Journal of Functional Analysis 270 (2016) 4196–4212

Contents lists available at ScienceDirect

Journal of Functional Analysis

www.elsevier.com/locate/jfa

Optimal approximation of multivariate periodic

Sobolev functions in the sup-norm
Fernando Cobos a,1 , Thomas Kühn b,∗,1 , Winfried Sickel c
a
Departamento de Análisis Matemático, Facultad de Matemáticas, Plaza de las
Ciencias 3, 28040 Madrid, Spain
b
Mathematisches Institut, Universität Leipzig, Augustusplatz 10, 04109 Leipzig,
Germany
c
Mathematisches Institut, Friedrich-Schiller-Universität Jena, Ernst-Abbe-Platz 2,
07737 Jena, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Using tools from the theory of operator ideals and s-numbers,
Received 29 November 2014 we develop a general approach to transfer estimates for
Accepted 30 March 2016 L2 -approximation of Sobolev functions into estimates for
Available online 7 April 2016
L∞ -approximation, with precise control of all involved con-
Communicated by F. Barthe
stants. As an illustration, we derive some results for periodic
MSC: isotropic Sobolev spaces H s (Td ) and Sobolev spaces of dom-
s
46E35 inating mixed smoothness Hmix (Td ), always equipped with
41A25 natural norms. Some results for Lp -approximation (2 < p <
∞) are also obtained.
Keywords: © 2016 Elsevier Inc. All rights reserved.
Approximation numbers
Isotropic and mixed Sobolev spaces
Wiener algebra
d-Dependence of the constants

* Corresponding author.
E-mail addresses: [email protected] (F. Cobos), [email protected] (T. Kühn),
[email protected] (W. Sickel).
1
Supported in part by the Spanish Ministerio de Economía y Competitividad (MTM2013-42220-P).

http://dx.doi.org/10.1016/j.jfa.2016.03.018
0022-1236/© 2016 Elsevier Inc. All rights reserved.
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4197

1. Introduction

Nowadays there is an increasing interest in the study of linear approximation (ap-

proximation numbers) in the context of Sobolev spaces. This is motivated by the fact
that a number of problems in finance and quantum chemistry are modelled on function
spaces on high-dimensional domains, and often the functions to be approximated have
some Sobolev regularity (see, for example, [23]).
In the papers [7,8], T. Ullrich and two of the present authors studied linear approxi-
mation of functions in the isotropic periodic Sobolev spaces H s (Td ) and in the smaller
s
spaces Hmix (Td ) of dominating mixed smoothness, where Td is the d-dimensional torus,
see also the paper by Dũng and Ullrich [4] in this context (but these authors used different
norms). The error was measured in L2 (Td ), with particular emphasis on the constants
and their dependence on the dimension d. In [7,8] the exact decay rate of the constants
as d → ∞ was found, which turned out to be polynomial in d for the isotropic spaces,
and super-exponential in d for the mixed spaces.
In the present paper we deal with the same approximation problems, but now the
error is measured in the sup-norm. For this aim, we develop a general method which
allows to transfer results on L2 -approximation into results on L∞ -approximation. Our
approach is based on tools from the theory of operator ideals and s-numbers, in the sense
of the monographs by Pietsch [13,15]. In particular, we work with absolutely 2-summing
operators.
We consider spaces of periodic functions on the d-dimensional torus, whose norms
are weighted 2 -sums of Fourier coefficients. In addition to the correct rate of decay of
the approximation numbers, we also obtain very precise information on the “hidden”
constants, especially their dependence on the dimension d. As an illustration, we apply
our general method to isotropic Sobolev spaces H s (Td ) and Sobolev spaces Hmix s
(Td )
of dominating mixed smoothness. Our estimates allow to control the dependence of the
constants on the dimension d, the smoothness s and the particular norm used in the
Sobolev space. We also show that L∞ (Td ) can be replaced by the much smaller Wiener
algebra A(Td ), without changing the associated approximation numbers. This is not
only surprising from a theoretical point of view, but also of some practical use, since the
Fourier-theoretical characterization of the Wiener algebra simplifies the computation of
approximation numbers (or any other s-numbers).
Sobolev and Besov spaces of dominating mixed smoothness are representing first at-
tempts in approximation theory to deal with high dimensions. These spaces are much
smaller than their isotropic counterparts, they have attracted a lot of interest in ap-
proximation theory since the early 1960s (mainly in Russia), and later also in IBC
(information-based complexity), we refer to the monographs [18,9–11].
The plan of the paper is simple. In Section 2 we recall some basic notions from
function spaces and operator theory, Section 3 contains our main abstract results, and
in the final Section 4 we apply these results to isotropic Sobolev spaces and Sobolev
4198 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

spaces of dominating mixed smoothness. Moreover, we establish some upper estimates

for approximation in the Lp -norm, 2 < p < ∞.

