Journal of Scientific Computing (2020) 82:54

https://doi.org/10.1007/s10915-020-01157-5

Discrete Prolate Spheroidal Wave Functions: Further Spectral

Analysis and Some Related Applications

Mourad Boulsane1 · NourElHouda Bourguiba1 · Abderrazek Karoui1

Received: 22 May 2019 / Revised: 4 February 2020 / Accepted: 7 February 2020 /

Published online: 15 February 2020
© Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract  
For fixed W ∈ 0, 21 and positive integer N ≥ 1, the discrete prolate spheroidal wave
functions (DPSWFs), denoted by Uk,W N , 0 ≤ k ≤ N − 1 form the set of eigenfunctions

of the positive and finite rank integral operator Q N ,W , defined on L 2 (−1/2, 1/2), with
sin(N π(x−y))
kernel K N (x, y) = sin(π(x−y)) 1[−W ,W ] (y). It is well known that the DPSWFs have a wide
range of signal processing applications. These applications rely heavily on the properties of
the DPSWFs as well as the behaviour of their eigenvalues  λk,N (W ). In his pioneer work
(Slepian in Bell Syst. Tech. J. 57: 1371–1430,1978) D. Slepian has given the properties
of the DPSWFs, their asymptotic approximations as well as the asymptotic behaviour and
asymptotic decay rate of these eigenvalues. In this work, we give further properties as well as
new non-asymptotic decay rates of the spectrum of the operator Q N ,W . In particular, we show
that each eigenvalue 
λk,N (W ) is up to a small constant bounded above by the corresponding
eigenvalue, associated with the classical prolate spheroidal wave functions (PSWFs). Then,
based on the well established results concerning the behaviour of the eigenvalues associated
with the PSWFs, we extend these results to the eigenvalues  λk,N (W ). Also, we show that
the DPSWFs can be used for the approximation of classical band-limited functions, as well
as those functions belonging to periodic Sobolev spaces. Finally, we provide the reader with
some numerical examples that illustrate the different results of this work.

Keywords Eigenvalues and eigenfunctions · Discrete prolate spheroidal wave functions and
sequences · Eigenvalues distribution and decay rate

Mathematics Subject Classification Primary 42A38 · 15B52. Secondary 60F10 · 60B20

This work is supported by the Tunisian DGRST research Grant UR 13ES47 and PHC-Utique research
Project 20G1503.

B Abderrazek Karoui
abderrazek.karoui@fsb.rnu.tn
1 Department of Mathematics, Faculty of Sciences of Bizerte, University of Carthage, Jarzouna 7021,
Tunisia

123
54 Page 2 of 19 Journal of Scientific Computing (2020) 82:54

1 Introduction

A breakthrough in the theory and the construction of the discrete prolate spheroidal wave
functions is due to D. Slepian [1], who has studied most of the properties, the numerical
computations, as well as the asymptotic  behaviours
 of the DPSWFs and their associated
eigenvalues. Note that for fixed W ∈ 0, 21 and integers N0 ∈ N, N ≥ 1, the DPSWFs are
characterized as the amplitude spectra (Fourier series) of index-limited complex sequences
with index support [[N0 , N0 + N − 1]] = {N0 , . . . , N0 + N − 1}, that are most concentrated
in the interval (−W , W ). For the sake of simplicity of the notations and without loss of
generality, we will only consider the N0 = 0 in this work. As it will described later on,
the DPSWFs are closely related to their associated Discrete Prolate Spheroidal Sequences
(DPSSs). These DPSSs are infinite sequences in 2 (C) with amplitude spectra supported in
[−W , W ], with coefficients mostly concentrated in the index range [[0, . . . , N − 1]]. The
DPSSs have been successfully used in various classical as well as fairly recent applications
from the signal processing area. To cite but a few, the prediction of white noise random
samples of discrete signals with bandwidth W0 [1], the DPSSs based scheme for compressive
sensing [2], parametric waveform and detection of extended targets [3] and fast algorithms
for Fourier extension [4], etc.  N
It has been shown in [1], that the different DPSWFs Uk,W )0≤k≤N −1 , are the N eigen-
W ,N , given by
functions of a finite rank integral operator Q

W ,N (Uk,W
W sin(N π(x − y)) N
Q N
)(x) = Uk,W (y)dy = 
λk,N (W )Uk,W
N
(x),
−W sin(π(x − y))
x ∈ [−1/2, 1/2]. (1)

Here, 1 >  λ0,N (W ) > 

λ1,N (W ) > · · · >  λ N −1,N (W ) is the sequence of the associated
eigenvalues, arranged in the decreasing order. As it has been shown in [1], the previous
eigenproblem is a consequence of the characterization of the DPSWFs as solutions of an
energy maximization problem. That is among the space S N of all sequences x = (xn )n ∈

N
l2 (C) with elements indexed on [[0, N ]], so that their amplitude spectra 
x (t) = xk e2iπ kt ,
k=0
find those sequences with amplitude spectra most concentrated on (−W , W ), that is to solve
the maximization problem


x (t)2L 2 (−W ,W )
U = arg max . (2)
x ∈S N 
x (t)2L 2 (−1/2,1/2)

The N DPSWFs are periodic of period 1 if N is odd and period 2 if N is even. Moreover,
they satisfy the following interesting bi-orthogonality property
 W  1/2
N N 
Uk,W (x)U j,W (x)d x = λk,N (W )δk j , N
Uk,W (x)U j,W
N (x)d x = δ ,
kj
−W −1/2
k, j = 0, . . . , N − 1. (3)
N
From [1], the DPSWFs are related to the DPSSs by the following rule. Let Vk,W =
(k) (k)
(v0 , . . . , v N −1 )T , k = 0, . . . , N − 1 be the N vectors obtained by truncating the DPSSs to
the index set [[0, N −1]]. Then, these truncated DPSSs are the N eigenvectors of the Toeplitz
matrix

123
Journal of Scientific Computing (2020) 82:54 Page 3 of 19 54


sin(2π(n − m)W )
ρ N ,W = . (4)
π(n − m) n,m=0,...,N −1

Moreover, we have

N −1
1, k even;
N
Uk,W (x) = k vn(k) e−iπ(N −1−2n)x , k = (5)
i, k odd.
n=0

Note that the matrix ρ N ,W has the same spectrum as the integral operator W ,N , that is
N
 N Q 
the DPSWFs Uk,W 0≤k≤N −1
and the corresponding truncated DPSS’s V k,W 0≤k≤N −1 are
associated with same eigenvalues  λk,N (W ).
Also, it is interesting to note that the spectrum associated with the DPSWFs has some
surprising similarities with the spectrum associated with the classical PSWFs. These PSWFs
have been introduced and greatly investigated since the early 1960’s, by D. Slepian and his
co-authors H. Landau and H. Pollak, see [1,5,6]. We recall that for a given real number c > 0,
called the bandwidth, the PSWFs (ψn,c (·))n≥0 constitute an orthonormal basis of L 2 (−1, 1),
an orthogonal system of L 2 (R) and an orthogonal basis of the Paley-Wiener space Bc , given
by

