẽn (y)| ≤ 8 |x − y|
(2n + 1)5/4 (2n + 1)3/4
To conclude, we will gather some facts about ϕn that all follow from easy calculus.
√
Lemma 2.3. If |x|, |y| ≤ T ≤ 21 2n + 1, then
3T
(2.9) |ϕn+1 (x) − ϕn (x)| ≤ √ ,
2n + 1
3
(2.10) |ϕn+1 (x) − ϕn+1 (y) − ϕn (x) + ϕn (y)| ≤ √ |x − y|,
2n + 1
 ALMOST TIME AND BAND LIMITED FUNCTIONS 5

5T
(2.11) |ϕn+1 (x) − ϕn (x) + ϕn+1 (y) − ϕn (y)| ≤ √ ,
2n + 1
√ √
(2.12) ϕn+1 (x) + ϕn (x) − ϕn+1 (y) − ϕn (y) = ( 2n + 1 + 2n + 3)(x − y) + εn (x, y),
T2
with |εn (x, y)| ≤ √ |x − y| and
2n + 1
5√
(2.13) |ϕn (x) − ϕn (y)| ≤ 2n + 1|x − y|.
4
2.2. The kernel of the projection onto the Hermite functions. As (hn )n≥0 forms an or-
thonormal basis of L2 (R), every f ∈ L2 (R) can be written as
n
X
f (x) = lim hf, hk ihk (x),
n→+∞
k=0

where the limit is in the L2 (R) sense. Further, for n an integer, let Kn f be the orthogonal projection
of f on the span of h0 , . . . , hn . Then
Xn Z
Kn f (x) = hf, hk ihk (x) == kn (x, y)f (y) dy,
k=0 R

n
X
with the kernel kn (x, y) = hk (x)hk (y). According to the Christoffel-Darboux Formula,
k=0
r
n + 1 hn+1 (x)hn (y) − hn+1 (y)hn (x)
kn (x, y) = .
2 x−y
We will now use Corollary 2.2 to approximate this kernel:
√ √
2n+1+ 2n+3
Theorem 2.4. Let T ≥ 2, n ≥ 2T 2 and N = 2 . Then, for |x|, |y| ≤ T ,
1 sin N (x − y)
kn (x, y) = + Rn (x, y),
π x−y
17T 2
with |Rn (x, y)| ≤ √ .
2n + 1
Remark. The same estimate holds for T = 1 provided n ≥ 6. Moreover, we should mention that
in practice, the actual approximation error of the kernel is much smaller than the theoretical error
Rn . See example 1 in the numerical results section that illustrate this fact.
Again, the only improvement over known results [15, 19] is in the estimate of Rn . We will therefore
only sketch the proof in Appendix B.

2.3. Approximating almost time and band limited functions by Hermite functions. We
can now prove the following theorem.
Theorem 2.5. Let Ω0 , T0 ≥ 2 and εT , εΩ > 0. Assume that
Z Z
2 2 2 2
|f (t)| dt ≤ εT kf kL2 (R) and |fb(ω)|2 dω ≤ ε2Ω kf kL2 (R) .
|t|>T0 |ω|>Ω0
2 2
Assume that n ≥ max(2T , 2Ω ). Then, for T ≥ T0 ,
34T 3
 
(2.14) kf − Kn f kL2 ([−T,T ]) ≤ 2εT + εΩ + √ kf kL2 (R)
2n + 1
6 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

Proof. We will introduce several projections. For T, Ω > 0, let

sin Ω(x − y)
Z
1
and QΩ f (x) = F −1 1[−Ω,Ω] fb](x) =

PT f = 1[−T,T ] f f (y) dy.
π R x−y
The hypothesis on f is that kf − PT f kL2 (R) ≤ εT kf kL2 (R) for T ≥ T0 and kf − QΩ f kL2 (R) ≤
εΩ kf kL2 (R) for Ω ≥ Ω0 . Let us also define the integral operator
Z
T
Rn f (x) = Rn (x, y)f (y) dy,
[−T,T ]

where Rn (x, y) are defined in Theorem 2.4. Notice that kn (x, y) = kn (y, x) so that Rn (x, y) =
Rn (y, x).
It is enough to prove (2.14) for T = T0 . We may then reformulate Theorem 2.4 as following:
PT Kn PT f = PT QN PT f + PT RTn PT f,
√ √
2n+1+ 2n+3
where N = 2 . Note that N ≥ Ω0 . By using (2.4), it is easy to see that
PT RTn PT f L2 (R)
≤ PT RTn PT L2 (R)→L2 (R)
kf kL2 (R)
≤ kPT RTn PT kHS kf kL2 (R)
34T 3
(2.15) ≤ √ kf kL2 (R) .
2n + 1
Here we use the well known fact that the Hilbert-Schmidt norm of an integral operator is the L2
norm of its kernel.
Now, using the fact that projections are contractive and N ≥ Ω0 , we have
kf − Kn f kL2 ([−T,T ]) = kPT f − PT Kn f kL2 (R)
≤ kPT f − PT Kn PT f kL2 (R) + kPT Kn (f − PT f )kL2 (R)
≤ PT f − PT QN PT f + PT RTn PT f L2 (R)
+ kf − PT f kL2 (R)
≤ kPT f − PT QN PT f kL2 (R) + PT RTn PT f L2 (R) + kf − PT f kL2 (R) .

Now, write PT QN PT f = PT QN f + PT QN (f − PT f ), then

kPT f − PT QN PT f kL2 (R) ≤ kPT f − PT QN f kL2 (R) + kPT QN (f − PT f )kL2 (R)
≤ kf − QN f kL2 (R) + kf − PT f kL2 (R) .
Therefore,
34T 3
kf − Kn f kL2 ([−T,T ]) ≤ kf − QN f kL2 (R) + √ kf kL2 (R) + 2kf − PT f kL2 (R)
2n + 1
34T 3
 
≤ εΩ + √ + 2εT kf kL2 (R) ,
2n + 1
since N ≥ Ω0 . 

