Credal Bayesian Networks

• Credal Bayesian Networks are graphical models using convex sets of conditionals instead of precise probability tables to represent uncertainty.
• They leverage a directed acyclic graph structure to support worst-case analysis and robust decision-making under both epistemic and aleatory uncertainty.
• Inference and optimization techniques such as multilinear programming and variable elimination are employed to manage non-convex constraints in CrBNs.

A Credal Bayesian Network (CrBN) is a graphical probabilistic model defined on a directed acyclic graph (DAG), in which local probabilistic uncertainty is represented not as precise conditional probability tables (CPTs), but by convex sets of conditional distributions termed credal sets. CrBNs generalize Bayesian networks (BNs) to allow for imprecise, indeterminate, or qualitative probability assessments, supporting robust modeling under epistemic or aleatory uncertainty and supporting worst-case analysis in the presence of distributional ambiguity.

1. Mathematical Structure and Semantics

A CrBN is specified by a DAG $G = (V, E)$ over discrete random variables $V_1, \ldots, V_n$, with each node $V_i$ and each parent configuration $u_i$ associated to a closed convex set $C_{V_i|u_i}$ of admissible row-vectors (credible conditionals) in the probability simplex of $\mathrm{Dom}(V_i)$.

Local Specification

For each node and parent configuration,

$$C_{V_i|u_i} = \left\{ \theta_{.\mid u_i} : A^{(i,u_i)}\theta_{.\mid u_i} \leq b^{(i,u_i)},\ \sum_v \theta_{v|u_i} = 1,\ \theta_{v|u_i} \ge 0\right\}$$

where $A^{(i,u_i)}$ and $b^{(i,u_i)}$ encode linear constraints or interval bounds.

Global Semantics: Strong Extension

Under strong extension, the global joint credal set is the convex hull of all product-form distributions constructed from extreme points of each local set: $$\mathcal{E}(G,\mathbb{K}) = \mathrm{conv}\left\{ p(v_1,\ldots,v_n) = \prod_{i=1}^n q_i(v_i|pa_i) : q_i(.|u_i)\in \mathrm{ext}(C_{V_i|u_i})\ \forall i,u_i \right\}$$ This structure generalizes BNs (recovered when all $V_1, \ldots, V_n$0 are singletons).

Irrelevance and Independence Notions

• Strong independence: Each extreme point of the joint credal set factorizes per the DAG, corresponding to stochastic independence at the distribution level.
• Epistemic irrelevance: A weaker, set-level irrelevance property where, for each variable $V_1, \ldots, V_n$1, conditioning on non-parent non-descendant variables does not alter the conditional credal set for $V_1, \ldots, V_n$2 given parents. This is asymmetric and may not satisfy standard graphoid axioms (Maua et al., 2013, Bock, 2017).

2. Inference and Optimization

Marginal and Conditional Bounds

For event $V_1, \ldots, V_n$3 (e.g., an assignment to variables), lower and upper probabilities are defined as: $V_1, \ldots, V_n$4 Computation typically reduces to non-convex multilinear programming due to the product-form constraints (Cozman et al., 2012).

Algorithms

• Exact Multilinear Programming (MP) and Reformulation–Linearization (RL): Global optimization over local constraint variables and auxiliary variables, with the non-convexities handled via Sherali–Tuncbilek reformulations (Cozman et al., 2012).
• Variable Elimination and Message Passing: For binary polytrees, exact polynomial-time solutions leveraging interval message propagation (the 2U engine) (Cozman et al., 2012, Rocha et al., 2012).
• Constraint Relaxation on Circuits: Compilation of the BN structure into an arithmetic circuit or sum–product network, followed by local LPs/QPs and constraint relaxation provides linear-time upper bounds on maximum marginal probabilities (MARₘₐₓ) (Wijk et al., 2022).
• Approximation Schemes: For general topologies, approaches include partial evaluation with cutsets, loopy interval propagation (L2U), and linear-programming-based relaxation schemes (Cozman et al., 2012, Andrade et al., 2012).

Maximum Marginal Probability (MARₘₐₓ):

Given event $V_1, \ldots, V_n$5, MARₘₐₓ is: $V_1, \ldots, V_n$6 which can be cast as a constrained optimization over an induced probabilistic circuit (Wijk et al., 2022).

