Greedy algorithms and best \(m\)-term approximation with respect to biorthogonal systems (Q5943952)

Let \(X\) be a real separable Banach space, \(\Phi\subset X\) a minimal, normalized, dense system in \(X\), furthermore, let \(\Psi\subset X'\) be a biorthogonal system to the \(\Phi\), \(\widehat f\) be the coefficient sequence for any \(f\in X\) with respect to the \(\Psi\). The notation \(c\geq c'\) means that \(|c_k|\geq|c_k'|\), \(k\in\mathbb N\). By the definition \(m\)-term polynomials are polynomials with \(\leq m\) non-zero coefficients, i.e. \[ \Sigma_m=\bigcup_{\#\Lambda\leq m}V_{\Lambda}, \] where \(V_{\Lambda}=\{f\in X:\widehat f_k=0\), \(k\notin\Lambda\}\), \(\Lambda\subset\mathbb N\); \(\mathbf 1_{\Lambda}:=\sum_{k\in\Lambda}\varphi_k\in V_{\Lambda}.\) The author is interested in the comparison between the best \(m\)-term approximation \[ \sigma_m(f)_X=\inf_{g\in\Sigma_m}\|f-g\|_X \] and the worst case behaviour of the greedy algorithm \(G_m\), i.e. \(G_mf=\sum_{k\in\Lambda_m}c_k\varphi_k,\) where \(\Lambda_m\) is a collection of indices of the \(m\) greatest coefficients (in absolute value) among \(\widehat f_k\), \(k\in\mathbb N.\) More precisely the asymptotic value of \[ \delta{X,\Phi}(m):=\sup_{f\in X}\frac{\|f-G_mf\|_X}{\sigma_m(f)_X} \] is investigated. The main result is the following: \[ \delta_{X,\Phi}(m)\leq 1+2A_{X,\Phi}(m)+B_{X,\Phi}(m), \text{ and }\delta_{X,\Phi}(m)\sim\max\bigl(A_{X,\Phi}(m),B_{X,\Phi}(m)\bigr), \] where \[ A_{X,\Phi}(m)=\sup_{g_{\Lambda}\in\Sigma_m}\frac{ \|g_{\Lambda}\|_{X,\Phi;2}}{\|g_{\Lambda}\|_{X,\Phi;1}},\quad B_{X,\Phi}(m)=\sup_{\Lambda'\cap\Lambda''=\emptyset, 1\leq\#\Lambda'=\#\Lambda''\leq m}\frac{\|{\mathbf 1}_{\Lambda'}\|_{X,\Phi;2}}{\|{\mathbf 1}_{\Lambda''}\|_{X,\Phi;1}}, \] \[ \|f\|_{X,\Phi;1}=\inf_{g\in X;\widehat g\geq\widehat f}\|g\|_X,\quad \|f\|_{X,\Phi;2}=\sup_{g\in X; \widehat g\leq\widehat f}\|g\|_X. \] Applications to the Haar system in different spaces (\(L_{\infty}\), \(L_1\), BMO, the dyadic BMO) are given.