Optimal Algorithms for Numerical Integration: Recent Results and Open Problems
Extracted source text ยท arXiv 2307.06787
%%%%%%%%%%%%%%%%%%%% novak.tex %%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% modified root file for "contribution" to a contributed volume
%
% I used author.tex as a template for my own input.
%
%%%%%%%%%%%%%%%%%%%% Erich Novak %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\documentclass[graybox,footinfo]{svmult}
\smartqed
\usepackage{mathptmx} % selects Times Roman as basic font
\usepackage{helvet} % selects Helvetica as sans-serif font
\usepackage{courier} % selects Courier as typewriter font
\usepackage{type1cm} % activate if the above 3 fonts are
% not available on your system
\usepackage{graphicx} % standard LaTeX graphics tool
% when including figure files
\usepackage{array,colortbl}
\usepackage{amsmath,amsfonts,amssymb,bm} % no amsthm, Springer defines Theorem, Lemma, etc themselves
\usepackage{microtype} % good font tricks
\usepackage[colorlinks=true,linkcolor=black,citecolor=black,urlcolor=black]{hyperref}
\urlstyle{same}
\usepackage{bookmark}
\pdfstringdefDisableCommands{\def\and{, }}
\makeatletter % to avoid hyperref warnings:
\providecommand*{\toclevel@author}{999}
\providecommand*{\toclevel@title}{0}
\makeatother
\usepackage{color}
% \usepackage[notcite,notref]{showkeys}
% \usepackage{showidx} % makes the index entries visible on each page
\usepackage{hyperref} % gives nice possibilities if you check the
% pdf on a screen
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% \include{common-macros} Oder alternativ:
\newcommand{\N}{{\mathbb{N}}} % natural numbers {1, 2, ...}
\newcommand{\R}{{\mathbb{R}}} % reals
\newcommand{\rd}{\,\mathrm{d}} % differential symbol
\newcommand{\norm}[1]{\left\Vert#1\right\Vert}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage{citeref} % generates page numbers, where an item is cited
%%%%%%%%%% OWN COMMANDS %%%%%%%%%%
\newcommand\novakJAT{{J. Approx. Th.}}
\newcommand\novakNM{{Numer. Math.}}
\newcommand\novakCA{{Constr. Approx. }}
\newcommand\novakJC{{J. Complexity }}
\newcommand\novakMC{{Math. Comp. }}
\newcommand\novakUSSR{{USSR Comp. Math. Math. Phys. }}
\newcommand{\novakchange}[1]{{}{#1}}
% \newcommand{\novakchange}[1]{{\color{blue}{#1}}}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% \title*{Numerical Integration and its Complexity: \\ \smallskip
\title*{Optimal Algorithms for Numerical Integration: \\ \smallskip
Recent Results and Open Problems}
% if the original contribution title is too long you may use the following
% short title
\titlerunning{Numerical Integration and its Complexity}
\author{Erich Novak}
% Use \authorrunning{Short Title} for an abbreviated version of
% your contribution title if the original one is too long
\institute{Erich Novak
\at Mathematisches Institut, University Jena,
Ernst-Abbe-Platz 2, D-07743 Jena, Germany\\
\email{erich.novak@uni-jena.de}}
\maketitle
\abstract{
We present recent
results on optimal algorithms for numerical integration
and several open problems.
The paper has \novakchange{six} parts:
\begin{enumerate}
\item \novakchange{Introduction}
\item Lower Bounds
\item Universality
\item General Domains
\item iid Information
\item Concluding Remarks
\end{enumerate}
}
\section{Introduction}
We study the problem of numerical integration, i.e., of approximating
the integral
\begin{equation}\label{novakeq01}
S_d(f) = \int_{D_d} f(x) \rd x
\end{equation}
over an open subset $D_d\subset \R^d$ of
Lebesgue measure $\lambda^d(D_d)=1$ for integrable functions
$f\colon D_d\to\R$.
The main interest is on the behavior of the minimal number of function values
that are needed in the worst case setting
to achieve an error of at most $\varepsilon>0$.
Note that classical examples of domains $D_d$ are the unit cube $[0,1]^d$ and
the normalized Euclidean ball (with volume 1),
which are closed. However, we work with their interiors
for definedness of certain derivatives.
We state the problem. Let $F_d$ be a class of % continuous
integrable functions $f\colon D_d\to\R$. For~\mbox{$f \in F_d$},
we approximate the integral $S_d(f)$, see~\eqref{novakeq01}, by algorithms
of the form
$$
A_{n}(f)=\phi_{n}(f(x_1),f(x_2),\dots,f(x_n)),
$$
where $x_j \in D_d$ can be chosen adaptively and $\phi_{n}\colon \R^n\to
\R$ is an arbitrary mapping. Adaption means that the selection of $x_j$
may depend on the already computed values $f(x_1),f(x_2),\dots,f(x_{j-1})$.
We define $N\colon F_d \to \R^n$ by
$N(f) = (f(x_1), \dots , f(x_n))$.
The (worst case) error of the algorithm $A_{n}$ is defined by % \footnote{Some
% authors still try to assess and compare different algorithms
% without specifying a class $F_d$ of inputs. This is hopeless -- in any part
% of computational mathematics.}
$$
e(A_{n}, F_d )=\sup_{f \in F_d}|S_d(f)-A_{n}(f)|,
$$
the optimal errors are given by
$$
e(n, F_d) = \inf_{A_n} \, e(A_n) .
$$
We minimize over all choices of $n$ adaptive sample points $x_j$ and
mappings $\phi_{n}$.
The information complexity $n(\varepsilon,F_d)$ is
the minimal number of function values
which is needed to guarantee that the error is
at most $\varepsilon$, i.e.,
$$
n(\varepsilon,F_d)=\min\{n \mid \exists\ A_{n}\ \mbox{such that}\
e(A_{n})\le\varepsilon\}.
$$
Here we minimize $n$ over all choices of adaptive sample points $x_j$ and
mappings $\phi_{n}$.
It is essential to specify a class $F_d$ in order to
assess and compare different algorithms,
and to do a complexity analysis.
It was proved by Smolyak and Bakhvalov
that as long as the class $F_d$
is convex and symmetric
we may restrict the minimization of $e(A_n)$ by
considering only nonadaptive
choices of $x_j$ and linear mappings $\phi_{n}$,
i.e., it is enough to consider $A_n$
of the form
\begin{equation} \label{novakeq03}
A_n(f) = \sum_{i=1}^n a_i f(x_i)
\end{equation}
and, moreover,
\begin{equation} \label{novak1234}
e(n, F_d) = %\operatorname*{\vphantom{p}inf}_{x_1,\ldots,x_n}
\inf_{x_1,\ldots,x_n}
\sup_{\substack{f \in F_d \\ N(f) = 0}}
S_d(f) \, .
\end{equation}
\begin{remark}
a) For a proof of this statement see,
for example, \cite[Theorem~4.7]{novakNW08}. This result clearly helps
to find optimal algorithms and to prove
complexity results.
b) A~linear algorithm~$A_n$ is called a quasi Monte Carlo (QMC) algorithm
if~\mbox{$a_i = 1/n$} for all~$i$ and is called a positive
quadrature formula if~\mbox{$a_i > 0$} for all~$i$.
c) More on the optimality of linear algorithms
and on the power of adaption can be found in
\cite{novakCW04,novakNo96,novakNW08,novakTWW88,novakTW80}.
There are important classes of functions that are \emph{not}
convex and symmetric,
see
\cite{novakCDHHZ,novakGo21,novakKNR19,novakNR96,novakPl15,novakPW09}.
\novakchange{In this paper} we mainly
consider convex and symmetric $F_d$
and then
we can use formula \eqref{novak1234} for $e(n,F_d)$.
\novakchange{
d) So far we spoke about deterministic algorithms and one
may ask whether adaption helps for randomized algorithms,
again under the assumption that $F$ is convex and symmetric.
This question was open for many years and just recently
solved by Stefan Heinrich~\cite{novakHe24}:
For \emph{some} classes $F$ adaption helps a lot in the randomized setting.
}
\qed
\end{remark}
\section{Lower Bounds}
The \emph{optimal rate of convergence} plays an important role
in numerical analysis.
We start with a classical result of Bakhvalov~\cite{novakBa59}
for the class
$$
F^k_d = \{ f\colon [0,1]^d \to \R \mid
\Vert D^\alpha f \Vert_\infty \le 1, \
|\alpha| \le k \},
$$
where $k \in \N$ and $|\alpha| = \sum_{i=1}^d \alpha_i $ for
$\alpha \in \N_0^d$ and $D^\alpha f$ denotes the respective
partial derivative.
For two sequences $a_n$ and $b_n$ of positive numbers we write
$a_n \asymp b_n$ if there are positive numbers $c$ and $C$ such
that
$c < a_n/b_n < C $ for all $n \in \N$;
we write $a_n \approx b_n$ if $\lim_{n \to \infty} \frac{a_n}{b_n} = 1$.
Bakhvalov proved that
\begin{equation} \label{novak59}
e(n, F_d^k) \asymp n^{-k/d} .
\end{equation}
Observe that we cannot conclude anything on
$n(\varepsilon,F_d^k)$ if $\varepsilon$ is fixed and $d$ is large, since
\eqref{novak59} contains \emph{hidden factors} that depend on
$k$ and $d$. Actually the lower bound was of the form
$$
e(n, F_d^k) \ge c_{d,k} n^{-k/d} ,
$$
where the $c_{d,k}$ decrease with $d \to \infty$ and tend to zero.
% TECHNIQUE OF BUMP FUNCTIONS.
% NEW RESULTS?
The result of Bakhvalov was recently improved and now we know that
$n(\varepsilon,F_d^k)$ is of the order $(c_k d/\varepsilon)^{d/k}$.
A simplified version of the (upper and lower) bounds reads
as follows.
\begin{theorem} \cite{novakHNUW17}. \
\label{novakT1}
\begin{equation}
\min(1/2, c_{k,1} d n^{-k/d} ) \le e(n, F_d^k)
\le \min (1, c_{k,2} d n^{-k/d} ) ,
\end{equation}
where, for the upper bound, we assume that $n=m^d$, $m \in \N$.
\end{theorem}
The lower bounds in \cite{novakHNUW17} are
explicit and they hold for all domains $D_d$ with volume 1.
For $k=2$ we have
\begin{equation}
n(\varepsilon,F_d^2( D_d) ) \ge \frac{2}{2+d} \left( \frac{1}{533}
\frac{d}{d+2} \frac{d}{\varepsilon}
\right)^{d/2} ,
\end{equation}
for all $\varepsilon \in (0, d/(d+2))$.
The classes $F_d^k(D_d)$ are defined as the $F_d^k$, the cube $[0,1]^d$ is now replacd
by a general domain $D_d$.
By the \emph{curse of dimensionality} we mean that
$n(\varepsilon,F_d)$ is exponentially large in $d$.
That is, there are positive numbers $c$, $\varepsilon_0$ and $\gamma$ such that
\begin{equation} \label{novakcurse}
n(\varepsilon,F_d) \ge c \, (1+\gamma)^d \quad
\mbox{for all} \quad \varepsilon \le \varepsilon_0 \quad \mbox{and infinitely many}
\quad d\in \N.
\end{equation}
If, on the other hand,
$n(\varepsilon,F_d)$ is bounded by a polynomial in $d$ and $\varepsilon^{-1}$
then we say that the problem is \emph{polynomially tractable}.
If
$n(\varepsilon,F_d)$ is bounded by a polynomial in $\varepsilon^{-1}$ alone, i.e.,
$n(\varepsilon,F_d) \le C \varepsilon^{-\alpha}$ for $\varepsilon < 1$
with $C$ and $\alpha$ independent of $d$,
then we say that the problem is \emph{strongly polynomially tractable}.
\begin{remark}
a) On the proof of Theorem~\ref{novakT1}:
For the upper bound it is
enough to consider standard product rules that are universal,
they do not depend on the smoothness $k$.
See Section~\ref{novaks3} for more on universality.
New are the lower bounds.
b)
It is often said that product rules are bad in higher dimension,
but of course this depends on the function classes. For the classes $F_d^k$
product rules are almost optimal.
c)
It clearly follows from Theorem~\ref{novakT1}
that the curse of dimensionality holds for the classes $F_d^k$.
It seems that these classes are too big and/or the error criterion
is too strong.
\end{remark}
\noindent
{\bf Open Problem:} \
For which $D_d$ do we have upper bounds of the form
\[
e(n, F_d^k (D_d))
\le c_{k} d n^{-k/d} ?
\]
Construct algorithms that achieve this bound for such
domains.
\smallskip
We also discuss randomized (or Monte Carlo) algorithms
and,
for the purpose of this paper,
it might be enough that a randomized algorithm $A$
is a random variable $(A^\omega)_{\omega \in \Omega}$
with a random element $\omega$ where,
for each fixed $\omega$, the algorithm $A^\omega$ is a
(deterministic) algorithm as before.
We denote by $\mu$ the distribution of the $\omega$.
In addition one needs rather weak measurability assumptions,
see also the textbook \cite{novakMNR09}.
Let $\bar n(f,\omega)$ be the number of function values
used for fixed $\omega$ and $f$.
The number
$$
\tilde n(A) = \sup_{f \in F_d} \int_\Omega \bar n(f, \omega) \rd\mu (\omega)
$$
is called the \emph{cardinality} of the randomized algorithm $A$ and
% $$
% e^{\rm ran} (A,F_d) = \sup_{f \in F_d} \left( \int_\Omega^*
% \Vert S(f) - \phi_\omega (N_\omega (f))
% \Vert^2 \rd\mu (\omega) \right)^{1/2}
% $$
\[
e^{\rm ran} (A,F_d) = \sup_{f \in F_d} \int_\Omega^*
\vert S(f) - \phi_\omega (N_\omega (f))
\vert \rd\mu (\omega)
\]
is the \emph{error} of $A$.