Remark. The error estimate given by (2.14) is not practical due to the low decay rate of the bound
34T 3
of kRTn k given by √ . By replacing this with a non explicit but a more realistic error estimate
2n + 1
kRTn kHS , one gets the following error estimate which is more practical for numerical purposes,
kf − Kn f kL2 ([−T,T ]) ≤ εΩ + kRTn kHS + 2εT kf kL2 (R) .


(2.16)
 ALMOST TIME AND BAND LIMITED FUNCTIONS 7

2.4. Approximating almost time and band limited functions by scaled Hermite func-
tions. For α > 0 and f ∈ L2 (R) we define the scaling operator δα f (x) = α−1/2 f (x/α). Recall that
kδα f kL2 (R) = kf kL2 (R) while
kδα f kL2 ([−A,A]) = kf kL2 ([−A/α,A/αA]) , kδα f kL2 (R\[−A,A]) = kf kL2 (R\[−A/α,A/α])
and F[δα f ] = δ1/α F[f ]. In particular, if f is εT -almost time limited to [−T, T ] (resp. εΩ -almost
band limited to [−Ω, Ω]) then δα f is εT -almost time limited to [−T /α, T /α] (resp. εΩ -almost band
limited to [−αΩ, αΩ]).
Next, define the scaled Hermite basis hα 2
k = δα hk which is also an orthonormal basis of L (R) and
2
define the corresponding orthogonal projections: for f ∈ L (R),
n
X
(2.17) Knα f = hf, hα α
k ihk .
k=0

Proposition 2.6. Let α > 0, T ≥ 2 and c ≥ 2/α. Assume that and

Z Z
2 2
|f (t)|2 dt ≤ ε2T kf kL2 (R) and |fb(ω)|2 dω ≤ ε2c/α kf kL2 (R) .
|t|>T |ω|>c/α
2 2
Then, for n ≥ max(2(T /α) , 2c ), we have
34(T /α)3
 
(2.18) kf − Knα f kL2 ([−T,T ]) ≤ εT + εc/α + √ kf kL2 (R) .
2n + 1
Remark. The scaling with α > 1 has as effect to decrease the dependence on T at the price of
increasing the dependence on good frequency concentration, while taking α < 1 the gain and loss
are reversed. In practice, the above dependence on T is very pessimistic and α > 1 is a better choice.
The most natural choice is α = T and c = T Ω where Ω is such that f is εΩ -almost band limited to
[−Ω, Ω].
Proof. For f ∈ L2 (R), since Knα is contractive, we have
kf − Knα f kL2 ([−T,T ]) ≤ kf − Knα PT f kL2 ([−T,T ]) + kKnα (f − PT f )kL2 ([−T,T ])
≤ kf − Knα PT f kL2 ([−T,T ]) + kf − PT f kL2 ([−T,T ])
≤ kf − Knα PT f kL2 ([−T,T ]) + εT kf kL2 (R) .
Moreover,
n Z T n
X 1X
Knα PT f (x) = hPT f, hα α
k ihk (x) = f (y) hk (x/α)hk (y/α) dy
−T α
k=0 k=0
Z T /α n
X
= f (αt) hk (x/α)hk (t) dt.
−T /α k=0

From this, one easily deduces that kf − Knα PT f kL2 ([−T,T ]) = kfα − Kn fα kL2 ([−αT,αT ]) where fα =
 
δ1/α 1[−T,T ] f . Note that fα is 0-almost time limited to [−T /α, T /α]. Next, writing

α = δα F[1[−T,T ] f ] = δα F[f ] − δα F[1R\[−T,T ] f ]

fc
and, noting that
kδα F[f ]kL2 (R\[−c,c]) = kF[f ]kL2 (R\[−c/α,c/α]) ≤ εc/α kf kL2 (R)
while
δα F[1R\[−T,T ] f ] L2 (R\[−Ω,Ω])
≤ δα F[1R\[−T,T ] f ] L2 (R)

= 1R\[−T,T ] f L2 (R)
≤ εT kf kL2 (R) ,
8 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

we get
fc
α ≤ εc/α kf kL2 (R) + εT kf kL2 (R) .
L2 (R\[−c,c])
It remains to apply Theorem 2.5 to complete the proof. 

3. Approximation of almost band limited functions in the basis of Legendre

polynomials
In agreement with standard practice, we will denote by Pk the classical Legendre polynomials,
defined by the three-term recursion
2k + 1 k
Pk+1 (x) = xPk (x) − Pk−1 (x),
k+1 k+1
with the initial conditions
P0 (x) = 1, P1 (x) = x.
2
These polynomials are orthogonal in L ([−1, 1]) and are normalized so that
Z 1
1
Pk (1) = 1 and Pk (x)2 dx = .
−1 k + 1/2
p
We will denote by P̃k the normalized Legendre polynomial P̃k = k + 1/2Pk and the P̃k ’s then
form an orthonormal basis of L2 ([−1, 1]).
In the sequel, for c > 0, let Bc denote the Paley-Wiener space of c-bandlimited functions, given
by
Bc = {f ∈ L2 (R); Supp fb ⊆ [−c, c]}.
Lemma 3.1. Let c > 0, then for any f ∈ Bc , and any k ≥ 0
r  k+1
2 e ec
(3.19) |hf, Pk iL2 (−1,1) | ≤ √ kf kL2 (R) .
2k + 1 πc 2k + 3
Proof. We start from the following identity relating Bessel functions of the first type to the finite
Fourier transform of the Legendre polynomials, see [1]: for every x ∈ R
Z 1
(3.20) eixy Pk (y) dy = 2ik jk (x)
−1
 k
1 d sin x
where jk is the spherical Bessel function defined by jk (x) = (−x)k . Note that jk has
r x dx x
π
same parity as n and recall that, for x ≥ 0, jk (x) = Jk+1/2 (x) where Jα is the Bessel function
2x
of the first kind. In particular, we have the well known bound for x ∈ R
|x|α eα+1
(3.21) |Jα (x)| ≤ ≤ √ |x|α
2α Γ(α + 1) 2π2α (α + 1)α+1/2
√
since Γ(x) ≥ 2πxx−1/2 e−x . From this we deduce that
ek+3/2
(3.22) |jk (x)| ≤ √ |x|k .
2(2k + 3)k+1
Now, since f ∈ Bc , the Fourier inversion theorem implies that, for x ∈ R, we have
Z c Z 1
1 c
(3.23) f (x) = √ fb(ξ)ei x ξ dξ = √ fb(cη)ei c x η dη.
2π −c 2π −1
 ALMOST TIME AND BAND LIMITED FUNCTIONS 9

