Multiplayer Games under Partial Information

• Multiplayer games under partial information are strategic models where players make decisions with incomplete observations of opponents’ actions, types, or states, altering equilibrium outcomes.
• These models span diverse frameworks—including networked noncooperative, extensive-form, and stochastic games—to address communication constraints and computational complexity.
• Algorithmic paradigms such as ADMM, variational inequalities, and reinforcement learning are adapted to manage limited information, ensuring convergence and actionable strategic insights.

Multiplayer games under partial information are strategic models in which at least one player acts without full observation of opponents’ actions, private types, budgets, states, or histories. Across recent research, the topic spans noncooperative games over communication networks where each player has only partial-decision information, stochastic games with incomplete channel knowledge, imperfect-information extensive-form games in sequence form, graph games with partial observation, bidding games with partially-observable budgets, stochastic games with asymmetric common and private information, and differential games with observation delay (Salehisadaghiani et al., 2017, A et al., 2016, Ganzfried, 28 Jun 2026, Kartik et al., 2019, Avni et al., 2022, Zhao et al., 23 Jun 2026). This body of work treats partial information not as a single restriction but as a family of information structures that alter equilibrium definitions, value concepts, algorithm design, and computational complexity.

1. Information structures and model classes

A first major class consists of multiplayer noncooperative games in which each player chooses an action $x_i \in \Omega_i$ and optimizes a local cost $J_i(x_i,x_{-i})$, but does not observe the full action profile. In "Distributed Nash Equilibrium Seeking under Partial-Decision Information via the Alternating Direction Method of Multipliers" (Salehisadaghiani et al., 2017), each player maintains a local estimate of the whole action profile, $x^i=(x_i^i,x_{-i}^i)$, and exchanges estimates over an undirected connected communication graph $G_c(V,E)$ that need not coincide with the cost or interference coupling graph. In "Distributed Algorithms for Complete and Partial Information Games on Interference Channels" (A et al., 2016), the information structure is stratified into complete information, incident-link knowledge, and direct-link-only knowledge, with stationary power policies defined on progressively smaller observation spaces.

A second class is imperfect-information extensive-form play. In "Projected Exploitability Descent for Nash Equilibrium Computation in Multiplayer Imperfect-Information Games" (Ganzfried, 28 Jun 2026) and "Quadratic Programming Approach for Nash Equilibrium Computation in Multiplayer Imperfect-Information Games" (Ganzfried, 30 Sep 2025), players observe only their information sets, and strategies are represented in sequence form under perfect recall. Persistent imperfect information also appears in DAG-structured stochastic games where private types persist across the entire game and are updated via Bayes’ rule along public histories (Ganzfried, 2020).

A third class concerns asymmetric and signal-based observation. In "Zero-sum Stochastic Games with Asymmetric Information" (Kartik et al., 2019), each player’s information decomposes as $I_t^i=C_t \cup P_t^i$, with common information $C_t$ and private information $P_t^i$. In "Zero-Sum Stochastic Games with Partial Information and Average Payoff" (Saha, 2014), both players observe a public signal process and act on a common posterior belief over the hidden state. In "Nash equilibria with partial monitoring; Computation and Lemke-Howson algorithm" (Perchet, 2013), players do not observe opponents’ actions directly and instead receive messages that induce uncertainty correspondences $\Phi$ and $\Psi$.

Other models specialize partial information to domain structure. In "Bidding Graph Games with Partially-Observable Budgets" (Avni et al., 2022), one player’s initial budget is drawn from a known distribution and is not fully observed by the opponent. In "Acquisition Games with Partial-Asymmetric Information" (Kavitha et al., 2019), agents continuously control Poisson search clocks but do not observe the acquisition status of others. In "Multiplayer Reach-Avoid Differential Games with Defender-Side Information Delay" (Zhao et al., 23 Jun 2026), each defender observes other agents’ states with a constant time delay. In "Games with a Weak Adversary" (Chatterjee et al., 2014), partial observation is organized by an information hierarchy, while "On Hyperproperty Verification, Quantifier Alternations, and Games under Partial Information" (Beutner et al., 4 Oct 2025) constructs multiplayer games whose observation relations mirror quantifier order and therefore satisfy hierarchical information.

2. Equilibrium concepts and mathematical representations

For noncooperative action-profile games, the basic equilibrium notion remains Nash equilibrium. In the networked setting of (Salehisadaghiani et al., 2017), $x^*$ is a Nash equilibrium if, for all players $J_i(x_i,x_{-i})$0,

$J_i(x_i,x_{-i})$1

Under convex feasible sets and suitable regularity, the equilibrium admits a variational inequality characterization

$J_i(x_i,x_{-i})$2

where $J_i(x_i,x_{-i})$3 is the pseudo-gradient and $J_i(x_i,x_{-i})$4 is the normal-cone vector to $J_i(x_i,x_{-i})$5. The interference-channel games of (A et al., 2016) use the same VI template in complete, partial, and incomplete information regimes, with affine VI structure under full state knowledge and non-affine gradient-based VI structure when policies depend only on incident or direct links.

