When reasoning under uncertainty, classical probability forces us to assign a single, precise number to the likelihood of an event. But in the real world, we often lack the data to justify such precision. The Dempster-Shafer theory of evidence offers an alternative by introducing intervals bounded by belief and plausibility. For years, this framework was frequently viewed as an ad-hoc heuristic separate from classical probability.
However, foundational work established that belief functions are not a departure from classical probability, but a direct consequence of it under incomplete information. By interpreting belief functions as the lower envelopes of a family of standard probability distributions, researchers bridged the gap between these two paradigms.
The Two Views of Belief
To understand why this works, we must distinguish between two ways of interpreting a belief function. As clarified by Joseph Halpern et al. in Two Views of Belief: Belief as Generalized Probability and Belief as Evidence (aaai 1990) and Two Views of Belief: Belief as Generalized Probability and Belief as Evidence (Artif. Intell. 1992), the two views are:
- Belief as Evidence: The traditional view where weights are assigned directly to sets of propositions and combined using Dempster's rule of combination.
- Belief as Generalized Probability: The view that a single true probability distribution exists, but we only know that it belongs to a certain family (or set) of candidate distributions.
Under the generalized probability view, the belief in an event $A$ represents the absolute minimum (lower bound) probability assigned to $A$ across all distributions in that family. Conversely, the plausibility of $A$ represents the upper bound.
This lower envelope interpretation is deeply rooted in classical measure theory. As shown by Joseph Halpern et al. in Uncertainty, Belief, and Probability and Uncertainty, belief, and probability, a belief function is mathematically equivalent to an inner measure induced by a probability space containing nonmeasurable sets. When an event cannot be precisely measured, its true probability is bounded from below by the inner measure (belief) and from above by the outer measure (plausibility).
The Mechanics of Lower Envelope Conditioning
This shift in perspective fundamentally changes how we update our knowledge when new evidence arrives.
In the traditional evidence-based view, updating is performed using Dempster's rule of conditioning. However, the authors of Combination and Propagation of Uncertainty with Belief Functions - A Reexamination demonstrated that Dempster's rule is highly sensitive to minor perturbations in input—where assigning a near-zero probability can yield radically different results than an absolute zero. While some frameworks like the Transferable Belief Model justify this via specialization matrices (as explored by Frank Klawonn et al. in The Dynamic of Belief in the Transferable Belief Model and Specialization-Generalization Matrices), it can lead to counterintuitive paradoxes when applied to probability bounds.
If we instead treat a belief function as a lower envelope of a family of probabilities, standard Dempster conditioning no longer applies. Instead, when we learn a new fact $B$, we must update every single probability distribution in our family using standard Bayesian conditioning. The new conditional belief is then the lower envelope of these conditional probabilities.
In A new approach to updating beliefs, Ronald Fagin et al. formalize this process, showing that conditioning a belief function as a lower envelope yields a clean, closed-form expression:
$$Bel(A|B) = \frac{Bel(A \cap B)}{Bel(A \cap B) + Pl(A^c \cap B)}$$
(Note: This matches the structural formulation Bel(A|B) = Bel(A ∩ B) / (Bel(A ∩ B) + Pl(A ∩ B)) where the denominator balances the absolute certainty of the intersection against its competing plausibility, ensuring the updated result remains a valid belief function without suffering from the instabilities of Dempster's rule).
Connecting Probability, Possibility, and Logic
Grounding belief functions as lower envelopes provides a unified mathematical language that connects several distinct uncertainty frameworks:
- Possibility Theory: When evidence is structured as a hierarchy of nested sets, Didier Dubois et al. showed in Combination and Propagation of Uncertainty with Belief Functions - A Reexamination that the plausibility function becomes mathematically identical to a possibility measure. To bridge these frameworks for decision-making, Phan Hong Giang et al. introduced partially consonant belief functions in Decision Making with Partially Consonant Belief Functions, creating a representation consistent with classical statistical principles like the likelihood principle.
- Algebraic Extensions: The lower envelope interpretation is robust to structural generalizations. Chunlai Zhou demonstrated in Belief Functions on Distributive Lattices that belief functions defined on distributive lattices remain equivalent to inner probability measures. To handle vague or conflicting expert knowledge, Thomas Augustin 0001 et al. in Generalized basic probability assignments extended this by calculating lower envelopes over entire sets of basic probability assignments.
- Logic Programming: This structural closure makes the lower envelope perspective highly effective for computational logic. In The Structure and Complexity of Credal Semantics, Fabio Cozman et al. showed that the lower probability of an event in a logic program with cycles and negation can be represented as an infinitely monotone Choquet capacity. Because this is mathematically equivalent to a belief function, it ensures the set of consistent distributions is closed and convex, allowing researchers to apply powerful optimization techniques from imprecise probability theory.
Key Foundational Work
Uncertainty, Belief, and Probability Ronald (1989) — contributed 3 concept elements to this direction.
Uncertainty, Belief, and Probability (1989) — contributed 2 concept elements to this direction.
Combination and Propagation of Uncertainty with Belief Functions - A Reexamination (1985) — contributed 2 concept elements to this direction.