Information Set Nash Equilibrium (ISNE)

• ISNE is a local equilibrium metric that normalizes counterfactual regret at each information set in two-player zero-sum extensive-form games with imperfect information.
• It guides adaptive perturbation by updating the epsilon value once the maximum local regret falls below a threshold, thereby refining approximate extensive-form perfect equilibria.
• ISNE addresses limitations of global exploitability measures by revealing stability issues in low-probability, off-equilibrium branches, ensuring robust equilibrium quality.

Searching arXiv for the cited ISNE-related papers and closely related equilibrium literature. Information Set Nash Equilibrium (ISNE) denotes, in the specific terminology of "Last-Iterate Convergence in Adaptive Regret Minimization for Approximate Extensive-Form Perfect Equilibrium," a local information-set-level equilibrium diagnostic for two-player zero-sum imperfect-information extensive-form games with perfect recall, defined in an $\epsilon$-perturbed game and used to guide the computation of approximate Extensive-Form Perfect Equilibria (EFPEs) (Ren et al., 11 Aug 2025). In that usage, ISNE is not the terminal refinement concept; rather, it is a max-information-set-regret criterion that tests whether the current strategy is sufficiently stable at every information set to justify shrinking the perturbation parameter. The term is not standard across the broader Nash-equilibrium literature represented here: several adjacent works analyze sequential equilibrium existence, solver-dependent equilibrium selection, distributed Nash seeking under informational constraints, or sequential equilibrium-like notions under imperfect information, but do not define ISNE as such (Roux, 2013, Leal, 26 Jun 2026, Salehisadaghiani et al., 2017, Bianchi et al., 2020, Arabneydi et al., 2019).

1. Origin and formal setting

The principal formal source for ISNE is the 2025 paper on adaptive regret minimization for approximate EFPE in two-player zero-sum extensive-form games (EFGs) with imperfect information and perfect recall (Ren et al., 11 Aug 2025). The game model uses histories $H$, terminal histories $Z$, a player function $P$, information partitions $\mathcal{I}_i$, behavioral strategies $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$, and sequence-form strategies $\boldsymbol{q}_i$. Behavioral and sequence-form representations are linked by

$$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$

ISNE is introduced inside the perturbed-game framework used to approximate EFPE. The perturbed game is

$$\Gamma^\epsilon = (\mathcal{Q}_1^\epsilon,\mathcal{Q}_2^\epsilon,\boldsymbol{U}),$$

with behavioral lower bounds

$$\sum_{a \in A(I)} \boldsymbol{x}_i[Ia] = 1, \quad \boldsymbol{x}_i[Ia] \geq \epsilon.$$

An $H$0-EFPE is simply a Nash equilibrium of this perturbed game. The motivation is that ordinary Nash equilibrium in imperfect-information EFGs can be too weak on off-equilibrium branches, because information sets that are rarely reached may remain poorly trained. Perturbations enforce minimum action probabilities, but fixed-$H$1 computation creates a trade-off: large $H$2 eases training in low-reach parts of the tree while degrading approximation to the unperturbed game, whereas small $H$3 improves approximation but can make optimization in deep low-probability branches slow and unstable.

Within that framework, ISNE is introduced as a criterion for local strategic stability at the level where imperfect-information pathologies occur: the information set. A plausible implication is that the definition is intended to capture the quality of a current perturbed-game solution more finely than a single global exploitability statistic.

2. Formal definition

The paper first defines an information-set-local regret quantity. For player $H$4 and information set $H$5,

$H$6

Here $H$7 is the vector of action values assuming all players reach $H$8. The defining distinction from ordinary counterfactual regret is explicit: information set regret assumes the opponent reaches $H$9 with probability $Z$0. The exact relation given in the paper is

$Z$1

where $Z$2 is the usual counterfactual regret and $Z$3 is the opponent’s reach probability for node $Z$4.

This normalization is the mathematical core of ISNE. Counterfactual regret can become numerically small merely because an information set is rarely reached; $Z$5 removes that damping factor. The paper states the local equilibrium interpretation directly: strategy stability at $Z$6 is proportional to the maximum regret $Z$7, with an equilibrium achieved when $Z$8. Equivalently, for action $Z$9,

$P$0

so $P$1 means that no pure action at $P$2 locally improves on the current mixed action under the conditional continuation values $P$3.

The global ISNE residual aggregates these local quantities over both players and all information sets: $P$4 A strategy $P$5 is a $P$6-ISNE if

$P$7

Accordingly, ISNE is not a separate equilibrium object attached to the entire extensive form in the same way that Nash equilibrium or EFPE are; it is a max-norm residual built from local information-set deviation advantages. The exact logic is hierarchical: define local information-set regret, take its positive maximum at each information set, then the maximum over all information sets and both players, and compare the result to a tolerance $P$8.