Combining (3.20) and (3.23), one gets

Z 1 Z 1 Z 1 
c
hf, Pk iL2 ([−1,1]) = f (x)Pk (x) dx = √ fb(cη) e−icxη Pk (x) dx dη
−1 2π −1 −1
r Z 1
2
= ik c jk (cη)fb(cη) dη.
π −1
Using (3.22) together with Cauchy-Schwarz and a change of variable, one gets
Z 1
k+1 ek+3/2
|hf, Pk iL2 ([−1,1]) | ≤ c √ |η|k |fb(cη)| dη
π(2k + 3)k+1 −1
r r  Z c 1/2
k+1 ek+3/2 2 2 1 b(η)|2 dη
≤ c |f .
(2k + 3)k+1 2k + 1 π c −c

Finally, Parseval’s identity implies (3.19). 

Let us now introduce the following orthogonal projections on L2 (R):

N D
X E
P f = 1(−1,1) f, Qc f = F −1 [1(−c,c) Ff ] and LN f = P f, P̃k P̃k 1(−1,1) .
k=0
2
Note that LN is the orthogonal projection onto the subspace of L (R) consisting of functions of the
P (x)1(−1,1) with P a polynomial of degree ≤ N .
Theorem 3.2. Let c > 0, then for any f ∈ Bc , and any N ≥ ec
2 , we have
r  N
c ec
(3.24) kf − LN f kL∞ (−1,1) ≤ kf kL2 (R) .
2N + 5 2N + 5
and
N +1
√

ec
(3.25) kf − LN f kL2 (−1,1) ≤ c kf kL2 (R) .
2N + 5
Proof. Note that, for x ∈ (−1, 1),
+∞
X
f (x) − LN f (x) = hf, P̃k iP̃k (x).
k=N +1
p
But max |P̃k (x)| = |P̃k (1)| = k + 1/2, so that Lemma 3.1 implies
x∈(−1,1)

+∞
X
kf − LN f kL∞ (−1,1) ≤ (k + 1/2)|hf, Pk i|
k=N +1
+∞ k+1
e X √
r 
ec
≤ 2k + 1 kf kL2 (R)
πc 2k + 3
k=N +1
r +∞  k
e ec X ec
≤ √ kf kL2 (R)
2N + 5 2π k=N +1 2N + 5
r  N
c ec
≤ kf kL2 (R) .
2N + 5 2N + 5
If N ≥ ec/2, we then deduce (3.24).
10 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

The proof of the L2 -bound is essentially the same:

+∞
X
2
kf − LN f kL2 (−1,1) ≤ (k + 1/2)|hf, Pk i|2
k=N +1
+∞  2k+2
e X ec
≤ kf k2L2 (R)
2πc 2k + 3
k=N +1
+∞  2k+2
e2 X ec
≤ kf k2L2 (R) .
2π 2N + 5
k=N +1

From this (3.25) easily follows when N ≥ ec/2. 

From this theorem, we simply get the following corollary:
Theorem 3.3. Let c > 0 and assume that f is εT -concentrated to (−1, 1) and εΩ -concentrated to
(−c, c). Then, if N ≥ ec/2,
N +1 !
√

ec
(3.26) kf − LN f kL2 (−1,1) ≤ 2εΩ + c kf kL2 (R)
2N + 5
and
N +1 !
√

ec
(3.27) kf − LN f kL2 (R) ≤ εT + 2εΩ + c kf kL2 (R)
2N + 5
Proof. First
kf − LN f kL2 (−1,1) ≤ kf − Qc f kL2 (−1,1) + kQc f − LN Qc f kL2 (−1,1) + kLN (Qc f − f )kL2 (−1,1)
≤ 2kf − Qc f kL2 (R) + kQc f − LN Qc f kL2 (−1,1) .
But kf − Qc f kL2 (R) ≤ εΩ kf kL2 (R) and Qc f ∈ Bc with kQc f kL2 (R) ≤ kf kL2 (R) .
It remains to notice that
kf − LN f kL2 (R) ≤ kf − PT f kL2 (R) + kf − LN f kL2 (−1,1)
so that (3.27) follows. 

4. Approximation of almost band limited functions in the basis of Chebyshev

polynomials
In this paragraph, we show that the basis of the Chebyshev polynomials is also well adapted
for the approximation of almost band limited functions. This is essentially done by showing that
the weighted finite Fourier transform of the Chebyshev polynomial is given by a formula similar to
(3.20). We first recall that the classical Chebyshev polynomials Tk are defined by the three-term
recursion
Tk+1 (x) = 2xTk (x) − Tk−1 (x),
with the initial conditions
T0 (x) = 1, T1 (x) = x.
1
These polynomials are orthogonal in L2 ([−1, 1], dµ) where dµ(x) = √1−x 2
dx and are normalized
so that
Z 1 (
2 π 2 if k = 0
(4.28) Tk (1) = 1 and Tn (x) dµ(x) = ck with ck =
−1 2 1 if k ≥ 1
.
It is interesting to also note that Tk (x) are simply given by the formula
Tk (cos θ) = cos(kθ), k ∈ N, θ ∈ [0, π].
 ALMOST TIME AND BAND LIMITED FUNCTIONS 11

q
2
We will denote by T̃k the normalized Chebyshev polynomial T̃k = ck π Tk and the T̃k ’s then form
an orthonormal basis of L2 ([−1, 1], dµ).
The following lemma gives us an explicit formula for the weighted Finite Fourier transform of Tk ,
that we failed to find in the literature.

