Quantal-Response Feedback: Quantum and Strategic Control

• Quantal-response feedback is a closed-loop control system that updates actions and rates based on stochastic signals rather than deterministic best responses.
• It is applied across fields such as mesoscopic quantum transport (e.g., freezing of cumulants) and game theory (via logit-QRE models) to achieve fixed-point equilibrium.
• Its practical impact includes improved empirical identification, robust exploitation strategies, and enhanced design of dynamic, adaptive systems.

Quantal-response feedback denotes a family of closed-loop response structures in which actions, controls, or rate parameters are updated from an observed signal through a stochastic or measurement-conditioned response law rather than an exact deterministic best response. In the materials considered here, the phrase spans several distinct but mathematically related usages: counting-conditioned closed-loop control in mesoscopic quantum transport, fixed-point feedback between expected payoffs and noisy choice probabilities in game theory, and measurement-conditioned quantum-control protocols that recycle quantum outcomes into subsequent controls (Brandes, 2010, Pogorelskiy et al., 2016, Porotti et al., 2022).

1. Common response architecture

A common formal pattern is the composition

$$\text{observed signal} \;\to\; \text{response map} \;\to\; \text{updated dynamics or action probabilities}.$$

In mesoscopic transport, the observed signal is the counting error

$q_n(t)\equiv I_0 t-n,$

and feedback enters through a multiplicative response function

$f(q_n(t)),\qquad f(0)=1,$

with weak-feedback studies emphasizing the linear choice

$f(x)=1+gx$

(Brandes, 2010).

In game-theoretic quantal response, the observed signal is typically an expected-payoff vector. Structural QRE models action $j$ for player $i$ as the maximizer of

$u_{ij}(\mathbf p)+\varepsilon_{ij},$

so the induced choice probability is

$$\pi_{ij}(\mathbf p)=P\!\left(j=\arg\max_{j'}\{u_{ij'}(\mathbf p)+\varepsilon_{ij'}\}\right),$$

and equilibrium imposes the fixed-point condition

$$\pi_{ij}^*=\pi_{ij}(\boldsymbol{\pi}^*),\qquad \forall i,j$$

(Pogorelskiy et al., 2016). In canonical normal-form notation, a quantal opponent may be written as

$$QR(\sigma,a^k)=\frac{q(u_2(\sigma,a^k))}{\sum_{a^i\in A_2}q(u_2(\sigma,a^i))},$$

with the logit specification

$q_n(t)\equiv I_0 t-n,$0

(Milec et al., 2020).

The same architecture appears in entropy-regularized Markovian models. In leader-follower Markov games, the follower’s quantal response has Boltzmann form

$q_n(t)\equiv I_0 t-n,$1

while in traffic Markov games the equilibrium policy is

$q_n(t)\equiv I_0 t-n,$2

(Chen et al., 2023, Pham et al., 9 Jan 2026).

2. Quantum transport and measurement-conditioned control

In mesoscopic quantum transport, quantal-response feedback is a closed-loop, counting-based control scheme in which transport parameters respond continuously to the stochastic transfer record. In the high-bias, unidirectional, Born-Markov regime, the dynamics is described by an $q_n(t)\equiv I_0 t-n,$3-resolved master equation,

$q_n(t)\equiv I_0 t-n,$4

and, after Fourier transformation,

$q_n(t)\equiv I_0 t-n,$5

For a tunnel junction with linear feedback, the first two cumulants are

$q_n(t)\equiv I_0 t-n,$6

so the mean continues to move linearly while the variance saturates at

$q_n(t)\equiv I_0 t-n,$7

More generally,

$q_n(t)\equiv I_0 t-n,$8

which is the paper’s “freezing” of the full counting statistics. For homogeneous feedback, the frozen cumulants still encode the original no-feedback transport statistics, yielding explicit reconstruction formulas such as

$q_n(t)\equiv I_0 t-n,$9

(Brandes, 2010).

