Operator decomposable measures and stochastic difference equation

Abstract

We consider the following convolution equation or equivalently stochastic difference equation $$\lam _k = \mu _k*\phi (\lam _{k-1}), k \in \Z \eqno (1) $$ for a given bi-sequence $(\mu _k)$ of probability measures on $\R ^d$ and a linear map $\phi $ on $\R ^d$. We study the solutions of equation (1) by realizing the process $(\mu _k)$ as a measure on $(\R ^d)^\Z$ and rewriting the stochastic difference equation as $\lam = \mu *\tau (\lam )$-any such measure $\lam$ on $(\R ^d)^\Z$ is known as $\tau$-decomposable measure with co-factor $\mu$-where $\tau$ is a suitable weighted shift operator on $(\R ^d)^\Z$. This enables one to study the solutions of (1) in the settings of $\tau$-decomposable measures. A solution $(\lam _k)$ of (1) will be called a fundamental solution if any solution of (1) can be written as $\lam _k*\phi ^k(\rho )$ for some probability measure $\rho $ on $\R ^d$. Motivated by the splitting/factorization theorems for operator decomposable measures, we address the question of existence of fundamental solutions when a solution exists and answer affirmatively via a one-one correspondence between fundamental solutions of (1) and strongly $\tau$-decomposable measures on $(\R ^d)^\Z$ with co-factor $\mu$. We also prove that fundamental solutions are extremal solutions and vice versa. We provide a necessary and sufficient condition in terms of a logarithmic moment condition for the existence of a (fundamental) solution when the noise process is stationary and when the noise process has independent $\ell _p$-paths.