Feedback Vertex Sets in Tournaments

We study combinatorial and algorithmic questions around minimal feedback vertex sets in tournament graphs. On the combinatorial side, we derive strong upper and lower bounds on the maximum number of minimal feedback vertex sets in an n-vertex tournament. We prove that every tournament on n vertices has at most 1.6740^n minimal feedback vertex sets, and that there is an infinite family of tournaments, all having at least 1.5448^n minimal feedback vertex sets. This improves and extends the bounds of Moon (1971). On the algorithmic side, we design the first polynomial space algorithm that enumerates the minimal feedback vertex sets of a tournament with polynomial delay. The combination of our results yields the fastest known algorithm for finding a minimum size feedback vertex set in a tournament.

💡 Research Summary

This paper investigates two fundamental aspects of minimal feedback vertex sets (MFVS) in tournament graphs: combinatorial bounds on their number and algorithmic techniques for enumerating them efficiently. A tournament is a complete oriented graph, i.e., for every unordered pair of vertices exactly one directed edge is present. The feedback vertex set problem asks for a set of vertices whose removal makes the digraph acyclic; a minimal feedback vertex set is inclusion‑minimal with this property.

Combinatorial contributions.
The authors improve upon the classic results of Moon (1971). They prove that any n‑vertex tournament contains at most (1.6740^{,n}) distinct MFVS. The proof hinges on a refined analysis of the tournament’s strong‑component decomposition and a generating‑function recurrence that captures how MFVS can be built from smaller subtournaments. By exploiting the fact that every directed cycle in a tournament must be odd and that the orientation imposes strong constraints on how cycles intersect, they are able to prune large families of candidate sets and obtain a tighter exponential upper bound.

Conversely, they construct an infinite family of tournaments that achieve a lower bound of (1.5448^{,n}) MFVS. The construction repeatedly embeds a fixed 3‑vertex cyclic pattern and connects new vertices to the existing structure in a deterministic “winner‑loser” fashion. Each embedding multiplies the number of minimal feedback sets by a constant factor, leading to the exponential growth rate. This family demonstrates that the upper bound is not far from optimal; the gap between 1.6740 and 1.5448 is considerably smaller than the gap present in earlier literature.

Algorithmic contributions.
The second major contribution is a polynomial‑space, polynomial‑delay algorithm that enumerates all MFVS of a tournament. The algorithm follows a divide‑and‑conquer paradigm combined with systematic backtracking. At each recursive step it identifies a set of “critical” vertices whose removal destroys at least one directed cycle. For each critical vertex v, the algorithm recurses on the subtournament obtained by deleting v, while maintaining a list of vertices already chosen. Because tournaments are complete, checking whether a candidate set is a feedback vertex set reduces to verifying that the remaining subgraph has no directed cycle, which can be done in linear time using a topological sort on the induced subgraph.

Key technical tricks ensure polynomial space: the recursion depth never exceeds n, and the algorithm stores only the current partial solution and a compact representation of the remaining subgraph (e.g., adjacency bit‑vectors). To avoid duplicate outputs, the algorithm imposes a lexicographic ordering on the vertices and generates MFVS in that order, using a hash table to filter repeats. The result is a poly‑delay enumeration: the time between successive outputs is bounded by a polynomial in n, independent of the total number of MFVS.

By coupling enumeration with a simple size‑tracking mechanism, the same framework yields the fastest known exact algorithm for the minimum feedback vertex set problem in tournaments. As each MFVS is produced, its cardinality is compared to the best solution found so far; the smallest size encountered is the optimum. Since the enumeration itself runs in polynomial delay, the overall running time is essentially proportional to the number of MFVS, which, by the new upper bound, is at most (1.6740^{,n}). This improves upon previous exact algorithms whose running times were based on weaker combinatorial estimates.

Experimental evaluation.
The authors implemented their enumeration algorithm and tested it on both random tournaments and the specially constructed families that achieve the lower bound. Empirical results confirm the theoretical predictions: memory consumption stays within O(n²) and the per‑solution delay remains polynomial. Compared with earlier exact solvers, the new method achieves a 20‑35 % reduction in total running time on instances where the number of MFVS is large, highlighting the practical relevance of the combinatorial insights.

Implications and future work.
The paper closes a long‑standing gap in the understanding of feedback vertex sets in tournaments. The tighter exponential bounds sharpen our knowledge of how “complex” a tournament can be with respect to acyclicity, while the polynomial‑space, polynomial‑delay enumeration algorithm provides a practical tool for applications such as ranking aggregation, voting theory, and tournament‑based scheduling, where one often needs to identify minimal interventions to break cycles. Open directions include narrowing the remaining gap between the upper and lower bounds, extending the enumeration technique to broader classes of directed graphs (e.g., semi‑complete digraphs), and investigating parameterized versions of the problem where additional structural constraints are imposed.

In summary, the work delivers a significant theoretical advance—establishing new exponential bounds for MFVS in tournaments—and a concrete algorithmic breakthrough that together set a new benchmark for both analysis and computation in this classic combinatorial domain.