2. Preliminaries

First we fix some notation. For d ∈ N, x = (x1 , . . . , xd ) ∈ Rd and 0 < p < ∞ let

 1/p
d
|x|p = j=1 |xj |
p
, and for p = ∞ we set |x|∞ = max1≤j≤d |xj |. In what follows
T denotes the torus, i.e. T = [0, 2π] where the endpoints of the interval are identified,
and Td stands for the d-dimensional torus. We equip Td with the normalized Lebesgue
measure (2π)−d dx. Consequently, {eikx : k ∈ Zd } is an orthonormal basis in L2 (Td ),
d
where kx = j=1 kj xj .
The Fourier coefficients of a function f ∈ L1 (Td ) are defined as

f(k) = (2π)−d f (x)e−ikx dx , k ∈ Zd .
Td

For 0 < s < ∞ and 0 < r ≤ ∞ we denote by H s,r (Td ) the isotropic Sobolev space formed
by all f ∈ L2 (Td ) having a finite norm

 
d 1/2
|f(k)|2
2s/r
f |H s,r (Td ) = 1+ |kj |r
k∈Zd j=1

(with the usual modification if r = ∞). Clearly, for fixed s, all these norms are equivalent,
whence all spaces H s,r (Td ) with 0 < r ≤ ∞ coincide. The superscript r just indicates
which norm we are considering.
For integer smoothness s = m ∈ N, the most natural norms are those with r = 2 and
r = 2m. Indeed, let Dα f be the distributional derivative of f of order α = (α1 , . . . , αd ).
As shown in [7], one has

  
1 2 1/2
√ f |H m,2 (Td ) ≤ Dα f |L2 (Td ) ≤ f |H m,2 (Td ) .
m! |α| ≤m 1

Note that the equivalence constants depend only on the smoothness m, but not on the
dimension d.
If r = 2m, one has even equality

 d
∂mf 
2 1/2
f |H m,2m (Td ) = f |L2 (Td )2 + d
m L2 (T ) .
j=1
∂x j

s,r
The Sobolev space Hmix (Td ) of dominating mixed smoothness consists of all f ∈
d
L2 (T ) having a finite norm
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4199

 d
 1/2
|f(k)|2
s,r 2s/r
f |Hmix (Td ) = 1 + |kj |r
k∈Zd j=1

(see [8]). For s = m ∈ N and r = 2m, it turns out that

  1/2
m,2m
f |Hmix (Td ) = Dα f |L2 (Td )2 .
α∈{0,m}d

Let us give a short comment on the role played by the parameter r. Of course, for fixed
d and s, different r’s result in equivalent norms, but the equivalence constants depend
heavily on d. From our point of view it is interesting to see how these changes of the
norm influence the behaviour of the associated approximation numbers.
The n-th approximation number of a (bounded linear) operator T : X → Y between
Banach spaces is defined as

an (T ) = inf{T − A : rank A < n} ,

i.e. an (T ) is the optimal error of approximating T by operators of rank less than n.

We refer to [13,15] for properties of these numbers. Let us just recall that for compact
operators T between Hilbert spaces an (T ) coincides with the n-th singular number sn (T )
of T .
An operator T : X → Y is called absolutely 2-summing if there is a constant C > 0
such that for all n ∈ N and all x1 , . . . , xn ∈ X the inequality

n 1/2 
n 1/2
 T xj |Y 2 ≤C sup |xj , x |2 , (2.1)
j=1 x |X  ≤1 j=1

holds, where X  is the dual space of X. The 2-summing norm π2 (T ) is defined as the
infimum of all C > 0 satisfying (2.1). For more information we refer to [13, Chapter 17].
Later on we shall use the fact that for operators T : H → G between Hilbert spaces
H and G, the 2-summing norm is equal to the Hilbert–Schmidt norm, whence for any
orthonormal basis {ei : i ∈ I} of H it holds
 1/2
π2 (T : H → G) = T ei |G2 .
i∈I

Given two sequences (an ) and (bn ), we write an  bn if there is a constant c > 0 such
that an ≤ c bn for all n ∈ N. The weak equivalence an ∼ bn means that an  bn and
an  bn , and the strong equivalence an bn means lim abnn = 1.
n→∞

3. General results

In what follows Fd (w) always stands for a Hilbert space of integrable functions on the
d-dimensional torus Td such that
4200 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

