Quantal Response Equilibrium

Classic game theory is built on a convenient fiction: every player always picks the best available action. Real people do not. We make mistakes, forget options, act on hunches — and those small departures from perfection change the game.

Quantal Response Equilibrium (QRE) was introduced by Richard McKelvey and Thomas Palfrey in 1995 to capture exactly this. Instead of assuming players pick the best action with certainty, QRE says players pick better actions more often — but never with absolute certainty. The gap between "best" and "second-best" matters: if one action pays far more than another, players overwhelmingly choose it; if the payoffs are nearly equal, choices are nearly random.

The magic comes from the temperature parameter $\lambda$ (or its inverse, often called "noise"). At $\lambda = 0$, all actions look equally attractive and players randomize uniformly. As $\lambda \to \infty$, the model collapses back to Nash equilibrium — every player locks onto their best response with probability 1. In between lies a rich family of predictions that fit experimental data far better than Nash alone ever could.

QRE is now a workhorse of behavioral game theory and computational economics, used wherever real players — humans, animals, or imperfectly-tuned algorithms — must be modeled honestly.

The demo below shows a symmetric $2 \times 2$ coordination game. Each player chooses A or B. The Nash equilibrium predicts a sharp split, but drag the noise slider and watch what happens: as noise rises, the QRE strategy drifts away from Nash and toward a uniform mix.

<!-- {{c_intro}} -->
<div class="controls">
  <label for="noise" title="{{noise_title}}">{{noise_label}} &lambda; = <span id="lambdaVal">1.0</span></label>
  <input id="noise" type="range" min="0" max="5" step="0.1" value="1.0" aria-label="{{noise_label}}">
</div>
<div class="payoff-section">
  <p class="hint">{{payoff_hint}}</p>
  <table class="matrix" aria-label="{{matrix_aria}}">
    <thead>
      <tr><th></th><th>{{col_A}}</th><th>{{col_B}}</th></tr>
    </thead>
    <tbody>
      <tr>
        <td class="row-label">{{row_A}}</td>
        <td><input class="payoff-input" id="p00" type="number" value="3" min="-9" max="9" aria-label="{{cell_AA}}"></td>
        <td><input class="payoff-input" id="p01" type="number" value="0" min="-9" max="9" aria-label="{{cell_AB}}"></td>
      </tr>
      <tr>
        <td class="row-label">{{row_B}}</td>
        <td><input class="payoff-input" id="p10" type="number" value="0" min="-9" max="9" aria-label="{{cell_BA}}"></td>
        <td><input class="payoff-input" id="p11" type="number" value="2" min="-9" max="9" aria-label="{{cell_BB}}"></td>
      </tr>
    </tbody>
  </table>
</div>
<div class="result-section">
  <div class="bar-group">
    <span class="bar-label">{{prob_A_label}}</span>
    <div class="bar-track"><div id="barA" class="bar bar-a"></div></div>
    <span class="bar-pct" id="pctA">50%</span>
  </div>
  <div class="bar-group">
    <span class="bar-label">{{prob_B_label}}</span>
    <div class="bar-track"><div id="barB" class="bar bar-b"></div></div>
    <span class="bar-pct" id="pctB">50%</span>
  </div>
  <p class="status" id="status"></p>
</div>

/* {{c_layout}} */
* { box-sizing: border-box; margin: 0; padding: 0; }
body { font-family: system-ui, sans-serif; color: #222; padding: .8rem 1rem 1rem; }
.controls { display: flex; align-items: center; gap: .8rem; margin-bottom: .9rem; flex-wrap: wrap; }
.controls label { font-size: .95rem; font-weight: 600; white-space: nowrap; }
.controls input[type=range] { flex: 1; min-width: 140px; accent-color: #1d3557; }
.hint { font-size: .85rem; color: #555; margin-bottom: .5rem; line-height: 1.4; }
/* {{c_matrix}} */
.payoff-section { margin-bottom: .9rem; }
.matrix { border-collapse: collapse; }
.matrix th, .matrix td { padding: .3rem .55rem; text-align: center; }
.matrix th { font-size: .8rem; color: #555; }
.row-label { font-size: .8rem; color: #555; font-weight: 600; }
.payoff-input { width: 52px; text-align: center; font-size: 1rem; font-weight: 700;
                border: 1px solid #cdd9e3; border-radius: 6px; padding: .2rem; }
/* {{c_bars}} */
.result-section { margin-top: .4rem; }
.bar-group { display: flex; align-items: center; gap: .5rem; margin-bottom: .4rem; }
.bar-label { width: 80px; font-size: .85rem; font-weight: 600; }
.bar-track { flex: 1; height: 22px; background: #e8eef3; border-radius: 6px; overflow: hidden; }
.bar { height: 100%; border-radius: 6px; transition: width .25s ease; }
.bar-a { background: #1d3557; }
.bar-b { background: #457b9d; }
.bar-pct { width: 40px; text-align: right; font-size: .9rem; font-weight: 700; color: #1d3557; }
.status { margin-top: .6rem; font-size: .9rem; color: #444; min-height: 1.3em; line-height: 1.45; }

