Credal Set¶

Represent imprecise probabilistic belief by a set of admissible probability measures, deriving lower and upper expectations as envelopes while keeping convexity, closure, conditioning, and independence choices explicit.

Version: v3 · 2026-09-19 · History
Domain-specific #: 1584
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Imprecise Probability → Mathematics
Aliases: Set of probabilities, Set of probability measures, Credal state

Core Idea¶

A credal set \(K\) is a set of probability distributions or probability measures admitted as possible representations of uncertainty. Instead of selecting one precise distribution \(P\), an analyst retains every \(P\in K\) consistent with elicited bounds, partial information, robustness neighborhoods, or other coherence constraints. The set may contain finitely or countably additive probabilities depending on the theory. It is often convex and closed, but those are additional regularity choices rather than consequences of the word Η set.

For a bounded gamble or measurable function \(f\), the lower and upper expectations induced by \(K\) are envelope functionals: \(\underline E_K(f)=\inf_{P\in K}E_P[f]\) and \(\overline E_K(f)=\sup_{P\in K}E_P[f]\). In finite closed sets the extrema are attained; for closed convex polytopes, linear expectations can be optimized at extreme points. The interval expresses model imprecision, not ordinary sampling error around one estimated distribution. Walley's monograph develops coherent lower previsions and their relation to sets of linear previsions as a general framework for statistical reasoning under imprecision.^[1]

Convexification has a behavioral interpretation in many frameworks: if two precise distributions are each admissible, mixtures between them need not be excluded. Yet nonconvex sets can be meaningful when admissibility encodes a disjunction of structured models and mixtures would introduce models outside either structure. Closure supports attainment and limit stability but can also add boundary distributions. A reference-grade use therefore states whether the credal set is closed, convex, finitely generated, dominated, or specified by inequalities, and whether an envelope representation loses decision-relevant information.

Walley's later comparison of lower probabilities, capacities, belief functions, sets of measures, partial preference orderings, and desirable gambles warns against treating any one representation as universally sufficient.^[2] Credal networks add graphical factorization and independence commitments to local or global credal sets; Cozman distinguishes extensions and develops inference for those models.^[3] A credal set alone contains no graph and does not determine how to update, condition, combine evidence, or choose an action. Those operations require separate rules whose consequences can differ.

How would you explain it like I'm…

All the Guesses That Fit

Imagine a bag of red and blue marbles, and you don't know exactly how many of each are inside. Instead of picking one guess, you keep every guess that fits what you know. Then you can say the best and worst chance of pulling a red one.

Keeping Every Fitting Chance

Sometimes we don't know enough to say exactly how likely something is, like whether a strange coin lands heads. A credal set keeps every way of assigning chances that fits the clues we have, instead of pretending we know just one. Then, for any question, we can find the lowest and highest answer those possibilities give. The spread between them shows how much we don't know about the model itself.

Set of Possible Probability Models

In probability, you normally describe uncertainty with one probability distribution: one assignment of chances to outcomes. A Credal Set instead keeps a whole set of distributions, all the ones consistent with limited information, bounds someone gave you, or other constraints. For any quantity you care about, you can then compute a lowest expected value and a highest expected value across the set. That gap is different from ordinary statistical error around one estimate: it shows that the model itself is not pinned down. The set might or might not include all the in-between mixtures of its members; that is an extra choice, not part of the definition. And the set by itself does not tell you how to update on new evidence or which action to take.

A Credal Set K is a set of probability measures admitted as possible representations of uncertainty, used in imprecise-probability theory instead of a single distribution P. It collects every P consistent with elicited bounds, partial information, robustness neighborhoods, or coherence constraints. For a bounded function f, it induces a lower expectation (the infimum of E_P[f] over K) and an upper expectation (the supremum); these envelope functionals express model imprecision, not sampling variability. Credal sets are often taken to be convex and closed, but those are extra modeling choices: convexity has a behavioral rationale when mixtures of admissible models should also be admissible, while a nonconvex set can make sense when it encodes a disjunction of structured models. Closure guarantees extrema are attained but can add boundary distributions, and for closed convex polytopes linear expectations are optimized at extreme points. Walley developed coherent lower previsions and their relation to sets of linear previsions, and cautioned that no single representation (lower probabilities, belief functions, sets of measures, desirable gambles) is universally sufficient. Credal networks add graph structure and independence assumptions on top; a bare credal set contains no graph and does not fix conditioning, combination, or decision rules.

