How Much Does the Size of a Mistake Matter? Kernel Bandwidth as an Equilibrium Selector

When a game has several equilibria, each a stable outcome no player would unilaterally leave, theory does not say which one people will actually reach. The most common behavioral fix is the quantal response equilibrium (QRE). In a QRE, players choose actions with probabilities that rise smoothly with expected payoff, controlled by one "rationality" parameter. That model assumes something that is rarely stated: a player's reluctance to take an action grows linearly with how much it could cost them. This note proposes to replace that linear treatment of regret (the gap between what you got and what you could have gotten against the opponent's actual move) with a Gaussian kernel of regret. The kernel is the same saturating similarity measure that robust signal processing calls correntropy. Its width, the bandwidth $\sigma$, then becomes an interpretable and estimable dial. At one end, only whether you were wrong matters. At the other end, large mistakes are penalized quadratically. Standard QRE is recovered as a point on the linear-regret path in between.

The payoff of the change is concrete. In coordination games with multiple equilibria, turning $\sigma$ continuously moves the selection threshold between equilibria. In a stag-hunt example worked out below, it sweeps from "magnitude-blind" selection through risk dominance to a squared-loss extreme. We propose to characterize this family theoretically. We then fit it to a large corpus of human one-shot play, 75 games from six experimental datasets, and test whether a single dimensionless bandwidth predicts which equilibrium people choose better than temperature-only models, prospect-theoretic payoff distortions, and level-$k$ reasoning models.

Core Claim

Let $r_i(a_i, a_{-i}) = \max_b u_i(b, a_{-i}) - u_i(a_i, a_{-i})$ be player $i$'s ex-post regret. Here $u_i$ is $i$'s payoff, $a_i$ is $i$'s action and $a_{-i}$ is the other players' actions. Regret is normalized by the game's payoff range so that it is unitless. We define the correntropic value of an action against beliefs $\pi_{-i}$ about the opponents as

$$V_\sigma(a_i;\pi_{-i}) = \mathbb{E}_{a_{-i}\sim\pi_{-i}}\left[\exp\left(-\frac{r_i(a_i,a_{-i})^2}{2\sigma^2}\right)\right],$$

and the correntropic response equilibrium (CRE) as a fixed point in which each player chooses $\pi_i(a_i) \propto \exp(\lambda V_\sigma(a_i;\pi_{-i}))$, with $\lambda$ a precision parameter. Existence follows from Brouwer's theorem, because the response map is continuous on the product of simplices.

The model has two limits. As $\sigma \to 0$, $V_\sigma$ becomes the probability that $a_i$ turns out to be a best response, and the size of a mistake stops mattering. As $\sigma \to \infty$, a second-order expansion gives $V_\sigma \approx 1 - \mathbb{E}[r_i^2]/2\sigma^2$, which is expected squared-regret minimization. Ordinary logit QRE corresponds to linear regret. Because $\mathbb{E}[u_i(a_i,\cdot)] = \mathbb{E}[\max_b u_i(b,\cdot)] - \mathbb{E}[r_i(a_i,\cdot)]$ and the first term does not depend on $a_i$, logit on expected payoff is the same as logit on negative expected linear regret.

A stag hunt makes the selection effect concrete. Take payoffs of 4 for mutual stag, 0 for stag against hare, and 3 for hare regardless of the opponent. If the opponent plays stag with probability $q$, the indifference threshold between the two actions under the kernel value is

$$q^*(\sigma) = \frac{1-e^{-9/2\sigma^2}}{\left(1-e^{-9/2\sigma^2}\right)+\left(1-e^{-1/2\sigma^2}\right)}.$$

This threshold moves from $1/2$ as $\sigma\to 0$ to $0.9$ as $\sigma\to\infty$. Along the way it passes the risk-dominance threshold of $0.75$ that linear regret gives. So the bandwidth is a genuine equilibrium-selection parameter: it changes which equilibrium has the larger basin. It does something temperature cannot do, which is to reshape how regret magnitudes are weighted rather than rescale all of them uniformly.

The empirical claim is that human initial play reveals a finite, interior bandwidth that is stable across datasets. If so, "how much does the size of a mistake matter" is a measurable behavioral constant, and it resolves part of the selection ambiguity that Nash and QRE leave open.