By $\int^*$ we denote the upper integral. For $n \in \N$, define
$$
e^{\rm ran} (n, F_d) = \inf \{ e^{\rm ran} (A, F_d) \, : \, \tilde n(A) \le n \} .
$$
If $A\colon F_d \to G$ is a deterministic algorithm
then $A$ can also be treated as a randomized algorithm
with respect to a Dirac (atomic) measure $\mu$.
In this sense we can say that deterministic algorithms are special
randomized algorithms. Hence the inequality
\begin{equation}
e^{\rm ran} (n, F_d) \le e (n, F_d)
\end{equation}
is trivial.
\begin{remark}
The error numbers $e(n,F)$ and $e^{\rm ran} (n,F)$ can be bounded from above
by the Kolmogorov widths of $F$ in the space
$L_\infty$ and $L_2$, respectively.
This is an old result from \cite{novakNo86} that got some attention
recently since similar upper bounds could be proved
for several approximation problems
(optimal recovery in $L_p$ spaces with deterministic or randomized
algorithms).
\end{remark}
\begin{remark}
There is another error criterion for randomized algorithms with two parameters:
Small error $\varepsilon$ with large probability $1- \delta$.
The median of mean algorithm is often used, in particular for small $\delta >0$,
see also
\cite{novakGL22,novakGS22,novakHJ19,novakKNR19,novakKR19,novakPO22,novakPO22}.
\end{remark}
Bakhvalov~\cite{novakBa59} found the optimal rate of convergence
for the numbers $e^{\rm ran} (n, F_d)$
already in 1959 for the classes
$
F^k_d = \{ f\colon [0,1]^d \to \R \mid
\Vert D^\alpha f \Vert_\infty \le 1, \
|\alpha| \le k \}
$
that we already defined.
He proved
\begin{equation}
e^{\rm ran} (n, F_d^k) \asymp n^{-k/d-1/2} .
\end{equation}
\begin{remark}
A~proof of the \emph{upper bound} can be given with a technique
that is often called \emph{separation of the main part}
or also \emph{control variates}.
For $n=2m$ use $m$~function values to construct a good $L_2$ approximation
$f_m$ of $f \in F_d^k$ by a deterministic algorithm.
The optimal rate of convergence is
$$
\Vert f - f_m \Vert_2 \asymp m^{-k/d} .
$$
Then use the unbiased estimator
$$
A_n^\omega (f) = S_d (f_m) + \frac{1}{m} \sum_{i=1}^m (f-f_m)(X_i)
$$
with iid random variables $X_i$ that are uniformly distributed
on $[0,1]^d$.
See, for example, \cite{novakMNR09,novakNo88} for more details.
An algorithm of Krieg and Novak~\cite{novakKN17} is
universal, it works for every $k$, see Section~\ref{novaks3}.
% See also Section~5.
%
% We add in passing that the optimal order of convergence can be obtained for
% many function spaces (Besov spaces, Triebel-Lizorkin spaces)
% and for arbitrary bounded Lipschitz domains
% $D_d \subset \R^d$; see \cite{novakNT06}, where the approximation problem is
% studied.
\end{remark}
To obtain an explicit randomized
algorithm with the optimal rate of convergence
one needs a random number generator for the set $D_d$.
If it is not possible to obtain
efficiently random samples from the uniform distribution
on $D_d$ one can work with Markov chain Monte
Carlo (MCMC) methods, see
Rudolf~\cite{novakRu12}.
\smallskip
\noindent
{\bf Open Problem: } \
Study, similarly as in Theorem~\ref{novakT1}, the dependence on the dimension
for randomized algorithms.
Using results of Krieg~\cite{novakKr19} one can prove upper bounds
of the form $d^t \, n^{-k/d-1/2}$.
One may take $t=(k+1)/2$, but the optimal (smallest) $t$ is not known.
\smallskip
Now we study error bounds for quadrature formulas and assume
that the integrand is from a Hilbert space $F$ of
real valued functions defined
on a set~$D$.
We assume that function evaluation is continuous and hence are dealing
with a reproducing kernel Hilbert space (RKHS) $F$ with a kernel $K$.
We want to compute $S(f)$ for $f \in F$, where $S$ is a
continuous
linear functional,
hence
$S (f) =\langle f , h \rangle$
for some $h\in F$.
Vyb{\'\i}ral~\cite{novakVy20}
recently solved old open problems on
positive definite matrices by improving
the Schur product theorem.
This yields new lower bounds for the complexity
of such integration problems,
see
\cite{novakHKNV21} \novakchange{and Krieg and
Vyb{\'\i}ral~\cite{novakKV23}.}
A main result of
\cite{novakVy20}
is the following.
Assume that $M \in \R^{n \times n}$ is positive semi-definite.
Then also
\begin{equation}
M \circ M - \frac{1}{n} ({\mathop{\mathrm{diag}}} M)({\mathop{\mathrm{diag}}} M)^T
\end{equation}
is positive semi-definite.
Here $M \circ M$ denotes the pointwise or Schur product
and
${\mathop{\mathrm{diag}}} M=(M_{1,1},\dots,M_{n,n})^T$ refers to
the diagonal entries of $M$
whenever $M \in \R^{n\times n}$.
%
% Moreover, if $A,B\in\R^{n\times n}$ are two symmetric matrices,
% we write $A\succeq B$ if $A-B$ is positive semi-definite.
% The Schur product of $A$ and $B$
% is the matrix $A\circ B$ with
% \[
% (A\circ B)_{i,j}=A_{i,j}B_{i,j}\quad\text{for}\quad i,j\le n.
% \]
The classical Schur product theorem states that
the Schur product of two positive semi-definite matrices
is again positive semi-definite.
The result of
Vyb{\'\i}ral~\cite{novakVy20}
improves this if the two matrices coincide.
Now we use that lower
bounds for the worst case error of quadrature formulas are equivalent
to the statement that certain matrices are positive semi-definite:
The inequality
\[
e(n, F)^2 \ge \Vert S \Vert^2 - \alpha^{-1}
\]
for an $\alpha >0$ is equivalent to the fact that
\[
(K(x_i,x_k) - \alpha h(x_i)h(x_k))_{i,k}
\]
is positive semi-definite for all $x_1, x_2, \dots , x_n$.
We are mainly interested in
\emph{tensor product problems}. We will therefore assume that
$H_i$ is a RKHS
on a domain $D_i$ with kernel $K_i$ for all $i\le d$
and that $F_d$ is the tensor product of these spaces.
That is, $F_d$ is a RKHS on $\mathbf D_d=D_1\times\cdots\times D_d$
with reproducing kernel
\[
\mathbf K_d: \mathbf D_d\times \mathbf D_d
\to \R,\quad \mathbf K_d(x,y)=\prod_{i=1}^d K_i(x^i,y^i).
\]
If $h_i \in H_i$ and $\tilde S_i(f)=\langle f,h_i \rangle$
for $f\in H_i$,
we will denote by $\mathbf h_d$
the tensor product of the functions $h_i$, i.e.,
$$
\mathbf h_d(t)=(h_1\otimes \dots\otimes h_d)(t)=h_1(t^1)\cdot\ldots\cdot h_d(t^d),
\quad t=(t^1,\dots,t^d)\in \mathbf D_d. %
$$
We study the tensor product functional
$S_d=\langle\cdot, \mathbf h_d \rangle$ on
the unit ball of $F_d$.
Before we state a general result, we present the example that initiated
this research, \novakchange{see \cite{novakNo99}}.
Here $H_i$ is the three-dimensional RKHS
on $D_i=[0,1]$ with the three orthonormal functions
$b_1=1$, $b_2=\cos(2\pi x)$ and $b_3=\sin(2\pi x)$,
and we consider the usual integral over $[0,1]^d$.
\novakchange{Vyb{\'\i}ral~\cite{novakVy20} proved,}
for this example,
\[
e(n,F_d)^2 \ge 1 - n 2^{-d}
\]
and positive quadrature formulas are optimal.
Of course this implies the curse of dimensionality:
To achieve the modest error $\frac{1}{2} \sqrt{2}$
we need $2^{d-1}$ function values.
A more general result is this:
\begin{theorem} \cite{novakHKNV21}. \ \label{novakT2}
For all $i\le d$, let $H_i$ be a RKHS and let $\tilde S_i$ be a bounded linear
functional on $H_i$
with unit norm and nonnegative representer~$h_i$.
Assume that there are functions $f_i$ and $g_i$ in $H_i$ and a number $\alpha_i\in (0,1]$ such that
$(h_i,f_i,g_i)$ is orthonormal in $H_i$ and $\alpha_i h_i=\sqrt{f_i^2+g_i^2}$.
Then the tensor product problem $S_d=\tilde S_1\otimes\hdots\otimes \tilde S_d$ %
satisfies for all $n\in\N$ that
\[
e(n,F_d)^2 \ge 1-n\prod_{i=1}^d(1+\alpha_i^2)^{-1}.
\]
\end{theorem}
In particular, if all the $\alpha_i$'s are
equal to some $\alpha>0$ and we want to achieve $e(n, F_d )\le \varepsilon$
for some $0<\varepsilon<1$,
we obtain
\[
n(\varepsilon , F_d )\ge (1-\varepsilon^2)(1+\alpha^2)^d.
\]
This implies the curse of dimensionality.
As an application, we may use this result to obtain lower bounds
for the complexity of the integration
problem on Korobov spaces with increasing smoothness.
These lower bounds complement existing upper bounds from~\cite[Section 10.7.4]{novakNW10}.
We add in passing that lower bounds of this form are known and much easier
to prove for quadrature formulas that only have positive weights,
see Theorem 10.2 of \cite{novakNW10}.
Using these results
one can also prove
new lower bounds
for rather classical Sobolev spaces
of periodic univariate functions
with small (logarithmic) smoothness,
see~\cite{novakHKNV22}:
The integration error
behaves worse than the approximation numbers
for these spaces.
\novakchange{Multivariate periodic functions and proof techniques
for lower bounds are studied in the recent paper
by
Krieg and
Vyb{\'\i}ral~\cite{novakKV23}.}
Lower bounds were studied in \cite{novakNW01} with
the technique of \emph{decomposable kernels}.
This technique is rather general as long as we consider finite smoothness.
The technique does not work, however, for analytic functions.
In contrast, the new approach
also works for polynomials and other analytic functions.
% One result of % \cite{novakHKNV21}
% is the following.
\smallskip
\noindent
{\bf Open Problem: } \
The lower bounds of \cite{novakHKNV21} are for Hilbert spaces.
Do similar lower bounds exist for Banach spaces?
\section{Universality} \label{novaks3}
Let us start with the midpoint rule
\[
A_n (f) = \frac{1}{n} \sum_{i=1}^n f \left( \frac{2i-1}{2n} \right)
\]
for the classes $F^k_1$ defined above.
The worst case error of $A_n$ is $\frac{1}{4n}$ for $k=1$ and
the algorithm %E (more exactly, the sequence $(A_n)_n$ of algorithms)
is optimal.
For $k \ge 2$ the worst case error is
$\frac{1}{24n^2}$, a worst case function is
$f(x) = \frac{1}{2} (x-\frac{1}{2})^2$.
For $k=2$
the algorithm is still almost optimal,
we obtain the optimal rate of convergence.
For $k \ge 3$ the algorithm is bad, the algorithm cannot use
the high smoothness of $f \in F^k_1$.
In contrast to this, the Gaussian rules, or the Clenshaw Curtis method,
are \emph{universal} in the sense that we obtain the optimal
rate $n^{-k}$ of convergence for any $k$.
We would like to have more good properties
and consider randomized algorithms that are good
in the randomized as well as the deterministic
worst case setting.
%E Kann man fuer k=2 Gauss, CC, optimal und midpoint rule vergleichen?
We start with the simplest case: Periodic functions on $[0,1]$,
hence $d=1$.
Then
a very simple algorithm
has many optimality properties:
Pick $n \in \{ m, m+1, m+2, \dots , 2m \}$ randomly
and then take a randomly shifted midpoint rule
with $n$ function values and equal weights.
This
algorithm, but without the random shift, was studied by Bakhvalov~\cite{novakBa61}.
% Welche Eigenschaften hat dieser Alg.? Bakhvalov zeigt die
% optimale Ordnung fuer Sobolev-Hilbertraeume. Bei Hoelder verliert er was.
In the case $d=1$ with non-periodic functions
on $[0,1]$ one can perform a periodization
(for example by adding a suitable polynomial) and thus can keep
the optimal order of convergence but loses some constants.
As an alternative, one can take $n=2m$ and first define, via interpolation at the Chebyshev
points, a polynomial $f_m$ of degree smaller than $m$. Then we use
separation of the main part together with the simplest Monte Carlo
method for $(f-f_m)$ to obtain an almost optimal
algorithm for each class $F^k_1$.
This algorithm is, asymptotically for large $n$, slightly
worse (with a logarithmic loss), but possibly better for
small $n$.
\smallskip
\noindent
{\bf Open Problem: } \
Study and compare these algorithms (for $d=1$) in more detail.
\smallskip
A randomized algorithm
\novakchange{$A_n= (A_n^\omega)_{\omega \in \Omega}$} of the form
\[
A_n^\omega (f) = \sum_{j=1}^n a_j(\omega) f(x_j(\omega))
\]
for $f \in L_1([0,1]^d)$
and the integral
$
S_d(f) = \int_{[0,1]^d} f(x) \rd x
$
was
introduced
by Krieg and Novak in \cite{novakKN17}.