3. Role in EFPE computation

In the paper’s algorithmic architecture, ISNE serves as the trigger for adaptive perturbation during RTCFR, a reward-transformation variant of counterfactual regret minimization designed for last-iterate convergence in fixed perturbed games (Ren et al., 11 Aug 2025). The algorithm is organized into $P$9 reward-transformation saddle-point phases (RT-BSPPs), each of length $\mathcal{I}_i$0, with strategy and reference strategy transferred across phases: $\mathcal{I}_i$1

At the beginning of each phase, the perturbation and ISNE tolerance are updated when the current strategy satisfies the ISNE threshold: $\mathcal{I}_i$2 followed by updating the basis matrix $\mathcal{I}_i$3 using the new $\mathcal{I}_i$4 for all $\mathcal{I}_i$5. The prose description states that the method “uses $\mathcal{I}_i$6-ISNE as a threshold to dynamically reduce the perturbation $\mathcal{I}_i$7,” and that within each RT-BSPP it monitors $\mathcal{I}_i$8 and reduces $\mathcal{I}_i$9 and $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$0 when $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$1. The pseudocode instead checks $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$2 at the start of each RT-BSPP. This suggests that ISNE is intended as a current-iterate, and more specifically a last-iterate, diagnostic rather than an average-iterate statistic.

The functional role of ISNE is therefore continuation in perturbation space. RTCFR attempts to solve a fixed $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$3-perturbed game; ISNE measures whether that current problem has been solved well enough locally at every information set; once it has, the perturbation is decreased, tightening the approximation to EFPE. The basis matrices $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$4, which map simplex coordinates into the perturbed strategy simplex, are then recomputed. In this construction, local equilibrium quality controls the schedule by which the algorithm moves from coarse perturbations toward the $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$5 regime required by EFPE.

4. Relation to Nash equilibrium, EFPE, and exploitability

The paper is explicit that the final target remains EFPE, not ISNE (Ren et al., 11 Aug 2025). The conceptual hierarchy is:

• NE: the global no-profitable-deviation condition in the original game.
• $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$6-EFPE: a Nash equilibrium of the $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$7-perturbed game $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$8.
• EFPE: the limit of $\boldsymbol{x}_i(I)\in \Delta^{A(I)}$9-EFPE as $\boldsymbol{q}_i$0.
• $\boldsymbol{q}_i$1-ISNE: a local max-information-set-regret condition in a fixed perturbed game.

ISNE is introduced because exploitability and standard counterfactual-regret diagnostics may fail to detect severe local disequilibrium at information sets with very small reach probability. The paper discusses the relation to exploitability qualitatively rather than through a theorem, stating that

$\boldsymbol{q}_i$2

typically holds in the $\boldsymbol{q}_i$3-perturbed game $\boldsymbol{q}_i$4. The paper does not present this as a formal bound. The intended point is that a strategy can have low global exploitability while still exhibiting poor local prescriptions in low-reach branches, whereas ISNE directly tests those branches by conditioning on arrival at the information set.

The paper also states that “Reducing $\boldsymbol{q}_i$5 yields a more precise $\boldsymbol{q}_i$6-EFPE, with an exact $\boldsymbol{q}_i$7-EFPE achieved as $\boldsymbol{q}_i$8.” In the adaptive-perturbation discussion, it further argues informally that RTCFR’s asymptotic last-iterate convergence in a fixed perturbed game implies that $\boldsymbol{q}_i$9 vanishes as $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$0, so a $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$1-ISNE converges to an $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$2-EFPE as $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$3. Thus, ISNE functions as an approximation metric for perturbed-game equilibrium quality, not as a replacement for equilibrium refinement.

A further interpretive claim in the paper is that $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$4-ISNE captures subgame perfection in imperfect-information games, analogously to subgame-perfect equilibrium in perfect-information games, conditioned on state distributions within information sets. This should be read as an analogy rather than an identity claim. The paper’s formal definition remains a local regret residual; it does not redefine EFPE or introduce a separate belief-consistency requirement.

5. Relation to adjacent literatures

The term ISNE is not used uniformly across the neighboring equilibrium literature in the information set provided here. The 2013 paper "Infinite sequential Nash equilibrium" studies pure-strategy Nash equilibrium in infinite sequential games of perfect information, with strategies defined on full histories $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$5, not on information sets; it therefore provides background on sequential extensive-form Nash existence but not a source for information-set-based equilibrium notions (Roux, 2013). Its strategy profile $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$6 presupposes perfect observability of history, and the paper explicitly does not model imperfect information.