Where This Sits

The kernel side supplies the mechanism. In robust estimation, the maximum correntropy criterion uses a Gaussian kernel of the error whose bandwidth interpolates between squared-error behavior at large width and an $L_0$-like count of misses at small width. Robust Learning with Kernel Mean p-Power Error Loss makes this interpolation explicit for regression and PCA, and its kernel-power family gives us a natural ablation, a power of the kernelized regret. The same literature also constrains the design. Modal Principal Component Analysis and Maximum Correntropy Unscented Filter show that small bandwidths produce flat, zero-gradient plateaus and many spurious optima. We expect the game-theoretic analogue: CRE multiplicity grows as $\sigma$ shrinks, and learning dynamics freeze when every action's regret is far outside the kernel window. Adaptive maximum correntropy based robust CKF with variational Bayesian for covariance estimation addresses filter freeze with a data-driven lower bound on bandwidth, and we borrow that idea for the learning experiments.

On the game side, three lines of work select equilibria by adding a smoothing term. Exploration-Exploitation in Multi-Agent Competition: Convergence with Bounded Rationality proves that entropy-regularized Q-learning converges to a unique QRE in weighted zero-sum polymatrix games, with the $T\to 0$ trace picking one Nash equilibrium. Maximum Entropy Correlated Equilibria picks the highest-entropy correlated equilibrium and notes that it can be socially poor. How Irrationality Shapes Nash Equilibria: A Prospect-Theoretic Perspective shows that concave prospect-theory value functions collapse continua of equilibria to symmetric points in aggregative games. All three act on payoff levels or on the entropy of action probabilities. None of them changes how regret magnitudes are weighted, and that is the one thing we change.

Adversarially Robust Kernel Smoothing is the closest use of bandwidth as a robustness knob, running from empirical risk minimization to worst-case optimization. It is a single-learner method and does not address strategic multiplicity. The behavioral benchmark and our strongest competitor come from Beyond Equilibrium: Predicting Human Behavior in Normal-Form Games. That paper assembles the ALL6 corpus and finds that quantal level-$k$ (QLk) beats QRE, Poisson cognitive hierarchy and Nash. It also reports that humans chose actions Nash assigns zero probability in 83% of games, which is the gap any selection model must close.

Research Program

The theoretical part has three goals. First, for $2\times 2$ games with two pure equilibria, derive the threshold map $\sigma \mapsto q^*(\sigma)$ in general. We want to establish when it is monotone and prove that the $\sigma\to 0$ limit is magnitude-blind (threshold $1/2$ whenever both equilibria are strict). We also want to identify the bandwidth at which CRE matches risk dominance. Second, bound multiplicity. We aim to show that for $\lambda$ small relative to $\sigma$ the response map is a contraction, so the CRE is unique. We expect the number of CRE to grow as $\sigma$ shrinks with $\lambda$ fixed; this is the game-theoretic form of the plateau problem documented in Modal Principal Component Analysis. Third, follow the CRE branch from the unique large-$\sigma$ solution as $\sigma$ decreases. This defines a bandwidth-tracing selector, analogous to the temperature trace in Exploration-Exploitation in Multi-Agent Competition: Convergence with Bounded Rationality, and we characterize where the two traces end at different equilibria.

The empirical part fits CRE, and a correntropic level-$k$ variant (CLk, in which level-$k$ players respond to level-$(k-1)$ beliefs using $V_\sigma$), to ALL6. We use a single bandwidth shared across all games. This is possible because regrets are range-normalized, and it keeps the model from buying fit with a free parameter per game. A secondary study runs discrete-time correntropic learning dynamics on randomly generated general-sum matrix games and small games from OpenSpiel: A Framework for Reinforcement Learning in Games. There we measure, as a function of $\sigma$, how often learning converges, which equilibrium it reaches, and whether an adaptive bandwidth floor prevents freezing.

Baselines

The naive reference is uniform random play, together with Nash equilibrium given an $\epsilon$-mixture with uniform so that likelihoods are finite, averaging over equilibria when there are several. This fixes the floor that any model must clear.

Logit QRE with a fitted precision is the direct competitor, because it is exactly the linear-regret member of our family. Any gain from CRE over it can therefore be attributed to the shape of the regret kernel. Poisson cognitive hierarchy is included as reported in the ALL6 comparison. The strongest existing method is QLk from Beyond Equilibrium: Predicting Human Behavior in Normal-Form Games, which that paper reports as significantly outperforming QRE, Poisson-CH and Nash on ALL6. The file gives only the ordering and no held-out likelihood value, so there is no reported result found on Lacuna for QLk's held-out log-likelihood on ALL6. We will therefore re-implement QLk and reproduce its ordering over QRE before any comparison counts.