\novakchange{We call this algorithm the ``randomized
Frolov algorithm''}.
To state its properties, we define the norms
\[
\Vert f \Vert^2_{H^k_d} = \sum_{\Vert \alpha \Vert_1 \le k} \Vert
D^\alpha f \Vert^2_{L_2([0,1]^d)}
\]
and
\[
\Vert f \Vert^2_{H^{k,{{\rm mix}}}_d } = \sum_{\Vert \alpha \Vert_\infty \le k} \Vert
D^\alpha f \Vert^2_{L_2([0,1]^d)} \, ,
\]
see also
M.~Ullrich~\cite{novakUl17} and
Krieg~\cite{novakKr19a}.
\begin{theorem} \cite{novakKN17,novakUl17}. \ \label{novakT3}
% There are positive constants $c, c_i$ such that
% the following holds for all $n \in \N$
% with $n \ge c$.
For all $f \in L_1 ([0,1]^d)$ the
\novakchange{randomized Frolov} algorithm is \emph{unbiased}, i.e.,
\begin{equation}
{\mathbb{E}} (A_n(f)) = S_d(f) .
\end{equation}
For classical smoothness $k>d/2$
\[
e^{\rm ran} (A_n, H^{k}_d ) \asymp n^{-k/d-1/2} \quad \text{and} \quad
e (A_n, H^k_d ) \asymp n^{-k/d} .
\]
For mixed smoothness $k \in \N$
\[
e^{\rm ran} (A_n, H^{k,{{\rm mix}}}_d ) \asymp n^{-k-1/2} \quad \text{and} \quad
e (A_n, H^{k,{{\rm mix}}}_d ) \asymp n^{-k} (\log n )^{(d-1)/2} .
\]
All these orders of convergence are optimal.
\end{theorem}
\begin{remark}
The algorithm from \cite{novakKN17} has some nice theoretical
properties\footnote{In particular, the optimal order for
$e^{\rm ran} (A_n, H^{k,{{\rm mix}}}_d )$, observe that there are no
log terms, is a new result of
Mario Ullrich~\cite{novakUl17}.}
but also severe shortcomings:
it is difficult to implement in higher dimensions and
its error is large for moderate values of $n$.
Hence there is a lot of motivation to
look for
other approaches to universality.
a) The Smolyak algorithm: As long as one only considers deterministic algorithms
and error bounds
(and the dimension is small), one can take the algorithm from~\cite{novakNR96b},
see~\cite{novakNR97}.
The algorithm is a combination of the Smolyak (sparse grids) algorithm
with the Clenshaw Curtis quadrature formula.
For deterministic and stochastic error bounds
one may use the algorithm of~\cite{novakBNR99}
(Smolyak algorithm with interpolation at the Chebyshev points)
for a
``separation of the main part'' or control variate and combine it
with the simplest Monte Carlo method.
As far as I know, this algorithm was never really propagated
and studied.
Compared to the algorithm of \cite{novakKN17},
it has slightly worse asymptotic error bounds (with a logarithmic loss),
but is better for moderate values of $n$ and is easier to implement.
Other randomized versions of the Smolyak algorithm were studied
by Gnewuch and Wnuk~\cite{novakGW20}.
b) Lattice rules and (also randomized) quasi-Monte Carlo methods
based on digital nets
have a great potential for universality.
We only cite a few papers, together
with the excellent books by
Dick and Pillichshammer~\cite{novakDP08}
and
Dick, Kritzer and Pillichshammer~\cite{novakDKP22}, where the reader
can find more details:
\cite{novakDG20},
\cite{novakDGS21},
\cite{novakDP13},
\cite{novakEKNO21},
\cite{novakGD15},
\cite{novakGS19},
\cite{novakGS22},
\cite{novakGSY16},
\cite{novakKKNU19},
\novakchange{\cite{novakKNW23}}.
See also other papers by Dick, Goda, Kritzer, Kuo, Nuyens, Pillichshammer, Suzuki
and their coworkers.
\end{remark}
\noindent
{\bf Open Problem: } \
Of course one should study universality for
tractability questions, not only for the rate of convergence.
Some of the cited papers already contain results in this direction.
See, in particular, Chapter~11 of Dick, Kritzer and Pillichshammer~\cite{novakDKP22}
and
the papers
\cite{novakDGS21,novakGS22,novakKKNU19,novakKNW23}.
\smallskip
\noindent
{\bf Open Problem: } \
So far only the domain $[0,1]^d$ was studied for
the construction of universal algorithms.
One might speculate that universal algorithms also
exist for other domains.
\section{General Domains}
We usually assumed that the domain is a cube, $D_d=[0,1]^d$,
now we discuss results for
more general domains
$D_d$.
We study different concepts to measure the complexity.
\subsection*{Order of convergence}
It is known for many problems that the order of convergence
of optimal algorithms does not depend
on the domain $D_d \subset \R^d$.
Results for Lipschitz domains (and beyond)
and classical (isotropic) function spaces
can be found in
\cite{novakKNS21,novakNWW04,novakNT06,novakWSH22}.
The paper by Wenzel, Santin and Haasdonk~\cite{novakWSH22}
has results saying that the order of convergence cannot decrease
if $D_d$ is replaced by a smaller set.
Such results depend on the way how we define
Sobolev (and related) spaces:
Usually all definitions coincide if $D_d$ is a bounded Lipschitz
domain but they do not coincide for more general domains.
Function recovery and integration on manifolds are studied by
Brandolini, Choirat, Colzani, Gigante, Seri
and Travaglini~\cite{novakBCCGST14},
Ehler, Graef and Oates~\cite{novakEGO19}
and
Krieg and Sonnleitner~\cite{novakKS21}.
For spaces of functions
with dominating mixed smoothness many
problems are still open, see Triebel~\cite{novakTr19}.
\subsection*{Asymptotic constant and explicit uniform bounds}
If the rate of convergence does not depend on the domain
$D_d$, then we may ask about the asymptotic constants
and we may study
explicit uniform bounds, i.e.,
bounds that hold for all $n$ (number of pieces of information)
and all (normalized) domains.
We
present examples for which the following statements
are true:
\begin{enumerate}
\item
Also the asymptotic constant does not depend on the shape of $D_d$
or the imposed boundary values,
but only on the volume of the domain.
These results can be compared with results of
Weyl~\cite{novakWeyl} who
proved that the asymptotic constant for the size of the eigenvalues
of the Dirichlet Laplacian
and also of the Neumann Laplacian do not depend on the shape of $D_d$,
it only depends on the volume of $D_d$.
\item
There are explicit and uniform lower (or upper, respectively)
bounds for the error that are
only slightly smaller (or larger, respectively)
than the asymptotic error bound.
% By explicit we mean that the bounds hold for any $n$, not just asymptotically.
% By uniform we mean that they hold for all domains of a given size.
\end{enumerate}
General bounded Lipschitz domains are studied in
\cite{novakNWW04,novakNT06,novakTr08,novakTr10,novakVy06,novakVy07b}.
In many cases, the optimal order of the errors is of the form
$$
e (n, F_d) \asymp n^{-\alpha} \, (\log n)^\beta ,
$$
where $\alpha$ and $\beta$ do not depend on the domain $D_d$.
It is interesting to know also the exact asymptotic constant
$$
C: = \lim_{n\to \infty} e(n, F_d) \, n^{\alpha} \, (\log n)^{-\beta} ,
$$
if it exists.
The value of $C$ is known only in rare cases
(unless $d=1$, we do not discuss the univariate case
in detail),
and usually only for very special domains, like the cube, see
\cite{novakCKS16,novakKr18,novakKSU14,novakKSU15}.
We assume that $D_d \subset \R^d$ is bounded and a metric on $D_d$ is induced
by a norm in $\R^d$. We denote by $B$ the unit
ball (in $\R^d$) and also write $\Vert \cdot \Vert_B$ for the norm.
To simplify the formulas we consider
only Lipschitz functions.
We denote the respective \novakchange{set} by $F^B(D_d)$;
it contains all functions $f\colon D_d \to \R$ with
$$
|f(x) - f(y)| \le \Vert x - y \Vert_B.
$$
We study the problem of numerical integration
and assume that $D_d \subset \R^d$ is Jordan measurable with
$0 < \lambda^d (D_d) < \infty$.
Asymptotic formulas, even for more general (weighted) integration problems,
were proved by Chernaya~\cite{novakCh95} and by Gruber~\cite{novakGr04}.
We add an upper bound for the asymptotic constant $\xi_B$ and see that
it is very close to the lower bound, in particular for large $d$.
The explicit uniform lower bound is from \cite{novakNo20}.
\begin{theorem}
Assume that $D_d \subset \R^d$ is a bounded Jordan measurable
set with an interior point.
Then
\begin{equation}
e(n, F^B(D_d) ) \approx \xi_B \, \lambda^d(D_d)
\left( \frac{\lambda^d(D_d)}{\lambda^d(B)}\right)^{1/d}
\cdot n^{-1/d},
\end{equation}
where $\xi_B$ is a constant that depends on the norm and
\[
\frac{d}{d+1} \le \xi_B \le % \frac{d}{d+1} \Theta_B^{1/d} \le
\frac{d}{d+1} ( d \log d + d \log \log d + 5d )^{1/d} .
\]
Moreover,
\begin{equation}
\inf_{D_d} e(n, F^B(D_d) ) \cdot \lambda^d (D_d)^{-(d+1)/d} = \frac{d}{d+1} \lambda^d (B)^{-1/d}
\cdot n^{-1/d} ,
\end{equation}
the infimum is attained when $D_d$ is the disjoint union
of $n$ balls with equal radius.
\end{theorem}
A similar result for Sobolev spaces
$W^1_p (D_d) $ with $p>d$ was recently obtained by
Babenko, Babenko and Kovalenko~\cite{novakBBK21}.
We now discuss $C^k$-functions.
% We start with a result from \cite{HNUW17}.
We know from \cite{novakHNUW17}
the following statement
for the classes
$
F^k_d (D_d) = \{ f\colon D_d \to \R \mid
\Vert D^\alpha f \Vert_\infty \le 1, \
|\alpha| \le k \}
$ for all
domains $D_d$ with volume 1:
For all $k \in \N$ there exists a constant $c_k>0$ such that for all
$d,n \in \N$
$$
e ( n, F^k_d (D_d) ) \ge \min \{ 1/2, \, c_k \, d \, n^{-k/d} \} .
$$
% Observe that the constant $c_k$ does not depend on $D_d$ or $d$, we have a
% lower bound for all $D_d$ with volume 1.
\begin{remark}
1) The lower bound cannot be improved since for cubes we have a similar upper bound.
It would be good to know more on the constants $c_k$ and on extremal sets $D_d$,
where $e( n, F^k_d(D_d) )$ is small for given $k$, $d$ and $n$.
For these function spaces there cannot be a meaningful explicit uniform upper bound
since
$\sup_{D_d} e( n, F^k_d(D_d) ) =1$.
The supremum over $D_d$ makes sense if we impose boundary conditions
such as $f(x) = 0$ for $x \in \partial D_d $.
2)
It is known that
the weak order
$$
e( n, F^k_d(D_d) ) \asymp n^{-k/d}
$$
holds at least for every bounded Lipschitz domain.
This follows from more general results of \cite{novakKNS21,novakNWW04,novakNT06}.
We guess that more is true and the asymptotic constant
$$
\lim_{n \to \infty} e( n, F^k_d(D_d) ) \cdot \novakchange{n^{k/d}} = C_{D_d}
$$
does not depend on $D_d$ (for fixed $d$) if $D_d$ is Jordan measurable
with $\lambda^d(D_d)=1$;
possibly under some assumptions on the boundary of $D_d$.
\end{remark}
\begin{remark}
The norm
$$
\max_{|\beta|_1 \le k} \Vert D^\beta f \Vert_\infty
$$
might be reasonable when we consider a cube $D_d = [0,1]^d$
but it is not invariant with respect to rotation.
The function $f(x) = \sum x_i$ has a gradient with length
$d^{1/2}$ as $g(x) = d^{1/2} \cdot x_1$, but all
partial derivatives of $f$ are bounded by one.
Therefore we also consider an orthogonal invariant norm.
Since for this modified space $\tilde C^k(D_d)$
many problems are still open, we only discuss the case $k=2$
in the following.
\end{remark}
\begin{example}
Let us discuss the integration problem for the class
$$
\tilde C^2 (D_d) = \{ f\in C^2 (D_d) \mid
\Vert f \Vert_\infty \le 1, \,
{{\rm Lip}} (f) \le d^{-1/2}, \,
{{\rm Lip}} (D^\Theta f ) \le d^{-1} \},
$$
as in \cite{novakHNUW12,novakHNUW14b,novakHPU19}.
Here $D^\Theta f$ denotes any directional derivative in a direction
$\Theta \in S^{d-1}$ and
$$
{{\rm Lip}} (g) = \sup_{x,y\in D_d} \frac{|g(x)-g(y)|}{\Vert x-y\Vert_2} .
$$
Again we assume that $D_d \subset \R^d$ is an open set
with $\lambda^d(D_d) = 1$.
We conjecture that there exists a constant $C>0$ (independent on $d$ and $D_d$) such that
\begin{equation} \label{novak123}
e (n, \tilde C^2 (D_d) ) \ge C \cdot n^{-2/d} .
\end{equation}
\end{example}
Some comments are in order:
1) This would be another explicit uniform lower bound and, because of known upper bounds
for the cube, it certainly cannot be improved.
