Lattice rules for nonperiodic smooth integrands

# Lattice rules for nonperiodic smooth integrands

Josef Dick, Dirk Nuyens, Friedrich Pillichshammer
July 19, 2019
###### Abstract

The aim of this paper is to show that one can achieve convergence rates of for (and for arbitrarily small) for nonperiodic -smooth cosine series using lattice rules without random shifting. The smoothness of the functions can be measured by the decay rate of the cosine coefficients. For a specific choice of the parameters the cosine series space coincides with the unanchored Sobolev space of smoothness .

We study the embeddings of various reproducing kernel Hilbert spaces and numerical integration in the cosine series function space and show that by applying the so-called tent transformation to a lattice rule one can achieve the (almost) optimal rate of convergence of the integration error. The same holds true for symmetrized lattice rules for the tensor product of the direct sum of the Korobov space and cosine series space, but with a stronger dependence on the dimension in this case.

## 1 Introduction

Quasi-Monte Carlo (QMC) rules are equal weight quadrature rules

 1NN−1∑n=0f(xn)

which can be used to approximate integrals of the form

 ∫[0,1]sf(x)dx,

see [10, 23, 28] for more information. In QMC rules, the quadrature points are chosen according to some deterministic algorithm. One can show that the convergence rate of the integration error depends on the smoothness of the integrand and some property of the quadrature points.

There are several deterministic construction methods for the quadrature points. One such method yields so-called digital nets. These yield a convergence of the integration error of for functions of bounded variation [10, 23]. Higher order digital nets, using an interlacing factor of , yield a convergence rate of order for integrands with square integrable partial mixed derivatives of order in each variable [8, 9].

An alternative to digital nets are lattice rules [23, 28]. In this case, for a given positive integer one chooses a vector called the generating vector and defines the quadrature points by

 {ngN}for 0≤n

Here, for a real number , the braces denote the fractional part, i.e., modulo . For vectors, the fractional part is taken component wise. It is well known that there are generating vectors for lattice rules for which the integration error converges with order () for smoothness , but with the restriction that and its partial derivatives up to order in each variable have to be periodic [19, Theorem 18, p. 120] and [29, 30]. Fast computer search algorithms for such vectors are known from [25, 26]. Hence, in order to be able to benefit from the fast rate of convergence, one needs to apply a transformation which makes the integrand (and its partial derivatives) periodic. This can cause some problems though and is not always recommended [19, 22]. Since arbitrarily high rates of convergence can be obtained using digital nets for nonperiodic functions, the question arises whether this is also possible for lattice rules. Until now, lattice rules achieve a convergence rate of at most () for nonperiodic integrands (via estimates of the star-discrepancy [23]). If one applies the so-called tent transformation and a random shift, this rate of convergence can be improved to for any for the worst-case error in an unanchored Sobolev space of smoothness , see [15].

In this paper we present quadrature rules which achieve a convergence rate of order , , for nonperiodic functions with smoothness . The way we measure smoothness in this paper is slightly different from the setting used in, for instance, higher order digital nets [8, 9] though. We consider functions which can be represented by a cosine series. Note that every continuous function can be represented by a cosine series (see [18, Theorem 1] for this basis over ). This follows from the fact that the functions

 1,√2cos(πx),√2cos(π2x),√2cos(π3x),…

are -orthogonal and complete.

As mentioned above, the “smoothness” of the cosine series in our context is measured by the rate of decay of the cosine coefficients. To illustrate this, consider a one-dimensional function given by its cosine series

 f(x) = ˜fcos(0)+∞∑k=1˜fcos(k)√2cos(πkx) = ˜f(0)+∞∑k=1˜fcos(2k)√2cos(π2kx)+∞∑k=1˜fcos(2k−1)√2cos(π(2k−1)x) =: c+fper(x)+fnonper(x),

where

 ˜fcos(k):=⎧⎪ ⎪ ⎪ ⎪⎨⎪ ⎪ ⎪ ⎪⎩∫10f(x)dx for k=0,∫10f(x)√2cos(πkx)dx for k∈N. (1)

The sum over the even frequencies is a -periodic function over . If the coefficients decay with order then is -times differentiable in the classical sense. However, this does not apply to the sum over the odd coefficients. For instance, the cosine series for is given by

 −4π2∞∑k=1k odd1k2cos(πkx) (2)

and hence the odd coefficients converge with order only, although is infinitely times differentiable.