Structural Signature¶

Outcome space. A measurable or finite state space is declared.
Probability objects. Every member satisfies the selected probability axioms.
Membership criterion. Constraints determine which measures are admissible.
Nonemptiness. Coherent use requires at least one admissible probability model.
Set-level uncertainty. Imprecision is represented by multiplicity across probability models.
Lower envelope. Infima of member expectations or event probabilities supply conservative bounds.
Upper envelope. Suprema supply the paired upper bounds.
Regularity declaration. Convexity, closure, compactness, additivity, and domination are explicit.
Extreme-point representation. Finite convex sets may be summarized by vertices for linear queries.
Updating rule. Conditioning or revision is chosen rather than inferred from membership alone.
Independence semantics. Product, strong, epistemic, or other independence concepts are stated separately.
Decision rule. Maximin, maximality, E-admissibility, or another choice criterion is not hidden inside the set.

What It Is Not¶

Not a single probability distribution. Its identity is multiplicity of admissible models.
Not necessarily a confidence set. Frequentist coverage over repeated samples is a different guarantee.
Not automatically convex. Convexity is common and useful but must be declared.
Not merely an interval probability. A family of event intervals can lose joint constraints and may not uniquely determine one credal set.
Not a Dempster–Shafer belief function. That is a specific representation with additional structure.
Not a credal network. A network adds graph-based factorization and independence semantics.
Not automatically robust Bayes. Robust Bayesian analysis is one use and needs priors, likelihoods, and updating choices.
Not a decision rule. The set alone does not select an action under incomparable expected utilities.

Scope of Application¶

Credal sets are literal where uncertainty about the correct probability model is retained explicitly as a family rather than collapsed prematurely into one distribution.

Expert elicitation. Probability bounds or partial orderings define admissible measures.
Robust Bayesian analysis. Classes of priors or likelihoods propagate model sensitivity to posterior conclusions.
Incomplete data. Partial constraints preserve uncertainty not justified by observation.
Credal networks. Local conditional credal sets combine with graph semantics to form global models.
Decision analysis. Lower and upper expected utilities reveal robust dominance and unresolved choice.
Risk bounds. Worst- and best-case probabilities are computed over a defensible ambiguity set.
Probability boxes and interval models. Representation relationships can be studied through induced sets of measures.
Sensitivity analysis. Conclusions are checked across probability models rather than around one parameter estimate only.

Clarity¶

Define the outcome space, sigma-algebra if needed, additivity convention, and exact membership constraints. State whether the set is nonempty, closed, convex, compact, finitely generated, or dominated. Distinguish lower and upper envelopes using infimum/supremum unless attainment is justified. If extreme points are used, explain why optimizing the query over them suffices. Do not conflate epistemic imprecision with sampling variance, parameter confidence regions, or aleatory variability. Specify conditioning, independence, and combination rules before reporting posterior credal sets. For interval constraints, preserve cross-event coherence. Name the decision criterion separately and report when admissible distributions support conflicting rankings.

Manages Complexity¶

A complex uncertainty assessment may contain ranges, partial preferences, competing models, and weak evidence. A credal set represents all precise probability models consistent with those commitments, allowing queries to become optimization problems rather than arbitrary point selection. Convex geometry, extreme points, and linear programming can reduce computation. The representation can nevertheless grow exponentially; conditioning can create new constraints; independence notions diverge; and lower/upper summaries may hide which distributions attain extrema. The abstraction manages complexity by preserving model multiplicity and separating representation from update and choice, not by pretending ambiguity has disappeared.

Abstract Reasoning¶

Define the outcome space and whether probabilities are finite or countably additive.
Translate elicited bounds, moment conditions, likelihoods, or robustness assumptions into membership constraints.
Check coherence and nonemptiness of the resulting probability set.
Decide whether convexification and closure preserve the intended meaning.
Choose an explicit representation such as inequalities, vertices, generators, or a graphical factorization.
Compute lower and upper probabilities or expectations over all members.
Identify extremizing models and sensitivity to constraint changes.
Apply a declared conditioning and independence semantics when evidence arrives.
Use a named decision rule and report incomparability instead of selecting silently.
Preserve the constraints and update provenance with every reported envelope.

Knowledge Transfer¶

The strict parent is Set and Membership. A credal set is literally a collection whose elements are probability measures and whose defining work lies in an admissibility predicate. Set and Membership supplies the substrate-independent grouping relation; probability, convexity, lower envelopes, and updating form the domain accent. Probability is a close ingredient but the candidate is uncertainty about which probability model applies, not one probability assignment.