Bc = f ∈ L 2 (R), Support 
f ⊂ [−c, c] . (6)

Here, f denotes the Fourier transform of f ∈ L 2 (R). They are eigenfunctions of the Sinc-
kernel operator defined on L 2 (−1, 1) by
 1
sin c(x − y)
Qc (ψn,c )(x) = ψn,c (y) dy = λn (c)ψn,c (x), x ∈ [−1, 1]. (7)
−1 π(x − y)

Unlike the classical case, where there exist a rich literature on the behaviour and the decay
rates of the eigenvalues λn (c), see for example [5–10], the existing literature for the discrete
counterpart is still very limited. The main existing decay rate result for the  λk,N (W ) is an
asymptotic one and it goes back to [1], where it has been shown that for fixed W ∈ 0, 21

and ε ∈ 0, 2W 1
− 1), we have


λk,N (W ) ≤ C1 (W , ε)e−C2 (W ,ε)N , ∀ k ≥ 2N W (1 + ε) , N ≥ N1 (W , ε). (8)

Here, N1 (W , ε) ∈ N and C1 (W , ε), C2 (W , ε) are two positive constants. All these constants
depend on W , ε and do not have explicit expressions. The previous decay rate has been
recently generalized in [11] to the multiband DPSSs setting. Moreover, in this last reference
and by using some advanced matrix analysis techniques, the authors have given the following

 1 1  of the λk,N1 (W ). If W is an union of J pairwise disjoint intervals with W ⊂
distribution
− 2 , 2 and ε ∈ 0, 2 , then
2 2N −1
2
log(N − 1) + π 2 N −1
#{k : ε ≤ 
λk,N (W) ≤ 1 − ε} ≤ J π2
. (9)
ε(1 − ε)
In this work, for c = π N W and by comparing the Hilbert–Schmidt norms of the integral
W ,N and Qc , given by (1) and (7), we prove that for J = 1, we have
operators Q
W2
1
log(2N W ) + 0.45 − 23 W 2 + sin2 (2c)
#{k : ε ≤ 
λk,N (W ) ≤ 1 − ε} ≤ π2 6c2
,
ε(1 − ε)
c = π NW. (10)

123
54 Page 4 of 19 Journal of Scientific Computing (2020) 82:54

It can be easily checked that for J = 1, N ≥ 2 and π N W ≥ 1, our estimate (10) improves
W ,N is well
the estimate (9). Also, we prove that for sufficiently small W , the spectrum of Q
approximated in the 2 -norm by the spectrum
 of the Sinc-kernel operator Qc , c = π N W .
More precisely, for any N ≥ 1 and W ∈ 0, 21 , we prove that

−1
 21  

N
 2 4π 2
λk,N (W ) − λk (c) ≤W 3
, c = π NW. (11)
3 sin(2W π)
k=0

On the other hand, by taking advantage from a connection between the energy maximization
problems associated with the DPSWFs and the classical PSWFs, (ψn,c )n with c = π N W ,
we prove the following unexpected and important result relating the 
λn,W
N and the λn (c),

λn,N (W ) ≤ A W λn (c), 0 ≤ n ≤ N − 1, (12)
where
 2
π2 2π 2 1
≤ AW = − W2 ≤ 2. (13)
8 cos2 (π W ) 4
Thanks to the estimate (12), all the existing known asymptotic and non-asymptotic decay rates
for the λn (c) are transmitted to the 
λn,N (W ). For example, based on the recent non-asymptotic
estimates of the λn (c), given in [8], one concludes that for any 2 ≤ eπ 2 N W ≤ n ≤ N − 1,
we have
 2n+2 
λn,N (W ) ≤ 2 e−(2n+1) log eπ N W . (14)
Moreover, we show that there exists a constant η > 0, such that
n−2N W
 −η
λn,N (W ) ≤ 2e log(π N W )+5 , 2N W + log(π N W ) + 6 ≤ n ≤ π N W . (15)
Also, in this work, we investigate the following applications of the DPSWFs. We first
get an estimate for the unknown constant appearing in the Turàn-Nazarov concentration
inequality. Then, we check that there exists √ N0 ≥ [2N W ] − 1, such that the eigen-space
N (W ·)
spanned by the first N0 dilated DPSWFs W Uk,W 0≤k≤N −1
is approximated by the
eigen-space spanned by the corresponding classical ψk,c (·). Also, we check that the DPSWFs
are well adapted for the spectral approximation of functions from the periodic Sobolev space
s (−1/2, 1/2), s > 0.
H per
Finally, this work is organized as follows. In Sect. 2, we give a first fast decay rate estimate
for the eigenvalues associated with the DPSWFs. This estimate is based on the use of the
Min-Max Theorem for the characterization of the eigenvalues of a self-adjoint compact
operator. In Sect. 3, we study some interesting connections between the DPSWFs and their
corresponding classical PSWFs ψn,c (·), c = π N W . Based on these connections, we deduce
various results on the distribution and the decay rates of the eigenvalues  λn,N (W ). Section 4
is devoted to the previous proposed applications of the DPSWFs. In the last Sect. 5, we give
some numerical examples that illustrate the different results of this work.

2 Eigenvalues Decay Rate by Min-Max Theorem

In this section, we first give a fast decay rate estimate for the eigenvalues associated with
the DPSWFs. This decay estimate is obtained in a fairly easy way by using the Min-Max
Theorem. More precisely, by considering the finite rank and positive-definite integral operator

123
Journal of Scientific Computing (2020) 82:54 Page 5 of 19 54

QW ,N as an operator acting on the Hilbert space L 2 (−W , W ) and by using the Min-Max
theorem, one gets the following proposition that provides us with a partial result related to
W ,N . We should mention that the proof of this proposition
the decay rate of the spectrum of Q
mimics the techniques used in [8] for proving a similar result concerning a decay rate of the
λn (c), the eigenvalues associated with the classical PSWFs.

Proposition 1 For any real number 0 < W < 2

eπ and any integer N ≥ 2, we have
 n− 1
CW eπ W (N − 1) 2

λn,N (W ) ≤ √   ,
N − 1 log eπ W2n 2n
(N −1)
eπ W (N − 1)
< n ≤ N − 1, (16)
2
√  
where C W = 2W 2 + eπ2W .