A broader network-theoretic treatment of quantum transport represents each bidirectional lead by an input-output pair and embeds transport devices into the SLH quantum-feedback-network formalism. This extends scatterer-only descriptions by allowing emission, absorption, Bose or Fermi fields, and nonlinear component dynamics, while series interconnection in transport is handled by a Redheffer-star-product reduction rather than a purely unidirectional cascade (Gough, 2014).

Related quantum-control literature uses measurement-conditioned feedback in a different sense. FALQON assigns each new circuit parameter from the measured commutator expectation,

$f(q_n(t)),\qquad f(0)=1,$0

and feedback-GRAPE optimizes policies $f(q_n(t)),\qquad f(0)=1,$1 conditioned on a discrete or continuous measurement record $f(q_n(t)),\qquad f(0)=1,$2. For discrete measurements, the gradient of the expected return contains both a direct differentiation term and a log-likelihood correction,

$f(q_n(t)),\qquad f(0)=1,$3

This suggests a broader quantum meaning of quantal-response feedback as measurement-conditioned adaptation, even though these protocols are not QRE models (Magann et al., 2021, Porotti et al., 2022).

3. Static strategic feedback and equilibrium structure

In game theory, the central feedback loop is

$f(q_n(t)),\qquad f(0)=1,$4

Melo, Pogorelskiy, and Shum characterize structural QRE choice probabilities as gradients of convex social-surplus functions,

$f(q_n(t)),\qquad f(0)=1,$5

which implies cyclic monotonicity across related games: $f(q_n(t)),\qquad f(0)=1,$6 A central methodological point is that this is static equilibrium feedback across beliefs, expected payoffs, and mixed actions, not dynamic feedback over repeated periods (Pogorelskiy et al., 2016).

The same equilibrium can be written as an entropy-regularized saddle problem. In two-player zero-sum games,

$f(q_n(t)),\qquad f(0)=1,$7

with negative entropy $f(q_n(t)),\qquad f(0)=1,$8, yields the logit-QRE conditions

$f(q_n(t)),\qquad f(0)=1,$9

This variational-inequality viewpoint is important because the regularized operator $f(x)=1+gx$0 is strongly monotone, which underlies linear convergence results for first-order solvers in both normal-form and sequence-form extensive-form games (Sokota et al., 2022).

Inverse-design work gives a further static fixed-point perspective. In multiplayer matrix games with Gumbel perception noise, each player’s mixed action satisfies

$f(x)=1+gx$1

Under

$f(x)=1+gx$2

the QRE is unique. This makes quantal-response feedback a design variable: cost matrices can be inferred so that a target joint strategy becomes the unique noisy-response fixed point (Yu et al., 2022).

4. Dynamic, evolutionary, and reinforcement-learning formulations

Several papers replace the static fixed-point condition by an explicit feedback dynamic. In iterative Logit quantal response dynamics, players repeatedly apply the logit rule to opponents’ previous mixed strategies. QRE remain fixed points, but they are partitioned into stable QRE (SQREs) and unstable QRE (USQREs). In symmetric $f(x)=1+gx$3 games, mixed QRE are stable for small $f(x)=1+gx$4 and become unstable beyond a critical $f(x)=1+gx$5, while pure-QRE branches remain stable (Zhuang et al., 2013).

On graphs, noisy binary-choice games yield the self-consistent equations

$f(x)=1+gx$6

This makes quantal-response feedback explicitly topology dependent: neighbors’ mean actions generate local fields, local fields generate noisy binary choices, and those choices feed back through the graph. In the complete graph, the phase transition is governed by

$f(x)=1+gx$7

while in the annealed approximation on random graphs it becomes

$f(x)=1+gx$8

(Leonidov et al., 2019).

In reinforcement-learning settings, the feedback signal is no longer only a normal-form payoff vector. For Quantal Stackelberg Equilibrium in episodic Markov games, the follower reacts to a committed leader policy by solving an entropy-regularized control problem, and the leader learns the follower’s latent response model from observed actions rather than from observed follower rewards (Chen et al., 2023). EvoQRE extends this logic to safety-critical traffic simulation: entropy-regularized replicator dynamics and a two-timescale actor-critic scheme converge to Logit-QRE under weak monotonicity assumptions, with expected KL error

$f(x)=1+gx$9

(Pham et al., 9 Jan 2026).

