A survey of uncertainty principles and some signal processing applications (Q744462): Difference between revisions

As the authors write in the introduction ``Uncertainty inequalities ... express the impossibility of a function ... to be simultaneously \textit{sharply concentrated} in two different representations, provided the latter are \textit{incoherent enough}''. In a true tour de force the many facets of this fascinating subject are reviewed in this paper. The paper is organized in two parts, dealing with continuous and discrete signals, respectively. First the authors discuss variance based inequalities, which result from projecting the signal (or ``state'' in quantum mechanics) onto eigenspaces of non-commuting selfadjoint operators. If the spreading of these projections is measured by variances, the lower bound of their product is given by the well-known Robertson-Schrödinger-inequality [\textit{G. B. Folland} and \textit{A. Sitaram}, J. Fourier Anal. Appl. 3, No. 3, 207--238 (1997; Zbl 0885.42006)]. The most famous manifestation is the position-momentum-uncertainty inequality (the Heisenberg uncertainty principle) of a quantum mechanical particle. The counterpart in signal processing is the time-frequency-uncertainty inequality. In some detail the authors review how the affine case may be mapped to the above ``Heisenberg-situation'' and point out that functions leading to equality not necessarily minimize the associated product of variances. The discussion continues with entropy-based uncertainty inequalities [\textit{I. I. Hirschman} jun., Am. J. Math. 79, 152--156 (1957; Zbl 0079.35104)] and localization notions in phase space with emphasis on bilinear time-frequency densities associated with the signal \(f\). In particular short-time Fourier transforms \({\mathcal V}_g f(t,\nu)\) and ambiguity functions \({\mathcal A}_f (t,\nu)={\mathcal V}_f f(t,\nu)\) are treated. Inequalities bounding from below joint concentration properties of a signal \(f\) in time and frequency, respectively, are described [\textit{D. L. Donoho} and \textit{P. B. Stark}, SIAM J. Appl. Math. 49, No. 3, 906--931 (1989; Zbl 0689.42001)] and the authors point out that these results are useful for proving the convergence of the Gerchberg-Papoulis-algorithm, employed in signal inpainting. In the second part, \(f\) is a finite discrete signal and the authors deal with uncertainty / spreading, if these signals are represented in different orthonormal bases or frames, which are in a suitably defined sense incoherent. Also in this case inequalities may be formulated giving lower bounds for joint concentration properties of these representations and suitably defined entropies. Surprisingly, minimizers or equalizers of discrete uncertainty inequalities may behave rather different from their continuous counterparts. Whereas in the continuous cases described above usually Gaussians and variants thereof appear as equalizers, in the discrete case this role might be taken by unlocalized signals like periodic series of Kronecker deltas (''picket fence-signals''). Thus finite discretization implies subtleties usually associated with the fact that the corresponding signal spaces are tori rather than spaces with certain decay properties as in the continuous case. As the authors point out, discrete uncertainty inequalities associated with families of frames are useful for improving the performance of signal separation algorithms. Here the main idea is to represent a priori unknown signal constituents by sparse representations with respect to individual members of this family. It turns out that this task performs the better, the more incoherent these frames are. Summarizing, in this paper the authors really succeed in providing an exhaustive survey of uncertainty principles and hopefully many readers are encouraged to explore this fundamental research area.