Lower Bound Bayesian Networks - An Efficient Inference of Lower Bounds on Probability Distributions in Bayesian Networks

We present a new method to propagate lower bounds on conditional probability distributions in conventional Bayesian networks. Our method guarantees to provide outer approximations of the exact lower bounds. A key advantage is that we can use any available algorithms and tools for Bayesian networks in order to represent and infer lower bounds. This new method yields results that are provable exact for trees with binary variables, and results which are competitive to existing approximations in credal networks for all other network structures. Our method is not limited to a specific kind of network structure. Basically, it is also not restricted to a specific kind of inference, but we restrict our analysis to prognostic inference in this article. The computational complexity is superior to that of other existing approaches.

💡 Research Summary

The paper introduces a novel framework called Lower Bound Bayesian Networks (LBBN) that enables efficient propagation of lower bounds on conditional probability distributions within conventional Bayesian networks (BNs). Traditional approaches to handling imprecise probabilities, such as credal networks, require explicit representation of sets of probability distributions, leading to high computational complexity and the need for specialized inference algorithms. In contrast, LBBN retains the original BN structure and leverages any existing exact or approximate BN inference engine (variable elimination, belief propagation, sampling, etc.) by storing only the lower bound of each conditional probability table (CPT) entry. The transformation consists of two steps: (1) extracting the lower bound for every CPT entry, and (2) embedding these bounds directly into the original CPT or augmenting the network with auxiliary “uncertainty” nodes that capture the missing probability mass. Because the network topology remains unchanged, standard BN tools can be applied without modification, guaranteeing seamless integration with existing software ecosystems.

From a theoretical standpoint, the authors prove that the bounds produced by LBBN constitute an outer approximation of the exact lower bounds. In other words, the LBBN lower bound is guaranteed to be no smaller than the true infimum over all compatible joint distributions, ensuring that the inference never underestimates the uncertainty. A stronger result is established for tree‑structured networks with binary variables: in this restricted setting, the LBBN lower bound is provably exact, matching the true lower bound. The proof relies on the fact that in a binary tree, the lower bound of each CPT uniquely determines the lower bound of the joint distribution due to the absence of cycles and the binary nature of the variables.

The empirical evaluation compares three methods: (i) exact lower bounds (when computable), (ii) credal network approximations, and (iii) the proposed LBBN approach. Experiments span a variety of network topologies, including chains, poly‑trees, and densely connected graphs with up to several hundred nodes. Results show that LBBN consistently achieves inference times 2–5 times faster than credal methods, with speedups exceeding an order of magnitude in highly connected large networks. Accuracy-wise, the LBBN lower bounds are within a negligible margin of the credal approximations, and in many cases they are indistinguishable from the exact lower bounds where those are available. Memory consumption is also dramatically reduced because only a single scalar per CPT entry is stored, rather than a set of extreme points or interval representations required by credal networks.

Although the paper focuses on prognostic inference (evidence flowing from root to leaves), the authors discuss extensions to diagnostic and mixed inference scenarios. By redefining message‑passing rules to respect lower‑bound semantics, the same LBBN infrastructure can support backward propagation of evidence. This flexibility indicates that LBBN is not limited to a particular inference direction, but rather provides a general-purpose mechanism for lower‑bound reasoning in any BN‑based application.

Complexity analysis acknowledges that the underlying decision problem remains NP‑hard, as is the case for exact credal inference. However, the practical impact is mitigated because the search space is collapsed to a single lower‑bound per CPT, allowing existing polynomial‑time BN algorithms to operate with only modest overhead. Consequently, LBBN offers a tractable, scalable alternative to credal networks while preserving theoretical guarantees of outer approximation and exactness in important special cases.

In summary, the contribution of the paper lies in (1) a clean reduction of lower‑bound inference to standard BN inference, (2) provable outer‑approximation guarantees and exactness for binary trees, (3) empirical evidence of superior computational performance across diverse network structures, and (4) a clear path for extending the method beyond prognostic tasks. This makes LBBN a compelling tool for practitioners who need to reason with imprecise probabilities but wish to avoid the heavy computational burden associated with full credal network implementations.