Proof We first recall the Courant–Fischer–Weyl Min-Max variational principle concerning

the positive eigenvalues of a self-adjoint compact operator A acting on a Hilbert space H,
with eigenvalues arranged in the decreasing order λ0 ≥ λ1 ≥ · · · ≥ λn ≥ · · · . In this case,
we have

λn = min max < A f , f >H ,

f ∈Sn f ∈Sn⊥ , f H =1

W ,N , H =
where Sn is a subspace of H of dimension n. In our case, we have A = Q
L 2 (−W , W ). We consider the special case of

Sn = Span P 1 (x) , . . . , P
0 (x) , P n−1 (x)

and
 
f (x) = k (x) ∈ Sn⊥ ,
ak P  f 2L 2 (−W ,W ) = |ak |2 = 1.
k≥n k≥n

k (x) = 2k + 1  x 
Here, P Pk , where Pk is the usual Legendre polynomial of degree k
2W W
and satisfying Pk (1) = 1. Note that the Pk , k ≥ 0 form an orthonormal basis of L 2 (−W , W ).
The normalization constant follows from the fact that
 W  y  21   1  21 
2W
|Pk | dy
2
= W |Pk (y)| dy
2
= = h k,W . (17)
−W W −1 2k + 1
On the other hand, we have

  W sin(N π(x − y)) 
Q k (x) =
W ,N P Pk (y) dy
−W sin(π(x − y))
 W 
N −1 iπ(N −1−2 j)(x−y)  y
e
= Pk dy
−W j=0 h k,W W


N −1  1 e−i W π(N −1−2 j)y
=W eiπ(N −1−2 j)x Pk (y)dy. (18)
−1 h k,W
j=0

123
54 Page 6 of 19 Journal of Scientific Computing (2020) 82:54

Moreover, from [12], we have

 1 
k 2π
e Pk (y)dy = i
ixy
J 1 (x), x ∈ R. (19)
−1 x k+ 2
where Jα (·) is the Bessel function of the first type and order α > 1. Moreover, Jα (·) has the
following fast decay with respect to the parameter α,
 z α
 
|Jα (z)| ≤ 2
, z ∈ C. (20)
(α + 1)
Here, (·) is the Gamma function, that satisfies the following bounds, see [13]
x+ 1 x+ 1
√ x + 21 2
√ x + 21 2
2e ≤ (x + 1) ≤ 2π , x > −1. (21)
e e
From the previous inequality and (17), we deduce that
  1 e−i W π(N −1−2 j)y 
 
 Pk (y)dy 
−1 h k,W
   1
2k + 1 eW π|N − 1 − 2 j| k+ 2
≤ . (22)
W 2 |N − 1 − 2 j|2(k + 1) 2(k + 1)
Then, by using (18), (22) and the Minkowski’s inequality, one gets for k ≥ eπ W (N − 1)/2,
   
Q k (x) 2
W ,N P
L (−W ,W )
  k+ 1

N −1
iπ(N −1−2 j)x 2k + 1 eW π|N − 1 − 2 j| 2
≤ W e  L 2 (−W ,W )
W 2 |N − 1 − 2 j|2(k + 1) 2(k + 1)
j=0
  k
√ 
N −1
eπ W eW π|N − 1 − 2 j|
≤ 2W
2(k + 1) 2(k + 1)
j=0
 k+ 1
CW eW π(N − 1) 2 √  2 
≤ √ , CW = 2W 2 + .
N −1 2(k + 1) eπ W
The last inequality follows from the fact that for k ≥ eπ W (N − 1)/2, we have

N −1
(N − 1)k+1  2 
|N − 1 − 2 j|k ≤ (N − 1)k + ≤ (N − 1)k 1 + .
k+1 eπ W
j=0

Hence, for the previous f ∈ Sn⊥ , by using Hölder’s inequality and taking into account that
 f  L 2 (I ,ωW ) = 1, so that |ak | ≤ 1, for k ≥ n, one gets

|<Q W ,N f , f > L 2 (−W ,W ) | ≤ W ,N P
|ak | Q k (·)  L 2 (−W ,W )
k≥n

  k+ 1
CW eW π(N − 1) 2
≤ √ |ak |
N −1 2(k + 1)
k≥n

  W eπ(N − 1) k+ 2
1
CW
≤ √ . (23)
N − 1 k≥n 2(k + 1)

123
Journal of Scientific Computing (2020) 82:54 Page 7 of 19 54

The decay of the sequence appearing in the previous sum, allows us to compare this later
with its integral counterpart, that is
  W eπ(N − 1) k+ 2  +∞  
1
−(x+ 21 ) log eW2(x+1)
≤ e π(N −1)
dx
2(k + 1) n−1
k≥n
 +∞  
−(x+ 21 ) log eW π(N
2n
≤ e −1)
d x. (24)
n−1
Hence, by using (23) and (24), one concludes that
max W ,N f , f > L 2 (−W ,W )
<Q
f ∈Sn⊥ ,  f  L 2 (I ,ω
W )=1
   
CW − n− 21 log eW π(N
2n
≤ √ e −1)
. (25)
N − 1 log( eW π(N
2n
−1) )
To conclude for the proof of the proposition, it suffices to use the previous Courant-Fischer-
Weyl Min-Max variational principle. 

3 The Spectrum Associated with the DPSWFs: Behaviour and Decay

Rates

In the first part of this section, we estimate the Hilbert-Schmidt norms of the integral operators
QW ,N and Qc , given by (1) and (7), respectively. As a consequence, we compare the 2 -
norms of the spectra associated with the DPSWFs with parameters N , W and the spectrum
associated with the classical PSWFs (ψn,c )n wih c = π N W . Also, we give a fairly precise
estimate of the number of the eigenvalues  λk,N (W ) lying in the interval [ε, 1 − ε], where
ε ∈ (0, 1/2). In the second part, we use the energy maximization characterizations of the
DPSWFs and the PSWFs, and get an interesting upper bound of the eigenvalues  λn,N (W )
in terms of the eigenvalues λn (c), for 0 ≤ n ≤ N − 1. Hence, by using the well established
decay rates and behaviour of the λn (c), we deduce similar results for the  λn,N (W ), with
c = π N W . The following proposition provides us with an 2 -estimate of the spectrum of
QW ,N in terms of the spectrum of Qc . Note that since Q W ,N is of finite rank, which is not
the case for the operator Qc , then this 2 -estimate is done under the rule that  λk,N (W ) = 0,
whenever k ≥ N .
Proposition 2 Let W ∈ (0, 21 ) and let N ≥ 1 be an integer. Then, under the previous notation,
we have for c = π N W ,
∞
1  
  2 2 4π 2

λ( Q W ,N ) − λ(Qc )2 = 
λk,N (W ) − λk (c) ≤W 3
. (26)
3 sin(2W π)
k=0

Proof We first consider the operator QW ,c associated with the classical PSWFs that are
c−bandlimited and mostly concentrated on [−W , W ] and have bandwidth [−c, c]. These
PSWFs are solutions of the eigenvalue problem
 W
sin(c(x − y))
QW ,c (ψ)(x) = ψk,W (y)dy = λk,W (c)ψk,W (x). (27)
−W π(x − y)
It is well know that λk,W (c) = λk,1 (cW ) = λk (cW ), ∀ W > 0. It is common to write
λk,1 (c) = λk (c) and Q1,c = Qc , where this later is given by (7). For W ∈ (0, 21 ), we let

123
54 Page 8 of 19 Journal of Scientific Computing (2020) 82:54

c N = π N and c = π N W . Let  ·  H S denote the Hilbert–Schmidt operator norm, then we

have
W ,N − Q W ,c N 2H S
Q
 W  W  
sin(c N (x − y)) sin(c N (x − y)) 2
= − d xd y
−W −W sin(π(x − y)) (π(x − y))
 1 1  2
1 1
= W2 (sin(c(t − u)))2 − dudt. (28)
−1 −1 sin(W π(t − u)) (W π(t − u))
But for X = W π(t − u) ∈ [−2π W , 2π W ], we have
   
 X − sin X   
  ≤ |X |  X  ≤ W π 2W π . (29)
 X sin X  6 sin X 
 3 sin(2W π)
The last inequality is due to the fact that x  → x
sin x is increasing on [−2π W , 2π W ]. Conse-
quently, by using (28) and (29), one gets
 2
W ,N − Q W ,c N 2H S ≤ 4W 2 W 2 2π 2
Q . (30)
3 sin(2π W )
Finally, by using (27) and the previous equality together with Wielandt-Hoffman inequality,
one gets
∞
1  
  2 2 4π 2
λ( QW ,N ) − λ(Qc )2 = λk,N (W ) − λk (c) ≤ W3 . (31)
3 sin(2W π)
k=0


Next by comparing the Hilbert–Schmidt norms of the operators QW ,N and Qc , together
W ,N ) −  Q
with a precise estimate of T race( Q W ,N 2 , we get the following theorem,
HS
showing that the eigenvalues 
λk,W cluster around 1 and 0.

