Partial information decomposition for discrete target and continuous source random variables.

Area of Science:

• Information theory
• Network systems analysis
• Computational neuroscience

Background:

• Partial Information Decomposition (PID) quantifies complex interactions in network systems by analyzing mutual information (MI).
• Existing PID methods are primarily for discrete variables, with recent extensions to continuous systems.
• Current PID schemes for mixed discrete-continuous variables require data manipulation, potentially altering information content.

Purpose of the Study:

• To develop a new PID scheme for mixed discrete-continuous variables that avoids data manipulation.
• To accurately estimate the mutual information between a discrete target and continuous sources.
• To provide a robust tool for analyzing information flow in complex systems.

Main Methods:

• Introduced a PID scheme expressing MI as Kullback-Leibler divergence for discrete target states and continuous sources.
• Employed a nearest-neighbor strategy for estimating the Kullback-Leibler divergence.
• Validated the method on simulated mixed-variable systems and benchmark datasets.

Main Results:

• The proposed PID scheme effectively quantifies information decomposition in mixed variable systems.
• The method accurately estimates mutual information without altering original data.
• Demonstrated effectiveness in simulated environments and on established benchmark data.

Conclusions:

• The novel PID approach overcomes limitations of existing methods for mixed discrete-continuous variables.
• This technique offers a non-invasive way to analyze information flow in complex systems.
• Applications include sensory coding in neuroscience and feature selection in machine learning.