On the structure and interpolation properties of quasi shift-invariant spaces (Q681708)

The authors consider two problems about quasi shift-invariant subspaces \(V(\psi,\mathcal{X})\) of \(L^2(\mathbb{R})\) generated by a discrete sequence \(\mathcal{X}=\{x_j\}_{j\in\mathbb{Z}}\) of shifts \(\psi(\cdot-x_j)\) of a single generator, whose elements have the form \(f(x)=\sum c_j\, \psi(x-x_j) \) with coefficients \(\{c_j\}\in \ell^2(\mathbb{Z})\), as analogues of principal shift invariant spaces (\(\mathcal{X}=\mathbb{Z}\)). Problem 1 essentially asks for conditions on \(\mathcal{X}\) and \(\psi\) such that \(V(\psi,\mathcal{X})\) is a closed subspace of \(L^2(\mathbb{R})\) and \(\psi(\cdot-x_j)\) forms a basis for \(V(\psi,\mathcal{X})\). Problem 2 essentially asks when any \(f\in V(\psi,\mathcal{X})\) can be recovered from its interpolants in a family of quasi shift-invariant spaces \(\{V(\phi_\alpha,\mathcal{Y})\}_\alpha\). Two standing assumptions are made on \(\psi\): (A1) is that \(\psi\in L^2(\mathbb{R})\) is positive definite, and (A2) is that the Fourier transform \(\widehat\psi\) belongs to the Wiener amalgam space \(W(L^\infty,\ell^1)\) and \(\widehat\psi\) is bounded below on \(\mathbb{T}\), thus \(C_\psi=\sum_{j\in\mathbb{Z}}\|\widehat\psi(\xi+2\pi j)\|_\infty/\sigma_\psi <\infty\), where \(\sigma_\psi=\inf_{\xi\in\mathbb{T}}|\widehat\psi(\xi)|\). The sequence \(\mathcal{X}\) is complete interpolating (CIS) for the Paley-Wiener space \(\text{PW}_\pi\) if it contains the frequencies of a Riesz basis of exponentials for \(L^2[-\pi,\pi]\). If \(\psi\) satisfies (A1) and (A2) and \(\mathcal{X}\) is a CIS, then for each \(f\in V(\psi,\mathcal{X})\), \(\|\hat{f}\|_{L^2(\mathbb{T})}\) is equivalent to \(\|f\|_{L^2(\mathbb{R})}\) \ (Prop. 2). In the particular case \(\text{supp}\,\widehat\psi=\mathbb{T}\), one has \(V(\psi,\mathcal{X})=\text{PW}_\pi\) while, if \(\mathbb{T}\nsubseteq \text{supp}\,\hat\psi\), there is a second CIS \(\mathcal{Y}\) such that \(V(\psi,\mathcal{X})\neq V(\psi,\mathcal{Y})\) (Prop 3). Theorem 2 establishes that, if \(\phi,\psi\) satisfy (A1), (A2) and \(\mathcal{X}\) and \(\mathcal{Y}\) are CISs, then, for each \(f\in V(\psi,\mathcal{X})\), there is a unique sequence \(\{a_j\}\) such that the function \(\mathcal{I}_\phi^{\mathcal{Y}} f(x)=\sum_j a_j \phi(x-y_j)\) interpolates \(f(y_j)\): \(\mathcal{I}_\phi^{\mathcal{Y}} f(y_j)=f(y_j)\), and \(\mathcal{I}_\phi^{\mathcal{Y}} f\in L^2\cap C_0\). Theorem 3 then establishes that, if \(f=\sum_j c_j \psi(\cdot-x_j)\in V(\psi,\mathcal{X})\), then \(\|f\|_{L^2}\) is equivalent to \(\|\{c_j\}\|_{\ell^2}\). It is also established that \(\|\mathcal{I}_\phi^{\mathcal{Y}} f\|_{L^2}\) is equivalent to \(\|\widehat{f}\|_{L^2(\mathbb{T})}\). Section 5 provides recovery criteria. One assumes a family \(\phi_\alpha\) indexed by \(\alpha\in A\), each satisfying (A1) and (A2) with \(\sup_\alpha C_{\phi_\alpha}<\infty\) in (A2). Let \(A_{\mathcal{X}}^k(\sum c_j e^{-ix_j\cdot})(t)=\sum_{j} c_j e^{-i x_j (t+2\pi k)}\), \(t\in \mathbb{T}\), \(T_{\phi,k} g=\hat\phi(\cdot+2\pi k) g/\sigma_\psi\) and \(M_\phi g=(\sigma_\phi/\widehat\phi)g\). One assumes that \(\lim_{\alpha\to\infty} \sum_{k\neq 0}\|T_{\psi,k}A_{\mathcal{X}}^k (M_\psi-M_{\phi_\alpha})g\|_{L^2(\mathbb{T})}=0\) for all \(g\) and \(\lim_{\alpha\to\infty} \sum_{k\neq 0}\|(T_{\phi_\alpha,k}A_{\mathcal{Y}}^k-T_{\psi,k}A_{\mathcal{X}}^k M_{\phi_\alpha})g\|_{L^2(\mathbb{T})}=0\). Under these assumptions, it is proved (Thm. 4) that, for every \(f\in V(\psi,\mathcal{X})\), \(\lim_{\alpha\to\infty}\|f-\mathcal{I}_{\phi_\alpha}^{\mathcal{Y}}\|_{L^2(\mathbb{R})}=0\), where \(\mathcal{I}_{\phi_\alpha}^{\mathcal{Y}}\in V(\phi_\alpha,\mathcal{Y})\) is the unique \(\mathcal{Y}\)-interpolant of \(f\). It is shown that the convergence is uniform on \(V(\psi,\mathcal{X})\). Several specific examples of recovery are discussed, including regular interpolators (\(\text{supp}\widehat\psi=\mathbb{T}\)); convolution examples for bandlimited \(\psi\) (e.g., Poisson kernel, Gaussian kernel; multiquadrics; approximate identities), and nonbandlimited convolvers. Section 7 finishes with a discussion of new consequences in the case of cardinal functions in which \(\mathcal{X}=\mathbb{Z}\) and \(V(\phi)=V(L_\phi)\), where \(\sqrt{2\pi}\,\widehat{L}_\phi= \widehat{\phi}/(\sum \widehat\phi(\xi+2\pi j))\).