2) It is known that the integration problem for $\tilde C^2 (D_d)$ and
\emph{certain} $D_d$ suffers from the curse
of dimensionality. This is true if $D_d$ has a small radius, see \cite{novakHPU19} for the
best known results.
For example, the curse is known if $D_d$ is a $\ell_p^d$ ball and $p \ge 2$.
It is not known for $p$ balls and $p<2$.
3) It is easy to see that the lower bound \eqref{novak123} is true if
$D_d$ is a disjoint union of $n$ Euclidean balls of the same size.
But, of course, this domain is not extremal for the given norm.
4) Assume that the $x_i$ form a grid
and $D_d$ is a cube. Then, for $d=1$, one may take a
quadratic spline $f_1$ as a fooling function and for $d>1$ one can take a
fooling function of the form
$$
f_d(x) = \frac{1}{d} \sum_{i=1}^d f_1 (x^i) ,
$$
where $x= (x^1, \dots , x^d)$.
Hence we obtain exactly the lower bound \eqref{novak123} in this case.
Loosely speaking,
the conjecture says that arbitrary $D_d$ and function values
at arbitrary points $x_1, x_2, \dots , x_n$ do not lead to much better
error bounds than grids for cubes.
\section{iid Information}
So far we mainly studied (deterministic or randomized)
quadrature formulas
\[
A_n(f) = \sum_{i=1}^n a_i f(x_i)
\]
that are, in some way, optimal.
The underlying assumption is that we, when we design an algorithm,
can choose the points $x_i$ and
the weights $a_i$.
This is a standard assumption in numerical analysis and complexity
theory, but it is not always realistic.
In some parts of learning, data science and uncertainty quantification
it is more realistic to assume that the $x_i$ are random and realizations of
iid random variables,
see~\cite{novakBGKP21,novakSB14,novakSC08,novakZh20}.
Then we cannot influence the data stream
\[
(x_1, f(x_1), x_2, f(x_2), x_3, f(x_3), \dots )
\]
and, of course, the construction of ``optimal''
points $x_1, x_2, \dots , x_n$
does not make sense;
we simply have to use
the
information as it comes in.
A natural question to ask is
``What is the power of random information?'',
with two possible answers:
Depending on the assumptions and involved
function spaces,
random information might be almost as good as optimal
information -- or much worse.
In the framework of information-based complexity,
the power of iid information was recently studied and surveyed
in~\cite{novakHKNPUsurvey,novakHKNPU21a,novakHJ19,novakKNR19,novakSo22}.
I summarize
results of the paper
Krieg, Novak and Sonnleitner~\cite{novakKNS21}
where it is proved, formulated as a slogan, that
random information is quite often
(almost) as good as optimal information.
We study integration for functions
from the Sobolev space $W^s_p(D_d)$ and compare optimal
randomized (Monte Carlo) algorithms with algorithms
that can only use iid sample points, uniformly distributed on the domain.
The main result is that we obtain the same optimal rate
of convergence if we restrict ourselves to iid sampling.
Let $D_d \subset {\mathbb{R} }^d$ be open and bounded.
We assume that $D_d $ satisfies an interior cone condition.
We also study the problem of approximating
a function $f$ from the Sobolev space $W_p^s(D_d )$
in the $L_q(D_d )$-norm based on function values $f(x_j)$
on a finite set of sampling points $P=\{x_1,\hdots,x_n\}$.
This makes sense if $s > d/p$,
in which case $W_p^s(D_d)$ is compactly embedded
into the space of bounded continuous functions~$C_b(D_d )$,
and in the case $p=1$ and $d=s$.
There is a vast literature on the error for optimal sampling points.
It is known that the rate of convergence of the
worst case error of optimal
deterministic algorithms is
$
n^{-s/d}
$
for the integration problem
and
$
n^{-s/d+(1/p-1/q)_+}
$
for the approximation problem.
For general domains, we refer to Narcowich, Wendland and
Ward \cite{novakNWW04} as well as \cite{novakNT06}
and \cite{novakKNS21}.
Now we assume that the sampling points
are given to us as realizations
of independent random variables
which are uniformly distributed on the domain.
Then $P \subset D_d $ is a realization of a \novakchange{distribution} $\mathcal{P}$ and
also an algorithm is
a randomized or Monte Carlo algorithm and one may compare
the algorithm with optimal randomized algorithms
based on
the expected error
for inputs from the unit ball of $W_p^s(D_d )$.
The class of all randomized algorithms is very large
since one may construct the random variable $x_{k+1}$
based on the realization
$(x_1, f(x_1), \dots , x_k, f(x_k))$
in a sophisticated and complicated adaptive way.
Nevertheless,
using this weaker notion of error,
one cannot improve the rate
$
n^{-s/d+(1/p-1/q)_+}
$
for the approximation problem,
see Math\'e~\cite{novakMa91} and Heinrich~\cite{novakHe08}
as well as Heinrich~\cite{novakHe09a,novakHe09b} for other Sobolev embeddings.
For the integration problem one now obtains the improved
order
$n^{-s/d-1/2}$ if $p \ge 2$
and
$n^{-s/d-1+1/p}$ if $1 \le p <2$,
see Bakhvalov~\cite{novakBa62} and \cite{novakNo88}.
With algorithms
based on iid sampling
we still obtain the optimal order
of convergence for the approximation problem
unless $p=q=\infty$: In this limiting case
there is a logarithmic loss.
For the integration problem we obtain the optimal order
for all $p$.
We describe the results in more detail.
We assume that $D_d \subset {\mathbb{R} }^d$ is open and bounded
and satisfies an interior cone condition.
That is, there is some $r>0$ and $\theta \in (0,\pi]$
such that for every $x\in D_d $, we find a cone
\[
K(x):=\{x+\lambda y: y\in\mathbb{S}^{d-1}, \langle y,\xi(x)\rangle\ge\cos\theta,\lambda
\in [0,r] \}
\]
with apex $x$, direction $\xi(x)$ in the unit sphere $\mathbb{S}^{d-1}$,
opening angle $\theta$ and radius $r$
such that $K(x) \subset D_d $.
Here $\langle \cdot, \cdot \rangle$ denotes the standard inner product.
For $s\in {\mathbb{N} }$ and $1\le p \le \infty$
such that $s > d/p$ or $p=1$ and $s=d$, we consider the Sobolev space
\[
W_p^s(D_d ) \,:=\, \left\{ f \colon D_d \to {\mathbb{R} } \ \big\vert\ D^\alpha f
\in L_p(D_d ) \text{ for all } \alpha \in {\mathbb{N} }_0^d \text{ with } |\alpha| \le s \right\}
\]
with semi-norm
\[
|f|_{W_p^s(D_d )} \,:=\, \left(\sum_{|\alpha| = s} \Vert D^\alpha f \Vert_{L_p(D_d )}^p\right)^{1/p}
\]
and norm
\[
\Vert f\Vert_{W_p^s(D_d )} :=
\left(\sum_{|\alpha| \le s} \Vert D^\alpha f \Vert_{L_p(D_d )}^p\right)^{1/p},
\]
with the usual modification for $p= \infty$.
Note that $W_p^s(D_d )$ is continuously embedded into $C_b(D_d )$,
the space of bounded continuous functions
with the sup norm, and that the embedding is compact
in the case $s>d/p$, see e.g.\ Maz'ya \cite[Section~1.4]{novakMaz11}
for this fact for domains satisfying an interior cone condition.
We study the problem of $L_q(D_d )$-approximation ($1\le q \le \infty$)
on $W_p^s(D_d )$.
For a random operator $A\colon W_p^s(D_d ) \to L_q(D_d )$
we study the Monte Carlo error
\[
e^{\rm ran} \left(A,W_p^s(D_d ),L_q(D_d )\right)
\,:=\,
\sup_{\Vert f\Vert_{W_p^s(D_d )} \le 1}\, {\mathbb{E}}\ \left\Vert f - A(f) \right\Vert_{L_q(D_d )}.
\]
We
may use the notation ${\mathbb{E}}$ for any mapping if we take the upper integral
for the definition.
Given a distribution $\mathcal{P}$, we put
\[
e^{\rm ran} \left(\mathcal{P},W_p^s(D_d ),L_q(D_d )\right)
\,:=\,
\inf_A\, e\left(A,W_p^s(D_d ),L_q(D_d )\right),
\]
where the infimum is taken over all random operators
of the form $A(f)=\varphi(f|_P)$
with a random mapping $\varphi\colon {\mathbb{R} }^P \to L_q(D_d)$.
Note that $f|_P=\big(f(x)\big)_{x\in P}$ is the restriction
of $f$ to the point set $P$ and we use this as our information.
Here $P$ is a realization of $\mathcal{P}$.
This is the smallest Monte Carlo error that can be achieved with the
\novakchange{distribution} $\mathcal{P}$.
Moreover, we put
\[
e^{\rm ran} \left(n,W_p^s(D_d ),L_q(D_d )\right)
\,:=\,
\inf_{\mathcal{P}}\, e\left(\mathcal{P},W_p^s(D_d ),L_q(D_d )\right),
\]
where the infimum is taken over all random point sets
of cardinality at most~$n$.
This is the smallest Monte Carlo error that can be achieved with
$n$ optimally chosen random sampling points.
It is known, at least for special domains $D_d $, that
\begin{equation}
e^{\rm ran} \left(n,W_p^s(D_d ),L_q(D_d )\right) \,\asymp\, n^{-s/d+(1/p-1/q)_+},
\end{equation}
see Math\'e~\cite{novakMa91} and Remark~\ref{novakrem1}.
\begin{theorem}\label{novakthm:main}\cite{novakKNS21}. \
Let $D_d \subset {\mathbb{R} }^d$ be open and bounded, satisfying an interior cone condition
and let $1 \le p,q \le \infty$ and $s\in {\mathbb{N} }$ such that $s>d/p$ or $p=1$ and $s=d$.
For every $n\in {\mathbb{N} }$, let $\mathcal{P}_n$ be a set of $n$ independent and uniformly distributed points
on~$D_d $.
Then
\[
e^{\rm ran} \left(\mathcal{P}_n,W_p^s(D_d ),L_q(D_d )\right)
\,\asymp\,
\begin{cases}
\, \displaystyle \left(n / \log n \right)^{-s/d}
& \text{if } p=q=\infty,\\
\, n^{-s/d+(1/p-1/q)_+}
& \text{else}. \vphantom{\Big|}
\end{cases}
\]
\end{theorem}
%
This means that independent and uniformly distributed points
are (asymptotically) as good as optimally selected
(deterministic or random) sampling points
in all cases except $p=q=\infty$.
This answers the question for the power of independent uniformly distributed samples
with respect to the Monte Carlo error criterion.
On the other hand, one might also be interested
in a stronger uniform error criterion.
For a random operator $A\colon W^{s}_{p}(D_d)\to L_{q}(D_d)$
the uniform error can be defined by
\[
{e^{\rm unif}}\left(A,W^{s}_{p}(D_d),L_{q}(D_d)\right)
\,=\, {\mathbb{E}} \sup_{\norm{f}_{W^{s}_{p}(D_d)}\le 1}\norm{f-A(f)}_{L_{q}(D_d)}.
\]
Note that the order of the supremum and the expected value is interchanged.
Thus, while a small Monte Carlo error \novakchange{$e^{\rm ran}$} means
that for every individual function
the error is small with high probability,
a small uniform error ${e^{\rm unif}}$ means
that with high probability
the error is small for every function.
As before, given a \novakchange{distribution} $\mathcal{P}$, we put
\[
{e^{\rm unif}}\left(\mathcal{P},W_p^s(D_d ),L_q(D_d )\right)
\,:=\,
\inf_A\, {e^{\rm unif}}\left(A,W_p^s(D_d ),L_q(D_d )\right),
\]
where the infimum is taken over all random operators
of the form $A(f)=\varphi(f|_P)$
with a random mapping $\varphi\colon {\mathbb{R} }^P \to L_q(D_d )$.
Again $P$ is a realization of $\mathcal{P}$.
This is thus the smallest uniform error that can be achieved
with the \novakchange{distribution} $\mathcal{P}$.
Moreover, we put
\[
{e^{\rm unif}}\left(n,W_p^s(D_d ),L_q(D_d )\right)
\,:=\,
\inf_\mathcal{P}\, {e^{\rm unif}}\left(\mathcal{P},W_p^s(D_d ),L_q(D_d )\right),
\]
where the infimum is taken over all random point sets
of cardinality at most $n$.
It is known that
\begin{equation}
{e^{\rm unif}}\left(n,W_p^s(D_d),L_q(D_d )\right)
\,\asymp\, n^{-s/d+(1/p-1/q)_+},
\end{equation}
see again \cite{novakKNS21,novakNWW04,novakNT06}.
For the uniform error of independent uniformly distributed samples,
the following result has been obtained in Krieg and Sonnleitner~\cite{novakKS20}
for bounded convex domains.
See also Ehler, Graef and Oates~\cite{novakEGO19}
and the survey \cite{novakHKNPUsurvey} for earlier results in this direction.
\begin{theorem}\label{novakthm:main-det}\cite{novakKNS21}. \
Let $D_d \subset {\mathbb{R} }^d$ be open and bounded, satisfying an interior cone condition
and let $1 \le p,q \le \infty$ and $s\in {\mathbb{N} }$ such that $s>d/p$ or $p=1$ and $s=d$.
For every $n\in {\mathbb{N} }$, let $\mathcal{P}_n$ be a set of $n$ independent and uniformly
distributed points on~$D_d$.