// Code not found

Notice that even a little noise moves the equilibrium noticeably. At high noise the payoff matrix barely matters — both actions are chosen almost equally. This is the core insight: QRE predictions are smooth and continuous in $\lambda$, while Nash is a single sharp point.

The defining equation of the logit QRE looks deceptively simple. For a two-player normal-form game, a mixed strategy profile $(\sigma_1, \sigma_2)$ is a logit QRE at temperature $\lambda$ if each player's strategy satisfies:

$$\sigma_i(a) = \frac{e^{\lambda \cdot u_i(a,\, \sigma_{-i})}}{\sum_{b} e^{\lambda \cdot u_i(b,\, \sigma_{-i})}}$$

This is a softmax (logistic choice) applied to expected payoffs — the higher the expected payoff of action $a$, the larger its probability. The catch: each player's expected payoff depends on the other player's strategy, which is itself a softmax of the first player's strategy. The equation is a fixed point in the product of simplices.

Several important facts follow:

• Existence: a logit QRE exists for every finite normal-form game and every $\lambda \geq 0$ (proven by McKelvey and Palfrey via Kakutani's fixed-point theorem).
• Uniqueness: not guaranteed. Some games have multiple QRE branches at the same $\lambda$.
• Computation: finding a QRE is in PPAD (the same class as Nash equilibrium); no polynomial-time algorithm is known in general, though path-following methods (homotopy from $\lambda = 0$) work reliably in practice.
• Limit behavior: the limit as $\lambda \to \infty$ of the principal QRE branch converges to a Nash equilibrium, not necessarily the "most reasonable" one — another subtlety the model surfaces.

The model's strength is also its challenge: the smooth parameterization makes it empirically testable and theoretically tractable, but the fixed-point structure means computation is never trivial.

QRE has become the standard tool whenever experimentalists observe that players are directionally rational but not perfectly so:

• Lab experiments: in centipede games, ultimatum games, and public-goods games, Nash equilibrium predicts extreme outcomes (full defection, zero cooperation) that no one observes. QRE with moderate $\lambda$ reproduces the actual distribution of choices remarkably well.
• Auction theory: overbidding in first-price auctions is a classic puzzle. QRE explains it as noise: when the expected payoff difference between bids is small, players spread probability across nearby bids rather than locking in.
• Industrial organization: firms in oligopoly markets rarely play the sharp Nash prediction. QRE models give a continuous family of predictions that match observed price dispersions.
• Algorithmic game theory: when agents are machine-learning models or heuristic bots rather than perfectly rational optimizers, QRE is a more honest equilibrium concept than Nash.
• Evolutionary biology: the logit QRE has formal links to the replicator dynamics under mutation — strategies that perform better spread faster, but no strategy is ever fully eliminated.

Understanding QRE connects directly to why Nash equilibrium sometimes fails as a prediction tool, and it sits naturally alongside concepts like PAC learning in the broader study of bounded rationality.

Quantal Response Equilibrium does something elegant: it keeps the fixed-point logic of Nash equilibrium but replaces the all-or-nothing best-response rule with a smooth probabilistic one. Better actions are still chosen more often — the model is not randomness for its own sake — but the sharp discontinuity of perfect rationality dissolves into a curve parameterized by $\lambda$.

McKelvey and Palfrey gave us a model that is continuous, always exists, and fits data. Its weakness is also visible: uniqueness is not guaranteed, computation is PPAD-hard in general, and the right $\lambda$ for any real situation must be estimated from data. But those are tractable problems compared to the wholesale failure of Nash equilibrium to explain what real players actually do.

Next time you watch a market, an auction, or a game-show contestant deviate from the "rational" choice, consider that they may not be irrational — they may simply be playing a QRE at finite temperature, doing the best they can with the cognitive resources they have.