 1/2
f ∈ Fd (w) ⇐⇒ f |Fd (w) := w(k)2 |f(k)|2 < ∞. (3.1)
k∈Zd

Here w(k) > 0 , k ∈ Zd , are certain weights. Important examples of such spaces are the
s,r
Sobolev spaces H s,r (Td ) and Hmix (Td ) introduced in the previous section.
We shall also deal with the Wiener algebra A(Td ), which is the collection of all in-
tegrable functions on Td with absolutely convergent Fourier series. A(Td ) is a Banach
space with respect to the norm

f |A(Td ) = |f(k)| .
k∈Zd

Here, as well as in the context of Sobolev and Besov spaces, we shall make use of the
following convention: If the equivalence class of a measurable function f contains a
continuous representative, then we call f itself continuous and work with the continuous
representative.
Several necessary and sufficient conditions are known for a function to belong to
A(Td ) (see [6, Chapitre II]). We just recall a characterization that describes the Wiener
algebra as an approximation space. Let O0 = {0} and, for n ∈ N, let On be the set of all
trigonometric polynomials having at most n non-zero coefficients. According to a result
of Stečkin [16] (see also [14]), a function f belongs to A(Td ) if and only if

∞

n−1/2 inf{f − p|L2 (Td ) : p ∈ On } < ∞ .
n=1

Our first general result provides a necessary and sufficient condition on the weights
w(k) which guarantees the existence of continuous or, equivalently, compact embeddings
of Fd (w) in A(Td ), C(Td ) or L∞ (Td ).

Theorem 3.1. The following conditions are equivalent.

(i) Fd (w) → A(Td ) compactly

(ii) Fd (w) → A(Td ) boundedly
(iii) Fd (w) → C(Td ) compactly
(iv) Fd (w) → C(Td ) boundedly
(v) Fd (w) → L∞ (Td ) compactly
(vi) Fd (w) → L∞ (Td ) boundedly

(vii) w(k)−2 < ∞
k∈Zd

Proof. Due to the continuous embeddings A(Td ) → C(Td ) → L∞ (Td ), the implications
(i) ⇒ (ii) ⇒ (iv) ⇒ (vi) and (i) ⇒ (iii) ⇒ (v) ⇒ (vi) are trivial. So it remains to prove
(vi) ⇒ (vii) and (vii) ⇒ (i).
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4201

Step 1. Proof of (vi) ⇒ (vii):

The formal identity from L∞ (Td ) into L2 (Td ) is absolutely 2-summing, whence Id :
Fd (w) → L2 (Td ) is also 2-summing, and therefore it is Hilbert–Schmidt, because Fd (w)
and L2 (Td ) are Hilbert spaces. Let ϕk (x) = eikx /w(k). By definition of Fd (w), the set
{ϕk : k ∈ Zd } is a complete orthonormal system in Fd (w), and consequently we have

 1   2
2
= ϕk |L2 (Td )2 = π2 Id : Fd (w) → L2 (Td ) < ∞.
w(k)
k∈Zd d
k∈Z

Step 2. Proof of (vii) ⇒ (i):

Let A : Fd (w) → 2 (Zd ) be the operator defined by Af = (w(k)f(k))k∈Zd . Clearly A is
an isometry. For ξ = (ξk )k∈Zd we define Dξ = (ξk /w(k))k∈Zd . Using (vii) and Hölder’s
inequality, we see that D : 2 (Zd ) → 1 (Zd ) is bounded. Now, for ξ = (ξk ) ∈ 1 (Zd ),
 ikx
we set (Bξ)(x) = k∈Zd ξk e . The series that defines Bξ converges absolutely and
uniformly, whence the operator B : 1 (Zd ) → A(Td ) is bounded with norm one, and the
following commutative diagram holds.

Id
Fd (w) - A(Td )
6
A B
?
D -  (Zd )
2 (Zd ) 1

Let (σn )n∈N be the non-increasing rearrangement of (1/w(k))k∈Zd , which by assumption

belongs to 2 (Zd ), and consider the associated diagonal operator Dσ x = (σn xn )n∈N for
x = (xn )n∈N . The rearrangement of (1/w(k))k∈Zd into the sequence (σn )n∈N defines a
one-to-one correspondence between the index sets Zd and N, whence the multiplication
property of approximation numbers gives

an (D : 2 (Zd ) → 1 (Zd )) = an (Dσ : 2 → 1 ) .

The commutative diagram yields

an (Id : Fd (w) → A(Td )) ≤ A · an (D : 2 (Zd ) → 1 (Zd )) · B


∞ 1/2
= an (Dσ : 2 → 1 ) = σj2 ,
j=n

where the last equality follows from [13, Theorem 11.11.4]. This implies lim an (Id :
n→∞
Fd (w) → A(Td )) = 0, and thus the embedding Id : Fd (w) → A(Td ) is compact. The
proof is complete. 2
4202 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

In Step 2 of the preceding proof we have obtained the following inequality, which we
state for later use as a separate lemma.

Lemma 3.2. Let Fd (w) be given by weights w(k) satisfying w(k)−2 < ∞. Then we
k∈Zd
have for all n ∈ N


∞ 1/2
an (Id : Fd (w) → A(Td )) ≤ σj2 ,
j=n

where (σj )j∈N is the non-increasing rearrangement of (1/w(k))k∈Zd .