Then
\[
{e^{\rm unif}}\left(\mathcal{P}_n,W_p^s(D_d ),L_q(D_d )\right)
\,\asymp\,
\begin{cases}
\, \displaystyle \left( n / \log n \right)^{-s/d+1/p-1/q}
& \text{if } q\ge p,\\
\ n^{-s/d}
& \text{if } q<p. \vphantom{\Big|}
\end{cases}
\]
\end{theorem}
%
This means, with respect to the uniform error criterion,
independent and uniformly distributed samples are
as good as optimally chosen sampling points if and only if $q<p$.
Moreover, if we are bound to independent and uniformly distributed samples,
one can achieve the same rate for the uniform error as for the Monte Carlo error
if and only if $q<p$ or $p=q=\infty$.
In all other cases,
the Monte Carlo error criterion provides a speed-up in comparison
to the uniform error criterion.
All the upper bounds of Theorems~\ref{novakthm:main} and \ref{novakthm:main-det}
are achieved by the same algorithm which is based on a polynomial
reproducing map.
The algorithm works for all $p$ and $q$ and up to a given
smoothness $s$,
but should be simplified for an implementation.
Let us now turn to the integration problem
\[
S_d (f) = \int_{D_d } f(x) \, {\rm d} x\,.
\]
We use a similar notation.
For a random operator $A\colon W_p^s(D_d ) \to {\mathbb{R} }$
we define the Monte Carlo error
\[
e^{\rm ran} \left(A,W_p^s(D_d ), S_d \right)
\,:=\,
\sup_{\Vert f\Vert_{W_p^s(D_d )} \le 1}\, {\mathbb{E}}\ \left\vert S_d (f) - A(f) \right\vert.
\]
Given a \novakchange{distribution} $\mathcal{P}$, we put
\[
e^{\rm ran} \left(\mathcal{P},W_p^s(D_d ), S_d \right)
\,:=\,
\inf_A\, e\left(A,W_p^s(D_d ), S_d \right),
\]
where the infimum is taken over all random operators
of the form $A(f)=\varphi(f|_P)$
with a random mapping $\varphi\colon {\mathbb{R} }^P \to {\mathbb{R} } $. Again $P$ is a realization of $\mathcal{P}$.
This is the smallest Monte Carlo error that can be achieved with the
\novakchange{distribution} $\mathcal{P}$.
Moreover, we put
\[
e^{\rm ran} \left(n,W_p^s(D_d ), S_d \right)
\,:=\,
\inf_\mathcal{P}\, e\left(\mathcal{P},W_p^s(D_d ), S_d \right),
\]
where the infimum is taken over all random point sets
of cardinality at most $n$.
This is the smallest Monte Carlo error that can be achieved with
$n$ optimally chosen sampling points.
It is known, at least for special domains $D_d $, that
\[
e^{\rm ran} \left(n,W_p^s(D_d ), S_d \right) \,\asymp\, n^{-s/d+ (1/p-1/2)_+ - 1/2 },
\]
see Bakhvalov~\cite{novakBa59,novakBa62} and \cite{novakNo88}.
Again, we are interested in the smallest Monte Carlo error
which can be achieved with $n$ independent and uniformly distributed
sampling points. As a corollary to Theorem~\ref{novakthm:main} we obtain the following.
\begin{corollary} \cite{novakKNS21}. \
Let $D_d \subset {\mathbb{R} }^d$ be open and bounded, satisfying an interior cone condition
and let $1 \le p \le \infty$ and $s\in {\mathbb{N} }$ such that $s>d/p$
or $p=1$ and $s=d$.
For every $n\in {\mathbb{N} }$, let $\mathcal{P}_n$ be a set of $n$ independent and uniformly distributed points
on~$D_d $.
Then
\[
e^{\rm ran} \left(\mathcal{P}_n,W_p^s(D_d ), S_d \right)
\,\asymp\,
n^{-s/d+ (1/p-1/2)_+ -1/2 } .
\]
\end{corollary}
This means that, with respect to the Monte Carlo error,
independent and uniformly distributed points
are (asymptotically) as good as optimally selected
(deterministic or random) sampling points
in all cases.
This answers the question for the power of independent uniformly distributed samples
for numerical integration
with respect to the Monte Carlo error criterion.
Again, one might also be interested
in a stronger uniform error criterion.
For a random operator $A\colon W^{s}_{p}(D_d )\to {\mathbb{R} } $
the uniform error can be defined by
\[
{e^{\rm unif}}\left(A,W^{s}_{p}(D_d ), S_d \right)
\,=\, {\mathbb{E}} \sup_{\norm{f}_{W^{s}_{p}(D_d )}\le 1} \vert S_d (f) - A(f) \vert .
\]
Again the order of the supremum and the expected value is interchanged.
As before, given a \novakchange{distribution} $\mathcal{P}\subset D_d $, we put
\[
{e^{\rm unif}}\left(\mathcal{P},W_p^s(D_d ), S_d \right)
\,:=\,
\inf_A\, {e^{\rm unif}}\left(A,W_p^s(D_d ), S_d \right),
\]
where the infimum is taken over all random operators
of the form $A(f)=\varphi(f|_P)$
with a random mapping $\varphi\colon {\mathbb{R} }^P \to {\mathbb{R} } $. Here $P$ is a realization of $\mathcal{P}$.
This is thus the smallest uniform error that can be achieved
with the \novakchange{distribution} $\mathcal{P}$.
Moreover, we put
\[
{e^{\rm unif}}\left(n,W_p^s(D_d ), S_d \right)
\,:=\,
\inf_\mathcal{P}\, {e^{\rm unif}}\left(\mathcal{P},W_p^s(D_d ), S_d \right),
\]
where the infimum is taken over all random point sets
of cardinality at most $n$.
For the uniform error of independent uniformly distributed samples,
the following result has been obtained in Krieg and Sonnleitner~\cite{novakKS20}
for bounded convex domains.
\begin{corollary} \cite{novakKNS21}. \
Let $D_d \subset {\mathbb{R} }^d$ be open and bounded, satisfying an interior cone condition
and let $1 \le p \le \infty$ and $s\in {\mathbb{N} }$ such that $s>d/p$
or $p=1$ and $s=d$.
For every $n\in {\mathbb{N} }$, let $\mathcal{P}_n$ be a set of $n$ independent and uniformly distributed points
on~$D_d $.
Then
\[
{e^{\rm unif}}\left(\mathcal{P}_n,W_p^s(D_d ), S_d \right)
\,\asymp\,
\begin{cases}
\, \displaystyle \left( n / \log n \right)^{-s/d}
& \text{if } p=1 ,\\
\, n^{-s/d}
& \text{if } p>1 . \vphantom{\bigg|}
\end{cases}
\]
\end{corollary}
This means, with respect to the uniform error criterion,
independent and uniformly distributed samples are
as good as optimally chosen sampling points
if and only if $p>1$.
We also notice that for independent and uniformly distributed samples,
the Monte Carlo error criterion provides a speed-up in comparison
to the uniform error criterion for all $p$.
\begin{remark} \label{novakrem1}
Different authors study different domains and possibly even
different function spaces $W^s_p(D_d )$.
All the standard definitions of the Sobolev spaces
coincide if $D_d $ is a bounded Lipschitz domain;
if the domain is not Lipschitz then different texts
possibly use different spaces.
If the Sobolev space on $D_d $ is defined by restriction
of the functions from $W^s_p({\mathbb{R} }^d)$ then one
obtains smaller spaces than the spaces defined above.
The stated results stay true for this altered definition
since the lower bounds
already hold for functions with compact support inside~$D_d $.
\end{remark}
\begin{remark}
% Our techniques
% can most likely
% be used to prove similar results
% for more general function spaces of isotropic smoothness
% like Triebel-Lizorkin or Besov spaces as well as Sobolev spaces on
% manifolds.
Another interesting family of spaces are function spaces
of mixed smoothness as surveyed in
D\~ung, Temlyakov and T.~Ullrich~\cite{novakDTU16}.
Here we are still far from understanding
the power of
independent and uniformly distributed sampling points,
and even the power of optimal sampling points
is not known in all cases.
There are recent results in this direction for
the special case of $L_2$-approximation
on the Hilbert space
$ H^{k,{{\rm mix}}}(\mathbb T^d)$
of functions with mixed smoothness $k$ on the $d$-torus.
Namely, it is known that independent and uniformly distributed sampling points
are optimal with respect to the Monte Carlo error~\cite{novakKri19}
and optimal up to a logarithmic factor with respect to the
uniform error~\cite{novakDKU22,novakKU19,novakUll20}.
See also
\cite{novakBSU22,novakCD21,novakKU21,novakNSU20,novakNW12,novakPU21,novakTem20}
for related results.
This also implies that they are optimal up to logarithmic factors
for the problem of integration on
$ H^{k,{{\rm mix}}}(\mathbb T^d)$
with respect to both error criteria.
We do not know whether the logarithmic loss
can be avoided
with uniformly distributed samples,
as it is the case with isotropic smoothness.
This is another Open Problem.
Known optimal sampling points for the integration problem
on
$ H^{k,{{\rm mix}}}(\mathbb T^d)$
have a very particular structure,
see e.g.\ \cite[Section~8.5]{novakDTU16} and the references therein
for the uniform error
and \cite{novakKN17,novakUl17} for the Monte Carlo error.
\end{remark}
\noindent
{\bf Open Problem: } \
In this section we presented results on the optimal
order of convergence if we may only use iid sample points.
There are other ways to compare the power of different algorithms.
In particular, one should study ``iid tractability'',
i.e.,
tractability properties
of algorithms that are based on iid sample points,
see also~\cite{novakLW2021,novakNW12}.
We already know that tractability properties are
not related to optimal order of convergence,
so there is much room for new research.
\medskip
It would be good to know more about the difference between optimal
and iid sample points, depending on the problem, the dimension $d$
and the domain $D_d \subset {\mathbb{R} }^d$.
One possibility is to study the asymptotic constants, such as
\[
\lim_{n \to \infty}
e\left(\mathcal{P}_n,W_p^s(D_d ),L_q(D_d )\right)
\cdot
n^{s/d-(1/p-1/q)_+} .
\]
We conjecture that all these asymptotic constants exist and
possibly they do not depend on the shape of $D_d $,
only on its volume.
The asymptotic constants are known only in very rare cases, see \cite{novakNo20}, though.
Here we present an example from \cite{novakHKNPUsurvey}
that nicely shows the quality of iid samples.
We study $L_1$-approximation for functions from the class
$$
F_d = \{ f: [0,1]^d \to {\mathbb{R} } \mid
|f(x)-f(y)| \le d(x,y) \},
$$
with the maximum metric on the $d$-torus, i.e.,
$$
d(x,y) = \min_{k \in \mathbb{Z}^d} \Vert x+k-y \Vert_\infty .
$$
Then
\[
\lim_{n \to \infty}
{e^{\rm unif}}\left(\mathcal{P}_n,F_d, \novakchange{L_1([0,1]^d)} \right)
\cdot n^{1/d} =
\frac{1}{2} \, \Gamma (1+ 1/d) \approx \frac{1}{2} - \frac{\gamma}{2d}
\]
with the Euler number $\gamma \approx 0.577$, noting that $\Gamma'(1)=-\gamma$.
This compares very well with the error of optimal methods that, for $n=m^d$, equals
$\frac{d}{2d+2} \, n^{-1/d}$.
To achieve the same error $\varepsilon$ in high dimension
with random iid sample points, we have to multiply the number
of optimal sample points by roughly $\exp(1-\gamma)\approx 1.526$;
this factor is quite small and does not
increase with $d$.
\section{Concluding Remarks}
\subsubsection*{Tractability} \
We defined the curse of dimensionality and certain notions
of tractability above, see \eqref{novakcurse}.
Results till 2012 are summarized in the monographs~\cite{novakNW08,novakNW10,novakNW12}
and we presented a few recent results.
Tractability is a very active research area and we mention a few more important papers:
Stability properties are studied by Dick and Goda~\cite{novakDG20};
the randomized error is studied
by Dick, Goda and Suzuki~\cite{novakDGS21} and
by
Kritzer, Kuo, Nuyens and M.~Ullrich~\cite{novakKKNU19};
many (old and new) results can be found in the
recent book by Dick, Kritzer and Pillichshammer~\cite{novakDKP22},
see also the introduction by
Leobacher and Pillichshammer~\cite{novakLP14}.
Dick~\cite{novakDi14},
Goda~\cite{novakGo22,novakGo23}
\novakchange{and Krieg~\cite{novakK23}} found unweighted function spaces
where polynomial tractability holds;
the Sobolev space that goes with the star discrepancy is an earlier
example in this direction, see~\cite{novakHNWW99}.
\novakchange{The curse of dimensionality for the $L_p$-discrepancy holds
for $p$ of the form $p=2\ell/(2\ell-1)$ with
a natural number $\ell$, see~\cite{novakNP23}.}
The Clenshaw Curtis Smolyak algorithm,
introduced and studied
more than twenty years ago
in~\cite{novakBNR99,novakNR96b,novakNR97,novakNR99}
has tractability properties for very smooth functions,
see~\cite{novakHNU14}
and Xu~\cite{novakXu15}.
Traditionally, tractability studies deal with functions of
finite smoothness and then we can at most hope
for polynomial tractability.
An important new topic is weighted spaces and exponential tractability
where, usually, analytic functions are studied and one might expect a much
faster rate of convergence.
I invite the reader to study the papers
Kritzer and Wo\'zniakowski~\cite{novakKW19},
Ebert and Pillichshammer~\cite{novakEP21}
and Krieg, Siedlecki, M. Ullrich and Wo\'zniakowski~\cite{novakKSUW22},
where one can find more references.
Also the papers \cite{novakNUWZ18}
and
Vyb\'\i ral~\cite{novakVy14}
deal with very smooth functions.
\subsubsection*{Oscillatory integrals}
Oscillatory integrals are important for applications and are well studied in numerical analysis.