Theorem 1 For any ε ∈ (0, 1/2) and any W ∈ (0, 21 ), let

N (W , ε) = #{k ; ε < 
λk,N (W ) < 1 − ε},
then we have
W2
2 log(2N W ) + 0.45 − 3 W + 6c2 sin (2c)
1 2 2 2
N (W , ε) ≤ π , c = π NW. (32)
ε(1 − ε)
   
W ,N 2 =
W W sin(N π(x − y)) 2 x y
Proof Since  Q d xd y, then for t = , u= ,
HS
−W −W sin(π(x − y)) W W
we get
   2
W ,N 2H S = W 2
1 1 sin(π N W (u − x))
Q dud x.
−1 −1 sin(π W (u − x))
That is for c = π N W , we have
 1  1−x   
W ,N 2H S − Qc 2H S = W2 1
Q (sin(ct))2 − dt d x
−1 −1−x sin (π W t) t 2 π 2
2
 1  1−x 
= (sin(ct))2 h W (t)dt d x, (33)
−1 −1−x

123
Journal of Scientific Computing (2020) 82:54 Page 9 of 19 54

 
with h W (t) = W 2 1
sin2 (y)
− 1
y2
= W 2 g(y) and y ∈] − π, π[. We check that h W (t) ≥
W2
3 . In fact, g(·) is an even and increasing function on [0, π]. Note that straightforward
  3
computations give us g  (y) = (sin(y))
2
3 − cos(y) +
sin(y)
y . Since g  (y) ≥ 0 for y ∈
[ π2 , π] and
  
sin(y) 3 y 2 3  y2 y4  y4  y2   π
− cos(y) ≥ 1 − − 1− + ≥ 1− ≥ 0, y ∈ 0, ,
y 6 2 24 24 9 2
then, we have
1
g(y) ≥ inf g(y) = lim g(y) = . (34)
y∈[0,π ] y→0 3
By combining (33) and (34), one gets
  
W ,N 2H S − Qc 2H S ≥ W
2 1 1−x 1 − cos(2ct)
Q dt d x
3 −1 −1−x 2
 1
2W 2 W2
= − [sin(2ct)]1−x
−1−x d x
3 12c −1
2W 2 W2
− 2 (sin(2c))2 .
≥
3 6c
On the other hand, from the proof of Lemma 2 of [14], it can be easily checked that
2c 1  2c 
Qc 2H S ≥ − 2 log − 0.45. (35)
π π π
By combining the previous two inequalities, one gets
  2 2
QW ,N 2H S ≥ 2c − 1 log 2c − 0.45 + 2W − W (sin(2c))2 .
π π2 π 3 6c2
W ,N ) = 2N W = , then by using the previous inequality, one gets
Since T race( Q 2c
π
W ,N ) −  Q
T race( Q W ,N 2H S

N −1
= 
λk,N (W )(1 − 
λk,N (W ))
k=0
1  2c  2W 2 W2
≤ log + 0.45 − + (sin(2c))2 ,
π2 π 3 6c2
c = π NW. (36)
That is for c = π N W , we have

N −1
η(N , W ) = 
λk,N (W )(1 − 
λk,N (W ))
k=0
1 2W 2 W2
≤ log(2N W ) + 0.45 − + (sin(2c))2 . (37)
π2 3 6c2
Finally, since ∀ ε ∈ (0, 21 ) and x ∈ (ε, 1 − ε) , we have x(1 − x) ≥ ε(1 − ε) , then

N −1
ε(1 − ε)N (W , ε) ≤ λk,N (W )(1 − 
 λk,N (W )) ≤ η(N , W ).
k=0

123
54 Page 10 of 19 Journal of Scientific Computing (2020) 82:54

This conclude the proof of the theorem. 

Remark 1 We should mention that the upper bound given by (32) is sharper than the bound
given in [11] in the sense that
2 2 2N −1
2 log(N − 1) + π 2 N −1
 1
N (W , ε) < π , ∀ W ∈ 0, , N ≥ 1.
ε(1 − ε) 2
Remark 2 We should mention that our non-asymptotic estimate of N (W , ε), given by (32),
makes sense only if ε is not too small. Recently, in [15], the authors have given the following
asymptotic estimate of N (W , ε), which is valid for small values of ε,
   
8 15
N (W , ε) = log(8N + 12) log .
π2 ε
The following theorem is one of the main results of this work. It gives a fairly good
upper bound of each eigenvalue λn,N (W ) in terms the corresponding eigenvalue λn (c), with
c = π N W and 0 ≤ n ≤ N − 1. Consequently, various new upper bounds for the  λn,N (W )
are easily obtained from the well established bounds for the classical eigenvalues λn (c).

Theorem 2 Under the previous notation, for any integer N ≥ 1 and for any W ∈ (0, 1/2),
we have for c = N π W ,

λn,N (W ) ≤ A W λn (c), 0 ≤ n ≤ N − 1, (38)
where
 2
π2 2π 2 1
≤ AW = − W2 ≤ 2. (39)
8 cos2 (π W ) 4
Proof We first use a classical technique for the construction of a subspace of the classical
band-limited functions
B N π = { f ∈ L 2 (−π, π), Support 
f ⊆ [−π, (2N − 1)π]}.
This is done as follows. Let ϕ(·) ∈ L 2 (R), with Support 
ϕ ⊆ [−π, π] and let
 
VN ,ϕ = Span e2iπ kt ϕ(t), 0 ≤ k ≤ N − 1 .


