Analyzing Training Using Phase Transitions in Entropy---Part I: General Theory

TL;DR

This paper develops a theoretical framework for analyzing phase transitions in conditional entropy during training, providing bounds on mutual information for large-scale systems without restrictive assumptions.

Contribution

It introduces a general method to compute bounds on mutual information using derivatives of conditional entropy, applicable to diverse training-based algorithms.

Findings

Bound on mutual information derived from entropy derivatives

Applicable to non-Gaussian, nonlinear, large-scale systems

No need for explicit parameter estimation or worst-case noise assumptions

Abstract

We analyze phase transitions in the conditional entropy of a sequence caused by a change in the conditional variables. Such transitions happen, for example, when training to learn the parameters of a system, since the transition from the training phase to the data phase causes a discontinuous jump in the conditional entropy of the measured system response. For large-scale systems, we present a method of computing a bound on the mutual information obtained with one-shot training, and show that this bound can be calculated using the difference between two derivatives of a conditional entropy. The system model does not require Gaussianity or linearity in the parameters, and does not require worst-case noise approximations or explicit estimation of any unknown parameters. The model applies to a broad range of algorithms and methods in communication, signal processing, and machine learning that employ training as part of their operation.