There exist error bounds but only few complexity results (lower bounds and fitting upper
bounds). One reason might be that some authors do not fix a class $F$ of inputs and
then a complexity analysis cannot be done.
See \cite{novakNUW13,novakNUWZ17,novakZN19}
for complexity results in the univariate case and for further references.
The paper Wu, Graham, Ma and Zhang~\cite{novakWGMZ22}
contains interesting results for the multivariate case and
a Filon-Clenshaw-Curtis-Smolyak rule.
In \cite{novakZN19}
we study optimal quadrature formulas for
weighted integrals and integrands from the Sobolev space $H^1([0,1])$.
A particular case is the computation of Fourier coefficients,
where the oscillatory weight
is given by $\rho_k(x) = \exp(- 2 \pi i k x)$.
Here we study the question whether equidistant nodes are optimal
or not. We prove that this depends on $n$ and $k$:
equidistant nodes are \emph{optimal}
if $n \ge 2.7 |k| +1 $ but might be suboptimal for small $n$.
In particular, the equidistant nodes
$x_j = j/ |k|$ for $j=0, 1, \dots , |k|$, hence $n=|k|+1$,
are the \emph{worst} possible nodes
and do not give any useful information.
\subsubsection*{A final open problem}
Study the
approximation
of the integral
\begin{equation}\label{novakeq27}
S_d(f) = \int_{D_d} f(x) \varrho(x) \, \rd x
\end{equation}
over an open subset $D_d\subset \R^d$
for integrable functions
$f\colon D_d\to\R$.
We are interested in universal
randomized algorithms $(A_n^\omega)$ of the form
\[
A_n^\omega (f) = \sum_{j=1}^n a_j(\omega) f(x_j(\omega)) \, ,
\]
possibly based on iid uniformly distributed $x_j$.
Ideally, we would like to have the same optimal error bounds as
in Theorem~\ref{novakT3}.
It is not clear whether such algorithms exist, we know it
only for $\rho=1$ on $D_d=[0,1]^d$.
In addition, the optimal order of convergence
may be useless if the involved constants are huge.
It might be more practical
to relax the order of the error bounds and to allow
additional log factors.
The papers
\cite{novakKUV21,novakKNS21,novakKU21,novakMN18}
could be useful to reach this goal.
One idea would be to use exactness spaces
\[
V_K = {\rm span} ( \{ x^i \mid \prod_{j=1}^d (i_j +1) \le K \} )
\]
for $K=1,2, \dots $
and proceed similarly as in \cite{novakMN18}.
\medskip
I invite the reader to study also my earlier
survey paper \cite{novakNo16} where one may find more results and references.
% \bigskip
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{acknowledgement}
I thank the \novakchange{referee and the} following colleagues and friends for valuable
remarks:
David Krieg,
Peter Kritzer,
Fritz Pillichshammer,
Mathias Sonnleitner,
Mario Ullrich,
Jan Vyb{\'\i}ral
and Marcin Wnuk.
\end{acknowledgement}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% The bibliography
%
% BibTeX users please use
%\bibliographystyle{spmpsci}
%\bibliography{mybibfile}
% and then copy paste the contents of the .bbl file here for the final version.
%
% E.g.:
\begin{thebibliography}{99.}%
\bibitem{novakBBK21}
% V.~F.\ Babenko, Y.~V.\ Babenko and O,~V.\ Kovalenko.
Babenko, V.F., Babenko, Y.V., Kovalenko, O.V.:
On asymptotically optimal cubatures for multidimensional
Sobolev spaces.
Res. Math. {\bf 29}(2), 15--27, 2021.
\bibitem{novakBa59}
% N.~S.\ Bakhvalov.
Bakhvalov, N.S.:
\newblock On the approximate calculation of multiple integrals.
\newblock {Vestnik MGU, Ser. Math. Mech. Astron. Phys. Chem.}
{\bf 4}, 3--18, 1959, in Russian.
\newblock English translation: \novakJC {\bf 31}, 502--516, 2015.
\bibitem{novakBa61}
% N.~S.\ Bakhvalov.
Bakhvalov, N.S.:
An estimate of the mean remainder term in quadrature formulae.
Zh.\ Vych.\ Mat. {\bf 1}, No.~1, 64--77, 1961.
\bibitem{novakBa62}
% N.~S. Bakhvalov.
Bakhvalov, N.S.:
On the rate of convergence of indeterministic integration
processes within the functional classes $W^{(l)}_p$.
Theor. Probability Appl. {\bf 7}, 227, 1962.
\bibitem{novakBGKP21}
%J. Berner, Ph. Grohs, G. Kutyniok and Ph. Petersen.
Berner, J., Grohs, Ph., Kutyniok, G., Petersen, Ph.:
The modern mathematics of deep learning.
In: Mathematical Aspects of Deep Learning,
pp. 1--111,
Cambridge University Press, 2022,
arXiv 2105:04026.
% \bibitem{novakBa71}
% N.~S.\ Bakhvalov.
% \newblock On the optimality of linear methods for operator
% approximation in convex classes of functions.
% \newblock \novakUSSR 11:244--249, 1971.
\bibitem{novakBSU22}
% F. Bartel, M. Sch\"afer and T. Ullrich,
Bartel, F., Sch\"afer, M., Ullrich, T.:
Constructive subsampling of finite frames with applications
in optimal function recovery.
Appl. Comput. Harmonic Anal. {\bf 65}, 209--248, 2023,
arXiv 2202.12635.
\bibitem{novakBNR99}
% V.~Barthelmann, E.~Novak and K.~Ritter.
Barthelmann, V., Novak, E., Ritter, K.:
High dimensional polynomial interpolation
on sparse grids.
Adv. in Comput. Math. {\bf 12}, 273--288, 1999.
\bibitem{novakBCCGST14}
% L. Brandolini, Chr. Choirat, L. Colzani, G. Gigante, R. Seri and G. Travaglini,
Brandolini, L., Choirat, Chr., Colzani, L., Gigante, G., Seri, R., Travaglini, G.:
Quadrature rules and distribution of points
on manifolds,
Ann. Sc. Norm. Super. Pisa
Cl. Sci. {\bf (5), 13}, 889--923, 2014.
\bibitem{novakCh95}
% E. V. Chernaya,
Chernaya, E.V.:
On the optimization of weighted cubature formulae on certain classes of
continuous functions,
East J. Approx. {\bf 1}, 47--60, 1995.
\bibitem{novakCDHHZ}
% N.~Clancy, Y.~Ding, C.~Hamilton, F.~J.\ Hickernell, and Y.~Zhang.
Clancy, N., Ding, Y., Hamilton, C., Hickernell, F.J., Zhang, Y.:
\newblock The cost of deterministic, adaptive, automatic algorithms:
Cones, not balls.
\newblock \novakJC {\bf 30}, 21--45, 2014.
\bibitem{novakCKS16}
% F. Cobos, T. K\"uhn and W. Sickel,
Cobos, F., K\"uhn, Th., Sickel, W.:
Optimal approximation of multivariate periodic Sobolev functions in the
sup-norm,
J. Functional Analysis {\bf 270}, 4196--4212, 2016.
\bibitem{novakCD21}
% A.~Cohen and M.~Dolbeault.
Cohen, A., Dolbeault, M.:
\newblock Optimal pointwise sampling for $L^2$ approximation.
\novakJC {\bf 68}, 101602, 2022,
arXiv 2105.05545.
\bibitem{novakCW04}
% J.~Creutzig and P.~Wojtaszczyk.
Creutzig, J., Wojtaszczyk, P.:
\newblock Linear vs.\ nonlinear algorithms for linear problems.
\newblock \novakJC {\bf 20}, 807--820, 2004.
\bibitem{novakDi14}
% J.~Dick,
Dick, J.:
Numerical integration of H\"older continuous, absolutely convergent Fourier,
Fourier cosine, and Walsh series,
J. Approx. Th. {\bf 183}, 14--30, 2014.
\bibitem{novakDG20}
% J.~Dick and T.~Goda.
Dick, J., Goda, T.:
Stability of lattice rules and polynomial
lattice rules constructed by the component-by-component algorithm.
J. Comput. Appl. Math. {\bf 382}, 113062, 2021,
arXiv 1912.10651.
\bibitem{novakDGS21}
% J.~Dick, T.~Goda and K.~Suzuki.
Dick, J., Goda, T., Suzuki, K.:
Component-by-component construction
of randomized rank-1 lattice rules
achieving almost the optimal randomized error rate.
Math. Comp. {\bf 91}, 2771--2801, 2022,
arXiv 2109.11694.
\bibitem{novakDKP22}
% J.~Dick, P.~Kritzer and F.~Pillichshammer.
Dick, J., Kritzer, P., Pillichshammer, F.:
\newblock {Lattice Rules.
Numerical Integration, Approximation, and Discrepancy}.
\newblock Springer Series in Computational Mathematics 58,
Springer, 2022.
\bibitem{novakDP08}
% J.~Dick and F.~Pillichshammer.
Dick, J., Pillichshammer, F.:
\newblock {Digital Nets and Sequences: Discrepancy Theory and Quasi-Monte
Carlo Integration}.
\newblock Cambridge University Press, 2010.
\bibitem{novakDP13}
% J.~Dick and F.~Pillichshammer.
Dick, J., Pillichshammer, F.:
\newblock Discrepancy theory and quasi-Monte Carlo integration.
\newblock In W.~Chen, A.~Srivastav, and G.~Travaglini,
editors, {Panorama in Discrepancy Theory},
pages 539--619. Lecture Notes in Mathematics 2107, Springer-Verlag, 2014.
\bibitem{novakDKU22}
% M.~Dolbeault, D.~Krieg and M.~Ullrich,
Dolbeault, M., Krieg, D., Ullrich, M.:
A sharp upper bound for sampling numbers in $L_{2}$,
Appl. Comput. Harmonic Anal. {\bf 63}, 113--134, 2023,
arxiv 2204.12621.
\bibitem{novakDTU16}
% {D.} D\~ung, V.N.~Temlyakov and T.~Ullrich.
D\~ung, D., Temlyakov, V.N., Ullrich, T.:
\newblock {H}yperbolic {C}ross {A}pproximation.
\newblock {Advanced Courses in Mathematics - CRM Barcelona}.
Springer International Publishing, 2018.
\bibitem{novakEKNO21}
% A.~Ebert, P.~Kritzer, D.~Nuyens and O.~Osisiogu.
Ebert, A., Kritzer, P., Nuyens, D., Osisiogu, O.:
Digit-by-digit and component-by-component constructions of
lattice rules for periodic functions with unknown smoothness.
\novakJC {\bf 66}, 101555, 2021.
\bibitem{novakEP21}
% A. Ebert and F. Pillichshammer.
Ebert, A., Pillichshammer, F.:
Tractability of approximation in the weighted Korobov space
in the worst case setting --
a complete picture,
J. Complexity {\bf 67}, 101571, 2021.
\bibitem{novakEGO19}
% M.~Ehler, M.~Gr\"af and C.~J.\ Oates.
Ehler, M., Gr\"af, M., Oates, C.J.:
Optimal Monte Carlo integration on closed manifolds.
Statistics and Computing {\bf 29}, 1203--1214, 2019.
\bibitem{novakGW20}
% M.~Gnewuch and M.~Wnuk.
Gnewuch, M., Wnuk, M.:
Explicit error bounds for randomized Smolyak algorithms and an application
to infinite-dimensional integration.
J. Approx. Th. {\bf 251}, 105342, 2020.
\bibitem{novakGo21}
% M.~Go\'cwin.
Go\'cwin, M.:
On optimal adaptive quadratures for automatic integration.
Bit {\bf 61}, 411--439, 2021.
\bibitem{novakGo22}
% T.~Goda,
Goda, T.:
Polynomial tractability for integration in an unweighted function
space with absolutely convergent Fourier series,
arxiv 2210.06185.
\bibitem{novakGo23}
\novakchange{
% T.~Goda,
Goda, T.:
Strong tractability for multivariate integration in a
subspace of the Wiener algebra,
arxiv 2306.01541.
}
\bibitem{novakGD15}
% T.~Goda and J.~Dick.
Goda, T., Dick, J.:
Construction of interlaced scrambled
polynomial lattice rules of arbitrary high order.
Found. Comp. Math. {\bf 15}, 1245--1278,
arXiv 1301.6441.
\bibitem{novakGL22}
% T.~Goda and P.~L'Ecuyer,
Goda, T., L'Ecuyer, P.:
Construction-free median quasi-Monte Carlo rules for function
spaces with unspecified smoothness and general weights,
SIAM J. Sci. Comp. {\bf 44},
A2765--A2788,
arxiv 2201.09413.
\bibitem{novakGS19}
% T.~Goda and K.~Suzuki.
Goda, T., Suzuki, K.:
Recent advances in higher order quasi-Monte Carlo methods.
In:
Discrepancy Theory,
De Gruyter, 2020, p.~69--102.
\bibitem{novakGS22}
% T.~Goda and K.~Suzuki,
Goda, T., Suzuki, K.:
A universal median quasi-Monte Carlo integration,
arxiv 2209.13186.
\bibitem{novakGSY16}
% T.~Goda, K.~Suzuki and T.~Yoshiki.
Goda, T., Suzuki, K., Yoshiki, T.:
Optimal order quadrature error bounds for infinite-dimensional higher
order digital sequences,
Found. Comp. Math. {\bf 18}, 433--458,
arXiv 1603.08638.
\bibitem{novakGr04}
% P. M. Gruber,
Gruber, P.M.:
Optimum quantization and its applications,
Adv. Math. {\bf 186}, 456--497, 2004.
