Sharp estimates for martingale transforms with unbounded transforming sequences (Q6165202)

Given a martingale \(f = (f_n)_{n\geq 0}\) (with values in \(\mathbb{R}\) or, more generally, in a separable Hilbert space) and a predictable sequence \(v = (v_n)_{n\geq 0}\) with values, say, in \(\{-1,1\}\) one can define the martingale transform of \(f\) by \(v\). This is the martingale \(g = (g_n)_{n\geq 0}\) defined the conditions \(g_0 = v_0 f_0\), \(g_n - g_{n-1} = v_n (f_n - f_{n-1})\) for \(n\geq 1\). Burkholder proved that in such a setting for any \(1<p<\infty\) we have \[ \|g\|_p \leq C_p \|f\|_p, \] where \(C_p = \max\{p, p/(p-1) \} - 1\). This constant is the best possible and the estimate holds with the same constant if we assume that the transforming sequence \(v\) has values in \([-1,1]\) (or, even less restrictive, if we simply assume that the martingale \(g\) is differentially subordinate to \(f\)). Weak-type analogues have also been studied in the literature. In the paper under review the authors consider the situation when we have different \(L^p\)-norms on both sides of the above inequality and the transforming sequence \(v\) is \emph{not} necessarily bounded by \(1\). They prove that if \(g\) is a transform of \(f\) by a predictable sequence \(v\) and \(1<p,q,r < \infty\) satisfy \(1/p = 1/q + 1/r\), then \[ \|g\|_p \leq C_{p,q,r} \|f\|_q \|v^*\|_r. \] Here \(v^* = \sup_{n\geq 0} |v_n|\) is the maximal function of \(v\). The optimal constant \(C_{p,q,r}\) is identified: it is equal to \( p - 1\) if \(p>2\), \((q - 1)^{-1} \) if \(p<q<2\), and \(1\) otherwise. A weak-type analogue is also provided (here the expression for the optimal constant is more complicated and includes a value of a certain special function). Actually, the authors prove that the estimates hold with the same constants if we assume that \(f\), \(g\) are martingales, \(v\) is a predictable sequence, and \(g\) is differentially subordinate to \(v\cdot f\), i.e., \(|g_0| \leq |v_0| |f_0|\) and \(|g_n - g_{n-1}| \leq |v_n| |f_{n} - f_{n-1}|\) for \(n\geq 1\). A continuous-time version of this result is established as well. The proof is based on Burkholder's method: the estimates are deduced from the existence of a certain special function which satisfies some concavity and size restrictions. The Reader might also be interested in an upcoming article of Osękowski, where similar inequalities are studied with the \(L^q\)-norm replaced by the \(BMO\)-norm or under the additional assumption that the transforming sequence \(v\) is also a martingale [personal communication, work in preparation as of December 2023].