N −1 
N −1
That is if f ∈ VN ,ϕ , then f (t) = 
P(k)e 2iπ kt
ϕ(t). Here, 
P(k)e 2iπ kt
= P(e2iπ t ),
k=0 k=0
where P is a polynomial of degree N − 1. Since Support 
ϕ ⊆ [−π, π] and since

N

f (ξ ) =  ϕ (ξ − 2πk), then Support 
P(k) f ⊆ [−π, (2N − 1)π], that is f ∈ B N π . By
k=0
using Plancherel’s equality, one gets
N −1
1 2 1   2
 f 2L 2 (R) =  f  L 2 (R) = | P(k)| 
ϕ 2L 2 (R) .
2π 2π
k=0

Also, from Parseval’s equality, we have


N −1
 2 = P(e2iπ t )2 2
| P(k)| .
L (−1/2,1/2)
k=0

123
Journal of Scientific Computing (2020) 82:54 Page 11 of 19 54

By combining the previous two equalities, one gets

1
 f 2L 2 (R) = P(e2iπ t )2L 2 (−1/2,1/2) 
ϕ 2L 2 (R) , deg P ≤ N − 1.
2π
On the other hand, for W ∈ (0, 1/2), we have
 W  W
 f 2L 2 (−W ,W ) = |P(e2iπ t )|2 ϕ 2 (t) dt ≥ min |ϕ(t)|2 |P(e2iπ t )|2 dt.
−W t∈[−W ,W ] −W
Hence, for any f ∈ VN ,ϕ , we have

1  f 2L 2 (−W ,W ) P(e2iπ t )2L 2 (−W ,W )

≥ 2π . (40)
mint∈[−W ,W ] |ϕ 2 (t)|  f 2L 2 (R) P(e2iπ t )2L 2 (−1/2,1/2) 
ϕ 2L 2 (R)

ϕ (ξ ) = 1[−π,π ] (ξ ) cos(ξ/2), we have 

In particular, for  ϕ 2L 2 (R) = π. In this case, we have
⎧  
 π ⎪
⎪ cos(π x)
1
⎨ 2π 1 , x ∈ (−1/2, 1/2)
1 4 −x
2
ϕ(x) = ei xξ cos(ξ/2) dξ =
2π −π ⎪
⎪
⎩1
2 if x = ± 21 .
Then, we have
2
1 cos(W π)
min |ϕ(x)| = 2
.
x∈[−W ,W ] 4π 2 1
4 − W2
Hence, for this choice of ϕ and by using (40), one concludes that for any polynomial PN of
degree N − 1, we have

PN (e2iπ t )2L 2 (−W ,W ) 2 2  f N 2 2
4 −W
1
L (−W ,W )
≤ 2π 2
, (41)
PN (e 2iπ t ) 2
2 cos(W π)  f N 2 2
L (−1/2,1/2) L (R)

where, f N (t) = PN (e )ϕ(t) ∈ VN ,ϕ . Next, let S N be the subspace of sequences x =

2iπ t


N −1
(xn )n ∈ l 2 (C) with elements indexed on [[0, N − 1]], so that x (t) = xk e2iπ kt . Also,
k=0
we denote by sn , vn , the (n + 1)−dimensional subspace of S N and VN ,ϕ , respectively. Note
that the eigenvalues of the Sinc-kernel operator, are invariant under dilation of the time-
concentration interval and translation and dilation of the bandwidth concentration interval.
That is for τ, c > 0, we have λ(Qτ,c ) = λ(Q1,τ c ) = λ(Qτ c ) or equivalently λn,τ (c) =
λn,1 (τ c) = λn (τ c). By using the previous properties as well as the Min-Max characterisation
of this later, together with inequality (41), and the fact that VN ,ϕ is a subspace of B(N +1)π ,
one gets

x 2L 2 (−W ,W )  f 2L 2 (−W ,W )

λn,N (W ) = max min ≤ A W max min
Sn x ∈Sn \{0} 
x 2L 2 (−1,1) vn f ∈vn \{0}  f 2L 2 (R)
 f 2L 2 (−W ,W )
≤ A W max min
Un f ∈Un \{0}  f 2L 2 (R)
 f 2L 2 (−1,1)
≤ A W max min = A W λn (c).
Wn f ∈Wn \{0}  f 2L 2 (R)

123
54 Page 12 of 19 Journal of Scientific Computing (2020) 82:54

 2
2π 2 1
Here, A W = − W 2
, Un is an (n + 1)− dimensional subspace of B N π and
cos2 (π W ) 4
Wn is an (n + 1)−subspace of Bc = { f ∈ L 2 (R), Support  f ⊆ [−c, c]}, c = π N W . This
concludes the proof of the theorem. 

The previous theorem allows us to extend some known estimates for the classical λn (c)

  λn,N (W ). In [8], it has been shown that for any c > 0 and for any
to the eigenvalues
2
−(2n+1) log( ec (n+1))
n ≥ max 2, ec 2 , we have λn (c) ≤ e . By using the previous theorem,
together with the previous non-asymptotic
 2 N −1  estimate of the λn (c), one concludes that for any
N ≥ 3 and any W ∈ 0, eπ N , we have
eπ

2
λk,N (W ) ≤ 2e−(2k+1) log( eπ N W (k+1)) , 2≤ N W ≤ k ≤ N − 1. (42)
2

π +log(c)+6
Moreover, it has been shown in [8] that for any 2c

≤ n ≤ c, there exists a uniform
n − 2c
π
constant η > 0 such that λn (c) ≤ exp −η . This last estimate combined with
log(c) + 5
the previous theorem, give us the following similar estimate
n−2N W
 −η
λn,N (W ) ≤ 2e log(π N W )+5 , 2N W + log(π N W ) + 6 ≤ n ≤ π N W . (43)

It is interesting to note that the decay rate given by (42), provides us also with estimates for
the unknown constants C1 (W , ε), C2 (W , ε) appearing in the following asymptotic decay
rate, given in [1]

λk,N (W ) ≤ C1 (W , ε)e−C2 (W ,ε)N , ∀ k ≥ 2N W (1 + ε) , N ≥ N1 (W , ε). (44)
eπ −4
More precisely, by comparing (42) and (44), one concludes that for N ≥ 3, ε > 4 and
2 N −1
W ≤ eπ N , we have
 
4(1 + ε)
C1 (W , ε) ≤ 2, C2 (W , ε) ≥ 4W (1 + ε) log . (45)
eπ

4 Applications

In this section, we give two applications of the DPSWFs. The first application is related
to a lower bound estimate for the constant appearing in the Turàn-Nazarov concentration
inequality, see [16]. The second application deals with the quality of approximation by the
DPSWFs of bandlimited functions and functions from periodic Sobolev spaces.
Let us first recall the following Turàn-Nazarov type concentration inequality. Let T be the
unit circle and let μ be the Lebesgue measure on T, normalized so that μ(T) = 1, then for
every 0 ≤ q ≤ 2, every trigonometric polynomial


n+1
P(z) = ak z αk , ak ∈ C, αk ∈ N, z ∈ T,
k=1

and every measurable subset E ⊂ T, with μ(E) ≥ 13 , we have

P L q (E) ≥ e−A n μ(T\E) P L q (T) . (46)

123
Journal of Scientific Computing (2020) 82:54 Page 13 of 19 54

 N 
Here, A is a constant independent of q, E and n. Since, the DPSWFs Un,W n
are given by


N −1
N −1
N
Un,W (x) = n vkn (e−2iπ x ) 2 −k = PN (e−2iπ x ), x ∈ [−1/2, 1/2],
k=0

then by combining (2) and inequality (46) with q = 2 and E = (−W , W ) where 1/6 ≤
W < 1/2, one gets
 
