Frame paths and error bounds for sigma-delta quantization (Q869647): Difference between revisions

The problem of estimating a signal in a Hilbert space \(\mathcal{H}\) of finite dimension \(d\) using a frame with \(N\) elements and with frame coefficients restricted to a fixed finite set of values is considered. An improved upper bound for the (worst case) Euclidean norm of the error induced by the first-order sigma-delta encoding of frame coefficients is derived. The authors define a uniform frame path \(f\:[a,b]\to \mathbb{R}^d\) (or to \(\mathbb{C}^d\)) as a mapping taking values in the unit sphere and admitting infinitely many \(N\) such that the samples \(F_N=\{f(a+k(b-a)/N)\}_{k=1}^N\) of \(f\) form a uniform tight frame for \(\mathbb{R}^d\) with frame bound \(N/d\). Examples provided include harmonic frames corresponding to \(d=2k\) and \(f(t)=\sqrt{2\over d} (e^{2\pi i t}, e^{4\pi i t},\dots, e^{\pi i d t})\). It is shown that harmonic frame expansions with coefficients quantized to the nearest integer multiple of \(\delta>0\) provide a worst case error on the order of \(\pi\delta d^{3/2}/ {2N}\), with precise constants provided depending on the parity of \(d\). Errors are estimated in terms of the variation of \(f\) with respect to the sample points. This estimate improves previously known estimates for harmonic frames by a factor of \(\sqrt{d}\). Corresponding estimates are given for other frame paths. The first order sigma-delta frame quantizer is defined by \(Q_F(x)={d\over N}\sum_{j=1}^N q_j f_j\) where \(\{f_j\}\) denotes the (ordered) frame and the quantized coefficients \(q_j\) of \(x\) are defined inductively by setting \(u_0=0\) and \(q_j=Q(\langle x,f_{j}\rangle +u_{j-1})\) and \(u_j=\langle x,f_j\rangle-q_j+u_{j-1}\) where \(Q\) denotes the quantization or round-off operation. The relevance of the relative scaling of the parameters \(\delta, d\) and \(N\) is explained in terms of tradeoff between quantization error and frame size.