A sharp good-\(\lambda\) inequality with an application to Riesz transforms (Q1113524): Difference between revisions

Let (\({\mathcal F}_ t)_{t\geq 0}\) be the Brownian filtration and \(X_ 1,...,X_ m\) be random variables in \(L^ p\) such that E \(X_ i=0\) for every \(1\leq i\leq m\). Let also \(X_{it}=E(X_ i| {\mathcal F}_ t)\), \(Y_ t=(\sum^{m}_{i=1}X^ 2_{it})^{1/2}\), \(Y_*=\sup \{| Y_ t|\); \(t\geq 0\}\), \(<X_ i>_ t\) be the increasing process of the martingale \((X_{it})_{t\geq 0}\) and \(Z_ t=(\sum_{i\leq m}<X_ i>_ t)^{1/2}.\) The author proves that there exist two universal constants (i.e. independent of p and m) such that \[ (1)\quad (C_ 2/\sqrt{p}) \| Z_{\infty}\|_ p\leq \| Y_{\infty}\|_ p\leq C_ 1\sqrt{p}\| Z_{\infty}\|_ p. \] The result known so far has ``p'' instead of ``\(\sqrt{p}''\) in the left and right sides of (1). The main tool in the proof is the following nice good \(\lambda\)-inequality, generalizing a Burkholder's inequality (in the case \(m=1):\) \[ (2)\quad P(Y_*\geq 2\lambda,Z_{\infty}\leq \epsilon \lambda)\leq C_ 1 \exp (- C_ 2/\epsilon^ 2)P(Y_*\geq \lambda), \] where \(C_ 1\) and \(C_ 2\) are universal constants (here the novelty is the presence of the term exp(.) in the right side of (2), absolutely essential for the rest). Eventually the author investigates the power of (1) when dealing with martingale transforms and their projection on \(R^ n\).