Below we introduce a reproducing kernel Hilbert space based on cosine series, with the smoothness measured by the decay rate of the cosine coefficients. Although the smoothness of a cosine series measured by the differentiability of the series can be larger than the decay rate of the cosine coefficients suggests, the opposite can not happen. That is, we show that the reproducing kernel Hilbert space based on cosine series is embedded in the unanchored Sobolev space with the same value of the smoothness parameter. The case of smoothness provides an exception, since there the cosine series space and the unanchored Sobolev space coincide. Various reproducing kernel Hilbert spaces are introduced in Section 2 and their embeddings are studied in Section 3.

In this paper we present two methods which allow us to achieve a higher convergence rate for smoother nonperiodic functions using lattice rules, namely:

1. application of the tent transformation to the integration nodes;

2. symmetrization of the integration nodes.

The tent transformation, ,

 ϕ(x):=1−|2x−1|

is a Lebesgue measure preserving function. The idea of using this transformation in conjunction with a random shift for integration based on lattice rules comes from Hickernell [15] and was also used in [6] for digital nets. In contrast to these works, here we do not rely on a random element in our quadrature rules. We show that a tent transformed lattice rule achieves an integration error of order , for any , for functions belonging to a certain reproducing kernel Hilbert space of cosine series with smoothness parameter . This result follows by showing that the worst-case error in the cosine series space for a tent-transformed lattice rule is the same as the worst-case error in a Korobov space of smooth periodic functions using lattice rules. Thus all the results for integration in Korobov spaces using lattice rules [7, 12, 20, 25, 26] also apply for integration in the cosine series space using tent-transformed lattice rules. In particular for smoothness , this yields deterministic point sets for numerical integration in unanchored Sobolev spaces with the same tractability properties as for numerical integration in the Korobov space.

Furthermore, we also use symmetrized lattice rules. We show that these rules achieve the optimal order of convergence for integration of sums of products of cosine series and Fourier series. We apply the transformation to each possible set of coordinates separately, so that if we start off with points we get points (see Section 4.2). This symmetrization approach is also mentioned in [19, 28, 34] and is one of the symmetry groups applied in the construction of cubature formulae, see, e.g., [4, 13]. We prove that a lattice rule symmetrized this way achieves an integration error of order , , for functions belonging to a certain reproducing kernel Hilbert space of cosine series and Fourier series with smoothness parameter . The advantage of symmetrized lattice rules is that functions of the form , where is a nonnegative integer, are integrated exactly and hence only the smoothness of the periodic part determines the convergence rate. Note that the decay rate of the cosine series coefficients of the periodic part and of the Fourier series coefficients part coincides with the classical smoothness. Thus the problem with functions where the smoothness in terms of differentiability differs from the rate of decay of the cosine series is overcome using symmetrization. For instance, the function is integrated exactly using symmetrization. However, a disadvantage of the symmetrization compared to the tent transformation is that the number of function evaluations grows exponential in the dimension and therefore symmetrization is only useful in smaller dimensions.

For both methods, the rates of convergence we obtain are essentially optimal by an adaption of the lower bound of Bakhvalov [3], which is presented in Section 4.3.

In the next section we introduce four reproducing kernel Hilbert spaces, the unanchored Sobolev space, the Korobov space, the cosine series space and the sum of the cosine and Korobov space. Since the unanchored Sobolev space and the Korobov space are frequently studied in the literature, we study the relations among these four spaces in Section 3 to put our results into context. It is shown that the Korobov space and the cosine series space differ, but both are embedded in the sum of the cosine series and Korobov space, which is itself embedded in the unanchored Sobolev space. In Section 4 we study numerical integration in the cosine series space using tent-transformed lattice rules and numerical integration in the sum of the cosine series and Korobov space using symmetrized lattice rules. Numerical results are presented in Section 5 and a conclusion is presented in Section 6.

We write for the set of integers, for the set of positive integers and for the set of nonnegative integers. We also write . Furthermore, for we write .

