Periodic Nash Equilibria Overview

• Periodic Nash equilibria are defined as equilibrium objects that vary periodically, capturing time-dependent strategies in games with cyclic payoff structures.
• The concept spans multiple literatures—from periodic zero-sum games and LQ feedback laws to operator-induced periodic strategies—differentiating equilibrium paths from averaged equilibria.
• Periodic learning dynamics reveal practical implications in equilibrium recovery and selection, distinguishing algorithm-driven cyclic behavior from static Nash outcomes.

Searching arXiv for papers directly relevant to periodic Nash equilibria, periodic strategies, and solver-dependent equilibrium selection. Periodic Nash equilibria are not a single standardized equilibrium concept. In current arXiv-adjacent usage, the expression covers several technically distinct objects: a moving instantaneous Nash equilibrium path in periodically varying games; a periodic sequence of feedback laws that forms an infinite-horizon Nash equilibrium in dynamic games; periodic strategies defined by an operator distinct from best response and only sometimes coincident with Nash; and, in repeated-play learning, periodic trajectories whose averages recover a Nash equilibrium even though the cycle itself is not an equilibrium. A precise treatment therefore begins by separating periodicity of the equilibrium object from periodicity of the learning dynamics that approach, encode, or fail to approach equilibrium (Fujimoto et al., 2024).

1. Conceptual scope and nomenclature

The main literatures using closely related terminology do not agree on a single formal definition. Some papers study Nash equilibrium in an environment whose primitives vary periodically; some define periodicity as a separate non-cooperative solution concept; some obtain periodic Nash equilibria as cycles of a Riccati recursion; and some analyze periodic play trajectories whose ergodic averages identify Nash strategies rather than constituting Nash equilibria themselves (Salizzoni et al., 28 Aug 2025).

Literature	Periodic object	Relation to Nash
Periodic zero-sum games (Fujimoto et al., 2024)	Instantaneous equilibrium path $(x^*(t),y^*(t))$ induced by $U(t)$	Nash equilibrium moves over time
Periodic strategies (Oikonomou et al., 2013)	Fixed point or cycle of $\mathcal Q=\varphi_2\circ\varphi_1$	Distinct from Nash in general
Multi-player extensions (Oikonomou et al., 2020)	Periodic actions under compositions of $\varphi_{ij}$	Distinct solution concept; existence guaranteed in finite games
LQ games (Salizzoni et al., 28 Aug 2025)	Periodic orbit of the coupled Riccati map	Yields a genuine infinite-horizon periodic Nash equilibrium
Hedge-myopic repeated play (Guo et al., 2024)	Eventual cycle of stage strategies	One-period average recovers exact NE under special assumptions
Periodic double auctions (Manvi et al., 2023)	Nonstationary MPNE over auction rounds	Dynamic Nash equilibrium in a repeated market
Nash polytopes and solver dynamics (Leal, 26 Jun 2026)	Cycling belongs to solver trajectories, not to equilibrium itself	Clarifies what periodicity is not

A recurring source of confusion is the conflation of these notions. In periodically varying games, the equilibrium may itself be time-dependent. In static games with learning dynamics, by contrast, periodicity often belongs only to the path traced by the algorithm. In the periodic-strategies literature, periodicity is an operator-theoretic property that reverses the usual best-response logic. These are structurally different uses of the same adjective.

2. Periodic equilibrium paths in time-varying zero-sum games

In periodically varying zero-sum matrix games, the canonical object is not a static equilibrium point but an instantaneous equilibrium trajectory induced by a time-dependent payoff matrix $U(t)$ satisfying

$U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$

Mixed strategies $x$ and $y$ remain elements of simplices, and expected payoff is

$u(x,y):=x^{\rm T}Uy.$

When $U$ varies with $U(t)$0, best responses and Nash equilibrium generically move as well (Fujimoto et al., 2024).

The clearest explicit realization is the $U(t)$1 sinusoidal construction

$U(t)$2

for which the instantaneous equilibrium takes the form

$U(t)$3

This is the most direct sense in which the literature speaks of a periodic Nash equilibrium: a periodic equilibrium path, not a stationary profile. The learning dynamics studied in that setting are continuous-time gradient descent-ascent / Euclidean FTRL, and the vector field is centered on the moving equilibrium rather than on a fixed saddle.

The central phenomenon is synchronization between the game frequency $U(t)$4 and the intrinsic cycling frequency of the learning dynamics. In the explicit $U(t)$5 construction, when $U(t)$6, secular terms appear and the distance from the moving equilibrium grows linearly in time. In the more general smooth two-action periodic zero-sum model, the paper proves that when $U(t)$7, the time averages $U(t)$8 and $U(t)$9 diverge over time $\mathcal Q=\varphi_2\circ\varphi_1$0, whereas when $\mathcal Q=\varphi_2\circ\varphi_1$1, they converge (Fujimoto et al., 2024).