 N 
Un,W  2
λn,N (W ) =   L (−W ,W ) ≥ e−A(1−2W )(N −1) , ∀ 0 ≤ n ≤ N − 1. (47)
 N 
Un,W  2
L (−1/2,1/2)

In particular, for n = N − 1, W = 1
6 and by using the estimate (42), together with straight-
forward computations, one gets
 
2 12
A≥ log = 1.02. (48)
1 − 2/6 eπ
Concerning the quality of approximation of bandlimited functions by the DPSWFs, we
have a partial result. In fact, we check that under some conditions on W and N , there
exists
√ N0 ≥ [2N W ] − 1, such that the eigenspace spanned by the first N0 dilated DPSWFs
N (W ·) approximates the eigenspace spanned by the corresponding classical ψ (·).
W Uk,W k,c
For this purpose, we first recall the following result given by Theorem 3 of [17] and concerning
the approximation of eigenspaces spanned by a set of eigenfunctions of positive self-adjoint
Hilbert–Schmidt operator and its positive self-adjoint perturbed version. More precisely, if
A is such an operator with simple eigenvalues λ0 > λ1 > · · · > λk > · · · and if there exists
an integer D > 0 such that λ D > 0 and δ D = 21 (λ D − λ D+1 ). Moreover, if A + B is such a
perturbed operator satisfying the extra condition that B < δ D /2, then
B
π D (A) − π D (A + B) ≤ . (49)
δD
Here, π D (A) denotes the orthogonal projection over the space spanned by the first D eigen-
functions of the operator A. In the sequel, we let  L W ,N denote the operator defined on
L 2 (−1, 1) by
 1
 sin(π N W (x − y))
L W ,N ( f )(x) = f (y) dy, x ∈ (−1, 1). (50)
−1 sin(π W (x − y))
√ 
Then, it is easy to check that the N dilated DPSWFs N (W ·)
W Uk,W are eigenfunctions of
k
 
L W ,N , with the same eigenvalues λk,N (W ) as the usual DPSWFs. In the special case where
c = π N W , the operators A and A + B are given by Qc , and  L W ,N , respectively. We then
obtain the following proposition that provides us with an approximation of the eigenspaces
spanned by classical PSWFs and the corresponding DPSWFs.

Proposition 3 Let π K and  π K be the two projection operators on the spaces spanned by the
first K -eigenfunctions of the the operator Qc and L ,N , respectively. For any real b > logπ 3 ,
 W 
there exists cb > 0 such that for any (W , N ) ∈ 0, 21 × N, with
 
sin 2π W π 3  3 
cb ≤ π N W ≤ exp αb − 2 log 2 − , αb = 1− , (51)
W 3 b 32bπ 1 + eπ b

123
54 Page 14 of 19 Journal of Scientific Computing (2020) 82:54

there exists N0 ≥ [2N W ] such that

  π
4bπ log(π N W ) + 2 log 2 + b
π N0 − 
π N0  ≤ W 3
. (52)
3 sin(2π W ) 1− 3
1+eπ b

Proof We first recall the following asymptotic behaviour of the λn (c), borrowed from [10],
1
lim λn c,b (c) = ,
c→+∞ 1 + eπ b

2c 2b b
n c,b = + log 2 + log c , b ≥ 0, c > 0. (53)
π π π
log 3
Hence, by applying the previous estimate for the two fixed values of b = 0 and b > π ,
one concludes that there exists Cb > 0 such that for any c ≥ cb , we have
n c,b −1
 1 3 1
0< λk (c) − λk+1 (c) = λn c,0 (c) − λn c,b (c) ≤ − . (54)
2 2 1 + eπ b
k=n c,0

Note that from (53), we have

 
2c 2b b 2c b 2b
n c,b − n c,0 = + log 2 + log c − ≤ log c + log 2 + 1.
π π π π π π
Consequently, by using (54), one concludes there exists N0 ≥ n c,0 such that
 
λ N0 (c) − λ N0 +1 (c) 1 1 3 1
δN 0 = ≥ b − .
2 π log c + π log 2 + 1
2b 4 4 1 + eπ b
Next, we consider the special case of c = π N W , and the operators A and A + B are now
given by Qc , 
L W ,N , respectively. By using (30) and (51), one can easily check that
δN 0
B =  L W ,N − Qc  ≤  L W ,N − Qc  H S ≤ .
2
Hence by using (49) and (30), one gets the desired result (52). 

It is well known see for example [6,18], that if f ∈ Bc , where Bc is the space of bandlimited
functions, given by (6), then we have
 f − π N0 f  L 2 (−1,1) ≤ λ N0 (c) f  L 2 (R) . (55)
Here, π N0 is the orthogonal
 projection over the finite dimensional space spanned by first
classical PSWFs ψk,c (·) 0≤k≤N −1.
0

Remark 3 By combining (52) and the previous inequality, one gets the following partial result
concerning the quality of approximation of bandlimited functions by the dilated DPSWFs.
For c = π N W , b > logπ 3 and under condition (51), there exists N0 ≥ [2N W ] such that for
f ∈ Bc , we have
 
4bπ log(π N W ) + 2 log 2 + πb
f − π N0 f  L 2 (−1,1) ≤ W 3
3 sin(2π W ) 1 − 1+e3 π b
 f  L 2 (−1,1) + λ N0 (c) f  L 2 (R) . (56)
√
Here, 
π N0 is the orthogonal projection over the first N0 dilated DPSWFs W Uk,W
N (W ·). In

a similar manner, we may extend this quality of approximation by the dilated DPSWFs to

123
Journal of Scientific Computing (2020) 82:54 Page 15 of 19 54

the more general class of functions of almost time- and band-limited functions. For more
details on this class of functions, the reader is refereed to [5,19]. We leave the details of this
extension to the reader.
Next, we check that the DPSWFs are well adapted for the spectral approximation of
functions from the periodic Sobolev spaces. Note that for a given real number s > 0, the
periodic Sobolev space H pers (−1/2, 1/2) is defined by



s
H per (−1/2, 1/2) = f ∈ L 2 (−1/2, 1/2),  f 2H s = (1 + n 2 )s | fˆn |2 < +∞ ,
n∈Z

 1/2
where 
fn = f (x)e−2iπ nx d x.
−1/2

Lemma 1 let W ∈ (0, 21 ), N ∈ N such that π N W ≥ 1. Let s > 0 and let

f ∈ H per
s (−1/2, 1/2), then there exists a constant M > 0 such that for any integer

[2N W ] + log(π N W ) + 6 ≤ K ≤ N − 1, we have

!
4
f −
π K ( f ) L 2 (−W ,W ) ≤  f  H s + 
λ K ,N (W ) f  L 2 (−1/2,1/2) . (57)
(4 + N 2 )s/2

Here, 
π K is the orthogonal projection over the space spanned by the first K DPSWFs, with
the parameters W , N .

