Randomized signal processing with continuous frames (Q2066256)

The article concerns the uniform convergence of the Monte Carlo method on spreading signal processing to represent the coefficient space faithfully. The randomization of the quadratic approximation of the continuous operator frames consists in a global change of the input signal \(s\) in the feature space and in a modification of the feature coefficients term-by-term with respect to their values: the authors apply the linear bounded operator \(T\in L^2(G)\) to a pointwise non-linearity \(r\) via the first coefficient space representation \(V_f[s]\in \mathcal{H}\) to produce the end-to-end pipeline \(V^{\ast}_{f}T(r\circ V_f[s])\) which is synthesized back through the definition of one or two or \(K\) envelopes or sequences of envelopes \(\psi\) up to \(\epsilon\) together with their probability densities and distributions with input random sample variable \(g\) in the phase space \(G\). The input Monte Carlo PSI operator is based on \(K\) joint random samples using a bounded frame kernel \(R(g,g^\prime)\) which leaves \(g^\prime\) as most countably many sets of finite or zero measure to be the continuous integration output variable (for \(k=k^\prime\)). The study of the analysis of approximation errors leads to Markov and Bernstein types concentration of error results measuring the discrete stochastic signal processing outputs, computing bounds for estimated and estimating functions, namely finite Fourier series expansions in the phase space domain for each \(M\in \mathbb{N}\). Subsection 5.3 contains Theorem 29 as a very interesting result for applications in statistical learning and data sciences. It is fundamental the intervention of the Gramian operator \(Q_{f}^{K}\) to estimate the Monte Carlo synthesis and its interpretation for applications, in particular for polyphonic audio signals when the LTFT atoms are defined for an integer time dilation phase vocoder, up to the low and high frequency truncation, that is the LTFT is based on integrating LVD wavelet transforms and this is an example of some contexts where the uncertainty principle presents the needs to simplify computational problems with finite dimensions of Hilbert spaces. The authors prove in Appendix A the Hilbert space Bernstein inequality on \(K\) random vectors of \(\mathbb{C}^{d_j}\) for a fixed number \(j\) of independent random vectors. Appendix B reports a pseudo inverse of frame analysis and synthesis operators with properties.