This result sharply separates pointwise tracking from ergodic behavior. Away from synchronization, trajectories need not converge pointwise to $\mathcal Q=\varphi_2\circ\varphi_1$2; they may remain bounded and cyclic while time averages converge. Under synchronization, even that weaker convergence can fail. A plausible implication is that in periodic environments the equilibrium benchmark is intrinsically path-valued, and stability analysis must be resonance-aware rather than based solely on static minimax structure.

3. Periodic strategies as a distinct non-cooperative solution concept

A different usage originates in the periodic-strategies literature, where periodicity is not defined through time-varying payoffs but through an operator on action spaces. In a two-player finite strategic-form game, one introduces maps

$\mathcal Q=\varphi_2\circ\varphi_1$3

where $\mathcal Q=\varphi_2\circ\varphi_1$4 selects the opponent action maximizing player $\mathcal Q=\varphi_2\circ\varphi_1$5's payoff conditional on $\mathcal Q=\varphi_2\circ\varphi_1$6, and $\mathcal Q=\varphi_2\circ\varphi_1$7 selects the opponent action maximizing player $\mathcal Q=\varphi_2\circ\varphi_1$8's payoff conditional on $\mathcal Q=\varphi_2\circ\varphi_1$9. The induced self-maps are

$\varphi_{ij}$0

An action is periodic if, for some $\varphi_{ij}$1,

$\varphi_{ij}$2

or analogously under $\varphi_{ij}$3 (Oikonomou et al., 2013).

This reverses the Nash optimization logic. Nash equilibrium is based on mutual best response: each player optimizes with respect to their own strategy variable. Periodic strategies instead optimize with respect to the opponent's action. In mixed $\varphi_{ij}$4 games with expected payoff

$\varphi_{ij}$5

the contrast is encoded by the first-order conditions. Nash uses

$\varphi_{ij}$6

whereas periodic mixed strategies use

$\varphi_{ij}$7

The intended robustness property is likewise reversed: at mixed Nash equilibrium, a player's payoff is independent of their own action given the opponent's equilibrium mixing; in a periodic strategy, the player's payoff is constructed to be independent of the opponent's action (Oikonomou et al., 2013).

The multi-player generalization preserves this structure. In simultaneous $\varphi_{ij}$8-player strategic-form games, maps

$\varphi_{ij}$9

are composed across players, and an action $U(t)$0 is periodic if

$U(t)$1

The strongest general result in that line is an existence theorem: every finite action simultaneous $U(t)$2-player strategic form game contains at least one periodic action. The set of periodic strategies $U(t)$3 is also proved set stable under the action of the maps $U(t)$4. The same paper extends the framework to Bayesian games via ex-ante and interim strategic-form representations, and to an epistemic interpretation in which a rationalizable periodic action with periodicity number $U(t)$5 corresponds to $U(t)$6 types in the two-player case (Oikonomou et al., 2020).

In this literature, “periodic Nash equilibrium” is therefore not the primitive concept. The relevant object is a periodic strategy or periodic action, and a Nash equilibrium is periodic only in the special case that the Nash actions are fixed points of the periodicity maps. The relation is overlap, not equivalence.

4. Periodic feedback Nash equilibria in linear-quadratic games

A more direct and structurally rigorous notion of periodic Nash equilibrium appears in discrete-time general-sum linear quadratic games. The state dynamics are

$U(t)$7

with quadratic stage costs

$U(t)$8

Finite-horizon equilibria are characterized by a backward coupled Riccati recursion, and the key insight is to reinterpret that recursion as a nonlinear dynamical system

$U(t)$9

on tuples of positive definite matrices $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$0 (Salizzoni et al., 28 Aug 2025).

Within that viewpoint, stationary and periodic Nash equilibria are different invariant sets of the same induced map. Fixed points

$U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$1

are proved to coincide exactly with the stationary infinite-horizon linear feedback Nash equilibria. Moreover, any such stationary equilibrium can be reproduced in the finite-horizon problem by choosing terminal costs $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$2, in which case the recursion remains constant and the equilibrium gains are stationary at all stages (Salizzoni et al., 28 Aug 2025).

Periodic Nash equilibria arise when the Riccati map has a cycle. A cycle of length $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$3 is a collection $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$4 satisfying

$U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$5

If $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$6 denotes the associated gain tuple, then the sequence $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$7 repeated periodically in time is proved to be a Nash equilibrium of the infinite-horizon LQ game, provided the standard stabilizability assumption holds. The closed loop is stable over one full period in the sense that

$U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$8

even though an individual $U(t)=U(t+2\pi/\omega)\qquad \text{for all } t.$9 within the cycle need not be stable (Salizzoni et al., 28 Aug 2025).

This is a genuine periodic Nash equilibrium rather than a heuristic or learning artifact. The equilibrium object itself is periodic: a periodic sequence of linear feedback laws. The paper further reports three asymptotic regimes of the finite-horizon Riccati dynamical system—convergence to a stationary equilibrium, convergence to a periodic equilibrium, and bounded non-convergent behavior—thereby placing periodic Nash equilibria alongside stationary equilibria within one unified dynamical-systems framework.

