Convolution random sampling in multiply generated shift-invariant spaces of \(L^p(\mathbb{R}^d)\) (Q2218256): Difference between revisions

Denote by \(V^p(\Phi)=\{\sum_{k\in\mathbb{Z}^d} c_k^T \Phi(\cdot-k)$, $\{c_k\}\in (\ell^p(\mathbb{Z}^d))^r\}\) the subspace of \(L^p(\mathbb{R}^d)\), \(1<p<\infty\), generated by the integer shifts of \(\Phi=\{\varphi_1,\dots, \varphi_r\}\). Define \(V^p_N(\Phi)\) to consist of those elements of \(V^p(\Phi)\) such that \(c_k=0\) if \(|k|\geq N\) (note that \(|x|\) denotes the sup-norm here) and \(V^{p,\ast}_N(\Phi)\) the unit-norm elements of \(V^{p}_N(\Phi)\). For \(C_K=[-K,K]^d\), \(\psi\) supported in \(C_K\) and \(\delta\in (0,1)\) fixed, let \(V^p_K(\Phi)=\{f\in V^p(\Phi): \int_{C_K}|f\ast\psi(x)|^p\, dx\geq (1-\delta)\|f\ast\psi\|_p^p \}\). The generators \(\varphi_i\) are assumed to satisfy \(|\varphi_i(x)|\leq C /(1+|x|)^m\), \(m>d/p'\). It is assumed that there are constants \(m_p,m_p', M_p,M_p'\) such that \[ m_p\|\{c_k\}\|_{\ell^p}\leq \left \|\sum c_k^T\Phi(\cdot-k)\right\|_p\leq M_p \|\{c_k\}\|_{\ell^p} \] and similarly \[ m_p'\|\{c_k\}\|_{\ell^p}\leq \left \|\sum c_k^T\Phi\ast\psi(\cdot-k)\right \|_p\leq M_p' \|\{c_k\}\|_{\ell^p}. \] Let \(\rho(x)\) be a probability distribution such that there are constants \(0<C_{\rho,l}\leq \rho(x)\leq C_{\rho,u}\) on \(C_K\). Define random variables \(X_j(f)=|f\ast \psi(x_j)|^p-\int_{C_K}|f\ast\psi(x)|^p \rho(x) \, dx\) where \(X=\{x_j\}\) is drawn from \(\rho\) on \(C_K\). Let \(C_{\psi,1}=\|\psi\|_{L^1(C_K)}\). It is proved that for any \(\gamma>0\), \[ C_{\rho,l}\|f\ast\psi\|_{L^p(C_K)}^p-\gamma \|f\|_{L^p(\mathbb{R}^d)}^p\leq \frac{1}{n} \sum_{j=1}^n |f\ast\psi (x_j)|^p \leq C_{\rho,u}\|f\ast\psi\|_{L^p(C_K)}^p+\gamma \|f\|_{L^p(\mathbb{R}^d)}^p \] holds for all \(f\in V_N^p(\Phi)\) with probability at least \[ P(A,B,\gamma,n)=1-A\exp\Bigl(-B\frac{\gamma^2 n}{12C_{\rho,u}C_{\psi,1}^{2p} +2\gamma C_{\psi,1}^p}\Bigr) \] where \(A,B\) are constants that arise in an application of Bernstein's inequality. The main theorem is a consequence that states that if \(\epsilon_1\), \(\epsilon_2\) and \(\gamma\) are chosen appropriately in terms of all of the constants defined above, then for appropriate \(B_1,B_2\) defined in terms of the \(\epsilon_i\), \(\gamma\) and the same constants, one has \[ B_1\|f\|_{L^p(\mathbb{R}^n)}^p\leq \frac{1}{n}\sum_{j=1}^n |f\ast \psi(x_j)|^p\leq B_2\|f\|_{L^p(\mathbb{R}^n)}^p \] for all \(f\in V_K^p(\Phi)\) with the same probability as above. The remainder of the work addresses stable reconstruction in \(V_N^p(\Phi)\). One defines the \(p\)-condition number \(\kappa(\mathcal{L},p)=\max \frac{\|\mathcal{L}a\|_p}{\|a\|_p}/(\min \frac{\|\mathcal{L}a\|_p}{\|a\|_p})\) of an operator \(\mathcal{L}\) on \(V_N^p(\Phi)\) where the maxima and minima are taken separately over \(a\in(\ell^p [-N,N]^d)^r\). Define a matrix \(U\) of size \(n\times r(2N+1)^d\) with entries \(U_j(k)=[\varphi_i(\cdot-k) \ast\psi](x_j)\), \(j=1,\dots, n\) and \(|k|\leq N\). With constants defined as above and \(\alpha_p\) such that \(\|\sum_{|k|\leq N} c_k^T \Phi\ast\psi(\cdot-k)\|_{L^p(C_K)}\geq \alpha_p\|c\|_{\ell^p}\), it is proved that for any \(0<\gamma\leq C_{\rho,l}(\alpha_p/M_p)^p\) one has \(\kappa(U,p)^p\leq (C_{\rho,u}(M'_p)^p+\gamma M_p^p)/(C_{\rho,l}(\alpha_p)^p+\gamma M_p^p)\) with probability at least \(P(A,B,\gamma,n)\). Technical tools include Bernstein's inequality and an estimate of the covering number of \(V^{p,\ast}_N(\Phi)\) in the \(L^\infty\)-norm.