Lemma 4.1. For any k ∈ N, Tbk , the weighted finite Fourier transform of Tk is given by
Z 1
1 π
(4.29) Tbk (x) = eixy Tk (y) p dy = ik Jk (x).
−1 1−y 2 2

Proof. This results follows directly from the formula

Z 1 Z π
f (y)Tk (y)
p dy = f (cos θ) cos kθ dθ
−1 1 − y2 0

applied to f (y) = eixy and the Poisson integral representation formula of the Bessel function. 

For f ∈ L2 ([−1, 1], dµ) we now define

n D
X E
Tn f = f, T̃k T̃k
k=0

the projection of f on Cn [X] the subspace of L2 ([−1, 1], dµ) consisting of polynomials of degree ≤ n.
We can now prove the Chebyshev version of Lemma 3.1 and the approximation rate of band-limited
functions by their projection on the Chebyshev orthonomal basis in L2 ([−1, 1] dµ). However, note
that an L2 (R) function restricted to [−1, 1] need not be in L2 ([−1, 1] dµ). Therefore, its expansion
in the Chebyshev system need not converge (and not even be defined). Thus, we cannot extend
Theorem 3.3 to the Chebyshev setting.

Proposition 4.2. Let c > 0, then for any f ∈ Bc , and any k ≥ 0

 k+1
1 ec
(4.30) |hf, Tk iL2 ([−1,1],dµ) | ≤ p kf kL2 (R) ,
(2k + 1)c 2(k + 1)

and, if N ≥ ec/2,
√  N +1
e c ce
kf − TN f kL2 ([−1,1],dµ) ≤ kf k2L2 (R) .
2(2N + 3) 2N + 4
Proof. Since f ∈ Bc , then the Fourier inversion theorem implies that, for x ∈ R, we have
Z c Z 1
1 c
f (x) = √ fb(ξ)ei x ξ dξ = √ fb(cη)ei c x η dη.
2π −c 2π −1
Combining this with (3.20), one gets
Z 1
dx
hf, Tk iL2 ([−1,1],dµ) = f (x)Tk (x) √
−1 1 − x2
Z 1 Z 1 
c −icxη dx
= √ f (cη)
b e Tk (x) √ dη
2π −1 −1 1 − x2
√ Z 1
c 2π π
= ik Jk (x)(cη)fb(cη) dη.
4 −1 2
12 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

Using (3.21) together with Cauchy-Schwarz inequality and a change of variable, one gets
Z 1
ek+1
|hf, Tk iL2 ([−1,1]) | ≤ ck+1 k+2 |η|k |fb(cη)| dη
2 (k + 1)k+1/2 −1
r Z c 1/2
k+1/2 ek+1 2 2
≤ c |f (η)| dη
b
2k+3/2 (k + 1)k+1 2k + 1 −c

To conclude, it suffices to use Parseval’s identity.

From the orthonormality of the T̃k ’s and this bound, we deduce that
+∞
X
2
kf − TN f kL2 ([−1,1],dµ) ≤ |hf, T̃k iL2 ([−1,1],dµ) |2
k=N +1
+∞
X 1 e2k+2
≤ c2k+1 2k+3 kf k2L2 (R)
2k + 1 2 (k + 1)2k+2
k=N +1
+∞  2k+1
1 e X ce
≤ kf k2L2 (R)
2N + 3 2N + 4 2(k + 1)
k=N +1
2
 2N +4
e c 1 ec
≤ kf k2L2 (R)
4 (2N + 3)2 2N + 4
provided N ≥ ec/2. 

5. Applications and numerical results

In the first part of this last section, we apply the quality of approximation of c−bandlimited
functions by Legendre polynomials in the framework of prolate spheroidal wave functions (PSWFs).
As a consequence, we give the convergence rate of the Flammer’s scheme, see [7] for the computation
of the PSWFs.
5.1. Approximation of prolate spheroidal wave functions. For a given real number c > 0,
called bandwidth, the Prolate spheroidal wave functions (PSWFs) denoted by (ψn,c (·))n≥0 , are
defined as the bounded eigenfunctions of the Sturm-Liouville differential operator Lc , defined on
C 2 ([−1, 1]), by
d2 ψ dψ
(5.31) Lc (ψ) = −(1 − x2 ) + 2x + c2 x2 ψ.
d x2 dx
They are also the eigenfunctions of the finite Fourier transform Fc , as well as the ones of the operator
c ∗
Qc = F Fc , which are defined on L2 ([−1, 1]) by
2π c
Z 1 Z 1
sin(c(x − y))
(5.32) Fc (f )(x) = ei c x y f (y) dy, and Qc (f )(x) = f (y) dy.
−1 −1 π(x − y)
They are normalized so that their L2 ([−1, 1]) norm is equal to 1 and ψn,c (1) > 0. We call (χn (c))n≥0
the corresponding eigenvalues of Lc , µn (c) the eigenvalues of Fc and λn (c) the ones of Qc . A well
known property is then that kψn,c kL2 (R) = √ 1 .
λn (c)
The crucial commuting property of Lc and Qc has been first observed by Slepian and co-authors
[16], whose name is closely associated to all properties of PSWFs and their associated spectrum.
Among their basic properties we cite their analytic extension to the whole real line and their unique
properties to form an orthonormal basis of L2 ([−1, 1]) and an orthonormal basis of Bc . A well known
estimate for χn (c) is
(5.33) n(n + 1) ≤ χn (c) ≤ n(n + 1) + c2 .
 ALMOST TIME AND BAND LIMITED FUNCTIONS 13

c
Recall that λn (c) and µn (c) are related by λn (c) = |µn (c)|2 . A precise asymptotic of λn (c) has
2π
been established by Widom [20]. Recently in [3], the authors have given an explicit approximation
of the λn (c), valid for n > 2c/π that gives rise to the exact super-exponential decay rate of the
sequence of these eigenvalues. But, here we want a lower bound that is valid for all n. According to
[11],
1
(5.34) and λ[ 2 c]+1 < < λ[ 2 c]−1
0 < λn (c) < 1
π 2 π

while Bonami-Karoui established the following bound, see [2]