## 2 Reproducing kernel Hilbert spaces

In this section we introduce several reproducing kernel Hilbert spaces [2]. For a reproducing kernel we denote by the corresponding reproducing kernel Hilbert space with inner product and corresponding norm . For any we have and we have the reproducing property

 f(y)=⟨f,K(⋅,y)⟩for all y∈[0,1] and f∈H(K).

Further, the function is symmetric in its arguments and positive semi-definite.

For higher dimensions we consider tensor product spaces. The reproducing kernel is in this case given by

 Ks(x,y)=s∏j=1K(xj,yj),

where and . Again, the corresponding reproducing kernel Hilbert space is denoted by , the corresponding inner product is denoted by and the corresponding norm by .

For further information on reproducing kernel Hilbert spaces we refer to [2] (or [10, Chapter 2] for reproducing kernel Hilbert spaces in the context of numerical integration).

### 2.1 The unanchored Sobolev space

The unanchored Sobolev space is a reproducing kernel Hilbert space with reproducing kernel given by

 Ksob1,γ(x,y):=1+γB1(x)B1(y)+γB2(|x−y|)2,

where and are Bernoulli polynomials and is a real number. The inner product in this space is given by

 ⟨f,g⟩Ksob1,γ=∫10f(x)dx∫10g(x)dx+1γ∫10f′(x)g′(x)dx.

In higher dimensions we consider the tensor product space . The reproducing kernel is in this case given by

 Ksob1,γ,s(x,y):=s∏j=1Ksob1,γj(xj,yj),

where , and .

The reproducing kernel Hilbert space can be generalized to higher order smoothness. For , consider the reproducing kernel given by

 Ksobα,γ(x,y):=1+γα∑τ=1Bτ(x)Bτ(y)(τ!)2−(−1)αγB2α(|x−y|)(2α)!,

where is the Bernoulli polynomial of order . The inner product for this space is given by

 ⟨f,g⟩Ksobα,γ=∫10f(x)dx∫10g(x)dx+1γα−1∑τ=1∫10f(τ)(x)dx∫10g(τ)(x)dx+1γ∫10f(α)(x)g(α)(x)dx.

To obtain reproducing kernel Hilbert spaces for the domain , we consider again the tensor product of the one-dimensional spaces. This space has the reproducing kernel

 Ksobα,γ,s(x,y):=s∏j=1Ksobα,γj(xj,yj).

Quasi-Monte Carlo rules in which yield the optimal rate of convergence of order for any where studied in [9].

### 2.2 The Korobov space

For , and we define

 rα,γ(h):={1if h=0,γ|h|−2αif h≠0. (3)

For and , we set

 rα,γ,s(h):=s∏j=1rα,γj(hj).

The Korobov space is a reproducing kernel Hilbert space of Fourier series. The reproducing kernel for this space is given by

 Kkorα,γ(x,y):=∑h∈Zrα,γ(h)e2πih(x−y),

and in higher dimensions by

 Kkorα,γ,s(x,y):=s∏j=1Kkorα,γj(xj,yj)=∑h∈Zsrα,γ,s(h)e2πih⋅(x−y),

where the “” denotes the usual inner product in . Let the Fourier coefficient for a function be given by

 ˆf(h):=∫[0,1]sf(x)e−2πih⋅xdx.

Then the inner product in the reproducing kernel Hilbert space is given by

 ⟨f,g⟩Kkorα,γ,s=∑h∈Zsˆf(h)¯¯¯¯¯¯¯¯¯¯¯ˆg(h)r−1α,γ,s(h).

The corresponding norm is defined by .

### 2.3 The half-period cosine space

The half-period cosine space is a reproducing kernel Hilbert space of (half-period) cosine series with reproducing kernel

 Kcosα,γ(x,y):=1+∞∑k=1rα,γ(k)√2cos(πkx)√2cos(πky),

where is defined as in (3) and where and . The inner product is given by

 ⟨f,g⟩Kcosα,γ=∞∑k=0˜fcos(k)˜gcos(k)r−1α,γ(k),

where and are the cosine coefficients of and , respectively, as defined in (1).