A prospect-theory QRE applies a concave value function to payoffs relative to a reference point, following How Irrationality Shapes Nash Equilibria: A Prospect-Theoretic Perspective. It is the strongest alternative source of nonlinearity.

Three ablations isolate the key ingredient. The first, payoff saturation, applies the same Gaussian kernel to payoff shortfall from the game's maximum payoff instead of to regret, which tests whether the counterfactual comparison matters. The second, power regret, replaces the Gaussian with $|r|^p$, the kernel-free half of the family in Robust Learning with Kernel Mean p-Power Error Loss, which tests whether saturation matters beyond curvature. The third fixes $\sigma$ per game by Silverman's rule on the regret distribution instead of fitting one shared value.

For fairness, every model is fitted by maximum likelihood with the same optimizer and the same number of random restarts. All are evaluated by 10-fold cross-validation over observations, as in the ALL6 study, and additionally by leave-one-dataset-out. Parameter counts are reported. CRE has two parameters, like prospect-theory QRE. CLk has one more than QLk, and it is also compared at matched parameter count by tying precision across levels.

Beating a baseline means a higher mean held-out log-likelihood with a paired bootstrap interval over folds that excludes zero. For the selection question, it also means higher accuracy at predicting which pure equilibrium carries the modal choice in games with multiple equilibria.

Why not just…?

Why not just tune the QRE temperature? Temperature rescales all payoff differences by the same factor, so in the stag hunt it cannot move the $0.75$ selection threshold away from risk dominance. Only the bandwidth does. Because QRE is nested as linear regret, the experiment settles the question directly: if the fitted CRE does not beat QRE out of sample, or if the fitted bandwidth sits where the kernel mimics linear regret, the proposal fails.

Why not just use a prospect-theoretic or other concave payoff transform? This is the real threat, because any saturating transform could produce similar fits. The difference is what gets transformed. Prospect theory distorts payoff levels against a reference point, whereas CRE distorts the counterfactual comparison with the best alternative. The payoff-saturation ablation and prospect-theory QRE share CRE's functional form but not its argument, so they separate the two. If they match CRE on ALL6, we will report that regret framing adds nothing and recast the contribution as a characterization of kernelized payoff selectors.

Why not just use level-$k$ reasoning (QLk)? Level-$k$ models explain departures from Nash through shallow belief hierarchies, not through how mistakes are weighed. The two mechanisms are orthogonal, which is why CLk combines them. If CLk does not improve on QLk at matched parameter count, bandwidth adds nothing beyond iterated reasoning in initial play.

Why not just use maximum-entropy correlated equilibria or $\alpha$-Rank? These are computational selectors that guarantee a unique answer. Maximum Entropy Correlated Equilibria does so by maximizing entropy, and A Generalized Training Approach for Multiagent Learning does so through an evolutionary stationary distribution. Neither makes a behavioral claim about which equilibrium people choose, and both must still be justified by some criterion. We include the maximum-entropy correlated equilibrium's marginals as an extra reference in the selection-accuracy analysis, but they answer a different question.

Decisive Checks

The proposal lives or dies on four results. First, the shared bandwidth fitted on ALL6 must be interior and well-determined. The bootstrap interval must exclude both $\sigma\to\infty$ (squared regret) and the value at which the kernel reproduces linear regret. Second, CRE must beat QRE and CLk must beat QLk in held-out log-likelihood. Third, the regret-based kernel must beat the payoff-saturation ablation and prospect-theory QRE; otherwise the regret framing is not what matters. Fourth, a bandwidth fitted on five datasets must predict the sixth nearly as well as a bandwidth fitted on the sixth itself, which is the test that $\sigma$ is a behavioral constant rather than a per-study nuisance.

In the learning study, the prediction is that equilibrium multiplicity and freezing rise as $\sigma$ shrinks, following the plateau mechanism in robust filtering. It also predicts that an adaptive bandwidth floor modeled on Adaptive maximum correntropy based robust CKF with variational Bayesian for covariance estimation restores convergence without changing which equilibrium is selected. If the floor changes the selected equilibrium, the bandwidth trace is not a stable selector and we will report that as a negative result.