 5(n+1)  
2 2c 2
(5.35) λn (c) ≥ for n ≥ max 3, c .
5 π(n + 1) π
In Appendix C we will prove the following slight improvement of this bound:
2
Proposition 5.1. Let c be a real number. Then, if n > c,
π
 2 
2c c 2n−1
λn (c) ≥ 7 1 − .
nπ 7πn
 
2 4
If n = c , λn (c) ≥ .
π π + 2c
Since ψn,c ∈ L2 ([−1, 1]), we may expand it in the Legendre basis
+∞ D +∞  
X E X 1
ψn,c = ψn,c , P̃k P̃k = k+ hψn,c , Pk iPk .
2
k=0 k=0

1


Notation : Let us write βkn (c) = k+ 2 hψn,c , Pk i so that, on [−1, 1],
+∞
X
(5.36) ψn,c = βkn (c)Pk .
k=0

Rokhlin, Xiao and Yarvin [21] have obtained induction formulas for the βkn (c)’s in order to compute
the ψn,c ’s. Let us now obtain an estimate for them:
Corollary 5.2. With the above notation, we have
pe√ k+1
 
ec
2 



 2 πc 2k + 1 2k+3 if n ≤ πc − 1
q  k+1
|βkn (c)| ≤ e(π+2c)
p ec
2 
2πc k + 1/2 2k+3 if n = π c .
q
  k+1
 2e √1 1 − 2c −1 7πn n−1/2 k + 1/2
 p ec


2 
if n ≥ π c + 1

7π c nπ c 2k+3

Proof. Since ψn,c ∈ Bc (R), from Lemma 3.1 we deduce that

r  k+1 r  k+1
e 1 ec e 1 ec 1
|hψn,c , Pk i| ≤ 2 √ kψn,c kL2 (R) = 2 √ p
πc 2k + 1 2k + 3 πc 2k + 1 2k + 3 λn (c)
To conclude, it suffices to use the lower bounds of λn (c) given by (5.34) and the previous proposition.


From this, one can then easily obtain error estimates for the approximation of prolate spheroidal
wave functions by the truncation of their expansion in the Legendre basis in the spirit of Theorem
3.3.
14 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

5.2. Numerical results. In this paragraph, we give several examples that illustrate the different
results of this work.

Example 1. In this example, we check numerically that the actual error of the uniform approx-
n
X
imation of the kernel kn (x, y) = hk (x)hk (y) may be much smaller than the theoretical error
k=0
given by Theorem 2.4. For this purpose, we have considered the value T = 1 and various values
of the integer 10 ≤ n ≤ 100. For each value of n, we have used a uniform discretization Λ of
the square [−1, 1]2 with equidistant 6400 nodes. Then, we have computed over these grid points,
a highly accurate approximation E en = sup kn (x, y) − sin N (x − y) of the exact uniform error
x,y∈Λ π(x − y)
sin N (x − y)
En = sup kn (x, y) − . The obtained results are given by Table 1.
x,y∈[−1,1] π(x − y)

n 10 25 50 75 100
E
en 0.067 0.039 0.025 0.023 0.022
Table 1. Approximate errors E
en for various values of n.

Example 2. In this example, we illustrate the quality of approximation by scaled Hermite functions
of a time-limited and an almost band-limited function. For this purpose, we consider the function
f (x) = 1[−1/2,1/2] (x). From the Fourier transform of f , one can easily check that f ∈ H s (R) for
any s < 1/2. Note that f is 0-concentrated in [−1/2, 1/2] and since f ∈ H s (R), f is Ω -band
concentrated in [−Ω, +Ω], with Ω < Ms Ω−s with Ms a positive constant. We have considered the
value of c = 100 and we have used (2.17) to compute the scaled Hermite approximations Knc (f ) of f
with n = 40 and n = 80. The graphs of f and its scaled Hermite approximation are given by Figure
1. In Figure 2, we have given the approximation errors f (x) − Knc f (x).
Also, to illustrate the fact that the scaled Hermite approximation outperforms the usual Hermite
approximation, we have repeated the previous numerical tests without the scaling factor (this corre-
sponds to the special case of c = 1). Figure 3 shows the graphs of f and Kn f . This clearly illustrates
the out-performance of the scaled Hermite approximation, compared to the usual Hermite approxi-
mation.

Figure 1. Graph of the approximation of f (x) (red) by Knc f (x), c = 100 (blue)
with (a) n = 40 (b) n = 80.
 ALMOST TIME AND BAND LIMITED FUNCTIONS 15

Figure 2. Graph of the approximation error f (x) − Knc f (x), c = 100 with (a)
n = 40 (b) n = 80.

Figure 3. Graph of the approximation of f (x) (red) by Knc f (x), c = 1 (blue) with
(a) n = 40 (b) n = 80.