Next we show a lower estimate for these approximation numbers. Here we allow a
greater generality. Recall that sj (T ) denotes the j-th singular number of a compact
operator T between Hilbert spaces.

Lemma 3.3. Let H be a Hilbert space, let (Ω, Σ, ν) be a finite measure space, and id ν :
L∞ (ν) → L2 (ν) the formal identity. Then one has, for every bounded linear operator
T : H → L∞ (ν) and all n ∈ N, the estimate

1 
∞ 1/2
an (T ) ≥ sj (id ν T )2 .
ν(Ω) j=n

Proof. Let A : H → L∞ (ν) be an arbitrary operator with rank A < n. By the additivity
of singular numbers we get for all j ≥ n the inequality

sj (id ν T ) ≤ sn (id ν A) + sj+1−n (id ν (T − A)) = sj+1−n (id ν (T − A)) ,

where we took into account that sn (id ν A) = 0. Changing the running index j ≥ n to
k = j + 1 − n ≥ 1, this implies
∞
 ∞

sj (id ν T )2 ≤ sk (id ν (T − A))2 = π2 (id ν (T − A))2
j=n k=1

≤ π2 (id ν )2 · T − A2 = ν(Ω) · T − A2 .

Here we also used the well-known facts that

∞

π2 (S)2 = sk (S)2
k=1

for compact operators S acting between Hilbert spaces, see [3, Theorem 4.10], and

π2 (id ν : L∞ (ν) → L2 (ν)) = ν(Ω) ,

 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4203

see [3, Example (d), p. 40]. Passing to the infimum over all operators A of rank A < n,
we arrive at the desired inequality

1 
∞ 1/2
an (T ) ≥ sj (id ν T )2 . 2
ν(Ω) j=n

The arguments in the above proof are similar to the ones used by Osipenko and
Parfenov in [12]. We thank Heping Wang for pointing out this paper to us.
As an immediate consequence of the preceding two lemmata we obtain the following
result.

Theorem 3.4. Let Fd (w) be given by weights satisfying w(k)−2 < ∞, and let (σj )j∈N
k∈Zd
denote the non-increasing rearrangement of (1/w(k))k∈Zd . Moreover, let

Gd = A(Td ) or C(Td ) or L∞ (Td ) .

Then one has for all n ∈ N


∞ 1/2
an (Id : Fd (w) → Gd ) = σj2 . (3.2)
j=n

Proof. Recall that the torus Td is equipped with the normalized Lebesgue measure. In
view of the norm one embeddings

A(Td ) → C(Td ) → L∞ (Td )

and the multiplicativity of the approximation numbers, we get from Lemma 3.3 and
Lemma 3.2 the following chain of inequalities

∞ 1/2
σj2 ≤ an (Id : Fd (w) → L∞ (Td ))
j=n

≤ an (Id : Fd (w) → C(Td ))


∞ 1/2
≤ an (Id : Fd (w) → A(Td )) ≤ σj2 .
j=n

The proof is finished. 2

Remark 3.5. Since σj = aj (Id : Fd (w) → L2 (Td )), equation (3.2) gives the nice formula


∞ 1/2
an (Id : Fd (w) → L∞ (Td )) = aj (Id : Fd (w) → L2 (Td ))2 .
j=n

The same is true if we replace L∞ (Td ) with C(Td ) or A(Td ).

4204 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

4. Applications to Sobolev spaces

In this section we apply our general result Theorem 3.4 to various Sobolev spaces.

4.1. Embeddings of isotropic Sobolev spaces

First we consider the isotropic Sobolev spaces H s,r (Td ), introduced in Section 2, where
s > 0 and 0 < r ≤ ∞. Taking the weights

 
d
s/r
w(k) := ws,r (k) = 1 + |kj |r ,
j=1

we get Fd (w) = H s,r (Td ). Note that, for fixed d ∈ N and s > 0, all weights ws,r with
0 < r ≤ ∞ are equivalent.
For the non-increasing rearrangement (σn )n∈N of (1/ws,r (k))k∈Zd we have σn = an (Id :
H (Td ) → L2 (Td )). It is a classical fact that
s,r

d
H s,r (Td ) → C(Td ) ⇐⇒ s> .
2

However, it can be checked directly that

 1 d
<∞ ⇐⇒ s> . (4.1)
ws,r (k)2 2
k∈Zd

By Theorem 3.1 we conclude that the condition s > d/2 is necessary and sufficient for the
existence of an embedding of H s,r (Td ) into A(Td ) (or C(Td ) or L∞ (Td )). The following
result was shown in [7, Theorems 4.3, 4.11 and 4.14] for the special values r = 1, 2, 2s,
but the proof works for all 0 < r ≤ ∞.

