arXiv2023
Let $f:=(f_n)_{n\in \mathbb{Z}_+}$ and $g:=(g_n)_{n\in \mathbb{Z}_+}$ be two martingales related to the probability space $(\Omega,\mathcal F,\mathbb P)$ equipped with the filtration $(\mathcal F_n)_{n\in \mathbb{Z}_+}.$ Assume that $f$ is in the martingale Hardy space $H_1$ and $g$ is in its dual space, namely the martingale $\rm BMO.$ Then the semi-martingale $f\cdot g:=(f_ng_n)_{n\in \mathbb{Z}_+}$ may be written as the sum $$f\cdot g=G(f, g)+L( f,g).$$ Here $L( f,g):=(L( f,g)_n)_{n\in\mathbb{Z}_+}$ with $L( f,g)_n:=\sum_{k=0}^n(f_k-f_{k-1})(g_k-g_{k-1)})$ for any $n\in\mathbb{Z}_+$, where $f_{-1}:=0=:g_{-1}$. The authors prove that $L( f,g)$ is a process with bounded variation and limit in $L^1,$ while $G(f,g)$ belongs to the martingale Hardy-Orlicz space $H_{\log}$ associated with the Orlicz function $$\Phi(t):=\frac{t}{\log(e+t)},\quad \forall\, t\in[0,\infty).$$ The above bilinear decomposition $L^1+H_{\log}$ is sharp in the sense that, for particular martingales, the space $L^1+H_{\log}$ cannot be replaced by a smaller space having a larger dual. As an application, the authors characterize the largest subspace of $H_1$, denoted by $H^b_1$ with $b\in {\rm BMO}$, such that the commutators $[T, b]$ with classical sublinear operators $T$ are bounded from $H^b_1$ to $L^1$. This endpoint boundedness of commutators allow the authors to give more applications. On the one hand, in the martingale setting, the authors obtain the endpoint estimates of commutators for both martingale transforms and martingale fractional integrals. On the other hand, in harmonic analysis, the authors establish the endpoint estimates of commutators both for the dyadic Hilbert transform beyond doubling measures and for the maximal operator of Ces\`{a}ro means of Walsh--Fourier series.