On probability and moment inequalities for supermartingales and martingales (Q5907036): Difference between revisions

A new maximal inequality for discrete-time supermartingales is proved. Let \(S_k\), \(k\geq 1\), \(S_0=0\), be a supermartingale on a filtered probability space \((\Omega, {\mathcal F}, ({\mathcal F}_k), {\mathbb P})\). Put \(X_k=S_k-S_{k-1}\), \(\sigma_k^2= {\mathbb E}\{X_k^2\mid {\mathcal F}_{k-1}\}\), \(B_k^2=\sum_1^k \sigma_j^2\), \(\overline S_n =\max_{1\leq k \leq n} S_k\), \(\overline X_n =\max_{1\leq k \leq n} X_k\), \(Q(x) = {\mathbb P}(\overline X_n > x) + {\mathbb P}(B_n > x)\). The main result of the paper is the following inequality: For every \(y>0\) and \(0<\gamma\leq 1\) \[ {\mathbb P}(\overline S_n > y) < c(t,\gamma) \gamma^{-t} \int_0^y Q(\varepsilon_t u) u^{t-1}\,du, \] where either (a) \(t\) is an arbitrary number satisfying \(t\geq\max(e^6,e^4/\gamma^2)\), \(\varepsilon_t = (\ln\,t -2\ln\,\ln\,t)/2t\), and \(c(t,\gamma)=2e^{6\gamma t}/\gamma\), or (b) \(t>0\) is arbitrary, \(\varepsilon_t=\eta/t\), and \(c(t,\gamma)=t e^{3\eta e^{\eta+1}}/\eta e^{\eta+1}\). The author compares this result with some known results of similar type. In particular, he shows that his inequality does not follow from that obtained by \textit{E. Haeusler} [Z. Wahrscheinlichkeitstheorie Verw. Geb. 65, 523--534 (1984; Zbl 0539.60032)]. One of Burkholder's moment inequalities for martingales is easily deduced from the above inequality. Reviewer's remark: The paper contains a number of inaccuracies. In particular, in view of an obvious mistake in (25) we are able to prove the result in case (a) only under a slightly stronger assumption on \(t\), e.g. \(t\geq\max(e^7,e^4/\gamma^2)\). On the other hand, the inequality in (21) may fail if \(s(\alpha,\varepsilon)<1\), which is possible in case (b) if \(t<1\); in order the result to be true it seems necessary to modify \(c(t,\gamma)\) in such a case.