Lower Bound Bayesian Networks (LBBNs)

Transforms any CrBN with interval-valued CPTs into an augmented BN by appending an "ignorance" state per node; standard BN inference yields guaranteed outer approximations (and exactness for binary trees) (Andrade et al., 2012).

3. Computational Complexity

The inferential complexity of CrBNs is heavily dependent on the irrelevance concept and network structure:

Structure	Epistemic Irrelevance	Strong Independence
Markov chains/HMMs	P	P
Polytrees (binary)	P	P
Polytrees (general)	NP-hard	NP-hard
General DAGs	$V_1, \ldots, V_n$7-hard	$V_1, \ldots, V_n$8-hard

• For polytrees, even with ternary variables, computation in the strong extension is NP-hard (Maua et al., 2013, Rocha et al., 2012).
• General inference is NP$V_1, \ldots, V_n$9-complete, matching the standard MAP complexity in BNs (Cozman et al., 2012).
• Epistemic irrelevance yields polynomial-time algorithms only in linear structures or undirected trees, and is otherwise intractable (Maua et al., 2013, Bock et al., 2012, Bock, 2017).

4. Qualitative, Relational, and Logical Extensions

CrBNs can represent not only numeric constraints, but also qualitative (e.g., influences, synergies) and indeterminate assessments:

• Qualitative Constraints: Monotonicity and qualitative influences yield linear inequalities among CPT entries. E.g., "positive influence" constraints (Cozman et al., 2012).
• Relational/First-Order CrBNs: CrBN templates can be specified over first-order structures; each grounding induces local credal sets (Cozman et al., 2012).
• Logical Credal Networks (LCNs): Generalizes CrBNs to enable probability bounds on arbitrary (propositional, FOL) formulas, supports cycles, and aggregates overlapping assessments; semantics are given via an extended Markov property on a dependency structure induced by logical "stamps" (Qian et al., 2021).

5. Model Selection, Inference Heuristics, and Robustness Analysis

Maximum Entropy Models

• Sequential Maximum Entropy: To pick a unique representative, solve a series of local entropy maximizations per variable in topological order, respecting all local credal constraints. This construction preserves DAG structure and yields a product-form model in the precise case (Lukasiewicz, 2013).

Robustness Algorithms

• CUB (Credal Upper Bound): Compiles the CrBN into an arithmetic/sum-product network, performs local relaxations, and computes worst-case marginal upper bounds in linear time, yielding provable (often tight) outer bounds for large models (Wijk et al., 2022).
• Projection and Local Search: Project circuit-induced weights back to the space of consistent CPTs, improving lower bounds via greedy local updates (CLB) (Wijk et al., 2022).

Empirical Observations

• CUB and CLB outperform ApproxLP both in accuracy and scalability, uniquely handling "hard" credal-inference queries on large graphs (e.g., the hepar2 network) (Wijk et al., 2022).
• LBBN provides competitive accuracy and superior complexity versus traditional outer-approximation methods, being exact on binary trees (Andrade et al., 2012).

6. Applications, Limitations, and Open Questions

Applications include robust decision-making under uncertainty, safety-critical systems, reliability modeling with partial knowledge, qualitative expert systems, and fusion of multiple uncertain or qualitative information sources.

Limitations and research frontiers:

• Exact inference is generally intractable except for low-treewidth, binary, or epistemically-irrelevant tree structures.
• Approximate, scalable methods and tighter bounding schemes for general topologies are active areas of research (Maua et al., 2013).
• Open problems include FPTAS development for epistemic irrelevance, efficient lifted methods for relational/first-order CrBNs, and extending frameworks to continuous variables and hybrid models (Bock, 2017, Qian et al., 2021).

CrBNs provide a mathematically rigorous framework for distributionally robust modeling, explicitly quantifying the impact of parameteric uncertainty and supporting both interval-valued and logical/qualitative assessments, while remaining closely related to classical Bayesian and probabilistic logic networks (Cozman et al., 2012, Qian et al., 2021, Lukasiewicz, 2013, Wijk et al., 2022).