Proof We first note that if f N ∈ L 2 (−1/2, 1/2) is the function given by

[(N
−1)/2]
f N (x) = 
f k e2iπ kx , then we have
k=−[(N −1)/2]

+∞
 1
 f − f N 2L 2 (−1/2,1/2) = (1 + n 2 )s | fˆn |2
(1 + n 2 )s
|n|≥[(N +1)/2]
4
≤  f 2H s . (58)
(4 + N 2 )s

On the other hand, by using the expressions of the DPSWFs, given by (5), as well as their
double orthogonality property (3), on gets


N −1
f N (x) = βk Uk,W
N
(x), ∀x ∈ [−1/2, 1/2],
k=0

N −1 N (x)
Uk,W
f N (x) = αk ! , ∀x ∈ [−W , W ]. (59)
k=0 λk,N (W )
!
Since the previous two expansions coincide on [−W , W ], then we have αk = βk  λk,N (W ).
Moreover, by using the previous equality, together with Parseval’s equality and the decay of
the 
λk,N (W ), one gets

123
54 Page 16 of 19 Journal of Scientific Computing (2020) 82:54

Table 1 Illustration of  
W c = πNW W ,N )
λ(Qc ) − λ( Q 4π 2
W 3 3 sin(2W
Proposition 2 concerning the 2 π)
2 -error approximation of the
sequence ( λn,N (W ))n by the 0.05 9.42 1.10E−03 5.32E−03
sequence (λn (c))n for N = 60 0.10 18.85 4.15E−03 2.24E−02
and different values of W
0.15 28.27 9.20E−03 5.49E−02
0.20 37.70 1.65E−02 1.11E−01
0.25 47.12 2.64E−02 2.06E−01
0.30 56.55 3.98E−02 3.74E−01


N −1 
N −1
 fN − 
π K ( f N )2L 2 (−W ,W ) = |αk |2 = 
λk,N (W )|βk |2
k=K k=K

N −1
≤
λ K ,N (W ) |βk |2 ≤ 
λ K ,N (W ) f 2L 2 (−1/2,1/2) . (60)
k=K

Also, since 
π K is an orthogonal projection, then we have 
π K  ≤ 1. Hence, by using (58)
and (60), one gets

π K ( f ) L 2 (−W ,W ) ≤  f − f N  L 2 (−W ,W ) + 
f − π K ( f − f N ) L 2 (−W ,W )
+ f N − π K ( f N ) L 2 (−W ,W )
!
≤ 2 f − f N  L 2 (−W ,W ) +  λk,N (W ) f  L 2 (−1/2,1/2)
!
4
≤  f H s +  λk,N (W ) f  L 2 (−1/2,1/2) . (61)
(4 + N 2 )s/2

√ 
Remark 4 It is easy to check that by considering the dilated DPSWFs W Uk,WN (W ·)
k
and by considering the periodic extension of f ∈ H (−1, 1), s > 0, we get the following
s

approximation result of f by the first K dilated DPSWFs,

!
4
f − π K ( f ) L 2 (−1,1) ≤ s  f H s +
λ K ,N (W ) f  L 2 −1/(2W ),1/(2W ) . (62)
(4 + N 2 ) 2

5 Numerical Results

In this section, we give four examples that illustrate the different results of this work.

Example 1 In this first example, we give a numerical test that illustrates the result of Propo-
sition 2, implying in particular that each eigenvalue 
λk,N (W ) is well approximated by the
corresponding eigenvalue λk (c), c = π N W , associated with the classical PSWFs. Note that
the 
λn,N (W ) are computed as the eigenvalues of the Toeplitz matrix (4). The λn (c) are com-
puted with high precision by using the method given in [20]. In Table 1, we have listed the
actual 2 -approximation errors λ(Qc )− λ( QW ,N )2 , versus our theoretical upper bound

4π 2
(26) of Proposition 2 and given by W 3 3 sin(2W π ) .

123
Journal of Scientific Computing (2020) 82:54 Page 17 of 19 54

Fig. 1 a Graphs of the eigenvalues 

λn,N (W ) for the values of W = 0.1, 0.2, 0.3, 0.4 (from left to right)
 
and different values of n. b Graphs of the associated log  λn,N (W ) (in circles) versus the graphs of the
 
corresponding log λn (c) , c = π N W . (in boxes)

Example 2 In this second example, we illustrate Theorem 1 as well as the main result of this
work, given by inequality (38) of Theorem 2. This last result is an unexpected one, in the sense
that it gives a bound of each  λn,N (W ) in terms of the corresponding λn (c), c = π N W and
up to a small constant A W , given by (39). As a consequence, one gets the exponential decay
rates for the 
λn,N (W ), given by (43) when n is close to the plunge region around [2N W ], and
by (42) when n is sufficiently far from [2N W ]. For these purposes, we have considered the
value of N = 60 and the four values of W = 0.1, 0.2, 0.3, 0.4. In Fig. 1a, we have plotted
the graphs of the  λn,N (W ). This figure illustrates Theorem 1 in the sense that the sequence

λn,N (W ) clusters around 1 and 0. Moreover, the number of the  λn,N (W ) in the plunge region
of the spectrum follows the bound given by Theorem 1. Also to illustrate the exponential
decay rate of the 
λn,N (W ) as well as the main result of Theorem 2, we have
 plotted
 in Fig. 1b,
the graphs of log  λn,N (W ) versus the graphs of the corresponding log λn (c) , c = π N W .
Finally, to further illustrate Theorem 2, we have listed in Table 2, some values of  λn,N (W )
with N = 60, W = 0.1, 0.3 and different values of 0 ≤ n ≤ 50. For these values, we have

λn,N (W )
also listed the ratios , c = π N W , as well as the corresponding exact values of the
λn (c)
π2
constant A W , given by (39). As predicted by Theorem 2, we have ≈ 1.2337 ≤ A W ≤ 2.
8

Example 3 In this third example, we illustrate the quality of the spectral approximation of ban-
dlimited functions by the DPSWfs, as partially predicted by Proposition 3 and Remark 3. For
sin(αx)
this purpose, we have considered the α-bandlimited function f α defined by f α (x) =
αx
with α = 56 and the special values of W = 0.3 and N = 60, so that c = π N W = 56.55.
Then, we have computed  π K f α and π K f α , the orthogonal projections of f α over the
finite dimensional subspace
√ spanned by the orthonormal set of L 2 (−1, 1) given by the
dilated DPSWFs, ( W Uk,W (W ·))0≤k≤K −1 and the classical PSWFs (ψk,c (·))0≤k≤K −1 ,
N

for the values of K = 38, 48, 58. In Table 3, we have listed the approximation errors
 fα − π K f α  L 2 (−1,1) and  f α − π K f α  L 2 (−1,1) , with the values of K = 38, 48, 59. It
is interesting to note that these numerical results indicate that as the classical PSWFs, the
DPSWFs are well adapted for the spectral approximation of bandlimited functions over
compact intervals.

