The mathematical landscape of partial information decomposition: A comprehensive review of properties and measures

Alberto Liardi, Keenan J. A. Down, George Blackburne, Matteo Neri, Pedro A. M. Mediano

The mathematical landscape of partial information decomposition: A comprehensive review of properties and measures

Abstract: Partial Information Decomposition (PID) has become one of the most prominent information-theoretic frameworks for describing the structure and quality of information in complex systems. Despite its widespread utility, there exists no unique solution constraining precisely how a PID should be constructed, leading to a multiverse of different formalisms with different mathematical commitments. In this work, we provide a comprehensive overview of the mathematical landscape of PID. By integrating existing PID measures into a common language, we systematically examine all major approaches to the PID framework that have emerged so far, determining for each measure whether or not each known property holds. In addition, we derive a web of all known theorems mapping the relationships and incompatibilities between these properties, before also revealing some novel interdependency results. In doing so, we chart a brief history of the framework, promote a unified perspective for its discussions, and offer a path towards both theoretical refinement and informed empirical applications for the future of this powerful method.

Introduction

Classical Shannon information theory provides a robust foundation for quantifying the total information shared between variables. However, it often fails to capture the nuanced "higher-order" dependencies inherent in complex systems. For instance, while mutual information can quantify the total shared content between two sources and a target, it cannot distinguish between redundancy (information provided by both sources independently) and synergy (information available only when sources are considered together). Partial Information Decomposition (PID), introduced in 2010, addresses this by decomposing mutual information into redundant, unique, and synergistic components.

The Axiomatic Landscape

Since its inception, the PID field has evolved into a "multiverse" of competing mathematical definitions. Researchers have proposed various "axioms"—desirable mathematical properties that a robust PID measure should satisfy. These range from basic requirements like symmetry and monotonicity to more specialized constraints like the "Blackwell Property" or "Equivalence-class Invariance."

Table of PID properties and their formal descriptions.

The challenge lies in the fact that many of these properties are mutually exclusive. This paper systematically maps these properties, providing a clear taxonomy to help researchers navigate the philosophical trade-offs required when selecting a measure for specific scientific inquiries.

Evaluating PID Measures

To rigorously evaluate the 19 major PID measures identified, the authors utilized "standard candles" or "toy gates"—simple, well-defined systems that serve as benchmarks in information theory. One of the most critical benchmarks is the Two-Bit Copy (TBC) system.

Truth table for the Two-Bit Copy (TBC) system, a standard benchmark for PID measures.

In the TBC system, the target $Y$ is a simple concatenation of two independent source bits $X_1$ and $X_2$. A "good" PID measure should ideally reflect that there is no redundancy or synergy in this system, only unique information from each source. Other benchmarks include the XOR-Source-Copy and various Gaussian systems for continuous variables.

Key Insights and the "No-Go" Reality

The systematic verification of measures against properties reveals a fundamental "No-Go" reality: a "perfect" PID measure is mathematically impossible. For example, the authors prove that one cannot simultaneously satisfy Equivalence-class Invariance (the measure is invariant under data relabeling) and the Identity Property (redundancy equals mutual information in specific cases) in systems with three or more sources without encountering logical contradictions.

Measures tend to cluster into distinct philosophical families. Some prioritize Local Positivity, ensuring that information "atoms" are never negative, while others allow negative values, interpreting them as "misinformation" similar to negative values in pointwise mutual information. Furthermore, a significant divide exists between measures viewing redundancy as a statistical correlation versus those viewing it as a reflection of underlying causal mechanisms.

Conclusion and Future Directions

This comprehensive review provides a definitive matrix of measure-property compatibility, revealing that while the largest set of compatible properties contains 19 axioms, the most compliant existing measure ($I^{BROJA}$) only satisfies 16.

Future work must address the empirical performance of these measures on noisy, high-dimensional real-world data, such as fMRI or genomic sequences. Additionally, extending the PID framework to fully continuous, non-Gaussian systems remains a significant open challenge for the field. By providing this "atlas," the authors offer a path toward both theoretical refinement and more informed empirical applications of PID.

Summary updated

Abstract

Partial Information Decomposition (PID) has become one of the most prominent information-theoretic frameworks for describing the structure and quality of information in complex systems. Despite its widespread utility, there exists no unique solution constraining precisely how a PID should be constructed, leading to a multiverse of different formalisms with different mathematical commitments. In this work, we provide a comprehensive overview of the mathematical landscape of PID. By integrating existing PID measures into a common language, we systematically examine all major approaches to the PID framework that have emerged so far, determining for each measure whether or not each known property holds. In addition, we derive a web of all known theorems mapping the relationships and incompatibilities between these properties, before also revealing some novel interdependency results. In doing so, we chart a brief history of the framework, promote a unified perspective for its discussions, and offer a path towards both theoretical refinement and informed empirical applications for the future of this powerful method.
BibTeX
@misc{liardi2026mathematical,
  author = {Alberto Liardi and Keenan J. A. Down and George Blackburne and Matteo Neri and Pedro A. M. Mediano},
  title = {{The mathematical landscape of partial information decomposition: A comprehensive review of properties and measures}},
  year = {2026},
  eprint = {2603.06678},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06678},
}
Figure