What Is Evidence Theory?

Evidence theory, also known as the Dempster-Shafer theory of belief functions, is a mathematical framework for representing and reasoning under uncertainty when available information is incomplete or imprecise. It was introduced by Arthur Dempster in the 1960s as a means of generalizing classical probability, and later formalized and extended by Glenn Shafer in his 1976 monograph "A Mathematical Theory of Evidence." Unlike Bayesian probability, which requires a prior distribution to be fully specified before new data can update beliefs, evidence theory allows portions of the total belief to remain unassigned, representing genuine ignorance rather than forcing a uniform prior.

The framework organizes uncertainty over a finite set of mutually exclusive hypotheses called the frame of discernment. A basic probability assignment (BPA), sometimes called a mass function, distributes a unit of belief across subsets of that frame rather than only across individual hypotheses. This structure allows an agent to express, for instance, that it believes a system failure is due to either component A or component B without committing to a specific probability for each individually. Belief and plausibility functions derived from the BPA bound the true probability from below and above, providing an interval rather than a point estimate.

Belief Propagation and Evidence Combination

The Dempster combination rule is the central operation for merging evidence from independent sources. It aggregates two BPAs by forming their orthogonal sum, normalizing the result to exclude conflicting mass. The rule has been influential in sensor fusion, where readings from multiple instruments with different reliability profiles must be combined into a single estimate. Glenn Shafer's foundational paper on Dempster-Shafer theory at Berkeley provides a concise treatment of how belief functions relate to probability intervals and the conditions under which Dempster's rule applies. The related topic of belief propagation, developed in probabilistic graphical models, shares conceptual roots with evidence theory: both represent uncertainty over a set of variables and define structured rules for propagating that uncertainty through a graph or hierarchy.

Relationship to Probabilistic and Fuzzy Methods

Evidence theory occupies a position within a wider family of frameworks for reasoning under uncertainty that also includes classical Bayesian probability, fuzzy set theory, and possibility theory. Each framework makes different assumptions about the nature and source of uncertainty. Bayesian methods treat uncertainty as subjective degree of belief updated by Bayes' theorem, and they require probability mass to be assigned to individual hypotheses at every step. Fuzzy logic characterizes vague or graded membership in a class rather than incomplete information about which of several hypotheses is true. Evidence theory is appropriate when the source of uncertainty is insufficient evidence rather than inherent randomness, a distinction that matters in diagnostic and classification tasks where the information base itself is sparse. The Springer volume on classic works of the Dempster-Shafer theory of belief functions collects the core theoretical papers and traces the framework's development from Dempster's 1967 upper and lower probability bounds to contemporary applications.

Computational Aspects

Practical implementations of evidence theory face a combinatorial challenge: the number of subsets of a frame of discernment grows exponentially with the number of hypotheses, making direct computation of belief functions infeasible for large frames. Approximation methods, including inner approximations that restrict mass assignments to singletons or small subsets and Monte Carlo sampling techniques, reduce computational cost at the expense of exactness. ScienceDirect's overview of Dempster-Shafer theory summarizes computational strategies alongside the theory's applications in classification and decision support.

Applications

Evidence theory has applications in a range of fields, including:

• Multi-sensor fusion in robotics and autonomous vehicles
• Medical diagnosis systems combining laboratory results, imaging, and clinical observations
• Fault diagnosis in complex engineering systems where sensor data is incomplete
• Intrusion detection and cybersecurity risk assessment
• Target recognition in radar and sonar processing