5. Testing, identification, and design

The empirical content of quantal-response feedback depends strongly on what is treated as observable. Across a series of laboratory games, cyclic monotonicity inequalities provide a nonparametric test of structural QRE. The pooled data reject QRE, but at the individual level the hypothesis cannot be rejected for over half of the subjects, highlighting the role of heterogeneity and the difference between individual noisy response and pooled representative-agent behavior (Pogorelskiy et al., 2016).

In initial-play econometrics, standard QRE can misestimate preferences because it interprets all stochasticity as noisy strategic response. Augmenting QRE with a payoff-sensitive non-strategic component yields QRE+L0, operationalized in forms such as QRE-QL4. In the reported experiments, QRE-none achieves roughly $j$0–$j$1 relative error in recovering the value parameter, whereas QRE-QL4 reaches about $j$2–$j$3; by contrast, QRE-uniform can be substantially worse at low values. This suggests that quantal-response feedback is often empirically entangled with non-strategic, but still payoff-sensitive, response channels (Chui et al., 2022).

Identification results can also be stronger than equilibrium testing. In unknown finite games with recommendation mechanisms, the moderator observes whether recommended actions are followed or weakly improved upon. Under the paper’s quantal-response support model,

$j$4

generic games with no weakly dominated actions are learnable up to positive affine equivalence. The constructive recommendation complexity is

$j$5

and the online regret of the associated recommendation algorithm is

$j$6

(Alanqary et al., 19 Feb 2026).

Information design under quantal response changes the persuasion problem because signals now affect action probabilities smoothly rather than through a deterministic threshold. With binary receiver action,

$j$7

In state-independent sender-utility environments, an optimal censorship signaling scheme exists for any $j$8, and the fully rational optimal censorship scheme has robust approximation ratio at most $j$9 over $i$0. In state-dependent sender-utility environments, by contrast, even binary-state instances can have no signaling scheme with bounded worst-case approximation over $i$1 (Feng et al., 2022).

6. Robust exploitation, aggregation, and criticism

Quantal-response feedback is also used normatively, as a model to be exploited. In two-player games against quantal opponents, Quantal Nash Equilibrium and Quantal Stackelberg Equilibrium are distinct: QNE can be worse than Nash equilibrium against the same quantal opponent, while QSE directly maximizes payoff against the induced quantal response. Exact computation is hard—QNE is PPAD-hard in normal-form games, and QSE is NP-hard in imperfect-information extensive-form games—so scalable heuristics are used. In Goofspiel 7, the paper reports exploitability/gain pairs $i$2 for CFR-QR, $i$3 for RQR, and $i$4 for the Nash-oriented baseline, illustrating the exploitation–robustness tradeoff (Milec et al., 2020).

Aggregation theory gives a different robustness perspective. Under binary-state c.i.i.d. private signals and logistic quantal response,

$i$5

majority voting is minimax-regret optimal whenever $i$6. More strikingly, for all $i$7, there exist signal structures and finite $i$8 such that a group of quantal responders can outperform perfectly rational agents, because decision randomness encodes weak but informative signals lost in deterministic behavior (Huang et al., 14 Mar 2026).

The main theoretical criticism is that structural QRE may itself be too restrictive. When at least one player has three actions, no set of monotone structural QRE models can be both consistent with payoff monotonicity and well-specified for studying payoff-monotone behavior. This is the paper’s “paradox of monotone structural QRE”: a structural noisy-response foundation can exclude monotone behaviors that are precisely the target phenomenon (Velez et al., 2019).

Taken together, these literatures treat quantal-response feedback as a unifying but non-uniform idea: a closed loop in which stochastic response laws, counting-conditioned controls, or measurement-conditioned policies transform observed signals into future behavior. The technical meaning depends on domain, but the recurring themes are fixed-point self-consistency, regularization of exact best response, and the possibility that noise, rather than merely degrading performance, can reshape what feedback reveals and what robust control or aggregation can achieve.