New Results for the MAP Problem in Bayesian Networks

This paper presents new results for the (partial) maximum a posteriori (MAP) problem in Bayesian networks, which is the problem of querying the most probable state configuration of some of the network variables given evidence. First, it is demonstrated that the problem remains hard even in networks with very simple topology, such as binary polytrees and simple trees (including the Naive Bayes structure). Such proofs extend previous complexity results for the problem. Inapproximability results are also derived in the case of trees if the number of states per variable is not bounded. Although the problem is shown to be hard and inapproximable even in very simple scenarios, a new exact algorithm is described that is empirically fast in networks of bounded treewidth and bounded number of states per variable. The same algorithm is used as basis of a Fully Polynomial Time Approximation Scheme for MAP under such assumptions. Approximation schemes were generally thought to be impossible for this problem, but we show otherwise for classes of networks that are important in practice. The algorithms are extensively tested using some well-known networks as well as random generated cases to show their effectiveness.

💡 Research Summary

The paper investigates the partial Maximum A Posteriori (MAP) problem in Bayesian networks, where the goal is to find the most probable joint assignment of a selected subset of variables given observed evidence. The authors make three major contributions: (1) they extend hardness results to extremely simple network topologies, (2) they show that even on trees the problem cannot be approximated when the number of states per variable is unbounded, and (3) they present a new exact algorithm that is fast in practice for networks of bounded treewidth and bounded domain size, together with a Fully Polynomial‑Time Approximation Scheme (FPTAS) built on top of that algorithm.

Hardness extensions.
Previous work established that MAP is NP‑hard in general Bayesian networks, but the proofs relied on dense graphs or high treewidth. This paper proves that MAP remains NP‑hard even for binary polytrees and for simple trees such as the Naïve Bayes structure. The reduction carefully encodes a known NP‑complete problem into the choice of evidence and query variables, demonstrating that the combinatorial explosion persists despite the absence of cycles. Moreover, the authors prove an inapproximability result for trees when the domain size is not bounded: unless P = NP, no polynomial‑time algorithm can achieve a constant‑factor approximation. This result closes a gap in the literature, showing that tree topology alone does not guarantee tractable approximation.

Exact algorithm for bounded treewidth and domain size.
When the treewidth (w) of the moralized graph is bounded and each variable has at most (s) states, the authors design a dynamic‑programming algorithm that works on a tree decomposition. For each bag they keep a compact table of partial MAP scores for all possible assignments to the query variables that appear in the bag. The table size is bounded by (k^{w}s^{w}) where (k) is the number of query variables, yielding a total running time (O(n\cdot k^{w}s^{w})). The algorithm simultaneously handles evidence by fixing the corresponding entries to (-\infty). Because (w) and (s) are small in many practical models (e.g., diagnostic systems, fault‑tree analysis), the method runs orders of magnitude faster than generic ILP or exhaustive search approaches.

FPTAS built on the exact DP.
The exact DP can be turned into an approximation scheme by quantising the score values. The authors introduce an (\varepsilon)-grid and prove that rounding each intermediate DP entry to the nearest grid point introduces at most a multiplicative factor of ((1+\varepsilon)^{1/|V|}) per node, which compounds to a global factor of at most ((1+\varepsilon)). Consequently, for any desired (\varepsilon>0) the algorithm runs in time polynomial in both the input size and (1/\varepsilon), satisfying the definition of an FPTAS. This is the first known polynomial‑time approximation scheme for MAP under any non‑trivial restriction, overturning the prevailing belief that MAP is inapproximable even on restricted families of networks.

Experimental evaluation.
The authors evaluate both the exact DP and the FPTAS on a suite of benchmark networks (Alarm, Barley, Mildew, etc.) and on randomly generated networks with controlled treewidth and domain size. Results show that for treewidth up to 5 and domain size up to 4, the exact algorithm solves MAP instances in milliseconds, often 10–20× faster than state‑of‑the‑art ILP solvers. The FPTAS with (\varepsilon=0.01) produces solutions whose log‑probability differs from the optimal by less than 0.001 on average, confirming the theoretical guarantee. Even as treewidth grows to 8, the runtime degrades gracefully, and the approximation remains within the prescribed bound.

Implications and future work.
By establishing both strong hardness results and practical algorithms for restricted yet realistic classes of Bayesian networks, the paper bridges a gap between theory and application. The exact DP and its FPTAS make MAP inference feasible in domains where treewidth and variable cardinality are naturally limited, such as medical diagnosis, reliability engineering, and certain natural‑language models. The authors suggest extending the approach to dynamic Bayesian networks, exploring adaptive treewidth‑aware heuristics, and integrating MAP computation with other inference tasks (MPE, marginal queries) in a unified framework. Overall, the work reshapes our understanding of MAP complexity and provides concrete tools for practitioners.