We can generalize the reproducing kernel Hilbert space to the domain by setting

 Kcosα,γ,s(x,y):=s∏j=1Kcosα,γj(xj,yj),

where and . The inner product is then given by

 ⟨f,g⟩Kcosα,γ,s=∑k∈Ns0˜fcos(k)˜gcos(k)r−1α,γ,s(k),

where the multi-dimensional cosine coefficients for a function are given by

 ˜fcos(k):=∫[0,1]sf(x)2|k|0/2s∏j=1cos(πkjxj)dx,

where for we define to be the number of nonzero components in . The corresponding norm is defined by , in particular we have

 ∥f∥2Kcosα,γ,s =∑k∈Ns0|˜fcos(k)|2rα,γ,s(k)=∑h∈Zs2−|h|0|˜fcos(|h|)|2rα,γ,s(h),

where .

### 2.4 The sum of the Korobov space and the half-period cosine space

In this section we introduce the kernel

 Kkor+cosα,γ(x,y) := 12(Kkorα,γ(x,y)+Kcosα,γ(x,y)) := 12∑h∈Zrα,γ(h)e2πih(x−y)+12+∞∑k=1rα,γ(k)cos(πkx)cos(πky) := 1+γ∞∑k=1k−2α(cos(2πk(x−y))+cos(πkx)cos(πky)).

The space resulting from the sum of kernels is studied in [2, Part I, Section 6]. The norm in the reproducing kernel Hilbert space is then defined by

 ∥f∥2Kkor+cosα,γ=minf=fkor+fcos2(∥fkor∥2Kkorα,γ+∥fcos∥2Kcosα,γ),

where the minimum is taken over all functions and such that .

For dimensions we define the reproducing kernel by

 Kkor+cosα,γ,s(x,y):=s∏j=112(Kkorα,γj(xj,yj)+Kcosα,γj(xj,yj)).

Thus the space is the tensor product . For we define

 Kkor+cosα,γ,s,u(x,y):=2−s∏j∈uKkorα,γj(xj,yj)∏j∈[s]∖uKcosα,γj(xj,yj),

where as usual an empty product is considered to be one. This is a reproducing kernel for the space

with the inner product

where

 ˜fu,kor+cos(hu,k[s]∖u):=∫[0,1]sf(x)(√2)|k[s]∖u|0∏j∈ue−2πihjxj∏j∈[s]∖ucos(πkjxj)dx.

Clearly we have that

 Kkor+cosα,γ,s(x,y)=∑u⊆[s]Kkor+cosα,γ,s,u(x,y).

## 3 Embeddings

In this section we investigate the relationships between the spaces introduced above.

For two reproducing kernel Hilbert spaces and we say that is continuously embedded in if

 H(K1)⊆H(K2)

and if

 ∥f∥K2≤C∥f∥K1for all f∈H(K1)

for some constant independent of . We write

 H(K1)↪H(K2)

in this case.

On the other hand it is possible that is not a subset of , i.e., there is a function in which is not in . In this case we write .

If and we write

 H(K1)⇋H(K2).

### 3.1 The Korobov space and the unanchored Sobolev space

It is well known, see, e.g., [24, Appendix A], that the Korobov space is continuously embedded in the unanchored Sobolev space

 H(Kkorα,γ,s)↪H(Ksobα,γ,s).

Conversely, as is also well known, for instance the function , is in for all , but not in , since is not periodic. Thus

 H(Ksobα,γ,s)⊄H(Kkorα,γ,s).

We note that for , , we have that where denotes the rescaled sequence .

### 3.2 The half-period cosine space and the unanchored Sobolev space

We now consider the half-period cosine space and the unanchored Sobolev space. For there is a peculiarity.

###### Lemma 1.

We have

 Ksob1,γ(x,y)=1+γπ2∞∑k=11k22cos(kπx)cos(kπy)=Kcos1,γπ−2(x,y).

For completeness we include a short proof.

###### Proof.

In the following we calculate the cosine coefficients of . It is easy to check that we have for . Further we have .

We have

 B2(|x−y|)= B2({x−y})=12π2∑k∈Z∖{0}e2πik(x−y)k2 = ∞∑k=1cos(2πkx)cos(2πky)π2k2+∞∑k=1sin(2πkx)sin(2πky)π2k2.

Using (2) we obtain

 (x−12)(y−12)+B2(|x−y|)2= ∞∑k,l=116π4(2k−1)2(2l−1)2cos(π(2k−1)x)cos(π(2l−1)y) +∞∑k=1cos(2πkx)cos(2πky)2π2k2+∞∑k=1sin(2πkx)sin(2πky)2π2k2.

