Unique wavelet sign retrieval from samples without bandlimiting (Q6572988)

This paper deals with phase retrieval in the following sense: Given a real-valued signal, unique up to a global sign, i.e., \(f\in L^2(\mathbb{R})/\sim\) with \(f\sim g\) iff \(f=\pm g\), can it be reconstructed from the modulus of a suitable wavelet transform, i.e., from \(|\mathcal{W_\phi}f(b,a)|, \; (b,a)\in \mathbb{R}\times \mathbb{R}^+\)? \N\NThe paper gives a positive answer, provided \(\phi\) is chosen to be a real-valued Poisson wavelet. Apart from this continuous procedure, the paper gives also a positive answer, if the wavelet transform comes from a multiwavelet procedure with discretely sampled scalings and translations, i.e. from the transform \(f\mapsto |\mathcal{W_{\phi_k}}f(b_m,a_n)|\). Here \(k=1,2,3\) counts specially designed wavelets such that \(\phi_1\), \(\phi_2\), \(\phi_3\) are suitably chosen linear combinations of the Poisson wavelet and its Hilbert transform, \((b_m,a_n)\) denotes a carefully constructed hyperbolic lattice in translations and scalings. In contrast to earlier results, these reconstruction statements require no band limitation, neither with respect to the wavelet(s) nor to the signal. \N\NThe presented analysis belongs to phase retrieval problems, important for example in applications like blind source separation or audio synthesis. For a corresponding paper cf. [\textit{I. Waldspurger}, IEEE Trans. Inf. Theory 63, No. 5, 2993--3009 (2017; Zbl 1368.94024)]. The analytic continuation arguments given in the present paper are similar to the corresponding arguments in the IEEE paper.