Differentiable Nash Equilibria: An Accessible Guide

Plain-language explainer for Paper 315 (preprint forthcoming)

The question Nash equilibrium cannot answer

Canada is considering raising its carbon price by $15 per tonne. By how much does that shift the probability that the G20 climate coalition holds together?

Nash equilibrium theory cannot answer this question. It can tell you whether full cooperation is an equilibrium — but not how sensitively the equilibrium participation rate responds to a carbon price change. The best-response correspondence is discontinuous: an infinitesimal change in payoffs can shift an equilibrium discontinuously, and the gradient simply does not exist at the equilibrium.

This paper gives the gradient.

What the Quantal Response Equilibrium adds — and what it doesn’t

The Quantal Response Equilibrium (QRE) was introduced by McKelvey and Palfrey in 1995. It is well-known in behavioural game theory and has been applied to dozens of experimental datasets. The key idea: instead of each player playing a best response (the argmax of expected utility), each player plays a softmax response — a Boltzmann distribution over actions, with weight proportional to $\exp(\beta \cdot \text{expected utility})$. At high $\beta$ (rational players), the softmax concentrates on the best action. At $\beta \to \infty$, it collapses to the argmax exactly: the Nash equilibrium is recovered.

The QRE is smooth. But the existing QRE literature uses this smoothness in a specific way: to estimate $\beta$ from observed choice data, or to trace the path of equilibria as $\beta$ varies.

What nobody did before this paper: differentiate through the QRE fixed point with respect to structural parameters — tax rates, damage coefficients, transfer payments. That derivative is the object you need for policy design, not parameter estimation. It is given by the Implicit Function Theorem (IFT), applied to the smooth fixed-point equation that defines the QRE.

The IFT gradient is:

\[\frac{\partial \tilde{\sigma}^\beta}{\partial \theta} = \beta\, J_F^{-1} \cdot \mathrm{diag}(\tilde{\sigma}^\beta) \cdot (I - \mathbf{1}\tilde{\sigma}^{\beta\top}) \cdot \frac{\partial \mathbb{E}U}{\partial \theta}\]

where $J_F$ is the Jacobian of the QRE fixed-point equation and $\theta$ is any structural parameter. This formula is computable in a single backward pass through the QRE fixed point in JAX or PyTorch.

Three things this unlocks

1. Coalition stability gradients

In the climate coalition game, each country chooses to cooperate (invest in decarbonisation) or free-ride. The QRE participation probability $\tilde{\sigma}^\beta(C)$ is a smooth function of all payoff parameters. Its gradient with respect to a carbon price $\tau$ is the marginal coalition-stability gain of the policy: for each dollar per tonne of carbon price, how much does the participation probability rise?

The IFT formula (Proposition 3.2 of the paper) gives this gradient in closed form:

\[\frac{\partial \tilde{\sigma}^\beta(C)}{\partial \tau} = \frac{\beta\, \tilde{\sigma}^\beta(C)\,(1 - \tilde{\sigma}^\beta(C))} {1 - \beta J} \cdot e_i\]

where $e_i$ is the actor’s emission intensity and $J$ is the social interaction strength. The denominator $1 - \beta J$ is the key: it tells you how close the system is to a tipping point.

2. Optimal heterogeneous carbon prices via reverse stress testing

The paper formulates coalition stability as a reverse stress test (RST): given a coalition that is partially stable, find the minimum-cost carbon price schedule $\boldsymbol{\tau}^*$ that pushes every actor’s participation probability above a target threshold $p^\dagger = 85\%$.

This is exactly the same mathematical problem as supply-chain RST (Paper 314): minimise total cost subject to a survival constraint, using JAX gradient descent on a differentiable loss. The criticality vector $c_i = \partial \tilde{\sigma}^\beta_i / \partial \tau_i$ ranks actors by how efficiently a carbon price increase improves their participation — high $c_i$ means a small tax increase brings a large coalition-stability gain.

At the calibrated temperature $\beta^* = 3.2$ (entropy-matched to observed Paris Agreement NDC revision behaviour), the minimum-cost schedule is:

Actor type	$\Phi_i$ (yield surface)	Carbon price $\tau_i^*$	Criticality
Temperate industrial	7.2	$28.5/tCO₂	0.031
Global mean	17.7	$17.3/tCO₂	0.044
Tropical agricultural	29.2	$6.1/tCO₂	0.058
Sahel economy	32.1	$3.4/tCO₂	0.062
Small island state	33.1	$0	—

Small island states cooperate from private incentive alone — their local damage is so severe that the coalition is individually rational even without any carbon price. Temperate industrial economies need the highest carbon price to reach the participation threshold. The criticality vector shows where a marginal dollar of carbon pricing effort is most effective: tropical and Sahelian actors, who are closest to the participation threshold.