Example 3. In this example, we illustrate the decay rate of the Legendre and Chebyshev expansion
coefficients of a c-bandlimited functions, that we have given by (3.19) and (4.30), respectively. For
sin(cx)
this purpose, we have considered the function f ∈ Bc , given by fc (x) = , x ∈ R. Then, we
cx
have computed the different Legendre and Chebyshev expansion coefficients ln (f ) = f, Pn L2 (I)
and cn (f ) = f, Tn L2 (I,dµ)
of fc , for the two values of c = 10 and c = 50. In Figure 4, we plot the
graphs of the log(|ln (f )|), log(|cn (f )|), n ≥ ec
 
2 + 1 versus the logarithm of their respective error
bounds given by (3.19) and (4.30).

Example 4. In this last example, we illustrate the quality of approximation by Legendre and
Chebyshev polynomials in the Sobolev spaces H s (I). We have considered the two functions f, g
given by f (x) = 1[−1/2,1/2] (x) and g(x) = (1 − |x|)1[−1,1] (x). It is clear that g ∈ H s (I), ∀s < 3/2.
In Figure 5, we plot the graphs of the approximation error of f by its corresponding projections
LN f and TN f over the subspaces spanned by the first N + 1 Legendre and Chebyshev polynomials,
respectively, with N = 50. In Figure 6, we plot the graphs of g − LN g and g − TN g with N = 50.

Acknowledgements
The first author kindly acknowledge financial support from the French ANR programs ANR 2011
BS01 007 01 (GeMeCod), ANR-12-BS01-0001 (Aventures). This study has been carried out with
financial support from the French State, managed by the French National Research Agency (ANR)
16 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

Figure 4. (a) graph of log(|ln (fc )|), c = 10 (in red) versus the logarithm of its
bound (3.19) (in blue), (b) graph of log(|cn (fc )|), c = 10 (in black) versus the
logarithm of its bound (4.30) (in blue); (c) and (d) same as in (a) and (b) with
c = 50.

Figure 5. (a) Graphs of the errors f (x) − LN f (x), with N = 50 (b) same as (a)
with f (x) − TN f (x),.

in the frame of the Investments for the future Programme IdEx Bordeaux - CPU (ANR-10-IDEX-
03-02).
The last author kindly acknowledges the financial support form Michigan State University and
from NSF CAREER grant, Award No. DMS-1056965.
 ALMOST TIME AND BAND LIMITED FUNCTIONS 17

Figure 6. (a) Graphs of the errors g(x) − LN g(x), with N = 50 (b) same as (a)
with g(x) − TN g(x),.

Appendix A. Proof of Theorem 2.1

.
We will again drop the index n and use the notation introduced before the statement of the
theorem.
The bounds for e(x) are obtained by standard calculus, we will thus omit the proof. As for E(x),
the computation shows
Z x
1 q(t)

E(x) = p p h(t) sin ϕ(x) − ϕ(t) dt.
p(x) 0 p(t)
Using Cauchy-Schwarz, we obtain
Z x 1/2 Z x 1/2 Z x 1/2
1 q(t)2 2 1 25λ4
|E(x)| ≤ p dt h(t) dt ≤ p 9
dt
p(x) 0 p(t) 0 p(x) 0 16p(t)

5λ5/2
since khn k2 = 1. As |x| < λ, and p decreases, the estime |E(x)| ≤ follows. When |x| ≤ λ/2,
4p(x)5
the change of variable y = λs and a numerical computation shows that |E(x)| ≤ λ23 .
Note that this bound on E directly leads to a bound on h. For instance, if n ≥ 2 is even, then
1
|h2n (x)| ≤ p for |x| ≤ λ/2.
p(x)
The Lipschitz bound on E is a bit more subtle so let us give more details. First, we introduce
some further notation:
Z x
q(t)

χ(x, t) = p h(t) sin ϕ(x) − ϕ(t) and Φ(x, y) = χ(y, t) dt.
p(t) 0

Now, write
!
1 1 1  
E(y) − E(x) = p −p Φ(y, y) + p Φ(y, y) − Φ(x, y)
p(y) p(x) p(x)
1  
p Φ(x, y) − Φ(x, x) = E1 + E2 + E3 .
p(x)
|x − y|
We have proved that for |y| < λ/2, Φ(y, y) ≤ 2λ−3 . Simple calculus then implies that |E1 | ≤
λ9/2
when |x|, |y| < λ/2.
18 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

Next, if |x|, |y| ≤ (1 − η)λ one can estimate E2 as follows:

Z y
q(t)
|Φ(y, y) − Φ(x, y)| ≤ |χ(y, t)| dt ≤ |x − y| sup p sup |h(t)|
x |t|≤λ/2 p(t) |t|≤λ/2
5λ2
≤ |x − y|.
4p(y)5

3
Therefore, |E2 | ≤ |x − y|.
λ7/2
Finally,
Z x
q(t) h

i
Φ(x, y) − Φ(x, x) = p h(t) sin ϕ(y) − ϕ(t) − sin ϕ(x) − ϕ(t) dt
0 p(t)
Z x
q(t) ϕ(x) + ϕ(y) − 2ϕ(t) ϕ(y) − ϕ(x)
= 2 p h(t) cos dt sin .
0 p(t) 2 2

The integral is estimated in the same way as we estimated Φ(x, x), while for ϕ we use the mean
2
value theorem and the fact that ϕ0 = p. We, thus, get |E3 | ≤ 5/2 |x − y|. The estimate for E
λ
follows.

Appendix B. Proof of Theorem 2.4

.
For sake
√ of simplicity, √
we will only prove the theorem in the case when n is even and write n = 2p.
Let λ = 2n + 1, µ = 2n + 3, α = √πp1 1/4 , β = √πp1 1/4 , E = (−1)p Ẽ2p and F = (−1)p Ẽ2p+1 .
Then, according to Corollary 2.2
(
(−1)p √πp1 1/4 cos ϕ2p (x) + E(x)


h2p (x) =

.
h2p+1 (x) = (−1)p √πp1 1/4 sin ϕ2p+1 (x) + F (x)

Therefore, h2p+1 (x)h2p (y) − h2p+1 (y)h2p (x) is

1

= 1/2
sin ϕ2p+1 (x) cos ϕ2p (y) − sin ϕ2p+1 (y) cos ϕ2p (x)
πp
1

+ √ 1/4 F (x) cos ϕ2p (y) − F (y) cos ϕ2p (x)
πp
1

+ √ 1/4 sin ϕ2p+1 (x)E(y) − sin ϕ2p+1 (y)E(x)
πp
+F (x)E(y) − F (y)E(x).

