Displacement Convexity, A Useful Framework for the Study of Spatially Coupled Codes (1304.6026v2)

Abstract: Spatial coupling has recently emerged as a powerful paradigm to construct graphical models that work well under low-complexity message-passing algorithms. Although much progress has been made on the analysis of spatially coupled models under message passing, there is still room for improvement, both in terms of simplifying existing proofs as well as in terms of proving additional properties. We introduce one further tool for the analysis, namely the concept of displacement convexity. This concept plays a crucial role in the theory of optimal transport and, quite remarkably, it is also well suited for the analysis of spatially coupled systems. In cases where the concept applies, displacement convexity allows functionals of distributions which are not convex in the usual sense to be represented in an alternative form, so that they are convex with respect to the new parametrization. As a proof of concept we consider spatially coupled $(l,r)$-regular Gallager ensembles when transmission takes place over the binary erasure channel. We show that the potential function of the coupled system is displacement convex. Due to possible translational degrees of freedom convexity by itself falls short of establishing the uniqueness of the minimizing profile. For the spatially coupled $(l,r)$-regular system strict displacement convexity holds when a global translation degree of freedom is removed. Implications for the uniqueness of the minimizer and for solutions of the density evolution equation are discussed.