Numerical computation of the Rosenblatt distribution and applications

Nikolai N. Leonenko a and Andrey Pepelyshev a · 2025

Read Numerical computation of the Rosenblatt distribution and applications by Nikolai N. Leonenko a and Andrey Pepelyshev a (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-23337.txt.

Introduction

Leonenko and Pepelyshev present a paper from twenty twenty-five on how to compute the Rosenblatt distribution, a chance law tied to long-memory processes. This law arises as a limit for nonlinear functions of stationary Gaussian processes. Such processes stay linked over long spans. The team find new formulas for the law's characteristic function. They also pin down the eigenvalues of a linked integral operator with high accuracy. A fast algorithm then works out the law's density from those values.

Simulations with small samples show the law emerging in practice. Several functionals of long-memory Gaussian processes serve as test cases. The work makes a once-abstract limit law computable for users. Uses in the statistics of dependent data are the target.

Catalog record

CREATORNikolai N. Leonenko a and Andrey Pepelyshev a
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-23337

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Numerical computation of the Rosenblatt distribution and applications Nikolai N. Leonenko a and Andrey Pepelyshev a Abstract The Rosenblatt distribution plays a key role in the limit theorems for non-linear functionals of stationary Gaussian processes with long-range dependence. We derive new expressions for the characteristic function of the Rosenblatt distribution. Also we present a novel accurate approximation of all eigenvalues of the Riesz integral operator associated with the correlation function of the Gaussian process and propose an efficient algorithm for computation of the density of the Rosenblatt distribution. We perform Monte-Carlo simulation for small sample sizes to demonstrate the appearance of the Rosenblatt distribution for several functionals of stationary Gaussian processes with long-range dependence. 1 Introduction The phenomenon of long-range dependence (also called long memory) is one of exciting area of research in the probability theory for last few decades due to non-standard normalizations and non-Gaussian limiting distributions of nonlinear functionals [ 23 ] . The Rosenblatt distribution serves a significant role in the study of this phenomenon which occurs in economics, finance, hydrology, turbulence, cosmology and physics [ 12 , 6 ] . The Rosenblatt distribution was introduced in [ 25 ] and later it was investigated by many researchers, see [ 27 , 28 , 10 ] among many others. The Rosenblatt distribution appears as the limiting distribution of some popular functionals, see [ 12 , 7 , 24 ] and references therein. The known analytical form of the characteristic function of the Rosenblatt distribution contains a series which converges in a neighbourhood of zero, see [ 25 , 1 , 31 ] . It turns out that the Rosenblatt distribution is infinitely divisible [ 31 ] and self-decomposable, moreover, it belongs to the Thorin class [ 20 , 16 , 18 ] . The density of the Rosenblatt distribution exists and bounded, however, the closed analytical form of the density is unknown, see [ 31 , 20 , 16 , 18 ] . The Edgeworth expansion was used to approximate the density of the Rosenblatt distribution [ 30 ] . The numerical evaluation of the density of the Rosenblatt distribution was proposed in [ 31 ] . The present paper is organized as follows. In Section 2 we review the known facts on the Rosenblatt distribution. In Section 3 we provide the novel analytic forms of the characteristic function of the Rosenblatt distribution on the entire line and a compelling viewpoint on the structure of the Rosenblatt distribution. In Section 4 we propose a accurate approximation of all eigenvalues of the Riesz integral operator, which allows us to perform fast numerical computation of the density of the Rosenblatt