Core Concepts
This paper proposes an explicit formula for calculating the unique, redundant, and synergistic information atoms in a partial information decomposition of multivariate systems, which satisfies all the previously proposed axioms and properties.
Abstract
The paper addresses the problem of extending Shannon's mutual information to multivariate systems, which fails to capture the fine-grained interactions between the source and target variables. In 2010, Williams and Beer proposed the concept of Partial Information Decomposition (PID) to decompose the mutual information into unique, redundant, and synergistic information atoms. However, despite extensive efforts, a general formula satisfying all the proposed axioms and properties has not been found.
The key contributions of this paper are:
Introduction of the "do-operation", which adjusts the marginal distribution of the target variable while minimally impacting its connections with other variables. This operation is inspired by Judea Pearl's do-calculus in causal analysis.
Defining the unique information as the expectation of mutual information between the adjusted target variable and the source variable, given the other source variable(s). This definition satisfies all the proposed axioms and properties.
Deriving the definitions of the redundant and synergistic information atoms from the unique information definition, using the relationships established in the axioms.
Providing intuitive explanations for the role of the do-operation and the interpretation of the information atoms.
The paper proves that the proposed definitions satisfy all the axioms and properties, including nonnegativity, commutativity, additivity, and continuity. The unique information is shown to be bounded by the mutual information and the conditional entropy, and the synergistic information is nonnegative under the closed-system assumption.