\bibitem{novakHe08}
% S. Heinrich:
Heinrich, S.:
Randomized approximation of Sobolev embeddings.
In: Keller A., Heinrich S., Niederreiter H. (eds) Monte Carlo and Quasi-Monte Carlo Methods 2006,
pp. 445--459, 2008.
Springer, Berlin, Heidelberg.
\bibitem{novakHe09a}
% S. Heinrich,
Heinrich, S.:
Randomized approximation of Sobolev embeddings II,
\novakJC {\bf 25}, 455--472, 2009.
\bibitem{novakHe09b}
% S. Heinrich,
Heinrich, S.:
Randomized approximation of Sobolev embeddings III,
\novakJC {\bf 25}, 473--507, 2009.
\bibitem{novakHe24}
\novakchange{
Heinrich, S.:
Randomized complexity of mean computation and the
adaption problem.
Work in progress.
}
\bibitem{novakHNWW99}
% S.~Heinrich, E.~Novak, G.~W.\ Wasilkowski, and H.~Wo\'zniakowski.
Heinrich, S., Novak, E., Wasilkowski, G.W., Wo\'zniakowski, H.:
\newblock The inverse of the star-discrepancy depends linearly on the dimension.
\newblock {Acta Arithmetica} {\bf 96}, 279--302, 2001.
\bibitem{novakHKNPUsurvey}
% A.~Hinrichs, D.~Krieg, E.~Novak, J.~Prochno and M.~Ullrich.
Hinrichs, A., Krieg, D., Novak, E., Prochno, J., Ullrich, M.:
\newblock {On the power of random information}.
\newblock In F.~J. Hickernell and P.~Kritzer, editors, {Multivariate
Algorithms and Information-Based Complexity}, pages 43--64. De Gruyter,
Berlin/Boston, 2020.
\bibitem{novakHKNPU21a}
% A.~Hinrichs, D.~Krieg, E.~Novak, J.~Prochno and M.~Ullrich.
Hinrichs, A., Krieg, D., Novak, E., Prochno, J., Ullrich, M.:
Random sections of ellipsoids and the power of random information.
Trans. of the AMS {\bf 374}, 8691--8713, 2021.
\bibitem{novakHKNV21}
% A.~Hinrichs, D.~Krieg, E.~Novak and J.~Vyb\'{\i}ral.
Hinrichs, A., Krieg, D., Novak, E., Vyb\'{\i}ral, J.:
\newblock Lower bounds for the error of quadrature formulas for Hilbert
spaces.
\newblock \novakJC {\bf 65}, 101544, 2021.
\bibitem{novakHKNV22}
% A.~Hinrichs, D.~Krieg, E.~Novak and J.~Vyb\'{\i}ral.
Hinrichs, A., Krieg, D., Novak, E., Vyb\'{\i}ral, J.:
\newblock Lower bounds for integration and recovery in $L_2$.
spaces.
\novakJC {\bf 72}, 101662, 2022.
\bibitem{novakHNU14}
% A.~Hinrichs, E.~Novak, and M.~Ullrich.
Hinrichs, A., Novak, E., Ullrich, M.:
\newblock On weak tractability of the Clenshaw Curtis Smolyak
algorithm.
\newblock \novakJAT {\bf 183}, 31--44, 2014.
\bibitem{novakHNUW12}
% A.~Hinrichs, E.~Novak, M.~Ullrich, and H.~Wo\'zniakowski.
Hinrichs, A., Novak, E., Ullrich, M., Wo\'zniakowski, H.:
\newblock The curse of dimensionality for numerical integration
of smooth functions.
\newblock \novakMC {\bf 83}, 2853--2863, 2014.
% \bibitem{novakHNUW13}
% A.~Hinrichs, E.~Novak, M.~Ullrich, and H.~Wo\'zniakowski.
% \newblock The curse of dimensionality for numerical integration
% of smooth functions II.
% \newblock \novakJC 30:117--143, 2014.
% \bibitem{HNUW14a}
% A. Hinrichs, E. Novak, M. Ullrich and H. Wo\'zniakowski,
% The curse of dimensionality for numerical integration of smooth functions,
% Math. Comp. {\bf 83} (290), 2853--2863, 2014.
\bibitem{novakHNUW14b}
% A. Hinrichs, E. Novak, M. Ullrich and H. Wo\'zniakowski,
Hinrichs, A., Novak, E., Ullrich, M., Wo\'zniakowski, H.:
The curse of dimensionality for numerical integration of smooth functions II,
J. Complexity {\bf 30}, 117--143, 2014.
\bibitem{novakHNUW17}
% A.~Hinrichs, E.~Novak, M.~Ullrich, and H.~Wo\'zniakowski.
Hinrichs, A., Novak, E., Ullrich, M., Wo\'zniakowski, H.:
Product rules are optimal for numerical integration in
classical smoothness spaces.
\novakJC {\bf 38}, 39--49, 2017.
\bibitem{novakHPU19}
% A. Hinrichs, J. Prochno and M. Ullrich,
Hinrichs, A., Prochno, J., Ullrich, M.:
The curse of dimensionality for numerical integration on general domains,
J. Complexity {\bf 50}, 25--42, 2019.
\bibitem{novakHJ19}
% M. Huber and B. Jones:
Huber, M., Jones, B.:
Faster estimates of the mean of bounded random variables,
Math. and Comp. in Simulation {\bf 161}, 93--101, 2019.
\bibitem{novakKUV21}
% L.~K\"ammerer, T.~Ullrich and T.~Volkmer.
K\"ammerer, L., Ullrich, T., Volkmer, T.:
Worst-case recovery guarantees for least squares
approximation using random samples.
Constr. Approx. {\bf 54}, 295--352, 2021.
\bibitem{novakKr18}
% D. Krieg,
Krieg, D.:
Tensor power sequences and the approximation of tensor product operators,
J. Complexity {\bf 44}, 30--51, 2018.
% \bibitem{Kr19}
% D. Krieg,
% Uniform recovery of high-dimensional $C^r$-functions,
% J. Complexity {\bf 50}, 116--126, 2019.
\bibitem{novakKr19a}
% D.~Krieg.
Krieg, D.:
Algorithms and Complexity for some Multivariate Problems.
Dissertation, Jena, 2019.
arXiv 1905.01166.
\bibitem{novakKr19}
% D.~Krieg.
Krieg, D.:
Uniform recovery of high dimensional
$C^r$-functions.
J. Complexity {\bf 50}, 116--126, 2019.
\bibitem{novakKri19}
% D.~Krieg.
Krieg, D.:
\newblock Optimal Monte Carlo methods for $L_2$-approximation.
\newblock {Constr.~Approx.}, {\bf 49}, 385--403, 2019.
% \bibitem{KN17}
% D.~Krieg and E.~Novak.
% \newblock A universal algorithm for multivariate integration.
% \newblock {Found. Comput. Math.}, {\bf 17}, 895--916, 2017.
\bibitem{novakK23}
\novakchange{
% D.~Krieg.
Krieg, D.:
\newblock Tractability of sampling recovery on unweighted
function classes.
arXiv 2304.14169.
}
\bibitem{novakKN17}
% D.~Krieg and E.~Novak.
Krieg, D., Novak, E.:
\newblock A universal algorithm for multivariate integration.
\newblock Found. Comput. Math. {\bf 17}, 895--916, 2017.
\bibitem{novakKNS21}
% D.~Krieg, E.~Novak and M.~Sonnleitner.
Krieg, D., Novak, E., Sonnleitner, M.:
Recovery of Sobolev functions restricted to iid sampling,
Math. Comp. {\bf 91}, 2715--2738, 2022,
arXiv 2108.02055.
\bibitem{novakKSUW22}
% D. Krieg, P. Siedlecki, M. Ullrich and H. Wo\'zniakowski,
Krieg, D., Siedlecki, P., Ullrich, M., Wo\'zniakowski, H.:
Exponential tractability of $L_2$-approximation
with function values,
Adv. Comput. Math. {\bf 49}, 18, 2023,
arXiv 2205.04141.
\bibitem{novakKS20}
% D. Krieg and M. Sonnleitner:
Krieg, D., Sonnleitner, M.:
Random points are optimal for the approximation of Sobolev
functions,
IMA J. Num. Anal. {\bf 00}, 1--26, 2023,
arXiv 2009.11275.
\bibitem{novakKS21}
% D.~Krieg and M.~Sonnleitner,
Krieg, D., Sonnleitner, M.:
Function recovery on manifolds using scattered data,
arxiv 2109.04106,
\bibitem{novakKU19}
% D.~Krieg and M.~Ullrich.
Krieg, D., Ullrich, M.:
\newblock {Function values are enough for $L_2$-approximation}.
Found. of Comput. Math. {\bf 21}, 1141--1151, 2021,
arXiv 1905.02516.
% \bibitem{KU21}
% D.~Krieg and M.~Ullrich.
% \newblock {Function values are enough for $L_2$-approximation: Part~II}.
% \emph{J.~Complexity} 66, 2021, 101569.
\bibitem{novakKU21}
% D.~Krieg and M.~Ullrich.
Krieg, D., Ullrich, M.:
Function values are enough for $L_2$-approximation:
part II.
J. Complexity {\bf 66}, 101569, 2021.
\bibitem{novakKV23}
\novakchange{
% D.~Krieg and J.~Vyb\'{\i}ral.
Krieg, D., Vyb\'{\i}ral, J.:
\newblock New lower bounds for the integration of periodic functions.
J. Fourier Anal. Appl. {\bf 29}, 41, 2023,
arXiv 2302.02639.
}
\bibitem{novakKKNU19}
% P.~Kritzer, F.\,Y.~Kuo, D.~Nuyens and M.~Ullrich.
Kritzer, P., Kuo, F.Y., Nuyens, D., Ullrich, M.:
Lattice rules with random $n$ achieve nearly
the optimal $\mathcal{O} (n^{-\alpha-1/2})$ error independently
of the dimension.
J. Approx. Th. {\bf 240}, 96--113, 2019.
\bibitem{novakKW19}
% P. Kritzer and H. Wo\'zniakowski,
Kritzer, P., Wo\'zniakowski, H.:
Simple characterizations of exponential tractability for
linear multivariate problems,
J. Complexity {\bf 51}, 110--128, 2019.
\bibitem{novakKSU14}
% T. K\"uhn, W. Sickel and T. Ullrich,
K\"uhn, T., Sickel, W., Ullrich, T.:
Approximation numbers of Sobolev embeddings -- Sharp
constants and tractability,
J. Complexity {\bf 30}, 95--116, 2014.
\bibitem{novakKSU15}
% T. K\"uhn, W. Sickel and T. Ullrich,
K\"uhn, T., Sickel, W., Ullrich, T.:
Approximation of mixed order Sobolev functions on the $d$-torus: asymptotics, preasymptotics, and
$d$-dependence,
Constr. Approx. {\bf 42}, 353--398, 2015.
\bibitem{novakKNR19}
% R.~J.~Kunsch, E.~Novak and R.~Rudolf.
Kunsch, R.J., Novak, E., Rudolf, D.:
Solvable integration problems and optimal sample size selection.
J. Complexity {\bf 53}, 40--67, 2019.
\bibitem{novakKR19}
% R.J. Kunsch and D. Rudolf:
Kunsch, R.J., Rudolf, D.:
Optimal confidence for Monte Carlo integration of
smooth functions,
Adv. in Comput. Math. {\bf 45}, 3095--3122, 2019.
% \bibitem{novakKR18}
% R.~J.~Kunsch and R.~Rudolf,
% Optimal confidence for Monte Carlo integration of smooth functions,
% arxiv, 1809.09890,
% 2018.
% \bibitem{KNR19}
% R.J. Kunsch, E. Novak and D. Rudolf:
% Solvable integration problems and optimal sample size selection,
% J. Complexity 53, 40--67, 2019.
\bibitem{novakKNW23}
\novakchange{
% F.Y. Kuo, D. Nuyens and L. Wilkes.
Kuo, F.Y., Nuyens, D., Wilkes, L.:
Random-prime--fixed-vector randomised lattice-based algorithm
for high-dimensional integration.
arXiv 2304.10413.
}
\bibitem{novakLP14}
% G.~Leobacher and F.~Pillichshammer.
Leobacher, G., Pillichshammer, F.:
\newblock {Introduction to Quasi-Monte Carlo Integration and Applications}.
\newblock Springer-Verlag, 2014.
\bibitem{novakLW2021}
% Wanting Lu and Heping Wang,
Lu, Wanting, Wang, Heping:
On the power of standard information
for tractability for $L_2$-approximation
in the average case setting.
J. Complexity {\bf 70}, 101618, 2022.
\bibitem{novakMa91}
% P. Math\'e:
Math\'e, P.:
Random approximation of Sobolev embeddings,
J. Complexity {\bf 7}, 261--281, 1991.
\bibitem{novakMaz11}
Maz'ya, V.G.:
\newblock {Sobolev spaces with applications to elliptic partial differential equations}.
\newblock volume~342 of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin, 2011.
\bibitem{novakMN18}
% G.~Migliorati and F.~Nobile.
Migliorati, G., Nobile, F.:
Stable high-order randomized cubature formulae in arbitrary
dimension,
J. Approximation Th. {\bf 275}, 105706, 2022,
arXiv 1812.07761.
\bibitem{novakMNR09}
M\"uller-Gronbach, Th., Novak, E., Ritter, K.:
\newblock {Monte-Carlo-Algorithmen}.
\newblock Springer-Verlag, 2012.
\bibitem{novakNSU20}
% N.~Nagel, M.~Sch\"afer and T.~Ullrich.
Nagel, N., Sch\"afer, M., Ullrich, T.:
A new upper bound for sampling numbers.