distribution. In Section 5 we propose a time-efficient algorithm for simulation of the stationary Gaussian sequences with long-range dependence with the power correlation function and the Mittag-Lefler correlation function. Furthermore, we demonstrate small sample properties of four popular functionals for analysis of sequences of correlated random variables including their non-standard normalization and convergence to the Rosenblatt distribution. Specifically, we consider estimation of the mean in Section 5.3 , the correlation function in Section 5.4 , the sojourn functionals in Section 5.5 and path roughness in Section 5.6 . All these problems are essentials of the statistical analysis of stationary Gaussian sequences with long-range dependence. 2 Formal statement We consider the Rosenblatt distribution with shape parameter a , zero mean, unit variance and the characteristic function \phi(z)=\exp\left(\frac{1}{2}\sum_{k=2}(2iz)^{k}\frac{c_{a,k}\sigma_{a}^{k}}{k}\right),~a\in[0,1/2],~z\in S_{0}\subset\mathbb{R}, (1) where S_{0} is a small neighbourhood of zero, \sigma_{a}=\sqrt{(1-2a)(1-a)/2} and c_{a,k}=\int_{0}^{1}\cdots\int_{0}^{1}|x_{1}-x_{2}|^{-a}|x_{2}-x_{3}|^{-a}\cdots|x_{k-1}-x_{k}|^{-a}|x_{k}-x_{1}|^{-a}dx_{1}\cdots dx_{k}. In the rest of this section we describe the Rosenblatt distribution following [ 31 ] . The random variable V from the Rosenblatt distribution can be given as V=\sum_{n=1}^{\infty}\lambda_{a,n}(\varepsilon_{n}^{2}-1), (2) where \varepsilon_{n} are i.i.d. random variables from the standard normal distribution and \lambda_{a,1},\lambda_{a,2},\ldots are eigenvalues of the Riesz integral operator \tilde{K}_{a}:L^{2}(0,1)\to L^{2}(0,1) defined as (\tilde{K}_{a}f)(x)=\sigma_{a}\int_{0}^{1}|x-u|^{-a}f(u)du. The representation ( 2 ) means that V is a specific instance of second-order Wiener chaos [ 22 ] . These eigenvalues satisfy the relation \sum_{n=1}^{\infty}\lambda_{a,n}^{k}=\begin{cases}\infty,&k=1,\\ c_{a,k}\sigma_{a}^{k},&k=2,3,\ldots.\end{cases} Moreover, we have \displaystyle\sum_{n=1}^{\infty}\lambda_{a,n}^{2} \displaystyle= \displaystyle 1/2, \displaystyle\sum_{n=1}^{\infty}\lambda_{a,n}^{3} \displaystyle= \displaystyle\frac{2\sigma_{a}^{3}}{(1-a)(2-3a)}\beta(1-a,1-a), where \beta(u,v)=\int_{0}^{1}x^{u-1}(1-x)^{v-1}dx is the beta function. The Rosenblatt distribution is infinitely divisible with the Lévy density m(x)=\frac{1}{2x}\sum_{n=1}^{\infty}\exp\left(-\frac{x}{2\lambda_{a,n}}\right),~x>0, that shows self-decomposability of the Rosenblatt distribution [ 31 ] . 3 Main results The Laplace transform of V defined in ( 2 ) is given by \phi_{LT}(s)=\mathbb{E}(e^{-sV})=\exp\left(-\sum_{n=1}^{\infty}\left(\frac{1}{2}\ln(1+2\lambda_{a,n}s)-\lambda_{a,n}s\right)\right),~s>-\frac{1}{2\lambda_{a,1}}, see [ 1 ] . We note that the formal expansion of the sum in the above expression gives -\frac{1}{2}\sum_{n=1}^{\infty}\ln(1+2\lambda_{a,n}s)+s\sum_{n=1}^{\infty}\lambda_{a,n}, where the latter summand equals infinity. Using the classical Taylor expansion of the logarithm \ln(x)=\sum_{k=1}^{\infty}\frac{(-1)^{k-1}}{k}(x-1)^{k},~x\in(0,2), which does not converge for all positive x , we obtain \phi_{LT}(s)=\exp\left(\frac{1}{2}\sum_{k=2}^{\infty}\frac{(-2s)^{k}}{k}\left(\sum_{n=1}^{\infty}\lambda_{a,n}^{k}\right)\right),~|s|<\frac{1}{2\lambda_{a,1}}, (3) which is related to ( 1 ) as \phi(z)=\phi_{LT}(-iz) , z\in\mathbb{R} . We note that the expressions of the characteristic function ( 1 ) and the Laplace transform ( 3 ) cannot be used for numerical calculation because they are defined at a neighbourhood of zero. In next theorem, we derive new expressions using three other expansions of the logarithm which converge on the full domain (0,\infty) and one integral representation. In other words, we construct analytic continuations for the expressions ( 1 ) and ( 3 ). Theorem 3.1 The characteristic function \phi(z) of the Rosenblatt distribution is given by \phi(z)=\phi_{LT}(-iz) , where the Laplace transform \phi_{LT}(s) with s>-\frac{1}{2\lambda_{a,1}} admits (i) the representation using the domain-scaled expansion \displaystyle\ln(\phi_{LT}(s))=\sum_{n=1}^{\infty}\left(\frac{(\lambda_{a,n}s)^{2}}{1+\lambda_{a,n}s}-\sum_{k=2}^{\infty}\frac{1}{2k-1}\left(\frac{\lambda_{a,n}s}{1+\lambda_{a,n}s}\right)^{2k-1}\right), (ii) the representation using the Ramanujan expansion \displaystyle\ln(\phi_{LT}(s))=\sum_{n=1}^{\infty}\lambda_{a,n}s\ln(1+2\lambda_{a,n}s)\sum_{k=1}^{\infty}\frac{1}{2^{k}(1+(1+2\lambda_{a,n}s)^{2^{-k}})}, (4) (iii) the representation using the Ramanujan-Bradley expansion \displaystyle\ln(\phi_{LT}(s))=\frac{1}{2}\sum_{n=1}^{\infty}\sum_{k=1}^{\infty}2^{k-1}\left((1+2\lambda_{a,n}s)^{2^{-k}}-1\right)^{2}, (5) (iv) the representation using the integral form \ln(\phi_{LT}(s))=\int_{0}^{s}\sum_{n=1}^{\infty}\frac{2\lambda_{a,n}^{2}u}{2\lambda_{a,n}u+1}du. The proof of Theorem 3.1 is deferred to the online supplement. We note that the evaluation of the characteristic function of the Rosenblatt distribution requires computation of the infinite sums with a special care because numerical computation of \sum_{i=1}^{\infty}\lambda_{a,n}^{2} is problematic for a\in(0.35,0.5) due to very slow convergence of the series \sum_{i=1}^{\infty}n^{2a-2} . To overcome difficulties of numerical computation, we propose to find smallest integer M_{\epsilon} such that \sum_{n=M_{\epsilon}+1}^{\infty}\lambda_{a,n}^{k}<\epsilon for k=3,4,\ldots , where \epsilon is a small positive number, for example, \epsilon=0.0001 . Then we can write the random variable V defined in ( 2 ) in the form V=W+Z , where W=\sum_{n=1}^{M_{\epsilon}}\lambda_{a,n}(\varepsilon_{n}^{2}-1),~Z=\sum_{n=M_{\epsilon}+1}^{\infty}\lambda_{a,n}(\varepsilon_{n}^{2}-1). Due to the choice of M_{\epsilon} , the random variable Z has approximately the normal distribution with mean zero and variance \sigma^{2}_{\epsilon}=1-2\sum_{n=1}^{M_{\epsilon}}\lambda_{a,n}^{2}=2\sum_{n=M_{\epsilon}+1}^{\infty}\lambda_{a,n}^{2}. Indeed, the characteristic function of Z has the form \displaystyle\phi_{Z}(z) \displaystyle= \displaystyle\exp\left(-\sum_{n=M_{\epsilon}+1}^{\infty}\left(\frac{1}{2}\ln(1-2\lambda_{a,n}iz)+\lambda_{a,n}iz\right)\right) \displaystyle\stackrel{{\scriptstyle|z|<\frac{1}{2\lambda_{a,M_{\epsilon}}}}}{{=}} \displaystyle\exp\left(\frac{1}{2}\sum_{k=2}^{\infty}\frac{(2iz)^{k}}{k}\left(\sum_{n=M_{\epsilon}+1}^{\infty}\lambda_{a,n}^{k}\right)\right) \displaystyle\approx \displaystyle\exp\left(-z^{2}\left(\sum_{n=M_{\epsilon}+1}^{\infty}\lambda_{a,n}^{2}\right)+\frac{4}{3}i^{3}z^{3}\epsilon\right) \displaystyle\approx \displaystyle\exp\left(-z^{2}\sigma^{2}_{\epsilon}/2\right). Moreover, we have \max_{|z|\leq\frac{1}{2\lambda_{a,M_{\epsilon}}}}\Big|\phi_{Z}(z)-\exp\left(-z^{2}\sigma^{2}_{\epsilon}/2\right)\Big|\to 0\text{~as~}\epsilon\to 0. This argument provides a clear constructive view on the Rosenblatt distribution. Specifically, the random variable from the Rosenblatt distribution can be simulated as V_{\epsilon}=\sigma_{\epsilon}\varepsilon_{0}+\sum_{n=1}^{M_{\epsilon}}\lambda_{a,n}(\varepsilon_{n}^{2}-1), where \varepsilon_{n} are i.i.d. random variables from the standard normal distribution. The density of the Rosenblatt distribution can be computed by the inverse Fourier transform of the characteristic function \phi_{\epsilon}(z)=\exp\left(-z^{2}\sigma^{2}_{\epsilon}/2-\sum_{n=1}^{M_{\epsilon}}\left(\frac{1}{2}\ln(1-2\lambda_{a,n}iz)+\lambda_{a,n}iz\right)\right). Applying the Stein method for the random variable V_{\epsilon} and taking the limit as \epsilon\to 0 , we obtain the following characterizing identity, which follows from [ 2 , Ch. 3] . Theorem 3.2 Let X be a random variable with zero mean. The Stein equation \mathbb{E}Xf(X)=\mathbb{E}\int_{0}^{\infty}\Big(f(X+x)-f(X)\Big)\frac{1}{2}\sum_{n=1}^{\infty}\exp\left(-\frac{x}{2\lambda_{a,n}}\right)dx (6) holds for all bounded Lipschitz function f(\cdot) if and only if X has the Rosenblatt distribution with parameter a . Let us use the Stein equation ( 6 ) with f(x)=x , we refer to [ 2 ,

Read the complete text