5. Periodic learning dynamics and equilibrium recovery

Another strand studies periodicity not as equilibrium itself but as a computational signature of equilibrium. In the Hedge-myopic system for a two-player zero-sum repeated game, player $x$0 uses Hedge / multiplicative weights,

$x$1

while player $x$2 plays a deterministic pure myopic best response

$x$3

Under a rational payoff matrix and an interior Nash equilibrium, the paper introduces the KL-shifted quantity

$x$4

and proves that $x$5 is bounded, that the Hedge state can take only finitely many values, and hence that the deterministic trajectory is eventually periodic after finitely many steps (Guo et al., 2024).

The crucial point is that the resulting cycle is generally not the Nash equilibrium. Rather, the cycle encodes equilibrium information. Once the trajectory enters a period $x$6,

$x$7

and the one-period average of player $x$8's actions is exactly a Nash equilibrium strategy: $x$9 The paper therefore proposes an asymmetric HBR paradigm: detect the cycle, average over one period, and recover an exact equilibrium strategy in the rational interior-NE case (Guo et al., 2024).

For general games the exact periodicity theorem no longer holds, but the same framework yields an approximate equilibrium guarantee. With

$y$0

the time-averaged strategy profile is proved to be a

$y$1

This places periodic dynamics in an intermediate role: neither purely pathological nor itself equilibrium, but algorithmically exploitable.

6. Markov perfect Nash equilibrium in periodic double auctions

In finite-horizon periodic double auctions, periodicity is institutional and temporal rather than operator-theoretic or spectral. The environment consists of $y$2 auction rounds, buyers with residual procurement requirements, a known composite supply curve in each round, and terminal outside procurement at price $y$3. The resulting model is a finite-horizon Markov game with state

$y$4

where $y$5 denotes residual demands and $y$6 the current residual supply curve (Manvi et al., 2023).

The equilibrium concept is Markov perfect Nash equilibrium. Strategies condition only on the current state and remaining time, and the paper derives an analytical deterministic MPNE for the case in which each buyer is allowed to make one bid per round. The construction is explicitly nonstationary: it depends on round $y$7, on the residual-demand ranking, and on the number of rounds left through quantities such as $y$8 and $y$9. Under the monotonicity lemma $u(x,y):=x^{\rm T}Uy.$0 when the supply curve does not change across rounds, together with the balancing-cost condition

$u(x,y):=x^{\rm T}Uy.$1

the paper proves that no deterministic unilateral deviation lowers continuation cost, hence the constructed policy is an MPNE (Manvi et al., 2023).

This is not a periodic Nash equilibrium in the sense of a recurring infinite-horizon cycle. It is instead a finite-horizon, period-indexed dynamic Nash equilibrium in a repeated auction institution. Its relevance lies in showing how “periodic” market environments naturally lead from static Nash analysis to dynamic, state-contingent equilibrium objects.

7. Static equilibrium polytopes, cycling, and common misconceptions

A final distinction is essential: periodicity of trajectories does not by itself define a new equilibrium concept. In finite two-player zero-sum games with non-singleton Nash sets, the equilibrium object may be a convex face or polytope $u(x,y):=x^{\rm T}Uy.$2, all of whose members share the same minimax value

$u(x,y):=x^{\rm T}Uy.$3

The relevant question is then equilibrium selection, not periodic equilibrium. In that setting, regularized last-iterate methods such as R-NaD and magnetic mirror descent empirically select the maximum-entropy member—or more generally the information projection of their reference policy onto $u(x,y):=x^{\rm T}Uy.$4—whereas regret-averaging methods such as CFR and CFR$u(x,y):=x^{\rm T}Uy.$5 drift toward lower-entropy faces. The paper is explicit that it is not about periodic Nash equilibrium as a formal equilibrium concept (Leal, 26 Jun 2026).

The conceptual payoff for the theory of periodic Nash equilibria is negative but important. The paper distinguishes a static selected point in the Nash polytope from cyclic or recurrent solver trajectories. Periodicity may appear in unregularized or insufficiently regularized learning dynamics as limit cycling or Poincaré-type recurrence, but the selected Nash point—when convergence occurs—is still a static element of $u(x,y):=x^{\rm T}Uy.$6. The same work further shows that solver dependence is algorithmic rather than seed-driven, and that moving-reference regularization changes the selected equilibrium without turning equilibrium itself into a periodic object (Leal, 26 Jun 2026).

This suggests a useful taxonomy. Periodic Nash equilibria, in the strongest sense, are equilibrium profiles that are themselves periodic objects, as in periodic zero-sum environments or LQ feedback cycles. Periodic strategies are a different non-cooperative solution concept defined through operator iteration. Periodic learning trajectories, finally, may either recover equilibrium by averaging or obstruct convergence altogether. Treating these as interchangeable obscures the technical structure of the subject.