Best $n$-term approximation of diagonal operators and application to function spaces with mixed smoothness

Nowadays, it is well understood that nonlinear methods of approximation and numerical methods derived from them often produce superior performance when compared with linear methods. In the last three decades there has been a great success in studying nonlinear approximation which was motivated by numerous applications such as numerical analysis, image processing, statistical learning as well as in the design of neural networks. We refer the reader to [13, 14, 15] for the development of nonlinear approximation and its application.

In the present paper we concentrate on a particular nonlinear method, the so-called best $n$-term approximation. Our particular interest is exact values of best $n$-term approximation of diagonal linear operators from $\ell_{p}(\mathbb{N})$ to $\ell_{q}(\mathbb{N})$. The exact values of approximation quantities of diagonal operators play an important role in high-dimensional approximation and particularly in studying tractability, see, e.g., [7, 24, 25, 26]. In this paper, the exact values of best $n$-term approximation of diagonal operators will be applied to get the asymptotic constants of best $n$-term approximation of function spaces with mixed smoothness by trigonometric system.

Let $X$, $Y$ be Banach spaces and $T$ a continuous linear operator from $X$ to $Y$. Let $\mathcal{D}$ be a given countable set in $Y$, called dictionary. For given $x\in X$ we consider the algorithm to approximate $Tx$ by a finite linear combination of elements contained in this dictionary. The error of this approximation is

We wish to approximate $Tx$ for all $x$ in the closed unit ball of $X$ with respect to the dictionary $\mathcal{D}$. This can be measured by the following benchmark quantity

In what follows, we shall call this quantity the best $n$-term approximation width.

Let $\ell_{p}(\mathbb{N})$, $0<p\leq\infty$, be the classical complex sequence space with the usual (quasi) norm. For $0<p,q\leq\infty$ and positive non-increasing sequence $\lambda=(\lambda_{k})_{k\in\mathbb{N}}$, consider the diagonal linear operator

from $\ell_{p}(\mathbb{N})$ to $\ell_{q}(\mathbb{N})$ and $\mathcal{E}=\{e_{k}:k\in\mathbb{N}\}$ where $e_{k}=(\delta_{k,j})_{j\in\mathbb{N}}$. We are concerned with the exact value of $\sigma_{n}(T_{\lambda},\mathcal{E})$. The first result in this direction was given by Stepanets [34] in the case $p=q$ with the condition $\lim_{k\to\infty}\lambda_{k}=0$. Later Stepanets generalized his result to the case $0<p\leq q<\infty$, see [35]. Under the same condition $\lim_{k\to\infty}\lambda_{k}=0$ but by different approach, Gensun and Lixin [17] also obtained exact value of $\sigma_{n}(T_{\lambda},\mathcal{E})$ in the case $p=q$. In this paper we give exact values of the best $n$-term approximation widths $\sigma_{n}(T_{\lambda},\mathcal{E})$, $n\in\mathbb{N}$, in all cases. We also show that the condition $\lim_{k\to\infty}\lambda_{k}=0$ in the case $0<p<q<\infty$ is not necessary. Our main result reads as follows. If $0<p<q<\infty$, then we have

where $n^{*}$ is the smallest integer $m>n$ such that

In the case $0<q<p<\infty$ and the series $\sum_{k=1}^{\infty}\lambda_{k}^{pq/(p-q)}$ converges, we get

where $n_{*}$ is the largest integer $m>n$ such that

The limiting cases $p=q$ or $p=\infty$ and/or $q=\infty$ are also obtained, see Theorem 2.1.

The above results will be applied to study best $n$-term approximation of embedding of function spaces with mixed smoothness by trigonometric system $\mathcal{T}^{d}:=\{e^{{\rm i}kx}:\,k\in\mathbb{Z}^{d}\}$ on the torus ${\mathbb{T}}^{d}$ of dimension $d$. Our motivation stems from high-dimensional approximation which has been the object of an intensive study recently. In many high-dimensional approximation problems when the high-dimensional signals or functions have appropriate mixed smoothness, one can apply efficiently approximation methods and sampling algorithms constructed on sparse grids to obtain tractability for algorithms or numerical methods. We refer the reader to the monographs [27, 28] for concepts of computation complexity and results on high dimensional problems. Survey on various aspects of high-dimensional approximation of functions having mixed smoothness can be found in the recent book [12].

There has been a numerous papers working on best $n$-term approximation of embeddings of function spaces with mixed smoothness by different dictionaries. For instance, Bazarkhanov [2], Dinh Dũng [8, 10, 11], Kashin and Temlyakov [23], Romanyuk [31, 32], Romanyuk and Romanyuk [33], Temlyakov [36, 37, 39, 40] worked on trigonometric system; Hansen and Sickel [1, 9, 19, 20] on wavelet system. For some recent contributions in this direction we refer to [3, 5, 41, 42]. Historical comments and further references for studies of best $n$-term approximation of function spaces with mixed smoothness can be found in the two recent books [12, Chapter 7] and [38, Chapter 9]. Let us emphasize here that all the above mentioned papers worked only on the asymptotic order of best $n$-term approximation widths of embeddings of function spaces with mixed smoothness. The asymptotic constants and pre-asymptotic estimates of this quantity have not been considered.

