Is subjective logic ordinary subjective probability?

No. It is a particular formalism with explicit opinion components and operators.

Can uncertainty be one?

Yes. A vacuous binary opinion has b = d = 0 and u = 1, projecting to its base rate.

Does uncertainty zero prove certainty in reality?

No. It is a property of the represented opinion, not a guarantee about the world.

AI Glossary / Statistics and probability / Advanced

Subjective logic

Definition

Subjective logic is a formal framework for reasoning with uncertain opinions, developed by Audun Jøsang. It represents both support for propositions and uncertainty arising from limited evidence, and provides operators for combining or transforming those representations. In a binary opinion, belief, disbelief, uncertainty, and a base rate are recorded separately. The framework makes some uncertainty explicit, but its outputs remain dependent on how evidence, assumptions, and relationships between sources are modeled.

What it means

A single probability can conceal how much evidence supports it. A projected probability of one-half might reflect strong balanced evidence or almost no evidence with a neutral base rate. Subjective logic can distinguish these situations through the opinion representation.

Subjective does not mean arbitrary reasoning. An observer's evidence and uncertainty are explicit. The uncertainty mass is not simply the probability that the proposition is false. Disbelief represents support against the proposition; uncertainty represents uncommitted mass in the opinion. Keeping those meanings separate is central to interpreting the numbers.

How it works

For a binary proposition, write an opinion as (b, d, u, a). Belief b, disbelief d, and uncertainty u are nonnegative and sum to one; the base rate a lies between zero and one. The projected probability is P = b + au.

Under a common evidence mapping with positive evidence r, negative evidence s, and prior weight W, set b = r / (r + s + W), d = s / (r + s + W), and u = W / (r + s + W). With suitable parameters, the corresponding beta distribution has shape parameters r + Wa and s + W(1 − a).

Choose operators according to the intended relationship, such as combining evidence, propagating trust, or reasoning conditionally. These are not interchangeable forms of averaging. The treatment of independent, dependent, conflicting, or duplicated evidence matters to the result.

Record the proposition, source provenance, base rate, evidence mapping, and chosen operator. Validate the resulting predictions or decisions against the intended application. Formal consistency of an opinion does not establish that the evidence was representative, reliable, or correctly interpreted.

Three practical examples

1. A trust-model researcher represents evidence for and against a source separately from the remaining uncertainty before applying a specified trust operator.

2. An AI system compares two predictions with the same projected probability but different evidence amounts, preserving information useful for review prioritization.

3. A sensor-fusion prototype records whether reports share an upstream source before deciding how their opinions should be combined.

Think of it this way

Think of three labeled containers: evidence supporting a claim, evidence opposing it, and uncommitted uncertainty. The base rate supplies a way to project the uncommitted portion into a probability. Two collections can produce the same final probability while containing very different amounts of unresolved uncertainty.

Worked example

For an original binary teaching example, choose prior weight W = 2 and base rate a = 0.5. Suppose positive evidence r = 6 and negative evidence s = 2. The denominator is 6 + 2 + 2 = 10, so b = 0.6, d = 0.2, and u = 0.2. These masses sum to one.

The projected probability is 0.6 + 0.5 × 0.2 = 0.7. The beta mapping gives shape parameters 6 + 2 × 0.5 = 7 and 2 + 2 × 0.5 = 3, whose mean is 7 / 10 = 0.7.

Now consider r = 1 and s = 1 with the same W and a. The opinion is (0.25, 0.25, 0.5, 0.5), projecting to 0.5. With r = 99 and s = 99, the opinion is (0.495, 0.495, 0.01, 0.5), also projecting to 0.5. Equal projected probabilities therefore do not imply equal uncertainty masses.

These are stipulated evidence mappings, not automatically valid conversions of arbitrary AI scores into trustworthy opinions. The counts require an appropriate evidence interpretation, and repeated copies of one report should not silently be treated as many independent observations.

Illustrative case study

In a fictional company, Alder Verification builds a prototype that combines reports about document validity. Two upstream tools often repeat the same source, but an early design counts their outputs as independent evidence. The team records shared provenance and revisits its fusion assumptions before using the combined opinions. It also displays uncertainty alongside projected probability so reviewers can distinguish sparse evidence from substantial balanced evidence. The framework improves transparency only when the evidence mapping and dependencies are represented honestly.

Common misconceptions and mistakes

Uncertainty mass is not the same as disbelief. Assigning mass to uncertainty does not assert that the proposition is false.

Two identical projected probabilities need not express the same evidence state. Conversely, low uncertainty mass is not a guarantee of correct information: a model can encode large amounts of biased or duplicated evidence.

Fusion operators are not interchangeable, and a single operator should not be applied blindly to every pair of sources. Base rates and prior weights also require interpretation. Subjective logic does not remove the need to define propositions precisely, evaluate calibration, or examine the consequences of acting on a mistaken conclusion.

Business use

Subjective logic can help make uncertainty and source assumptions explicit in trust systems, evidence aggregation, and AI decision support. Its practical attraction is the richer representation compared with a lone confidence score.

Businesses should establish what evidence units mean and how source dependence is tracked before adopting opinion calculations. Compare the approach with simpler probabilistic models and measure whether the additional representation improves decisions or review efficiency. Preserve provenance and explain the meaning of the displayed masses to users. A mathematically valid tuple should not be presented as an independently verified assessment of truth.

Check your understanding

Answer: An opinion has b = 0.3, d = 0.2, u = 0.5, and a = 0.4.

Is it normalized, and what is its projected probability?

Answer: The first three masses sum to one, and P = 0.3 + 0.4 × 0.5 = 0.5.

Related terms

Base rate
Subjective opinion
belief mass
epistemic uncertainty
beta distribution
evidence fusion
trust propagation
probabilistic logic

Learn next

Study binomial opinions and their beta-distribution mapping. Then examine the assumptions of particular fusion operators before combining dependent sources.

Sources

Audun Jøsang, Subjective Logic, UAI tutorial: https://www.auai.org/uai2016/tutorials_pres/subj_logic.pdf . The primary author material defines opinion components, projected probability, evidence mappings, and operator families. The numerical comparisons and fictional provenance case are original; no specific fusion algorithm is claimed to be implemented here.

Previous: Subject-verb agreement Next: Subjective probability