{Found.~Comput.~Math.} {\bf 22}, 445--468, 2020,
arXiv 2010.00327.
\bibitem{novakNWW04}
% F. J. Narcowich, J. D. Ward and H. Wendland,
Narcowich, F.J., Ward, J.D., Wendland, H.:
Sobolev bounds on functions with scattered zeros, with
applications to radial basis function surface fitting.
Math. Comp. {\bf 74}, 743--763, 2004.
\bibitem{novakNo86}
Novak, E.:
Quadrature and widths,
J. Approx. Th. {\bf 47}, 195--202, 1986.
\bibitem{novakNo88}
% E.~Novak.
Novak, E.:
Deterministic and Stochastic Error Bounds in
Numerical Analysis.
\newblock Lecture Notes in Mathematics 1349, Springer-Verlag, 1988.
\bibitem{novakNo96}
Novak, E.:
\newblock On the power of adaption.
\newblock \novakJC {\bf 12}, 199--237, 1996.
\bibitem{novakNo99}
\novakchange{
Novak, E.:
\newblock Intractability results for positive quadrature formulas
and extremal problems for trigonometric polynomials.
\novakJC {\bf 15}, 299--316, 1999.
}
\bibitem{novakNo16}
Novak, E.:
Some results on the complexity of numerical integration.
In: R.~Cools and D.~Nuyens (eds.), Monte Carlo and Quasi-Monte Carlo Methods,
Springer Proceedings in Math. and Statistics 163, 2016.
\bibitem{novakNo20}
Novak, E.:
Algorithms and complexity for functions on general
domains.
J. Complexity {\bf 61} (2020), 101458.
\bibitem{novakNP23}
\novakchange{
Novak, E., Pillichshammer, F.:
The curse of dimensionality for the $L_p$-discrepancy
with finite~$p$.
J. Complexity {\bf 79}, 101760, 2023,
arXiv 2303.01787.
}
\bibitem{novakNR96b}
% E.~Novak and K.~Ritter.
Novak, E., Ritter, K.:
\newblock High dimensional integration of smooth functions over cubes.
\newblock \novakNM {\bf 75}, 79--97, 1996.
\bibitem{novakNR97}
% E.~Novak and K.~Ritter.
Novak, E., Ritter, K.:
\newblock The curse of dimension and a
universal method for numerical integration.
\newblock In G.~N\"urnberger, J.~W.\ Schmidt, and G.~Walz,
editors, {Multivariate Approximation and Splines},
pages 177--188.
ISNM {\bf 125}, Birkh\"auser, 1997.
\bibitem{novakNR99}
% E.~Novak and K.~Ritter.
Novak, E., Ritter, K.:
\newblock Simple cubature formulas with high polynomial exactness.
\newblock \novakCA {\bf 15}, 499--522, 1999.
\bibitem{novakNR96}
Novak, I., Roschmann, I.:
Numerical integration of peak functions.
J. Complexity {\bf 12}, 358--379, 1996.
% \bibitem{novakNR14}
% E.~Novak and D.~Rudolf.
% \newblock Computation of expectations by Markov chain
% Monte Carlo methods.
% \newblock In S.~Dahlke et al., editors,
% \emph{Extraction of quantifiable information from complex systems}.
% Springer-Verlag, 2014.
\bibitem{novakNT06}
Novak, E., Triebel, H.:
Function spaces in Lipschitz domains and optimal rates of convergence
for sampling.
\newblock \novakCA {\bf 23}, 325--350, 2006.
\bibitem{novakNUW13}
% E.~Novak, M.~Ullrich, and H.~Wo\'zniakowski.
Novak, E., Ullrich, M., Wo\'zniakowski, H.:
\newblock Complexity of oscillatory integration
for univariate Sobolev spaces.
\newblock \novakJC {\bf 31}, 15--41, 2015.
\bibitem{novakNUWZ17}
% E.~Novak, M.~Ullrich, H.~Wo\'zniakowski and S.~Zhang.
Novak, E., Ullrich, M., Wo\'zniakowski, H., Zhang, S.:
\newblock Complexity of oscillatory integrals
on the real line.
\newblock Adv. Comput. Math. {\bf 43}, 537--553, 2017.
\bibitem{novakNUWZ18}
% E.~Novak, M.~Ullrich, H.~Wo\'zniakowski and S.~Zhang.
Novak, E., Ullrich, M., Wo\'zniakowski, H., Zhang, S.:
\newblock Reproducing kernels of Sobolev spaces on $\R^d$
and applications to embedding constants and tractability.
\newblock Anal. and Appl. {\bf 16}, 693--715, 2018.
\bibitem{novakNW01}
% E.~Novak and H.~Wo\'zniakowski.
Novak, E., Wo\'zniakowski, H.:
Intractability results for integration and discrepancy.
\novakJC {\bf 17}, 388--441, 2001.
\bibitem{novakNW08}
% E.~Novak and H.~Wo\'zniakowski.
Novak, E., Wo\'zniakowski, H.:
\newblock {Tractability of Multivariate Problems},
Volume I: Linear Information.
\newblock European Mathematical Society, 2008.
\bibitem{novakNW10}
% E.~Novak and H.~Wo\'zniakowski.
Novak, E., Wo\'zniakowski, H.:
\newblock {Tractability of Multivariate Problems},
Volume II: Standard Information for Functionals.
\newblock European Mathematical Society, 2010.
% \bibitem{novakNW11}
% E.~Novak and H.~Wo\'zniakowski.
% \newblock Lower bounds on the complexity for linear functionals
% in the randomized setting.
% \newblock \novakJC 27:1--22, 2011.
\bibitem{novakNW12}
% E.~Novak and H.~Wo\'zniakowski.
Novak, E., Wo\'zniakowski, H.:
\newblock {Tractability of Multivariate Problems},
Volume III: Standard Information for Operators.
\newblock European Mathematical Society, 2012.
\bibitem{novakPO22}
% Z.~Pan and A.~B.~Owen,
Pan, Z., Owen, A.B.:
Super-polynomial accuracy of one dimensional randomized nets using the median-of-means,
Math. Comp. {\bf 92}, 805--837, 2023.
% \bibitem{novakPO22a}
% Z.~Pan and A.~B.~Owen,
% Super-polynomial accuracy of multidimensional randomized nets using the median-of-means,
% arxiv, 2208.05078,
% 2022.
\bibitem{novakPl15}
Plaskota, L.:
Automatic integration using asymptotically optimal
adaptive Simpson quadrature.
Numer. Math. {\bf 131}, 173--198, 2015.
\bibitem{novakPW09}
Plaskota, L., Wasilkowski, G.W.:
The power of adaptive algorithms for functions with singularities.
J. Fixed Point Th. Appl. {\bf 6}, 227--248, 2009.
\bibitem{novakPU21}
Pozharska, K., Ullrich, T.:
A note on sampling recovery of multivariate functions
in the uniform norm.
SIAM J. Num. Anal. {\bf 60}, 1363--1384, 2022,
arXiv 2103.11124.
\bibitem{novakRu12}
Rudolf, D.:
Explicit error bounds for Markov chain Monte Carlo.
{Dissertationes Mathematicae} {\bf 485}, 2012.
% \bibitem{novakSW98}
% I.~H.\ Sloan and H.~Wo\'zniakowski.
% \newblock When are quasi-Monte Carlo algorithms efficient for high
% dimensional integrals?
% \newblock \novakJC 14:1--33, 1998.
% \bibitem{novakSm63}
% S.~A.\ Smolyak.
% \newblock Quadrature and interpolation formulas for tensor products
% of certain classes of functions.
% \newblock \emph{Doklady Akademy Nauk SSSR} 4:240--243, 1963.
\bibitem{novakSB14}
Shalev-Shwartz, Sh., Ben-David, Sh.:
Understanding Machine Learning.
Cambridge University Press, 2014.
\bibitem{novakSo22}
Sonnleitner, M.:
The power of random information for numerical approximation and integration,
Ph.~D. thesis, Passau,
arxiv 2209.07266.
\bibitem{novakSC08}
Steinwart, I., Christmann, A.:
Support Vector Machines.
Springer, New York, 2008.
\bibitem{novakTem20}
Temlyakov, V.:
On optimal recovery in $L_2$.
J.~Complexity {\bf 65}, 2021, 101545.
% \bibitem{novakSu79}
% A.~G.\ Sukharev.
% \newblock Optimal numerical integration formulas for some classes of functions.
% \newblock \emph{Soviet Mathematics Doklady} 20:472--475, 1979.
% \bibitem{novakTem03a}
% V.~N.\ Temlyakov.
% \newblock Cubature formulas, discrepancy, and nonlinear approximation.
% \newblock \novakJC 19:352--391, 2003.
\bibitem{novakTWW88}
Traub, J.F., Wasilkowski, G.W., Wo\'zniakowski, H.:
Information-Based Complexity.
\newblock Academic Press, 1988.
\bibitem{novakTW80}
Traub, J.F., Wo\'zniakowski, H.:
{A General Theory of Optimal Algorithms}.
\newblock Academic Press, 1980.
% \bibitem{Tr07}
% H. Triebel,
% Wavelet para-bases and sampling numbers in function spaces
% on domains,
% \JC {\bf 23}, 468--497, 2007.
\bibitem{novakTr08}
Triebel, H.:
{Function Spaces and Wavelets on Domains},
European Mathematical Society (EMS), Z\"urich, 2008.
\bibitem{novakTr10}
Triebel, H.:
{Bases in Function Spaces, Sampling, Discrepancy,
Numerical Integration},
European Mathematical Society (EMS), Z\"urich, 2010.
% \bibitem{novakTr10}
% H.~Triebel.
% \newblock \emph{Bases in Function Spaces, Sampling, Discrepancy,
% Numerical Integration}.
% \newblock European Mathematical Society, 2010.
\bibitem{novakTr19}
Triebel, H.:
{Function Spaces with Dominating Mixed Smoothness},
European Mathematical Society (EMS), Z\"urich, 2019.
% \bibitem{novakUl17}
% M.~Ullrich.
% A Monte Carlo method for integration of multivariate smooth functions.
% SIAM J. Numer. Math.
% 55, 1188--1200, 2017.
\bibitem{novakUl17}
Ullrich, M.:
A Monte Carlo method for integration of multivariate smooth functions.
SIAM J. Numer. Anal., {\bf 55}, 1188--1200, 2017.
\bibitem{novakUll20}
Ullrich, M.:
{On the worst-case error of least squares algorithms for
$L_2$-approximation with high probability}.
J. Complexity {\bf 60}, 2020.
% \bibitem{novakVa21}
% A.~A.\ Vasil'eva.
% Kolmogorov widths of weighted Sobolev classes on a
% multi-dimensional domain with conditions on the derivatives of order
% $r$ and zero.
% J. Approx. Th. 269, 105602, 2021.
\bibitem{novakVy06}
Vyb\'\i ral, J.:
Sampling numbers and function spaces,
\novakJC {\bf 23}, 773--792, 2007.
\bibitem{novakVy07b}
Vyb\'\i ral, J.:
Widths of embeddings in function spaces,
\novakJC {\bf 22}, 545--570, 2008.
\bibitem{novakVy14}
Vyb\'\i ral, J.:
Weak and quasi-polynomial tractability of approximation of infinitely
differentiable functions,
\novakJC {\bf 30}, 48--55, 2014.
% \bibitem{novakVy13}
% J.~Vyb{\'\i}ral.
% \newblock Weak and quasi-polynomial tractability of approximation
% of infinitely differentiable functions.
% \newblock \novakJC 30:48--55, 2014.
\bibitem{novakVy20}
Vyb{\'\i}ral, J.:
\newblock A variant of Schur's product theorem and
applications.
\newblock Adv. Math. {\bf 368}, 107140, 2020.
% \bibitem{novakWW95}
% G.~W.\ Wasilkowski and H.~Wo\'zniakowski.
% \newblock Explicit cost bounds of algorithms
% for multivariate tensor product problems.
% \newblock \novakJC 11:1--56, 1995.
% \bibitem{novakWW99}
% G.~W.\ Wasilkowski and H.~Wo\'zniakowski.
% \newblock Weighted tensor-product algorithms for linear multivariate problems.
% \newblock \novakJC 15:402--447, 1999.
\bibitem{novakWSH22}
Wenzel, T., Santin, G., Haasdonk, B.:
Stability of convergence rates: Kernel interpolation on non-Lipschitz domains,
arxiv 2203.12532.
\bibitem{novakWeyl}
Weyl, H.:
Das asymptotische Verteilungsgesetz der Eigenwerte linearer partieller
Differentialgleichungen,
Math. Ann. {\bf 71}, 441--479, 1912.
\bibitem{novakWGMZ22}
Wu, Z., Graham, I.G., Ma, D., Zhang, Z.:
A Filon-Clenshaw-Curtis-Smolyak rule for multi-dimensional
oscillatory integrals with application to a
UQ problem for the Helmholtz equation.
arXiv 2208.10078.
\bibitem{novakXu15}
Xu, Guiqiao:
On weak tractability of the Smolyak algorithm
for approximation problems,
J. Approx. Th. {\bf 192}, 347--361, 2015.
\bibitem{novakZh20}
Zhang, J.:
Modern Monte Carlo methods for efficient uncertainty quantification
and propagation: a survey.
WIREs Comput. Stat. {\bf 13}, 5, 1939--5108,
arXiv 2011.00680.
\bibitem{novakZN19}
Zhang, S., Novak, E.:
Optimal quadrature formulas for the Sobolev space $H^1$.
J. Scientific Computing {\bf 78}, 274--289, 2019.
\end{thebibliography}
\end{document}