The approximate computation of the definite integral of a function of several variables is one of the basic problems of numerical analysis. The problem is hard because of the so called “curse of dimensionality”. This curse consists of the following: applying an integration rule with N nodes to an univariate function, you will get an integration error, say, ε> 0. Applying
The approximate computation of the definite integral of a function of several variables is one of the basic problems of numerical analysis. The problem is hard because of the so called “curse of dimensionality”. This curse consists of the following: applying an integration rule with N nodes to an univariate function, you will get an integration error, say, ε> 0. Applying
In this dissertation we study the tractability of the information-based complexity n(ε,d) for d-variate function approximation problems. In the deterministic setting for many unweighted problems the curse of dimensionality holds, that means, for some fixed error tolerance ε>0 the complexity n(ε,d) grows exponentially in d. For integration problems one can usually break the curse with the standard Monte Carlo method. For function approximation problems, however, similar effects of randomization have been unknown so far. The thesis contains results on three more or less stand-alone topics. For an extended five page abstract, see the section "Introduction and Results". Chapter 2 is concerned with lower bounds for the Monte Carlo error for general linear problems via Bernstein numbers. This technique is applied to the L∞-approximation of certain classes of C∞-functions, where it turns out that randomization does not affect the tractability classification of the problem. Chapter 3 studies the L∞-approximation of functions from Hilbert spaces with methods that may use arbitrary linear functionals as information. For certain classes of periodic functions from unweighted periodic tensor product spaces, in particular Korobov spaces, we observe the curse of dimensionality in the deterministic setting, while with randomized methods we achieve polynomial tractability. Chapter 4 deals with the L1-approximation of monotone functions via function values. It is known that this problem suffers from the curse in the deterministic setting. An improved lower bound shows that the problem is still intractable in the randomized setting. However, Monte Carlo breaks the curse, in detail, for any fixed error tolerance ε>0 the complexity n(ε,d) grows exponentially in d only.
In this talk I will show the beautiful simplicity of approximating high-dimensional integrals using lattice rules. In current engineering applications it is relatively easy to come up with very high and ultra high dimensional integrals, which might be expectations, that depend on hundreds, thousands, or even an infinite number of variables. As the minimal non-trivial tensor product rule, constructed from a 2-point rule, in 100 dimensions already uses 2^{100} = 1267650600228229401496703205376 ≈ 10^{30} function evaluations—which would take 10^{24} seconds at 1 function evaluation per microsecond and we remember the age of the universe to be only approximately 10^{17} seconds—we all know why to use Monte Carlo or quasi-Monte Carlo methods where we can freely choose the number of function evaluations independent of the number of dimensions. While plain vanilla Monte Carlo methods are very robust and easy to apply, in the case of quasi-Monte Carlo methods a lot more details can pop up. I will first introduce the most elegant—in my biased opinion, which I will defend—of all quasi-Monte Carlo methods: the rank-1 lattice rule. This method is defined by one single integer generating vector together with the total number of points in the rule. Approximating a multivariate integral by this method can be done in one line of Matlab code! As one is often interested in successive approximations I will next introduce a rank-1 lattice sequence. Maybe surprisingly, also here, one simple line of Matlab code is sufficient to calculate a segment of a lattice sequence (with a special property). While for Monte Carlo integration the usual assumption is L2 integrability, leading to an error bound of O(N^{−1/2}). For quasi-Monte Carlo methods we generally work with more involved function space settings, leading to error bounds of the form O(N^{−α}) with α ≥ 1/2. I will show the standard function space settings and dive into the weighted function spaces which are being used in current theory to be able to make the error bounds independent of the number of dimensions. We will gloss over the Korobov space of periodic smooth functions, the unanchored Sobolev space with first mixed derivatives, and the cosine space which is a special space of smooth non-periodic functions. All of which are spaces of so-called mixed dominating smoothness, which in terms of mixed derivatives would mean ∂^{k_1+···+k_s} /(∂x_1^{k_1} · · · ∂x_s^{k_s}) exists for all (k_1, . . . , k_s) ≤ (α, . . . , α). For non-periodic function spaces lattice rules and sequences can only reach O(N^{−1}) convergence. To arrive at higher-order convergence of O(N^{−α}) for non-periodic functions we dive into higher-order Sobolev spaces (of mixed dominating smoothness) in combination with digital nets, but here disguised as a lattice rule, namely polynomial lattice rules. Here, all arithmetic is over a finite field (we will choose the one of order 2) and our lattice rules now take the form of a lattice of polynomials which we map back to points in the unit cube. Higher order can now be obtained by means of the interlacing method from Josef Dick. Although more complicated, again, given the generating matrices of the digital net (as a bunch of integers), and two lines for a magical de Bruijn sequence, we can approximate an integral by such a digital sequence with a single line of (complicated) Matlab code. While Matlab/Octave one-liners might be fun, and somehow instructive, I will also point you to my online resources where code is available to (a) construct good lattice rules, lattice sequences and polynomial lattice rules (for product, order-dependent, product-and-order-dependent aka POD, and smoothness-driven product-and-order dependent aka SPOD) weighted spaces, and (b) generate points given the generating vector or matrices. Armed with this software you will be able to attack your own problems in no time. This talk is based upon the results of collaborations with many colleagues.
