Scaling limits for sample autocovariance operators of Hilbert space-valued linear processes

TL;DR

This paper investigates the asymptotic behavior of sample autocovariance operators for Hilbert space-valued linear processes with long-range dependence, revealing new limit distributions including Gaussian and non-Gaussian types.

Contribution

It introduces the first weak convergence results to double Wiener-Itô integrals in infinite-dimensional spaces for such processes, expanding the understanding of long-range dependence effects.

Findings

Limiting objects are Gaussian or non-Gaussian Hilbert space-valued variables.

Established conditions for convergence of sample autocovariance operators.

Generalized finite-dimensional Hermite process convergence to infinite dimensions.

Abstract

This article considers linear processes with values in a separable Hilbert space exhibiting long-range dependence. The scaling limits for the sample autocovariance operators at different time lags are investigated in the topology of their respective Hilbert spaces. Distinguishing two different regimes of long-range dependence, the limiting object is either a Hilbert space-valued Gaussian or a Hilbert space-valued non-Gaussian random variable. The latter can be represented as a unitary transformation of double Wiener-It\^o integrals with sample paths in a function space. This work is the first to show weak convergence to such double stochastic integrals with sample paths in infinite dimensions. The result generalizes the well known convergence to a Hermite process in finite dimensions, introducing a new domain of attraction for probability measures in Hilbert spaces. The key technical contributions include the introduction of double Wiener-It\^o integrals with values in a function space and with dependent integrators, as well as establishing sufficient conditions for their existence as limits of sample autocovariance operators.