Minimum Entropy Orientations

We study graph orientations that minimize the entropy of the in-degree sequence. The problem of finding such an orientation is an interesting special case of the minimum entropy set cover problem previously studied by Halperin and Karp [Theoret. Comput. Sci., 2005] and by the current authors [Algorithmica, to appear]. We prove that the minimum entropy orientation problem is NP-hard even if the graph is planar, and that there exists a simple linear-time algorithm that returns an approximate solution with an additive error guarantee of 1 bit. This improves on the only previously known algorithm which has an additive error guarantee of log_2 e bits (approx. 1.4427 bits).

💡 Research Summary

The paper investigates the problem of orienting the edges of an undirected graph so that the entropy of the resulting indegree distribution is as small as possible. Formally, given a graph G=(V,E) with m=|E|, an orientation ~G assigns a direction to each edge; the indegree of a vertex v in ~G is denoted ρ_{~G}(v). The indegree distribution is p(v)=ρ_{~G}(v)/m and its entropy is H(p)=−∑_{v∈V} p(v)log₂p(v). The Minimum Entropy Orientation (MINEO) problem asks for an orientation that minimizes H(p).

The authors first establish the computational hardness of MINEO. By reducing from a planar, 3‑regular version of the 1‑in‑3 SAT problem (known to be NP‑complete), they construct a planar graph G whose optimal orientation encodes an exact cover of a set system derived from the SAT instance. The construction uses a gadget (Figure 2) for each element of the set system; each gadget forces a specific indegree pattern (e.g., (4,3,3,1,0,0)) in any optimal orientation. Lemma 1 shows that if one indegree sequence dominates another (in the sense of majorization), the corresponding entropy is strictly smaller. Using this “strict dominance” property, they prove that an optimal orientation exists with the prescribed pattern if and only if the original exact‑cover instance is solvable. Consequently, finding a minimum‑entropy orientation of a planar graph is NP‑hard. The proof also implies that any orientation problem possessing the strict dominance property is NP‑hard on planar graphs.

Next, the paper presents a simple linear‑time approximation algorithm with an additive error of at most one bit. The algorithm computes a “biased” orientation: for each edge (u,v), if deg(u)>deg(v) the edge is oriented from u to v; ties are broken arbitrarily. Let p̂ be the indegree distribution of this biased orientation. The authors prove (Theorem 2) that H(p̂) ≤ OPT(G)+1, where OPT(G) is the optimal entropy. The proof rewrites OPT(G) as log m minus the average of log ρ_{~G}(v) over all arcs, observes that ρ_{~G}(v) ≤ deg(v), and then relates H(p̂) to OPT(G) via the Kullback‑Leibler divergence D(p̂‖d) where d(v)=deg(v)/(2m) is the normalized degree distribution. Since D(p̂‖d)≥0, the inequality follows. Moreover, the biased orientation coincides with the optimal solution of a linear program that minimizes ∑_v –p(v)log deg(v)/m over the polytope P(G) of all indegree distributions; this LP can be solved by a greedy procedure that exactly builds the biased orientation. Hence the algorithm runs in O(|V|+|E|) time and achieves the claimed guarantee. The bound is tight (e.g., for cycles the bias orientation’s entropy is exactly OPT+1).

The authors also discuss structural properties of MINEO. Unlike the Minimum Sum Vertex Cover (MINSVC) problem, MINEO is “linear”: the optimal orientation of a disjoint union of graphs is simply the union of the optimal orientations of the components. This makes the problem more amenable to decomposition and analysis. They further describe the polytope P(G) of feasible indegree distributions, noting that its vertices correspond to acyclic orientations, and that the biased orientation can be interpreted as solving a cost‑minimization LP with costs c(v)=−log(deg(v)/m).

In summary, the paper makes two main contributions: (1) it proves that the Minimum Entropy Orientation problem remains NP‑hard even on planar graphs, using a reduction that leverages the strict dominance property of entropy; (2) it provides a linear‑time algorithm that produces a 1‑bit additive approximation, improving upon the previously known log₂e≈1.44‑bit guarantee. These results advance the theoretical understanding of entropy‑based graph orientation problems and have potential applications in areas such as computational biology (e.g., haplotype assembly) where entropy minimization is a natural objective.