Proposition 4.1. Fix d ∈ N and s > 0. Then, for all 0 < r ≤ ∞, it holds

lim ns/d an (Id : H s,r (Td ) → L2 (Td )) = vol(Brd )s/d .

n→∞

Here Brd is the unit ball in Rd with respect to the (quasi)-norm | · |r .

Remark 4.2. (i) One can rephrase Proposition 4.1 as a strong equivalence

an (Id : H s,r (Td ) → L2 (Td )) vol(Brd )s/d n−s/d .

The importance of this result is that it provides asymptotically optimal constants, for
arbitrary fixed d, s, r.
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4205

(ii) We comment on the influence of the parameter r. The formula

 d  Γ(1 + 1/r)d
vol(Brd ) := vol x ∈ Rd : |xj |r ≤ 1 = 2d ,
j=1
Γ(1 + d/r)

for the volume of the unit ball Brd is well-known, see, e.g., Wang [22], where

∞
Γ(1 + x) = tx e−t dt , x > 0,
0

denotes the Gamma function. Using Stirling’s formula

√  x x
Γ(1 + x) 2πx as x → ∞
e

we obtain
 1/r  1/2d  1/r
d 2πd d
Γ(1 + d/r) 1/d
as d → ∞ .
er r er

Hence, for fixed s and r, and setting c(r) := 2Γ(1 + 1/r)(er)1/r , we have

vol(Brd )s/d c(r)s d−s/r as d → ∞ .

Theorem 4.3. Let d ∈ N, s > d/2 and 0 < r ≤ ∞. Then


d
lim n s/d−1/2
an (Id : H s,r
(T ) → L∞ (T )) =
d d
vol(Brd )s/d . (4.2)
n→∞ 2s − d

Proof. By Theorem 3.4 we have


∞ 1/2
ns/d−1/2 an (Id : H s,r (Td ) → L∞ (Td )) = ns/d−1/2 σj2 ,
j=n

where σj = aj (Id : H s,r (Td ) → L2 (Td )). From Proposition 4.1 we know that for any
ε > 0 there exists a natural number n0 (ε) such that

|ns/d σn − vol(Brd )s/d | < ε for all n ≥ n0 (ε).

This yields, if n ≥ n0 (ε),

∞
 ∞
 ∞

 2  2
vol(Brd )s/d − ε j −2s/d ≤ σj2 ≤ vol(Brd )s/d + ε j −2s/d .
j=n j=n j=n
4206 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

Comparing the series with an integral,

∞ ∞ ∞
n1−2s/d dx dx (n − 1)1−2s/d
= ≤ j −2s/d ≤ = ,
2s/d − 1 x 2s/d
j=n
x2s/d 2s/d − 1
n n−1

we obtain
 2 d
vol(Brd )s/d − ε ≤ n2s/d−1 an (Id : H s,r (Td ) → L∞ (Td ))2
2s − d
 2s/d−1
 2 d n
≤ vol(Brd )s/d + ε .
2s − d n − 1

Since this is true for all ε > 0 and sufficiently large n, the claim follows by letting first
n → ∞ and then ε → 0. 2

Remark 4.4. (i) Rephrasing Theorem 4.3 in terms of the strong equivalences

d
an (Id : H s,r
(T ) → L∞ (T ))
d d
n1/2−s/d
vol(Brd )s/d
2s − d

d
n1/2 an (Id : H s,r (Td ) → L2 (Td ))
2s − d

we see that, compared to L2 -approximation, the rate of L∞ -approximation

 is worse by
d
a factor n1/2 . Moreover, for the constant we need the correction factor 2s−d .
(ii) Of course, the asymptotic behaviour an (Id : H s,r (Td ) → L∞ (Td )) ∼ n of the 1/2−s/d

approximation numbers is known since some time, see e.g. the monograph by Temlyakov
[18, Theorem 1.4.2, Theorem 2.4.2], but only in the sense of weak equivalence, without
explicit constants. At the end of that book, some historic remarks on the periodic case
can be found. For the non-periodic case we refer to Edmunds and Triebel [5] and to
Vybíral [21]. The novelty of Theorem 4.3 is that it gives strong equivalence and provides
exact information about the asymptotically optimal constants and their dependence on
d, s and r.
(iii) Theorem 3.4 yields that relation (4.2) remains true if we replace L∞ (Td ) with A(Td )
or C(Td ).

After determining the asymptotic behaviour of an (Id : H s,r (Td ) → L∞ (Td )), we study
next single estimates between an (Id ) and n1/2−s/d . We need some preparation.

Proposition 4.5. Assume that σn := an (Id : Fd (w) → L2 (Td )) satisfies

An−α (log n)β ≤ σn ≤ Bn−α (log n)β for all n≥N

for some α > 1/2, β ≥ 0, 0 < A ≤ B < ∞ and N ∈ N. Then it follows that
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4207


(i) an+1 (Id : Fd (w) → L∞ (Td )) ≤ B 2
2α−1 n1/2−α (log n)β
for all n ≥ max(N, e4β/(2α−1) )
and 
(ii) an (Id : Fd (w) → L∞ (Td )) ≥ A 4α−2
1
n1/2−α (log 2n)β for all n ≥ N ≥ 2.