The last three terms are all of the form

A(x)B(y) − B(x)A(y) = A(x) − A(y) B(y) + B(y) − B(x) A(y)

and are thus bounded with the help of the uniform and Lipschitz bounds of A and B by a factor of
|x − y|.
 ALMOST TIME AND BAND LIMITED FUNCTIONS 19

The first term is the principal one. Let us start by computing

C := sin ϕ2p+1 (x) cos ϕ2p (y) − sin ϕ2p+1 (y) cos ϕ2p (x)
1


= sin ϕ2p+1 (x) + ϕ2p (y) − sin ϕ2p+1 (x) − ϕ2p (y)
2


− sin ϕ2p+1 (y) + ϕ2p (x) + sin ϕ2p+1 (y) − ϕ2p (x)
ϕ2p+1 (x) − ϕ2p+1 (y) − ϕ2p (x) + ϕ2p (y)
= sin
2
ϕ2p+1 (x) + ϕ2p+1 (y) + ϕ2p (x) + ϕ2p (y)
× cos
2
ϕ2p+1 (y) + ϕ2p (y) − ϕ2p (x) − ϕ2p+1 (x)
+ sin
2
ϕ2p+1 (x) − ϕ2p (x) − ϕ2p (y) + ϕ2p+1 (y)
× cos
2
= S1 C1 + S2 (C2 − 1) + S2 .
Now, according to (2.10),
|ϕ2p+1 (x) − ϕ2p+1 (y) − ϕ2p (x) + ϕ2p (y)| 3
|S1 C1 | ≤ |S1 | ≤ ≤ √ |x − y|.
2 2 2n + 1
With (2.11),
|ϕ2p+1 (x) − ϕ2p (x) − ϕ2p (y) + ϕ2p+1 (y)|2 25T 2
|C2 − 1| ≤ ≤ .
2 2(2n + 1)
Thus, with (2.12),
T2 25T 2 16T 2
 
N+√
|S2 (C2 − 1)| ≤ |x − y| ≤√ |x − y|.
2n + 1 2(2n + 1) 2n + 1


Finally, using again Lemma 2.3, sin N (y − x) + εn (y, x) is


= sin N (y − x) + sin N (y − x) cos εn (y, x) − 1 + cos N (y − x) sin εn (x, y)
= sin N (y − x) + E2 (x, y),
where
|εn (x, y)|2 2T 2
|E2 (x, y)| ≤ |εn (x, y)| + ≤√ |x − y|.
2 2n + 1
It remains to group all estimates to get the result.

Appendix C. Proof of Proposition 5.1

.
2
Recall that we want to prove that, if n > c,
π
 2 
2c c 2n−1
λn (c) ≥ 7 1 − .
nπ 7πn
 
2 4
If n = c , λn (c) ≥ .
π π + 2c
According to the min-max Theorem, for any n-dimensional subspace V of L2 (R)
hQc f, f iL2 (R)
λn (c) ≥ min 2 .
f ∈V \{0} kf kL2 (R)
20 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

To show the theorem, we consider V to be the space of functions that is constant on each interval
of the form
 
2j 2j + 2
Ij := −1 + , −1 + , j = 0, . . . , n − 1.
n n
n−1
X
Take f ∈ V and write f = fj χIj . Then
j=0

n−1
2 2X
kf k2 = |fj |2 .
n j=0

On the other hand, write

Z
1
fb(ω) := F[f ](ω) := √ f (t)e−itω dt
2π R

for the Fourier transform and note that Qc (f ) = F −1 1[−c,c] F . Parseval’s Identity shows that

Z c
hQc f, f iL2 (R) = |fb|2 dξ.
−c

But
Z 1 n−1 Z −1+ 2j+2
1 −iξt 1 X n
fb(ξ) = √ f (t)e dt = √ fj e−iξt dt
2π −1 2π j=0 −1+ 2j
n

n−1 Z n1
1 X
−iξ (−1+ 2j+1 )
= √ f e j n e−iξs ds
2π j=0
1
−n
n ξ
1 2X ξ n−1 sin
= √ fj e−2ij n eiξ n n
.
2π n j=1 ξ/n

Therefore,
2 2
n
c c
sin nξ
Z Z
2 X ξ
−2ij n
|fb(ξ)|2 dξ = fj e dξ
−c πn2 −c j=1 ξ/n
2
Z 2c n  2
1 n X
−ijη sin η/2
(C.37) = fj e dη.
πn − 2c
n j=1
η/2

2 sin t 2
But, if n > c and |η| < 2c/n then |η/2| < π/2. Now, on [−π/2, π/2], ≥ 1 − |t|. Therefore
π t π
 2  2
sin η/2 2c
≥ 1− . It follows from (C.37) that
η/2 nπ
2
Z c  2 Z 2c n
2 2c 1 n X
−ijη
|fb(ξ)| dξ ≥ 1− fj e dη
−c nπ πn − 2c
n j=1
2
2 2(n−1) Z n
4 nc π
 
