Optimal reconstruction systems for erasures and for the \(q\)-potential (Q837009): Difference between revisions

Given Hilbert spaces \(\mathcal{H}\) and \(\mathcal{K}\) of respective dimensions \(d\) and \(l\), the author defines an \((m,l,d)\)-reconstruction system to be a family of \(m\) linear operators \(V_i:\mathcal{H}\to\mathcal{K}\) such that the positive operator \(S=\sum_{i=1}^m V_i^\ast V_i\) is bounded and continuously invertible. If \(S=I_d\) then the system is called an \((m,l,d)\)-protocol. Such a system generalizes the notion of a frame (\(m=1\)) as well as \textit{B. G. Bodmann}'s protocols [Appl. Comput. Harmon. Anal. 22, No. 3, 274--285 (2007; Zbl 1193.42113)] in which each \(V_i^\ast V_i\) corresponds to a Parseval frame. The author studies the stability of such reconstruction systems under erasures. The starting point is a generalization of the Benedetto-Fickus frame potential [\textit{J. J. Benedetto} and \textit{M. Fickus}, Adv. Comput. Math. 18, No. 2--4, 357--385 (2003; Zbl 1028.42022)] to the notion of a \(q\)-potential of an \((m,l,d)\)-reconstruction system, namely the \(l\times l\)-matrix \[ {\mathbf{P}}_q(V)=\sum_{i,j=1}^m |V_i V_j^\ast|^2 \] corresponding to the partial trace of the Grammian \(V V^\ast\) with respect to the block representation of size \(ml\). The author quantifies a general notion of optimality of such a system and gives further details of what it means in the cases of uniformly weighted projections (u.w.p) in which each \(V_i^\ast V_i\) is a multiple of a projection. A model of encoding and decoding in terms of such reconstruction systems is given and a corresponding notion of packet loss is quantified in terms of a unitarily invariant norm. Stability under erasures is then quantified in terms of a norm inequality that becomes an equality in the case of u.w.p's. To define optimality, for \(\mathbf{x}\) and \(\mathbf{y}\in\mathbb{R}^l\) with all coordinates nonnegative, one says that \(\mathbf{x}\) is submajorized by \(\mathbf{y}\), denoted \(\mathbf{x}\preceq_w\mathbf{y}\), if each partial sum of the decreasing rearrangement of \(\mathbf{x}\) is bounded by the corresponding partial sum for \(\mathbf{y}\). Submajorization extends to self-adjoint matrices \(A\) and \(B\) by requiring that the spectrum of \(A\) pointwise dominates the spectrum of \(B\), also expressed as \(\mathbf{B}\preceq_w\mathbf{A}\). The optimality of general protocols is expressed as follows. Let \(\{V_i\}\) be an \((m,l,d)\)-reconstruction system such that \(\text{tr}\, (V^\ast V)=\sum_{i=1}^m \text{tr}\, (V_i^\ast V_i)\geq d\). Then \((d/l) I_l\preceq_w \mathbf{P}_q (V)\). Consequently, if \(\|\cdot\|\) is a unitarily invariant norm on \(l\times l\) matrices with symmetric gauge \(\psi\) then \(d \eta_\psi(l)\leq \| {\mathbf{P}}_q(V)\|\) and, if \(f:[0,\infty)\to \mathbb{R}\) is a convex increasing function with \(f(0)=0\) then \(l f(d/l)\leq \text{tr}\, (f({\mathbf{P}}_q(V)))\). Protocols are also characterized among the reconstruction systems in terms of special cases of these inequalities.