Examples¶

Canonical¶

For a Boolean event \(A\), suppose the only justified statement is \(0.3\leq P(A)\leq0.6\). The credal set contains all Bernoulli distributions with parameter \(p\in[0.3,0.6]\). For payoff \(f=10\mathbf1_A\), the lower expectation is \(3\) and the upper expectation is \(6\). The interval reflects imprecision about \(p\), not the variance of one Bernoulli random variable or a sampling confidence interval.

Mapped back: partial probability constraint → set of admissible precise models → lower/upper expectation envelope → preserved model imprecision.

Applied / In Practice¶

A classifier's class probabilities are constrained by several experts and calibration bounds. Their compatible distributions form a convex polytope. A decision is robust only if one action has greater expected utility at every vertex; otherwise the analyst reports incomparability or applies a separately justified criterion. If features are arranged in a credal network, the graph's independence and extension semantics are documented before inference.^[3]

Mapped back: expert and calibration constraints → convex credal polytope → vertex-wise utility comparison → robust choice or explicit incomparability.

Structural Tensions¶

Expressiveness vs. computation. Rich sets preserve ambiguity but make inference harder. Diagnostic: Which constraints materially affect the queried envelope?
Convexity vs. model identity. Mixtures simplify geometry but may add unintended models. Diagnostic: Does mixing two admissible models remain substantively admissible?
Envelope summary vs. joint structure. Event-wise intervals can hide dependencies. Diagnostic: Is the underlying set retained with the reported bounds?
Imprecision vs. indecision. Wide sets may leave actions incomparable. Diagnostic: Is additional information or an explicit decision rule needed?
Updating vs. dilation. Conditioning can widen uncertainty under some dependence assumptions. Diagnostic: Which conditioning and independence semantics generated the posterior?
Autonomous object vs. generic Set. Every collection is a set. Diagnostic: Are its members probability measures representing unresolved model choice and supporting envelope inference?

Structural–Framed Character¶

A set of admissible probability measures, membership constraints, and lower/upper envelope queries are structural. Outcome space, additivity, convexity, closure, generators, conditioning, independence, and decision rule are framed. A credal set represents uncertainty about probability models; it does not by itself quantify sampling confidence, supply causality, guarantee computational tractability, prescribe a unique update, or choose among actions.

Structural Core vs. Domain Accent¶

The transferable skeleton is Set and Membership under a predicate. The domain accent is probability-measure members, imprecise belief, convex geometry, lower and upper previsions, extreme models, conditioning, and robust decisions. Removing probability semantics yields a generic set; selecting one member yields a precise probability model; adding a graph and factorization yields a Credal Network.

This entry is a kind of Set and Membership.

Set and Membership is the broader abstraction, because a credal set groups probability measures by explicit admissibility constraints. The specialization is literal, and it does not collapse the member-level operation of Probability into the set-level representation of uncertainty.

Relationships to Other Abstractions¶

Current abstraction Credal Set Domain-specific

Parents (1) — outgoing typed relations

Credal Set is a kind of Set and Membership Prime

Set and Membership is the strict parent because a credal set groups probability measures by explicit admissibility constraints.

Hierarchy path (1) — routes to 1 parentless root

Credal Set → is a kind of Set and Membership

Neighborhood in Abstraction Space¶

Credal Set sits in a sparse region of the domain-specific corpus (81^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Foundations of Probability & Inference (29 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25

Not to Be Confused With¶

Probability Distribution. One precise member rather than a set of candidate models.
Confidence Set. A random inferential region with repeated-sampling coverage semantics.
Probability Box. A bound on cumulative distributions that induces a particular set of measures.
Belief Function. A Dempster–Shafer representation with focal-set structure.
Credal Network. A graphical model coupling credal sets with factorization and independence.
Robust Bayesian Analysis. A workflow propagating sets of priors or models through Bayesian updating.
Interval Arithmetic. Bounds numerical computation without necessarily representing probability-model ambiguity.

References¶

[1] Peter Walley, Statistical Reasoning with Imprecise Probabilities, Monographs on Statistics and Applied Probability 42 (Chapman & Hall, 1991), ISBN 978-0-412-28660-5. registry ↩

[2] Peter Walley, “Towards a Unified Theory of Imprecise Probability,” International Journal of Approximate Reasoning 24, nos. 2–3 (2000): 125–148, https://doi.org/10.1016/S0888-613X(00)00031-1. registry ↩

[3] Fabio G. Cozman, “Credal Networks,” Artificial Intelligence 120, no. 2 (2000): 199–233, https://doi.org/10.1016/S0004-3702(00)00029-1. registry ↩a ↩b Show verification details