We prove lower bounds on the worst-case error of numerical integration in tensor product spaces. The information complexity is the minimal number N of function evaluations that is necessary such that the N -th minimal error is less than a factor ε times the initial error, i.e., the error for N = 0 , where ε belongs to ( 0 , 1 ) . We are interested to which extent the information complexity depends on the number d of variables of the integrands. If the information complexity grows exponentially fast in d , then the integration problem is said to suffer from the curse of dimensionality. Under the assumption of the existence of a worst-case function for the uni-variate problem, we present two methods for providing lower bounds on the information complexity. The first method is based on a suitable decomposition of the worst-case function and can be seen as a generalization of the method of decomposable reproducing kernels. The second method, although only applicable for positive quadrature rules, does not require a suitable decomposition of the worst-case function. Rather, it is based on a spline approximation of the worst-case function and can be used for analytic functions. Several applications of both methods are presented.
Abstract. We present an extension of the convergence analysis for Richardson-extrapolated polynomial lattice rules from [J. Dick, T. Goda, and T. Yoshiki, SIAM J. Numer. Anal., 57 (2019), pp. 44–69] for high-dimensional, numerical integration to classes of integrand functions with quantified smoothness and quasi–Monte Carlo (QMC) integration rules with so-called smoothness-driven, product and order dependent (SPOD) weights. We establish in particular sufficient conditions for the existence of an asymptotic expansion of the QMC integration error with respect to suitable powers of N, the number of QMC integration nodes. We derive a dimension-separated criterion for a fast component-by-component (CBC) construction algorithm for the computation of the QMC generating vector with quadratic scaling with respect to the integration dimension s. We prove that the proposed QMC integration strategies (a) are free from the curse of dimensionality, (b) afford higher-order convergence rates subject to suitable summability conditions on the QMC weights, (c) allow for certain classes of high-dimensional integrand functions a computable, asymptotically exact numerical estimate of the QMC quadrature error, with reliability and efficiency independent of the dimension of the integration domain, and (d) accommodate fast, FFT-based matrix-vector multiplication from [J. Dick, F. Y. Kuo, Q. T. Le Gia, C. Schwab, SIAM J. Sci. Comput., 37 (2015), pp. A1436–A1450] when applied to parametric operator equations. The integration methods are applicable for large classes of many-parametric integrand functions with quantified parametric smoothness. We verify all hypotheses and present numerical examples arising from the Galerkin finite-element discretization of a model linear parametric elliptic PDE illustrating (a)–(d). We verify computationally the scaling of the fast CBC construction algorithm with SPOD QMC weights and examine the extrapolation-based a posteriori numerical estimation of the QMC quadrature error. We find in examples with parameter spaces of dimension s = 10, …, 128 that the extrapolation-based, computable QMC integration error indicator has an efficiency index between 0.9 and 1.1 for a moderate number N of QMC points.