A Gaussian upper bound for martingale small-ball probabilities

From MaRDI portal



Abstract: Consider a discrete-time martingale Xt taking values in a Hilbert space mathcalH. We show that if for some Lgeq1, the bounds mathbbEleft[|Xt+1−Xt|mathcalH2midXtight]=1 and |Xt+1−Xt|mathcalHleqL are satisfied for all times tgeq0, then there is a constant c=c(L) such that for 1leqRleqsqrtt, [mathbb{P}(|X_t|_{mathcal H} leq R mid X_0 = x_0) leq c frac{R}{sqrt{t}} e^{-|x_0|_{mathcal H}^2/(6 L^2 t)},.] Following [Lee-Peres, Ann. Probab. 2013], this has applications to diffusive estimates for random walks on vertex-transitive graphs.


The theme of this paper is the Gaussian behavior of martingales taking values on a Hilbert space. More specifically, let us consider such a martingale starting at 0 that has a.s. bounded increments (in terms of magnitude) but the conditional variance is bounded from below by 1. It has been proved that the probability that the magnitude of the \(t\)th element of the martingale sequence being smaller than 1 is at least \(t^{-\alpha}\) (up to a multiplicative constant) for \(0< \alpha < \frac12\). The authors prove that if the conditional variance is given by a deterministic sequence, then this is not the case. There, the scaling is of order \(1/t^{\frac12}\). This result is applied in the particular case of a random walk on an infinite but locally-finite, connected and vertex-transitive graph, where the notion of the distance is the graph distance from the starting vertex.











This page was built for publication: A Gaussian upper bound for martingale small-ball probabilities

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q504261)