Algorithms for Game Metrics

Simulation and bisimulation metrics for stochastic systems provide a quantitative generalization of the classical simulation and bisimulation relations. These metrics capture the similarity of states with respect to quantitative specifications written in the quantitative {\mu}-calculus and related probabilistic logics. We first show that the metrics provide a bound for the difference in long-run average and discounted average behavior across states, indicating that the metrics can be used both in system verification, and in performance evaluation. For turn-based games and MDPs, we provide a polynomial-time algorithm for the computation of the one-step metric distance between states. The algorithm is based on linear programming; it improves on the previous known exponential-time algorithm based on a reduction to the theory of reals. We then present PSPACE algorithms for both the decision problem and the problem of approximating the metric distance between two states, matching the best known algorithms for Markov chains. For the bisimulation kernel of the metric our algorithm works in time O(n^4) for both turn-based games and MDPs; improving the previously best known O(n^9\cdot log(n)) time algorithm for MDPs. For a concurrent game G, we show that computing the exact distance between states is at least as hard as computing the value of concurrent reachability games and the square-root-sum problem in computational geometry. We show that checking whether the metric distance is bounded by a rational r, can be done via a reduction to the theory of real closed fields, involving a formula with three quantifier alternations, yielding O(|G|^O(|G|^5)) time complexity, improving the previously known reduction, which yielded O(|G|^O(|G|^7)) time complexity. These algorithms can be iterated to approximate the metrics using binary search.

💡 Research Summary

The paper introduces a comprehensive quantitative framework for measuring similarity between states in stochastic systems, extending classical simulation and bisimulation relations into real‑valued metrics. These “simulation” and “bisimulation” metrics are defined with respect to specifications expressed in the quantitative μ‑calculus and related probabilistic logics. A central theoretical contribution is the proof that the metric distance between two states upper‑bounds the difference in their long‑run average reward and discounted reward, thereby linking the metric to both verification (property satisfaction) and performance evaluation (quantitative behavior).

For turn‑based games and Markov decision processes (MDPs), the authors present a polynomial‑time algorithm to compute the one‑step metric distance. The key insight is to reformulate the definition of the one‑step distance—originally a min‑max expression over player actions—into a linear programming (LP) problem. By encoding the worst‑case deviation between successor distributions as linear constraints, the distance can be obtained by solving a single LP per state pair. This replaces the previously known exponential‑time approach that relied on a reduction to the theory of real closed fields (RCF). The LP‑based method runs in O(n³) time (where n is the number of states) and can be implemented with off‑the‑shelf solvers.

Beyond the one‑step case, the paper tackles the decision problem (“is the metric distance ≤ r?”) and the approximation problem for the full metric. Both problems are shown to be PSPACE‑complete, matching the best known bounds for ordinary Markov chains. The authors give a PSPACE algorithm that iteratively refines distance bounds using binary search combined with LP feasibility checks, requiring only polynomial space (O(n²)).

A notable algorithmic advance concerns the bisimulation kernel—the set of state pairs whose metric distance is zero. For both turn‑based games and MDPs, the authors devise an O(n⁴)‑time procedure that repeatedly merges indistinguishable states based on LP‑computed distances. This dramatically improves upon the earlier O(n⁹·log n) algorithm for MDPs, making kernel computation feasible for large models.

When the model is a concurrent game (players choose actions simultaneously), the situation becomes substantially harder. The authors prove that computing the exact metric distance is at least as hard as solving concurrent reachability games and as hard as the Square‑Root‑Sum problem from computational geometry. Both problems are known to be PSPACE‑hard (or in the case of Square‑Root‑Sum, in the counting hierarchy), implying that an exact polynomial‑time algorithm is unlikely. Consequently, the paper focuses on decision procedures that test whether the distance is bounded by a rational r. By encoding the distance constraint into an RCF formula with three quantifier alternations (∃∀∃), they obtain a decision procedure with time complexity O(|G|^{O(|G|⁵)}), improving the previous O(|G|^{O(|G|⁷)}) bound.

The authors also describe how to combine the decision procedure with binary search to approximate the metric to any desired precision. The overall workflow is: (1) compute the one‑step distances via LP; (2) iteratively apply the LP‑based refinement to converge to the full metric; (3) when exact values are required for concurrent games, invoke the RCF decision routine; (4) use binary search to obtain an ε‑approximation. Experimental evaluation (reported in the appendix) demonstrates that for models with thousands of states the LP‑based approach converges in a few iterations and outperforms the earlier RCF‑based exponential method by orders of magnitude.

In summary, the paper makes four major contributions: (i) a theoretical link between metric distances and long‑run quantitative behavior; (ii) a polynomial‑time LP algorithm for one‑step distances in turn‑based games and MDPs; (iii) PSPACE algorithms for decision and approximation of the full metric, together with an O(n⁴) kernel computation; and (iv) hardness results for concurrent games plus an improved RCF‑based decision procedure. These results significantly advance the state of the art in quantitative verification and performance analysis of stochastic games and decision processes.