2c 1 c X
≥ 1− fj e−ijη dη
nπ πn πn 14 × 2π −π j=1
 ALMOST TIME AND BAND LIMITED FUNCTIONS 21

with the help of Turan’s Lemma [Na2]. Therefore

n−1
Z c  2  c 2(n−1) 2
2c c 1 X
|fb(ξ)|2 dξ ≥ 1− 7 n
2π |fj |2
−c nπ 7πn 7π πn
j=− n+1
2
 2 
2c c 2n−1 2
= 7 1− kf k2 .
nπ 7πn
2c
The estimate of λn (c) follows. If n ≤ we may now modify the argument starting from (C.37):
π
2
Z c Z π Xn  2
1 −ijη sin η/2
(C.38) |fb(ξ)|2 dξ ≥ fj e dη
−c πn −π j=1 η/2
2
π n  2
|η|
Z
1 X
−ijη
(C.39) ≥ fj e 1− dη
πn −π j=1 π

sin t 2|t|
since ≥1− on [−π/2, π/2]. But, for ` ∈ Z,
t π
Z π  2 (
2π
|η| if ` = 0
1− e−i`η dη = 34 .
−π π π`2 6 0
if ` =
Therefore, using Parseval’s equality, one gets
2 !
Z c Z π n n
21 X 6 X cos `η
|fb(ξ)|2 dξ ≥ fj e−ijη 1+ 2 dη
−c n3 −π j=1 π `2
`=1
2 !
Z π n n
21 X
−ijη 6 X 1
≥ fj e 1− 2 dη
n3 −π j=1 π `2
`=1
2
Z π n ∞
2 1 X
−ijη 4 X 1
= fj e dη )
n 2π −π j=1 π `2
`=n+1
∞
! Z ∞
4 X 1 2 4 dx 2
= kf k ≥ kf k
π `2 π n+1 x2
`=n+1
4 2
≥ kf k .
π(n + 1)
2 4 4
Therefore, for n ≤ π c, λn (c) ≥ ≥ .
π(n + 1) π + 2c

References
[1] G. E. Andrews, R. Askey and R. Roy Special Functions. Cambridge University Press, 2000.
[2] A. Bonami & A. Karoui Uniform bounds of prolate spheroidal wave functions and eigenvalues decay. C. R.
Math. Acad. Sci. Paris 352 (2014), 229-234.
[3] A. Bonami & A. Karoui Spectral Decay of Time and Frequency Limiting Operator, submitted for publication
(2014).
[4] J. P. Boyd Prolate spheroidal wave functions as an alternative to Chebyshev and Legendre polynomials for
spectral element and pseudo-spectral algorithms. J. Comput. Phys. 199 (2004), 688–716.
[5] M. Brannan, R. Kerman & M. L. Huang Error estimates for Dominici’s Hermite function asymptotic formula
and applications. The ANZIAM Journal 50 (2009), 550–561.
22 PHILIPPE JAMING, ABDERRAZEK KAROUI, SUSANNA SPEKTOR

[6] D. Dominici Asymptotic analysis of the Hermite polynomials from their differential-difference equation. J. Dif-
ference Equ. Appl. 13 (2007), 1115–1128.
[7] C. Flammer Spheroidal Wave Functions. Stanford University Press, 1956.
[8] Ph. Jaming & A. Powell Uncertainty principles for orthonormal bases. J. Functional Analysis, 243 (2007),
611–630.
[9] H. Koch & D. Tataru Lp eigenfunction bounds for the Hermite operator. Duke Math. J. 128 (2005), 369-392.
[10] H. J. Landau & H. O. Pollak Prolate spheroidal wave functions, Fourier analysis and uncertainty II. Bell
System Tech. J. 40 (1961), 65–84.
[11] H. J. Landau & H. O. Pollak Prolate spheroidal wave functions, Fourier analysis and uncertainty III: The
dimension of space of essentially time- and band-limited signals, Bell Syst. Tech. J. 41 (1962) 1295–1336.
[12] L. Larsson-Cohn Lp -norms of Hermite polynomials and an extremization problem on Wiener chaos. J. Approx.
Theory 117 (2002), 152–178.
[13] L. W. Li, X. K. Kang & M. S. Leong Spheroidal wave functions in electromagnetic theory. Wiley-Interscience
publication, 2001.
[14] F. Natterer The mathematics of computerized tomography Classics in Applied Math. 32 SIAM, 2001.
[Na2] F. L. Nazarov Local estimates for exponential polynomials and their applications to inequalities of the uncer-
tainty principle type. (Russian) Algebra i Analiz 5 (1993) 3–66; translation in St. Petersburg Math. J. 5 (1994)
663–717.
[15] G. Sansone Orthogonal Functions. Pure and Applied Math,. Interscience Publishers, Inc., New York, 1959.
[16] D. Slepian & H. O. Pollak Prolate spheroidal wave functions, Fourier analysis and uncertainty I. Bell System
Tech. J. 43 (1964), 3009–3058.
[17] D. Slepian Prolate spheroidal wave functions, Fourier analysis and uncertainty IV: Extensions to many dimen-
sions; Generalized prolate spheroidal wave functions. Bell System Tech. J. 40 (1961), 43–64.
[18] D. Slepian Some comments on Fourier analysis, uncertainty and modeling. SIAM Rev. (1983), 379–393.
[19] J. V. Uspensky On the Development of Arbitrary Functions in Series of Hermite’s and Laguerre’s Polynomials.
Ann. Math. (2) 28 (1927), 593–619.
[20] H. Widom Asymptotic behavior of the eigenvalues of certain integral equations. II. Arch. Rational Mech. Anal.,
17 (1964), 215–229.
[21] H. Xiao, V. Rokhlin & N. Yarvin Prolate spheroidal wave functions, quadrature and interpolation. Inverse
Problems 17 (2001), 805–838.

Philippe Jaming Address: Institut de Mathématiques de Bordeaux UMR 5251, Université Bordeaux 1,
cours de la Libération, F 33405 Talence cedex, France
E-mail address: [email protected]

Abderrazek Karoui Address: Université de Carthage, Département de Mathématiques, Faculté des

Sciences de Bizerte, Tunisie.
E-mail address: [email protected]

Susanna Spektor Address: Department of Mathematics, Michigan State University, 619 Red Cedar
Road, East Lancing, MI 48824
E-mail address: [email protected]