Can ordinary probability appear as a special case?

Yes. When all mass is assigned to singleton outcomes, belief and plausibility coincide with an additive probability measure.

How is complete ignorance represented?

In the standard finite setup, assigning all mass to the full frame leaves proper nonempty propositions with zero belief and full plausibility.

Does a wider belief - plausibility gap always mean worse data?

It signals less resolved support for that proposition under the model, but interpreting its practical cause requires examining the evidence construction.

AI Glossary / Graphs and symbolic AI / Advanced

Belief function

Definition

A belief function is a mathematical representation of evidential support used in Dempster - Shafer theory. It can assign support to sets of possible outcomes, allowing evidence to support a broad possibility without forcing that support onto one specific outcome.

What it means

Ordinary probability distributions assign probabilities to individual outcomes and their combinations. Belief-function models can represent evidence that narrows the possibilities without resolving them completely. For example, a report might support “the item is either red or blue” while offering no basis for choosing between those colors.

A basic mass assignment records how support is allocated to subsets of an explicitly defined set of possibilities, called the frame. The belief in a proposition combines masses that support it entirely. A related quantity, plausibility, includes masses that do not rule it out.

This is a technical meaning of belief, not a claim that a machine has personal convictions. It also differs from a Bayesian belief state or the belief-propagation algorithm, despite the shared word.

How it works

In the usual finite normalized formulation, nonnegative masses are assigned to subsets of the frame, sum to one and assign zero mass to the empty set. A mass assigned to a set supports that set as a whole; it is not automatically divided equally among its members.

To calculate Bel(A), add the masses assigned to subsets contained entirely in A. To calculate plausibility Pl(A), add masses assigned to subsets that intersect A. Equivalently, Pl(A) = 1 − Bel(not A). The gap between belief and plausibility reflects support that remains unresolved with respect to that proposition.

Combining evidence introduces further assumptions, including the suitability and independence of sources under the selected combination rule. Conflicting evidence requires careful interpretation. Belief and plausibility should not automatically be described as a statistical confidence interval or as bounds on an unknown real-world frequency. Their meaning depends on the evidential model.

Three practical examples

A sensor-fusion experiment represents a reading that narrows an object's type to two possibilities without choosing one arbitrarily.
A document-review tool records that evidence supports one of several document categories while leaving the exact classification unresolved.
A classroom uncertainty model compares complete ignorance with equal probabilities, showing why they are different representations.

Think of it this way

Think of placing evidence cards into labeled envelopes. An envelope marked “A or B” supports that combined possibility, but the cards inside have not been assigned specifically to A or specifically to B.

The analogy has limits. The numerical masses need a justified interpretation; they are not simply counts of arbitrary opinions. Combining envelopes from dependent sources can also exaggerate support if the underlying evidence is counted repeatedly.

Worked example

Calculating belief and plausibility

Let the frame contain three exclusive possibilities: red, blue and green. Use an invented mass assignment with 0.4 on {red}, 0.3 on {red, blue} and 0.3 on the full frame. These masses sum to one.

Bel(red) is 0.4 because only the singleton {red} lies entirely inside the proposition red. Pl(red) is 1.0 because every mass-bearing set intersects red. The model contains direct support of 0.4 for red and no mass that rules red out.

Bel(blue) is 0 because no mass is committed to blue alone. Pl(blue) is 0.6 from the {red, blue} and full-frame masses. Zero belief therefore does not mean blue has been disproved.

For “red or blue,” belief is 0.4 + 0.3 = 0.7 and plausibility is 1.0. None of these calculations requires splitting the ambiguous 0.3 equally between red and blue.

Illustrative case study

Fictional case: Cedar Sorting builds a prototype that combines two inspection reports. Developers initially treat the reports as independent evidence, although both were generated from the same underlying camera image.

The combined result appears more decisive than either report. A review identifies the shared source and revisits the evidence model before using the result operationally. The team also separates unresolved support from direct support for a particular item category in its display.

The lesson is that a sophisticated uncertainty calculus cannot repair an unjustified evidence assumption. Source relationships and the meaning of the masses need as much attention as the arithmetic.

Common misconceptions and mistakes

Zero belief means impossibility. A proposition can have no direct support while retaining substantial plausibility.
Mass on a set should always be divided equally. That introduces an additional assumption absent from the original mass assignment.
Belief and plausibility are automatically frequentist confidence limits. Their interpretation comes from the chosen evidence framework.
Repeating the same evidence creates independent confirmation. Dependence can lead to double counting.
A belief function is the same as a neural-network confidence score. The mathematical objects and their semantics differ.

Business use

Belief functions can be useful when a project needs to represent partial support and unresolved alternatives explicitly. They may appear in evidence fusion, classification research and decision-support models.

Before adoption, define the frame, explain how masses are obtained and document any rule used to combine sources. Test contradictory, missing and dependent evidence, not only consistent examples. Decide separately how the resulting representation informs action: evidential support does not by itself specify the cost of errors or the best operational choice. Compare the approach with other uncertainty models suited to the task.

Check your understanding

A model assigns all mass to {A, B}, with A and B the complete frame. Does it assert that A and B each have probability 0.5?

Answer: No. It leaves the support unresolved between them. Assigning equal singleton probabilities would be a different representation with an additional commitment.

Related terms

Probability
Dempster–Shafer theory
mass assignment
plausibility
frame of discernment
evidence fusion
ignorance
source dependence

Learn next

Create a three-outcome frame and assign masses to both singleton and multi-outcome sets. Calculate belief and plausibility for each outcome and one combined proposition. Compare all mass on the full frame with equal singleton probabilities. Study evidence-combination assumptions before applying a fusion rule.

Sources

Glenn Shafer, Perspectives on the Theory and Practice of Belief Functions, explains foundations, interpretation and evidence dependence: https://www.glennshafer.com/assets/downloads/articles/article40.pdf

AIMA glossary, belief function, provides the source-list context: https://github.com/aimacode/aima-glossary/blob/master/glossary.md?plain=1

The colour calculation and fictional sorting case are original illustrations of the standard finite normalized formulation.

Previous: Belief aggregation Next: Belief propagation