Let $0<s<\infty$ and $0<r\leq\infty$. This paper considers the best $n$-term approximation of embedding of Sobolev space with mixed smoothness $H^{s,r}_{\mathop{\rm mix}}({\mathbb{T}}^{d})$ on the torus ${\mathbb{T}}^{d}$ into either $L_{2}({\mathbb{T}}^{d})$ or Wiener space $\mathcal{A}({\mathbb{T}}^{d})$. In this context we will not only investigate the optimal order of the decay of the best $n$-term approximation widths but we will determine the asymptotic constant as well. This sheds some light not only on the dependence on $n$, but also on the dependence on $s,r$ and in particular on $d$. We have

and if $s>1/2$

In this paper we also obtain the asymptotic constants of best $n$-term approximation widths of embeddings of Sobolev spaces with mixed smoothness $H^{s,2}_{\mathop{\rm mix}}({\mathbb{T}}^{d})$ into the energy norm space $H^{1}({\mathbb{T}}^{d})$. Those embeddings are of particular importance with respect to the numerical solution of the Poisson equation, see [4]. In this case, with $s>1$ we get

where

The paper is organized as follows. In Section 2 we collect some properties of best $n$-term approximation widths and give exact values of best $n$-term approximation widths of diagonal operators. The next Section 3 is devoted to the study of asymptotic constants of best $n$-term approximation widths of embeddings of weighted classes $F_{\omega,p}({\mathbb{T}}^{d})$. These results will be used in final Section 4, where we deal with the particular family of weights associated to function spaces of dominating mixed smoothness.

This section is devoted to give exact values of the best $n$-term approximation widths of the diagonal operator defined in (1.1). Let $X,Y$ be Banach spaces, $T\in\mathcal{L}(X,Y)$, and $\mathcal{D}\subset Y$ a dictionary. By definition, it is clear that $(\sigma_{n}(T,\mathcal{D}))_{n\in\mathbb{N}}$ is a non-increasing sequence. If $W,Z$ are Banach spaces and $A\in{\mathcal{L}}(W,X)$, $B\in{\mathcal{L}}(Y,Z)$ then we have

A proof of this fact can be found in [5, Lemma 6.1]. For further properties of the best $n$-term approximation widths such as additivity, interpolation we refer the reader to [18, 5, 42]. In fact the best $n$-term approximation widths belong to the notion of pseudo $s$-numbers introduced by Pietsch, see [42].

Let $T_{\lambda}$ be the diagonal operator from $\ell_{p}(\mathbb{N})$ to $\ell_{q}(\mathbb{N})$ defined in (1.1). By definition we have

where $B_{p}$ is the closed unit ball of $\ell_{p}(\mathbb{N})$ and $\Gamma_{n}$ is an arbitrary subset of $\mathbb{N}$ with $n$ elements. In the following we give exact value of $\sigma_{n}(T_{\lambda},\mathcal{E})$ with $0<p,q\leq\infty$. The proof is mainly based on the exact values of $n$-term approximation widths of the diagonal operators from $\ell_{p}^{M}$ to $\ell_{q}^{M}$ obtained by Gao in [16]. Here $\ell_{p}^{M}$ stands for $\mathbb{C}^{M}$ equipped with the usual norm $\|\cdot\|_{\ell_{p}^{M}}$. For a vector $\lambda=(\lambda_{j})_{j=1}^{M}$ with $\lambda_{1}\geq\lambda_{2}\geq\ldots\geq\lambda_{M}>0$ the diagonal operator $T^{M}_{\lambda}$ from $\ell_{p}^{M}$ to $\ell_{q}^{M}$ is defined by $(\xi_{j})_{j=1}^{M}\mapsto(\lambda_{j}\xi_{j})_{j=1}^{M}$. Let $\mathcal{E}_{M}=\{e_{1},\ldots,e_{M}\}$ be the standard basis of $\mathbb{R}^{M}$. If $n\in\mathbb{N}$ and $n\leq M$ we have

where $B_{p}^{M}$ is the closed unit ball of $\ell_{p}^{M}$ and $\Gamma_{n}^{M}$ is an arbitrary subset of $\{1,\ldots,M\}$ with $n$ elements. For $\lambda=(\lambda_{j})_{j\in\mathbb{N}}$ we define $T^{M}_{\lambda}:=T^{M}_{\tilde{\lambda}}$ where $\tilde{\lambda}=(\lambda_{j})_{j=1}^{M}$. When $q=\infty$ the summation in (2.3) is replaced by supremum.

Note that if $\lambda=(\lambda_{j})_{j\in\mathbb{N}}$ satisfying $\lambda_{1}\geq\lambda_{2}\geq\ldots\geq\lambda_{M}>0$ and $\lambda_{j}=0$ for $j\geq M+1$, then

which were obtained in [16]. Therefore, we only consider the operator $T_{\lambda}$ where $\lambda=(\lambda_{j})_{j\in\mathbb{N}}$ is a positive sequence. Our main result in this section reads as follows.

As mentioned in Introduction, the exact values of $\sigma_{n}(T_{\lambda},\mathcal{E})$, $n\in\mathbb{N}$, in the case $0<p\leq q\leq\infty$ were obtained in [34, 35, 17] under the condition $\lim_{k\to\infty}\lambda_{k}=0$. To prove the above theorem, we need some auxiliary results.

We are now in position to prove Theorem 2.1.

Let ${\mathbb{T}}^{d}$ be the $d-$dimensional torus. We equip ${\mathbb{T}}^{d}$ with the probability measure $(2\pi)^{-d}\mathrm{d}x$. In this section we study the asymptotic constants of best $n$-term approximation widths of embeddings of the weighted function classes $F_{\omega,p}({\mathbb{T}}^{d})$ by trigonometric system $\mathcal{T}^{d}$. For a function $f\in L_{1}({\mathbb{T}}^{d})$, its Fourier coefficients are defined as