The 2026 paper on solver-dependent equilibrium selection in zero-sum Nash polytopes does not define or use ISNE either, but it is highly relevant to a different issue: multiple Nash equilibria in imperfect-information extensive-form games may imply different prescriptions at information sets, even when they share the same minimax value $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$7 (Leal, 26 Jun 2026). In that paper, behavioral strategies are defined over information sets, the equilibrium set $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$8 is a Nash polytope in sequence-form coordinates, and solver families select different members of that set. The connection to ISNE is therefore about information-set behavior and equilibrium selection, not about a local regret certificate.

The distributed Nash-seeking papers under partial-decision information analyze informational constraints of another kind. "Distributed Nash Equilibrium Seeking under Partial-Decision Information via the Alternating Direction Method of Multipliers" and "Nash equilibrium seeking under partial-decision information over directed communication networks" both compute the standard Nash equilibrium when players can only communicate locally and must maintain estimates of others’ actions; neither paper defines ISNE or an information-set-specific equilibrium concept (Salehisadaghiani et al., 2017, Bianchi et al., 2020). Their informational restrictions are algorithmic and network-theoretic rather than extensive-form information partitions.

By contrast, "Deep Nash and Sequential Mean-Field Equilibria in Cooperative and Non-cooperative Games with Imperfect Information Structures" is described as highly relevant to ISNE even though it does not use classical extensive-form terminology (Arabneydi et al., 2019). Its deep Nash equilibrium is an exact sequential equilibrium notion under imperfect information structures such as deep-state sharing, where strategies are defined on restricted observation histories like $$\boldsymbol{q}_i[Ia] = \prod_{I'a'\in \Sigma_i: I'a' \preceq Ia} \boldsymbol{x}_i[I'a'], \qquad \boldsymbol{x}_i[Ia] = \frac{\boldsymbol{q}_i[Ia]}{\boldsymbol{q}_i[pI]}.$$9 rather than on explicit information sets. This suggests a broader family of ISNE-like ideas in dynamic stochastic games: sequential rationality may be enforced on observation-based information states without adopting the classical extensive-form partition language.

6. Significance, limitations, and empirical role

The main significance of ISNE, in its defined sense, is diagnostic granularity. It evaluates local no-profitable-deviation conditions at every information set after renormalizing out opponent reach probability, so low-probability branches remain visible to the optimization process (Ren et al., 11 Aug 2025). This is especially important for EFPE computation, because perturbation-based refinements are designed precisely to make off-equilibrium prescriptions meaningful under trembles. Figure 1 in the paper is captioned as a comparison for NE computation by exploitability and EFPE computation by maximum information set regret, making clear that the empirical quantity labeled “Max Regret” is exactly the ISNE residual $$\Gamma^\epsilon = (\mathcal{Q}_1^\epsilon,\mathcal{Q}_2^\epsilon,\boldsymbol{U}),$$0. In the largest Liar’s Dice experiment, the adaptive method is reported to achieve “an ISNE with regret below $$\Gamma^\epsilon = (\mathcal{Q}_1^\epsilon,\mathcal{Q}_2^\epsilon,\boldsymbol{U}),$$1.”

The concept also has clear limitations. First, ISNE is introduced in a specific algorithmic setting: two-player zero-sum EFGs with imperfect information and perfect recall, under perturbation constraints $$\Gamma^\epsilon = (\mathcal{Q}_1^\epsilon,\mathcal{Q}_2^\epsilon,\boldsymbol{U}),$$2. Second, the paper does not prove a standalone theorem about ISNE as an independent equilibrium concept; its theoretical status is derivative of RTCFR convergence in fixed perturbed games. Third, the broader literature represented here does not treat ISNE as a standard universal term. Closely related works address perfect-information sequential Nash existence, distributed NE under local observability, solver-dependent equilibrium selection, or sequential imperfect-information equilibria under observation-based state summaries, but not the same object.

A plausible conclusion is that ISNE should be understood in two layers. In the narrow technical sense, it is the $$\Gamma^\epsilon = (\mathcal{Q}_1^\epsilon,\mathcal{Q}_2^\epsilon,\boldsymbol{U}),$$3-thresholded max-information-set-regret metric of (Ren et al., 11 Aug 2025). In a broader interpretive sense, it points toward a persistent concern in imperfect-information game solving: equilibrium quality is not exhausted by global exploitability, because local prescriptions at information sets—especially rarely reached ones—may matter both for equilibrium refinement and for downstream behavioral robustness.