A study of local approximations in information theory

[摘要] The intractability of many information theoretic problems arises from the meaningful but nonlinear definition of Kullback-Leibler (KL) divergence between two probability distributions. Local information theory addresses this issue by assuming all distributions of interest are perturbations of certain reference distributions, and then approximating KL divergence with a squared weighted Euclidean distance, thereby linearizing such problems. We show that large classes of statistical divergence measures, such as f-divergences and Bregman divergences, can be approximated in an analogous manner to local metrics which are very similar in form. We then capture the cost of making local approximations of KL divergence instead of using its global value. This is achieved by appropriately bounding the tightness of the Data Processing Inequality in the local and global scenarios. This task turns out to be equivalent to bounding the chordal slope of the hypercontractivity ribbon at infinity and the Hirschfeld-Gebelein-Renyi maximal correlation with each other. We derive such bounds for the discrete and finite, as well as the Gaussian regimes. An application of the local approximation technique is in understanding the large deviation behavior of sources and channels. We elucidate a source-channel decomposition of the large deviation characteristics of i.i.d. sources going through discrete memoryless channels. This is used to derive an additive Gaussian noise channel model for the local perturbations of probability distributions. We next shift our focus to infinite alphabet channels instead of discrete and finite channels. On this front, existing literature has demonstrated that the singular vectors of additive white Gaussian noise channels are Hermite polynomials, and the singular vectors of Poisson channels are Laguerre polynomials. We characterize the set of infinite alphabet channels whose singular value decompositions produce singular vectors that are orthogonal polynomials by providing equivalent conditions on the conditional moments. In doing so, we also unveil the elegant relationship between certain natural exponential families with quadratic variance functions, their conjugate priors, and their corresponding orthogonal polynomial singular vectors. Finally, we propose various related directions for future research in the hope that our work will beget more research concerning local approximation methods in information theory.

[发布机构] Massachusetts Institute of Technology