123
54 Page 18 of 19 Journal of Scientific Computing (2020) 82:54

Table 2 Illustration of Theorem 2 that provides upper bounds of the 

λn,N (W ) in terms of the associated λn (c)
for c = π N W , N = 60, W = 0.1, 0.3 and for different values of the integer n

λn,N (W )
W n 
λn,N (W ) AW
λn (c)

0.1 0 1.000000000E−00 1.00000E − 00 1.2570E − 00

10 9.254515507E−01 1.00231E−00
20 1.991723157E−01 6.11697E−01
30 9.092564385E−25 9.51536E−02
40 4.333274529E−40 1.40655E−03
50 1.706983238E−67 2.59396E−07
0.3 0 1.000000000E−00 1.00000E−00 1.4626E−00
10 1.000000000E−00 1.00000E−00
20 1.000000000E−00 1.00000E−00
30 1.000000000E−00 1.00009E−00
40 9.999999620E−01 3.34258E−01
50 1.091835223E−02 4.51384E−03

Table 3 The K  fα − 
π K f α  L 2 (−1,1)  f α − π K f α  L 2 (−1,1)
L 2 (−1, 1)−approximation errors
corresponding to Example 3 with
38 1.05E−03 1.22E−03
α = 56
48 8.60E−07 9.78E−09
58 9.11E−11 1.87E−10

Table 4 The L 2 (−1, 1)−approximation errors corresponding to Example 4

K W1 − 
π K W1  L 2 W1 − π K W1  L 2 W2 − 
π K W2  L 2 W2 − π K W2  L 2

38 1.67E−02 1.67E−02 1.43E−02 1.36E−02

48 1.47E−02 1.21E−02 1.86E−04 1.39E−04
58 8.62E−03 8.62E−03 6.12E−05 6.08E−05

Example 4 In this last example, we illustrate the quality of approximation of a function from
s (−1, 1), s > 0 by the dilated DPSWFs, as predicted by
the periodic Sobolev space H per
Lemma 1 and Remark 4. For this purpose, we have considered the Weierstrass function
 cos(2k x)
Ws (x) = , −1 ≤ x ≤ 1. (63)
2ks
k≥0

Note that Ws ∈ H per s− (−1, 1), ∀ 0 <  < s. We consider the special values of s = 1, 2
√
and the same set of dilated DPSWFs ( W Uk,W N (W ·))
0≤k≤N −1 , as the previous example
with W = 0.3 and N = 60. Then, we have computed the orthogonal projections  π K Ws ,
and π K Ws , over the subspace spanned by the first K dilated DPSWFs and the classical
PSWFs (ψk,c (·))0≤k≤K −1 , c = π N W , respectively. In Table 4, we have listed the differ-
ent approximation errors, generated by these two projections, with the different values of
K = 38, 48, 58. These numerical results indicate that the DPSWFs and the PSWFs based
approximation schemes have similar quality of approximation for this these Weierstrass func-
tions. This indicates that the DPSWFs are also well adapted for the approximation of functions

123
Journal of Scientific Computing (2020) 82:54 Page 19 of 19 54

from periodic Sobolev spaces. Hence, such a DPSWFs based approximation scheme can be
considered as an alternative to the classical PSWFs based scheme, that has been studied in
[18,21,22].

Acknowledgements The authors thank very much the anonymous referees for the valuable comments and
suggestions that helped them to improve the revised version of this work.

References
1. Slepian, D.: Prolate spheroidal wave functions, Fourier analysis and uncertainty-V: The discrete Case.
Bell Syst. Tech. J. 57, 1371–1430 (1978)
2. Davenport, M.A., Wakin, M.B.: Compressive sensing of analog signals using discrete prolate spheroidal
sequences. Appl. Comput. Harmon. Anal. 33, 438–472 (2012)
3. Yin, F., Debes, C., Zoubir, A.M.: Parametric waveform design using discrete prolate spheroidal sequences
for enhanced detection of extended targets. IEEE Trans. Signal Process. 60(9), 4525–4536 (2012)
4. Adcock, B., Huybrechs, D.: On the numerical stability of Fourier extensions. Found. Comput. Math.
14(4), 635–687 (2014)
5. Landau, H.J., Pollak, H.O.: 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, 1295–1336 (1962)
6. Slepian, D., Pollak, H.O.: Prolate spheroidal wave functions, Fourier analysis and uncertainty I. Bell Syst.
Tech. J. 40, 43–64 (1961)
7. Bonami, A., Karoui, A.: Spectral decay of time and frequency limiting operator. Appl. Comput. Harmon.
Anal. 42, 1–20 (2017)
8. Bonami, A., Jaming, P., Karoui, A.: Non-asymptotic behaviour of the sinc-kernel operator and related
applications (2018). arXiv:1804.01257
9. Hogan, J.A., Lakey, J.D.: Duration and Bandwidth Limiting: Prolate Functions, Sampling, and Applica-
tions Applied and Numerical Harmonic Analysis Series. Birkhäuser, London (2013)
10. Slepian, D.: Some asymptotic expansions for prolate spheroidal wave functions. Stud. Appl. Math. 44(4),
99–140 (1965)
11. Zhu, Z., Wakin, M.B.: Approximating sampled sinusoids and multiband signals using multiband modu-
lated DPSS dictionaries. J. Fourier Anal. Appl. 23, 1263–1310 (2017)
12. Olver, F.W., Lozier, D.W., Boisvert, R.F., Clark, C.W.: NIST Handbook of Mathematical Functions.
Cambridge University Press, New York (2010)
13. Batir, N.: Inequalities for the gamma function. Arch. Math. 91, 554–563 (2008)
14. Bonami, A., Karoui, A.: Random discretization of the finite Fourier transform and related kernel random
matrices (2019). availbale at arxiv:1703.10459
15. Karnik, S., Zhu, Z., Wakin, M., Romberg, J., Davenport, M.: The fast Slepian transform. Appl. Comput.
Harmon. Anal. 46, 624–652 (2019)
16. Nazarov, F.L.: Complete version of Turàn’s lemma for trigonometric polynomials on the unit circumfer-
ence. In: Havin, V.P., Nikolski, N.K. (eds.) Complex Analysis, Operators, and Related Topics. Operator
Theory: Advances and Applications, vol. 113, pp. 239–246. Birkhäseur Verlag Basel, Switzerland (2000)
17. Zwald Laurent, L., Blanchard, G.: On the convergence of eigenspaces in kernel principal component
analysis. In: Advances in Neural Information Processing Systems 18: Proceedings of the 2005 Conference
(Neural Information Processing). MIT Press (2006)
18. Wang, L.L.: Analysis of spectral approximations using prolate spheroidal wave functions. Math. Comp.
79, 807–827 (2010)
19. Jaming, P., Karoui, A., Spektor, S.: The approximation of almost time- and band-limited functions by
their expansion in some orthogonal polynomials bases. J. Approx. Theory 212, 41–65 (2016)
20. Karoui, A., Moumni, T.: New efficient methods of computing the prolate spheroidal wave functions and
their corresponding eigenvalues. Appl. Comput. Harmon. Anal. 24(3), 269–289 (2008)
21. Bonami, A., Karoui, A.: Approximation in Sobolev spaces by prolate spheroidal wave functions. Appl.
Comput. Harmon. Anal. 42, 361–377 (2017)
22. Wang, L.L., Zhang, J.: A new generalization of the PSWFs with applications to spectral approximations
on quasi-uniform grids. Appl. Comput. Harmon. Anal. 29, 303–329 (2010)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

123