Proof. Step 1. Let us show (i). The upper estimate in Theorem 3.4 gives for n ≥ N

∞
(log j)2β
a2n+1 := an+1 (Id : Fd (w) → L∞ (Td ))2 ≤ B 2 . (4.3)
j=n+1
j 2α

Taking the derivative of the function f (x) = (log x)2β x−2α is easy to check that f is
decreasing for x ≥ eβ/α . Therefore, for n ≥ eβ/α , we can estimate the series (4.3) against
an integral and obtain

∞ ∞ 
(log x)2β log(nt) 2β dt
a2n+1 ≤B 2
dx = B 2 n1−2α (log n)2β . (4.4)
x2α log n t2α
n 1

Since 1 + x ≤ ex , it follows that

 log(nt) 2β  log t 2β log t 2β

= 1+ ≤ e log n 2β = t log n .
log n log n

If 2β/ log n ≤ α − 1/2, i.e. log n ≥ 2β/(α − 1/2) = 4β/(2α − 1) or n ≥ e4β/(2α−1) , we

get

∞  ∞ ∞
log(nt) 2β dt dt dt 2
≤ t α−1/2
= = .
log n t2α t2α tα+1/2 2α − 1
1 1 1

Together with (4.4), this yields the upper estimate for all n ≥ max(N, e4β/(2α−1) ).
Step 2. Now we turn to (ii). As above we conclude

∞ ∞ 
(log x)2β log(nt) 2β dt
a2n ≥ A2 dx = A2 n1−2α (log n)2β
x2α log n t2α
n 1
n
dt
≥ A2 n1−2α (log n)2β
t2α
1
1  1 
≥ A2 n1−2α (log n)2β 1 − 2α .
2α − 1 n

The last factor is bounded below by 1/2 if n ≥ 2. This finishes the proof. 2
4208 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

Equipped with this proposition, it is now easy to transfer the two-sided estimates of
an (Id : H s,r (Td ) → L2 (Td )) that have been obtained in [7] into two-sided estimates of
an (Id : H s,r (Td ) → L∞ (Td )). As an example we consider the case r = 2. The following
estimates have been shown in [7, Theorem 4.15].

Proposition 4.6. Let s > 0 and d ∈ N. Then we have

 32e s/2
an (Id : H s,2 (Td ) → L2 (Td )) ≤ n−s/d for n ≥ 9d ed/2
d
and
 1 s/2
an (Id : H s,2 (Td ) → L2 (Td )) ≥ n−s/d for n ≥ 11d ed/2 .
e(d + 2)

Together with Proposition 4.5, this immediately implies

Corollary 4.7. Let s > d/2 and d ∈ N. Then we have for n ≥ 9d ed/2

2d  32e s/2 1/2−s/d
an+1 (Id : H s,2
(T ) → L∞ (T )) ≤
d d
n
2s − d d

and for n ≥ 11d ed/2 it holds


4d  1 s/2
an (Id : H s,2
(T ) → L∞ (T )) ≥
d d
n1/2−s/d .
4s − d e(d + 2)

4.2. Embeddings of Sobolev spaces of dominating mixed smoothness

s,r
Next we focus our attention on Sobolev spaces Hmix (Td ) of dominating mixed smooth-
ness. The weights are now

d
mix
w(k) = ws,r (k) = (1 + |kj |r )s/r
j=1

where 0 < s < ∞ and 0 < r ≤ ∞. It is known since a long time that

s,r 1
Hmix (Td ) → C(Td ) ⇐⇒ s> .
2
However, this can be calculated also directly by checking
 1
mix (k)2
<∞ ⇐⇒ s > 1/2 . (4.5)
ws,r
k∈Zd

s,r
By Theorem 3.4 this guarantees also Hmix (Td ) → A(Td ).
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4209

The existence of the following limits was shown in [8, Theorem 4.3, Corollaries 4.4
and 4.7] for the special values r = 1, 2, 2s, but the proof works also for all other
0 < r ≤ ∞,
s,r
ns an (Id : Hmix (Td ) → L2 (Td ))  2d s
lim = . (4.6)
n→∞ (log n)(d−1)s (d − 1)!

For approximation in L∞ (Td ) we have the following result.

Theorem 4.8. Let d ∈ N, s > 1/2 and 0 < r ≤ ∞. Then

s,r
ns−1/2 an (Id : Hmix (Td ) → L∞ (Td )) 1  2d s
lim = √ .
n→∞ (log n)(d−1)s 2s − 1 (d − 1)!