This immediately implies that

 ∫10∫10Ksob1,γ√2cos(πmx)cos(πny)dxdy=0

if is even and is odd, or is odd and is even, or are even with . If for even , we obtain

 ∫10∫10Ksob1,γ(x,y)√2cos(πkx)√2cos(πky)dxdy=γπ2k2.

Now let be odd. We have and therefore

 ∫10∫10Ksob1,γ(x,y)√2cos(πmx)√2cos(πny)dxdy= 8π4m2n2+16π4∞∑k=11(4k2−m2)(4k2−n2).

For we have and further . Thus we obtain

Note that the cosine series for converges absolutely. Since the function is continuous, the cosine series converges to the function pointwise. This completes the proof. ∎

The above lemma and Mercer’s theorem also yield the eigenfunctions of the operator

 T(g)(y)=∫10Ksob1,γ(x,y)g(x)dx.

These are and the corresponding eigenvalues are .

###### Remark 1.

An analoguous result for a slightly different reproducing kernel Hilbert space was established in [33]. For this space one obtains the same set of eigenfunctions. For the anchored Sobolev space the eigenfunctions are slightly different and have been found in [32].

We thus find

 H(Ksob1,γ)=H(Kcos1,γπ−2)and∥f∥Ksob1,γ=∥f∥Kcos1,γπ−2for all f∈H(Ksob1,γ).

The same also applies for the higher dimensional tensor product space

 H(Ksob1,γ,s)=H(Kcos1,γπ−2,s)and∥f∥Ksob1,γ,s=∥f∥Kcos1,γπ−2,sfor all f∈H(Ksob1,γ,s),

where denotes the sequence . Thus

Now consider . The function belongs to for all . On the other hand we have

 x=12−4π2∞∑k=1k oddcos(πkx)k2.

Hence the function is not in for and therefore

 H(Ksobα,γ)⊄H(Kcosα,γ)for α∈N,α≥2.

Conversely, let , . Let be given by

 f(x)=˜fcos(0)+∞∑k=1˜fcos(k)√2cos(πkx)

with

 ∥f∥2Kcosα,γ=|˜fcos(0)|2+1γ∞∑k=1|˜fcos(k)|2|k|2α<∞.

Then for we have

 f(τ)(x)=∞∑k=1˜fcos(k)(−1)⌈τ/2⌉(kπ)τ√2ϕτ(πkx),

where for even and for odd and where for a real number , denotes the smallest integer bigger or equal to . Thus

 1γ∣∣∣∫10f(τ)(x)dx∣∣∣2≤1γ∫10|f(τ)(x)|2dx=π2τγ∞∑k=1|˜fcos(k)|2k2τ≤π2τ∥f∥2Kcosα,γ.

Thus

 ∥f∥Ksobα,γ≤(α∑τ=0π2τ)1/2∥f∥Kcosα,γ=(π2(α+1)−1π2−1)1/2∥f∥Kcosα,γ.

Thus we have

 H(Kcosα,γ)↪H(Ksobα,γ).

This result can be generalized to the tensor product space, thus

 H(Kcosα,γ,s)↪H(Ksobα,γ,s).

### 3.3 The half-period cosine space and the Korobov space

For the embedding results for the half-period cosine space and the Korobov space follow from the previous two subsections.

Let now . Let

 f(x)=sin(2πx).

Then for all . On the other hand we have for

 ˜fcos(k)=∫10sin(2πx)√2cos(πkx)dx=⎧⎨⎩4√2π(4−k2)if k% is odd,0otherwise.

Thus

 ∥f∥2Kcosα,γ=1γ∞∑\lx@stackrelk=1k odd|˜fcos(k)|2k2α=1γ32π2∞∑k=1(2k−1)2α((2k−1)2−4)2.

Thus we have for . Thus

 H(Kkorα,γ)⊄H(Kcosα,γ)for α≥3/2.

Conversely, let now

 f(x)=cos(πx).

Then for all . On the other hand we have for

 ˆfkor(h)=∫10cos(πx)e−2πihxdx=4ihπ(1−4h2).

For any we have

 ∥f∥2Kkorα,γ=1γ16π2∑h∈Z∖{0}|h|2αh2(4h2−1)2=∞,

which implies that