Uncertainty principles and optimally sparse wavelet transforms (Q2300761)

This paper defines a localization concept in the reference frame of generalized wavelet transforms, based on self-adjoint ``observables'', denoted with \(\check{T}\), and related one-parameter-group-representations \(U(\alpha)\). The notion observable is borrowed from quantum mechanics, in this paper it is connected to meaningful operators measuring localizations of a signal. By ``meaningful'' we address the fact that transforming the signal with \(U(\alpha)\) leads to a shift of the spectrum (the possible measurement values) of \(\check{T}\). To give an example, consider a signal \(s(x) \in L^2\) and the localization operator \(\check{T} s(x)=x\cdot s(x)\). Signal translations are represented by the one-parameter-group-representation \(U(\alpha)s(x)=s(x-\alpha)\). One easily verifies that \[U(\alpha)\check{T} U^\ast(\alpha)=\check{T}-\alpha Id\tag{1} .\] Thus indeed \(\check{T}\) is meaningful for measuring signal localization in time \(x\), since it transforms in the same way as the signal, when (time-)shifts are applied. Equation (1) is related to canonical commutation relations or (in unitary form) Weyl-relations in quantum mechanics; a first analysis of this relation with respect to variance-based uncertainty principles in signal processing has been presented in [\textit{R. Levie} et al., Adv. Comput. Math. 40, No. 3, 609--627 (2014; Zbl 1336.42028)]. The present paper extends this work considerably in several aspects. Starting with the Cartesian product of certain abelian one-parameter-groups a canonical observable-manifold-structure is defined, which extends the work above to non-group-phase spaces. As an example the continuous Curvelet transform is discussed [\textit{E. J. Candès} and \textit{D. L. Donoho}, Appl. Comput. Harmon. Anal. 19, No. 2, 162--197 (2005; Zbl 1086.42022)]. Further, a nested semi-direct product group construction \(G\) together with so called ``multi-observables'' \(\check{\mathcal T}=\check{T}^k_m\) is introduced with the following remarkable properties: -- The construction allows for the introduction of variance-based uncertainty measures, constructed from suitable linear combinations of individual variance measures of the family members of \(\check{\mathcal T}\), which are constant on group orbits. In this way a global uncertainty measure is defined, which characterizes the wavelet transform at hand. -- As examples the finite short-time-Fourier-transform, the continuous 1-\(d\)-wavelet-transform, the finite 1-\(d\)-wavelet-transform and the shearlet-transform are investigated. The finite versions are essential for numerical implementetations of the respective continuous versions. The paper concludes with a discussion of sparsity. Given canonical multi-observables together with a group \(G\) a signal \(s(t)\) is called sparse, if its transform \(S(g)\) may be written as \(S(g)=\sum_k c_k \delta(g^{-1}g_k)\), i.e., a ``delta train'' localized at the points \(g_k\) in phase space. Naturally, given the corresponding wavelet \(f\) and the wavelet transform \(V_f(\cdot)\) such a signal cannot belong to the range of a wavelet transform \(V_f\), since the range is a reproducing kernel subspace of \(L^2(G)\) with convolution kernel \(V_f(f)\). The quantity \(V_f(f)\) is the ambiguity function. Thus, when applying the composition \(V_f \circ V_f^\ast\) to \(S(g)\) (i.e., reconstruction followed by transformation), the delta-peaks at \(g_k\) will be blurred with the ambiguity function and the blurring will be small (i.e., sparsity will be preserved), if \(V_f(f)\) decreases rapidly. The paper concludes with estimates bounding the ambiguity decay with the global uncertainty measures mentioned above.