A New Framework for Decomposing Multivariate Information

Abstract

What are the distinct ways in which a set of predictor variables can provide information about a target variable? When does a variable provide unique information, when do variables share redundant information, and when do variables combine synergistically to provide complementary information? The redundancy lattice from the partial information decomposition of Williams and Beer provided a promising glimpse at the answer to these questions. However, this structure was constructed using a much-criticised measure of redundant information, and despite sustained research, no completely satisfactory replacement measure has been proposed. This thesis presents a new framework for information decomposition that is based upon the decomposition of pointwise mutual information rather than mutual information. The framework is derived in two separate ways. The first of these derivations is based upon a modified version of the original axiomatic approach taken by Williams and Beer. However, to overcome the difficulty associated with signed pointwise mutual information, the decomposition is applied separately to the unsigned entropic components of pointwise mutual information which are referred to as the specificity and ambiguity. This yields a separate redundancy lattice for each component. Based upon an operational interpretation of redundancy, measures of redundant specificity and redundant ambiguity are defined which enables one to evaluate the partial information atoms separately for each lattice. These separate atoms can then be recombined to yield the sought-after multivariate information decomposition. This framework is applied to canonical examples from the literature and the results and various properties of the decomposition are discussed. In particular, the pointwise decomposition using specificity and ambiguity is shown to satisfy a chain rule over target variables, which provides new insights into the so-called two-bit-copy example. The second approach begins by considering the distinct ways in which two marginal observers can share their information with the non-observing individual third party. Several novel measures of information content are introduced, namely the union, intersection and unique information contents. Next, the algebraic structure of these new measures of shared marginal information is explored, and it is shown that the structure of shared marginal information is that of a distributive lattice. Furthermore, by using the fundamental theorem of distributive lattices, it is shown that these new measures are isomorphic to a ring of sets. Finally, by combining this structure together with the semi-lattice of joint information, the redundancy lattice form partial information decomposition is found to be embedded within this larger algebraic structure. However, since this structure considers information contents, it is actually equivalent to the specificity lattice from the first derivation of pointwise partial information decomposition. The thesis then closes with a discussion about whether or not one should combine the information contents from the specificity and ambiguity lattices.