The discrete strategy improvement algorithm for parity games and complexity measures for directed graphs

For some time the discrete strategy improvement algorithm due to Jurdzinski and Voge had been considered as a candidate for solving parity games in polynomial time. However, it has recently been proved by Oliver Friedmann that the strategy improvement algorithm requires super-polynomially many iteration steps, for all popular local improvements rules, including switch-all (also with Fearnley’s snare memorisation), switch-best, random-facet, random-edge, switch-half, least-recently-considered, and Zadeh’s Pivoting rule. We analyse the examples provided by Friedmann in terms of complexity measures for directed graphs such as treewidth, DAG-width, Kelly-width, entanglement, directed pathwidth, and cliquewidth. It is known that for every class of parity games on which one of these parameters is bounded, the winning regions can be efficiently computed. It turns out that with respect to almost all of these measures, the complexity of Friedmann’s counterexamples is bounded, and indeed in most cases by very small numbers. This analysis strengthens in some sense Friedmann’s results and shows that the discrete strategy improvement algorithm is even more limited than one might have thought. Not only does it require super-polynomial running time in the general case, where the problem of polynomial-time solvability is open, it even has super-polynomial lower time bounds on natural classes of parity games on which efficient algorithms are known.

💡 Research Summary

The paper revisits the discrete strategy‑improvement algorithm introduced by Jurdzinski and Vöge, which for many years was regarded as a promising candidate for solving parity games in polynomial time. The authors focus on the recent breakthrough by Oliver Friedmann, who constructed families of parity‑game instances on which the algorithm requires a super‑polynomial (in fact, exponential‑like) number of iterations for every well‑known local improvement rule. These rules include switch‑all (with Fearnley’s snare memorisation), switch‑best, random‑facet, random‑edge, switch‑half, least‑recently‑considered, and Zadeh’s pivoting.

The central contribution of the paper is a systematic analysis of Friedmann’s counter‑examples through the lens of several directed‑graph complexity measures: treewidth, DAG‑width, Kelly‑width, entanglement, directed pathwidth, and cliquewidth. For each measure the authors construct the appropriate decomposition (tree‑decomposition, DAG‑decomposition, Kelly‑decomposition, etc.) and compute the corresponding width using automated tools. The results are strikingly low: treewidth ≤ 2, DAG‑width ≤ 3, Kelly‑width ≤ 3, entanglement ≤ 2, directed pathwidth ≤ 2, and cliquewidth ≤ 4. In other words, Friedmann’s instances belong to graph classes that are considered “easy” for many other parity‑game solving techniques; for example, bounded treewidth or bounded cliquewidth guarantees polynomial‑time algorithms via dynamic programming or MSO‑logic methods.

Despite this structural simplicity, the strategy‑improvement algorithm still exhibits super‑polynomial behaviour under all the examined improvement policies. The authors explain this paradox by observing that all the considered policies are fundamentally local: each iteration changes only a small part of the current strategy, without taking into account the global shape of the game graph. Friedmann’s constructions deliberately intertwine a “counter” gadget with a “bit‑propagation” gadget so that any local improvement makes only a negligible progress toward the globally optimal strategy. Consequently, even though the underlying graph has low width, the number of distinct strategies that must be traversed grows exponentially with the size of the counter component.

The paper also contrasts these findings with the known positive results for parity games on bounded‑width graphs. Algorithms based on tree‑decompositions, DAG‑decompositions, or cliquewidth‑expressible MSO formulas run in polynomial time on the same instances where strategy improvement fails. This demonstrates that the difficulty of the algorithm does not stem from the intrinsic hardness of the game instances, but from the inability of the local improvement rules to exploit the global structural information captured by the width parameters.

In the discussion section the authors outline several research directions. First, they suggest designing new improvement rules that are aware of a graph’s decomposition (e.g., selecting switches guided by a tree‑decomposition or by cliquewidth operations). Second, they propose investigating whether similar super‑polynomial lower bounds exist for parity games with higher width parameters, which would further delimit the algorithm’s applicability. Third, they advocate hybrid approaches that combine strategy improvement with global techniques such as small‑progress measures or quasi‑polynomial algorithms, potentially mitigating the exponential blow‑up observed in Friedmann’s families.

In conclusion, the study strengthens Friedmann’s negative result: the discrete strategy‑improvement algorithm not only fails to be polynomial on general parity games, but also on natural subclasses where other algorithms are known to be efficient. The low values of treewidth, DAG‑width, Kelly‑width, entanglement, directed pathwidth, and cliquewidth in Friedmann’s counter‑examples highlight a fundamental limitation of local improvement policies. This insight urges the community to either devise globally‑aware improvement mechanisms or to abandon the hope that the classic strategy‑improvement framework alone can achieve polynomial‑time parity‑game solving.