Proof. We can apply the same arguments as in the proof of Theorem 4.3, replacing
Proposition 4.1 by formula (4.6). Here one has to take into account that

∞ ∞ 
(log x)(d−1)2s log(nt) (d−1)2s dt
dx = n1−2s (log n)(d−1)2s
x2s log n t2s
n 1

see (4.4), and

∞  ∞
log(nt) (d−1)2s dt dt 1
lim = = . 2
n→∞ log n t2s t 2s 2s − 1
1 1

Remark 4.9. (i) We can express this also in terms of a strong equivalence,

s,r n s,r
an (Id : Hmix (Td ) → L∞ (T ))d
an (Id : Hmix (Td ) → L2 (Td )) .
2s − 1

Comparing L∞ -approximation with L2 -approximation, the asymptotic rate is again

worse by the factor n1/2 , as in the case of isotropic spaces. But now the value of the limit
√
does not depend on the parameter r, and the additional correction factor 1/ 2s − 1 is
independent of the dimension d, it depends only on the smoothness s. We find this quite
interesting.
(ii) As in case of the isotropic Sobolev spaces, for fixed s, d and r the asymptotic
s,r
behaviour of the approximation numbers an (Id : Hmix (Td ) → L∞ (Td )) as n → ∞ has
been known for more than 20 years, see Temlyakov [17,19], but without explicit constants.
The novelty of our results is that we can control the dependence of the constants on s,
d and r, so that we can even determine the optimal constants for n → ∞.

In [8] several other L2 -estimates are given for approximation numbers an with large n.
Our general Theorem 3.4 in combination with Proposition 4.5 allows also to transfer
4210 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212

each of these L2 -estimates into L∞ -estimates. We conclude this subsection with another
example which follows from [8, Theorem 4.15].

Corollary 4.10. Let d ∈ N and s > 1/2. Then we have

  √ s
s,2 2 (3 · 2)d (log n)(d−1)s
an (Id : Hmix (Td ) → L∞ (Td )) ≤
2s − 1 (d − 1)! ns−1/2
if n > max(27d , e4(d−1)s/(2s−1) ) , and
  s
1 5 (log(2n))(d−1)s
s,2
an (Id : Hmix (Td ) → L∞ (Td )) ≥ √
4s − 2 6 d!(1 + log 12)d ns−1/2
if n > (12 e2 )d .

Remark 4.11. Our approach also works if the target space L∞ (Td ) is replaced by cer-
tain Besov spaces of smoothness s = 0. The corresponding results can be found in the
extended arXiv-version of this article, see [2], and will be published elsewhere.

4.3. Approximation in the Lp -norm

Since we are able to control the behaviour of the approximation numbers an (Id :
Fd (w) → L∞ (Td )) and an (Id : Fd (w) → L2 (Td )), it is quite natural to ask if this can
also be done for approximation in the Lp -norm, 2 < p < ∞. Accordingly, we finish the
paper with some upper estimates in the Lp -norm.

Proposition 4.12. Let 2 < p < ∞, define r by 1/r = 1/2 − 1/p, assume that Fd (w) is

given by weights w(k) satisfying k∈Zd w(k)−r < ∞. Moreover, let (σj )j∈N denote the
non-increasing rearrangement of (1/w(k))k∈Zd . Then one has for all n ∈ N


∞ 1/r
an (Id : Fd (w) → Lp (Td )) ≤ σjr .
j=n

Proof. Let the operators A, B and D be defined as in the proof of Theorem 3.1. We
consider the following commutative diagram.

Id
Fd (w) - L (Td )
p

6
A B
? D
2 (Zd ) -   (Zd )
p
 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212 4211

To make sure that this diagram really makes sense, we have to check the mapping
properties of the operators B and D. Concerning B, we know already that

B : 2 (Zd ) → L2 (Td ) = B : 1 (Zd ) → L∞ (Td ) = 1 .

Complex interpolation of Lp -spaces (see e.g. [1, section 5.1] or [20, subsection 1.18.4])
yields

B : p (Zd ) → Lp (Td ) ≤ 1 ,

where the conjugate index p of p is defined by p1 + p1 = 1.


Now we turn to D. Since 1/2 + 1/r = 1 − 1/p = 1/p and 1/w(k) k∈Zd ∈ r (Zd ),
Hölder’s inequality implies that D maps 2 (Zd ) into p (Zd ). Moreover we have


∞ 1/r
an (D : 2 (Zd ) → p (Zd )) = σjr
j=n

where 1/r = 1/p − 1/2 = 1/2 − 1/p, see [13, Theorem 11.11]. By the multiplicativity of
the approximation numbers, the desired estimate follows. 2

The following technical lemma can be proved parallel to Proposition 4.5.

Lemma 4.13. Let 2 < p < ∞ and put 1/r = 1/2 − 1/p, let α > 1/r, C > 0 and N ∈ N.
Assume that

an (Id : Fd (w) → L2 (Td )) ≤ C n−α for all n ≥ N.