This is the heterogeneous carbon price as a gauge field — the same structure identified from the geometric side in Paper 311 (equation 6.5), now derived from first principles as the QRE reverse stress test solution.

3. The Schelling bifurcation as a computable early-warning signal

Thomas Schelling (1971) identified a social tipping point in residential segregation: above a critical threshold of same-group neighbours, a neighbourhood tips from integrated to segregated — and the tipping is hard to reverse. The same mathematical structure appears in climate coalitions: above a critical defection rate, cooperative participation collapses.

The QRE makes this precise. The participation fixed-point equation undergoes a pitchfork bifurcation at:

\[\beta_c = \frac{1}{J}, \qquad J = p^*(1-p^*) \cdot \frac{\partial \Delta U}{\partial p}\]

Below $\beta_c$: one stable QRE (the unique cooperative equilibrium). Above $\beta_c$: three fixed points emerge — full cooperation, full defection, and an unstable middle. The logit path branches.

The susceptibility:

\[\chi(\beta) = \frac{\beta\, p(1-p)}{1 - \beta J}\]

diverges as $\beta \to \beta_c$. This is directly observable from panel data: $\chi$ is the ratio of aggregate participation variance to individual-level shock variance. An empirical increase in $\chi$ before a major coalition event — a mass defection, a treaty collapse, a sudden surge in NDC ambition — is a measurable leading indicator of the approach to the tipping point.

The same formula $\chi = 1/(1 - \beta J)$ is the social multiplier of Brock and Durlauf (2001), the Keynesian beauty contest fragility of Morris and Shin (2002), and every ecological tipping point with a diffusive feedback mechanism. The QRE unifies them in a single computable object.

What is new vs. what is known

Aspect	Status
QRE definition and existence	McKelvey-Palfrey 1995 — classical
Logit path as $\beta$ varies	Turocy 2005 — classical
Calibrating $\beta$ from choice data	Wright & Leyton-Brown 2010 — classical
IFT gradient w.r.t. structural parameters $\theta$	This paper — 🆕
Coalition RST as a differentiable optimisation	This paper — 🆕
Criticality vector for heterogeneous carbon prices	This paper — 🆕
Schelling bifurcation as computable early-warning $\chi(\beta)$	This paper — 🆕
Connection to tropical limit and Nash recovery rate	This paper — new framing

The existing QRE literature treats $\beta$ as the object of interest. This paper treats $\theta$ — the structural parameters — as the object of interest, and uses $\beta$ as a fixed (calibrated) temperature. The IFT gradient with respect to $\theta$ is the policy gradient; nothing in the prior literature computes it.

The tropical limit: why the gradient is honest

A common concern with smooth relaxations of combinatorial objects is that the relaxation is a surrogate — it approximates the original, but is not the same object, and the gradient of the surrogate may not reflect the behaviour of the original.

The QRE avoids this by the tropical limit theorem (Theorem 6.1): as $\beta \to \infty$, the QRE converges to the Nash equilibrium at rate $O(\beta^{-1} \ln |A|)$, where $|A|$ is the number of action profiles. For the two-action participation game, this is $O(\beta^{-1} \ln 2)$: at $\beta^* = 3.2$, the QRE is within $\ln(2)/3.2 \approx 0.22$ in $\ell_1$ norm of the Nash equilibrium.

Calibrating $\beta^*$ from data automatically chooses the temperature at which this gap is comparable to the empirical noise in participation rates — meaning the QRE is not a surrogate but a statistically faithful extension of the Nash equilibrium.

The connection to supply-chain RST

The coalition stability RST (equation 3.3 of the paper) and the supply-chain RST of Paper 314 are mathematically identical:

\[\mathcal{L}(\boldsymbol{\tau}) = \sum_i \mathrm{ReLU}(p_i^\dagger - \tilde{\sigma}^\beta_i(C)) \cdot 10 + \lambda \sum_i \tau_i\]

In the supply chain: $\tilde{\sigma}^\beta_i$ is replaced by output capacity, $\tau_i$ by inventory buffer, and $p_i^\dagger$ by the minimum acceptable throughput. The criticality vector, gradient descent algorithm, and convergence analysis are identical.

This mathematical identity is not a coincidence. Both problems are instances of the same abstract structure: a smooth fixed-point object (QRE participation / SoftMin capacity), a survival constraint (coalition threshold / output threshold), and a minimum-cost intervention (carbon tax / buffer stock). The Thermal Economics programme (Paper 313) makes this structure explicit and provides the shared IFT infrastructure for both.