Multi-scale analysis and random walks boundaries (Q1801552)

The recently developed theory of wavelets and multiscale analysis can be related to methods used in probability theory of \(\mu\) boundaries of groups. The main example is given by the affine group acting on \(\mathbb{R}\). The purpose of this article is to present a general method for the approximation of functions defined on the \(\mu\) boundary of a group which is analogous to the analysis in wavelets bases. Let \(X_ n\) be the right random walk of law \(\mu\) on a topological metrisable group acting on a metric space \(E\) and \(\nu\) be an invariant probability measure on \(E\) that is such that \(\mu*\nu=\mu\). Then a martingale argument shows that almost surely the sequence \(X_ n(\omega)\nu\) weakly converges to some probability measure \(P(\omega,.)\) on \(E\). Let \(\tilde P(f)\) be the positive contraction from \(L^ q(\Omega,\mathbb{P})\) to \(L^ q(E,\nu)\) defined by \(\mathbb{E}(f(Pg))=\int_ Eg(x)\tilde Pf(x)d\nu(x)\) for any \(g\in L^ p(E,\nu)\), then the approximation operator \(T\) is given by \(T^ n(f)=\tilde P(Hf(X_ n))\) where the convolution operator \(H\) is defined by \(Hf(g)=\int_ Ef(g\cdot x)d\nu(x)\). In general one has that \(T^ n(f)\) converges to \(\tilde PPf\) in \(L^ p(E,\nu)\) for any \(f\in L^ p(E,\nu)\), \(1\leq p\leq\infty\), and pointwise whenever \(p>1\). In the case of a \(\mu\) boundary then \(\tilde PP=I\), then one gets the desired result. When the measure \(g\nu\) is absolutely continuous with respect to \(\nu\) for \(\mu\) almost all \(g\), then \(T\) is a kernel operator and one recovers the wavelet approximation in the particular case of the affine group \(G\) endowed with a discrete measure \(\mu\) (but there are a lot of other interesting examples).