In game theory, the Nash Equilibrium is the standard solution concept, but it is often not enough. Many games contain "unreasonable" equilibria that rely on players making threats they would never actually carry out. To filter these out, researchers use refinements like the Extensive-Form Perfect Equilibrium (EFPE) or the "trembling-hand" equilibrium. These concepts assume that players might make tiny, infinitesimal mistakes—a "tremble"—and a robust strategy must be a best response even as the probability of these mistakes approaches zero.
While theoretically elegant, these refinements push standard computational hardware to its breaking point. As Farina et al. note, modeling these mistakes introduces significant numerical instability. When the probability of an error becomes extremely small, the differences between strategy values can fall below the precision limits of standard 64-bit floating-point arithmetic.
The Precision Wall in Game Solvers
Most modern scientific computing relies on IEEE 754 floating-point numbers. However, Li et al. point out that exact equilibrium solvers frequently fail on large-scale games not because of time complexity, but because of numerical precision errors and overflows within their underlying linear programming subroutines.
This "precision wall" appears in several ways:
- Vanishing Probabilities: In stochastic games, transformations used to ensure the game eventually ends can create transition probabilities so small they effectively become zero in 64-bit representations. Kretinsky et al. found that even small models with fewer than 30 states can become unsolvable due to these infinitesimal probabilities.
- Double-Exponential Scaling: Certain game configurations require strategy probabilities to be double-exponentially small to reach an exact equilibrium. Asadi et al. and the authors argue that standard fixed-point notation simply cannot encode these values.
- Path Divergence: Algorithms that find equilibria by "following a path" through the solution space are particularly vulnerable. Gatti et al. explain that rounding errors can cause a solver to lose track of critical vertices, leading to total algorithm failure in sensitive game instances.
Symbolic and Arbitrary-Precision Solutions
To bypass these limits, researchers have turned to arbitrary-precision arithmetic and symbolic perturbations. Instead of using a specific small number like $10^{-15}$ to represent a player's mistake, Gatti et al. utilize symbolic variables. This allows the solver to maintain the exact logic of the "tremble" without ever rounding a value.
However, this precision comes at a cost. As Chatterjee et al. observe, requiring multi-precision arithmetic significantly increases the "bit-level complexity" of algorithms, making them slower than discrete, automata-based methods. Marchesi et al. describe a direct trade-off: as you increase the accuracy of the equilibrium approximation to satisfy theoretical requirements, the computational overhead rises sharply.
Stability Beyond Games
This challenge is not unique to game theory. In many-body physics, Streltsov et al. found that capturing degenerate states in Bose-Einstein condensates required high-precision arithmetic reaching up to 90 digits to ensure convergence.
In the context of game verification, tools like PRISM-games 3.0 are beginning to integrate these high-precision techniques. Kwiatkowska et al. suggest that the future of formal verification for multi-agent systems depends on moving away from standard floats toward arbitrary-precision representations to maintain numerical stability in complex, concurrent models.
Practical Mitigations
While arbitrary precision is the "gold standard" for correctness, it isn't always practical for massive games. Researchers have developed several hybrid strategies:
- Relative Distance Metrics: Asadi et al. propose using floating-point numbers combined with relative distance metrics to prevent rounding error accumulation when dealing with extremely small discount factors.
- Basis Stability Checks: Farina et al. introduced iterative approaches with numerical "basis stability" checks, allowing for significant speedups while maintaining the integrity of the equilibrium.
- Threshold-based Value Iteration: Allamigeon et al. proved that value iteration can remain algorithmically sound with finite precision, provided the precision stays above a specific threshold determined by the game's algebraic properties.
Ultimately, the shift toward infinitesimal-aware computation represents a transition from "good enough" numerical approximations to exact symbolic reasoning, ensuring that the strategies we compute are not just artifacts of rounding errors, but true mathematical equilibria.
Key Foundational Work
Computing an Extensive-Form Perfect Equilibrium in Two-Player Games (2011) — contributed 4 concept elements to this direction.
Exact ground state of finite Bose-Einstein condensates on a ring (2005) — contributed 1 concept elements to this direction.