Then it follows that

C
an+1 (Id : Fd (w) → Lp (Td )) ≤ n1/r−α for all n ≥ N.
(αr − 1)1/r

Combining Lemma 4.13, Proposition 4.12 and Proposition 4.6 we can supplement
Corollary 4.7 as follows.

Corollary 4.14. Let d ∈ N, 2 < p < ∞ and s > d 12 − 1
p . Then we have for n ≥ 9d ed/2
the estimate

1  32e s/2
an+1 (Id : H s,2 (Td ) → Lp (Td )) ≤ n1/2−1/p−s/d
(αr − 1)1/r d

where 1/r = 1/2 − 1/p and α = s/d.

4212 F. Cobos et al. / Journal of Functional Analysis 270 (2016) 4196–4212


Remark 4.15. (i) Note that the condition s > d 12 − 1
p is no restriction, since

s 1 1
H s,2 (Td ) → Lp (Td ) ⇐⇒ > − .
d 2 p

(ii) The upper bound gives the correct asymptotic rate in n (which is well-known), the
novelty is the explicit constant.
(iii) Corollary 4.14 is just one example of possible applications of Proposition 4.12. One
can also derive results involving Sobolev spaces of dominating mixed smoothness. This
could easily be done by introducing an additional log-term in the assumption of the
technical Lemma 4.13 and combining this with known estimates of an (Id : Hmix s
(Td ) →
L2 (Td )) from [8].

References

[1] J. Bergh, J. Löfström, Interpolation Spaces: An Introduction, Springer, Berlin – Heidelberg – New
York, 1976.
[2] F. Cobos, T. Kühn, W. Sickel, Optimal approximation of multivariate periodic Sobolev functions
in the sup-norm, ArXiv e-prints, arXiv:1505.02636 [math.FA], 2015.
[3] J. Diestel, H. Jarchow, A. Tonge, Absolutely Summing Operators, Cambridge Univ. Press, Cam-
bridge, 1995.
[4] D. Dũng, T. Ullrich, N -widths and ε-dimensions for high-dimensional approximations, Found. Com-
put. Math. 13 (2013) 965–1003.
[5] D.E. Edmunds, H. Triebel, Function Spaces, Entropy Numbers, Differential Operators, Cambridge
Univ. Press, Cambridge, 1996.
[6] J.-P. Kahane, Séries de Fourier absolument convergentes, Springer, Berlin, 1970.
[7] T. Kühn, W. Sickel, T. Ullrich, Approximation numbers of Sobolev embeddings – sharp constants
and tractability, J. Complexity 30 (2014) 95–116.
[8] T. Kühn, W. Sickel, T. Ullrich, Approximation of mixed order Sobolev functions on the d-torus:
asymptotics, preasymptotics and d-dependence, Constr. Approx. 42 (2015) 353–398.
[9] E. Novak, H. Woźniakowski, Tractability of Multivariate Problems. Volume I: Linear Information,
EMS, Zürich, 2008.
[10] E. Novak, H. Woźniakowski, Tractability of Multivariate Problems. Volume II: Standard Information
for Functionals, EMS, Zürich, 2010.
[11] E. Novak, H. Woźniakowski, Tractability of Multivariate Problems. Volume III: Standard Informa-
tion for Operators, EMS, Zürich, 2012.
[12] K.Yu. Osipenko, O.G. Parfenov, Ismagilov type theorems for linear, Gel’fand and Bernstein n-
widths, J. Complexity 11 (1995) 474–492.
[13] A. Pietsch, Operator Ideals, North-Holland, Amsterdam, 1980.
[14] A. Pietsch, Approximation spaces, J. Approx. Theory 32 (1981) 115–134.
[15] A. Pietsch, Eigenvalues and s-Numbers, Cambridge University Press, Cambridge, 1987.
[16] S.B. Stečkin, About the absolute convergence of orthogonal series (Russian), Dokl. Akad. Nauk
SSSR 102 (1955) 37–40.
[17] V.N. Temlyakov, Approximation of periodic functions of several variables by trigonometric polyno-
mials and widths of some classes of functions, Math. USSR, Izv. 27 (1986) 285–322.
[18] V.N. Temlyakov, Approximation of Periodic Functions, Nova Science, New York, 1993.
[19] V.N. Temlyakov, On approximate recovery of functions with bounded mixed derivative, J. Com-
plexity 9 (1993) 41–59.
[20] H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, North-Holland, Amster-
dam, 1978.
[21] J. Vybíral, Widths of embeddings in function spaces, J. Complexity 24 (2008) 545–570.
[22] X. Wang, Volumes of generalized unit balls, Math. Mag. 78 (5) (2005) 390–395.
[23] H. Yserentant, Regularity and Approximability of Electronic Wave Functions, Springer, Berlin, 2010.