In imperfect-information extensive-form games, sequence form is the standard compact representation. For player $J_i(x_i,x_{-i})$6, the feasible realization plans satisfy

$J_i(x_i,x_{-i})$7

and a profile $J_i(x_i,x_{-i})$8 is a Nash equilibrium if no player can improve by unilateral deviation (Ganzfried, 28 Jun 2026). This representation supports both approximate and exact equilibrium metrics. "Projected Exploitability Descent" (Ganzfried, 28 Jun 2026) defines multiplayer exploitability

$J_i(x_i,x_{-i})$9

and the smoother surrogate

$x^i=(x_i^i,x_{-i}^i)$0

By contrast, (Ganzfried, 30 Sep 2025) formulates exact equilibrium conditions as a nonlinear complementarity problem over realization plans, multipliers, and complementarity slacks.

Zero-sum stochastic games under asymmetric information replace Nash equilibrium by upper and lower values, saddle-point strategies, or belief-state dynamic programs. In (Kartik et al., 2019), the common-information belief is

$x^i=(x_i^i,x_{-i}^i)$1

and the expanded virtual game yields a dynamic program on $x^i=(x_i^i,x_{-i}^i)$2. In (Saha, 2014), the public posterior over the hidden state transforms a partially observable zero-sum game into an equivalent completely observable game on the belief space. In partially observable bidding games, however, a value under pure strategies need not exist: (Avni et al., 2022) defines lower and upper values separately and proves that they can differ even in a simple strongly-connected bowtie game.

Partial monitoring alters best-response geometry rather than the underlying solution concept. In (Perchet, 2013), messages induce uncertainty correspondences

$x^i=(x_i^i,x_{-i}^i)$3

and Nash equilibrium is defined through robust maximin best responses with respect to these uncertainty sets. This suggests that partial information changes equilibrium analysis not only by hiding state variables but also by changing the effective payoff correspondence itself.

3. Algorithmic paradigms

Distributed optimization methods under partial information often combine consensus enforcement with first-order updates. The inexact-ADMM method of (Salehisadaghiani et al., 2017) introduces local copies $x^i=(x_i^i,x_{-i}^i)$4, consensus constraints $x^i=(x_i^i,x_{-i}^i)$5 over communication edges, edge slack variables, and aggregated dual variables $x^i=(x_i^i,x_{-i}^i)$6. Each iteration requires one round of neighbor exchange of $x^i=(x_i^i,x_{-i}^i)$7, a dual ascent step,

$x^i=(x_i^i,x_{-i}^i)$8

a projected proximal action update for $x^i=(x_i^i,x_{-i}^i)$9, and an affine estimate update for $G_c(V,E)$0. The method uses constant step-sizes and achieves convergence under strong monotonicity of $G_c(V,E)$1, Lipschitz continuity of $G_c(V,E)$2 and $G_c(V,E)$3, and a graph-regularization parameter $G_c(V,E)$4 chosen above an explicit spectral threshold (Salehisadaghiani et al., 2017).

In stochastic interference-channel games, (A et al., 2016) develops a VI-based fixed-point framework

$G_c(V,E)$5

together with a two-phase heuristic. Phase 1 iterates projected better responses; Phase 2 minimizes the residual objective $G_c(V,E)$6 by projected steepest descent, with restart if the method stalls at a non-equilibrium local minimum. The same paper also derives lower bounds on utilities via water-filling-like policies based only on channel distributions and average power constraints, and gives a Bayesian learning procedure that converges to an $G_c(V,E)$7-Nash equilibrium in the direct-link-only regime.

Reinforcement-learning methods replace equilibrium solves by local temporal-difference updates. In (Medhin et al., 2023), each player observes the common state, its own action, its own reward, and the next state, but not the opponent’s action or strategy. The algorithm updates a marginal Q-vector,

$G_c(V,E)$8

and selects a greedy mixed action from $G_c(V,E)$9. The limiting strategy is proved to be a full-information Nash equilibrium.

For multiplayer imperfect-information extensive-form games, two distinct algorithmic lines are now visible. "Projected Exploitability Descent" (Ganzfried, 28 Jun 2026) performs projected subgradient descent on $I_t^i=C_t \cup P_t^i$0 over sequence-form polytopes, with Euclidean projections onto $I_t^i=C_t \cup P_t^i$1 and subgradients formed from sequence-form best responses. The paper also proposes FP-PED, which runs fictitious play for an initial burn-in period before switching to PED. "Quadratic Programming Approach for Nash Equilibrium Computation in Multiplayer Imperfect-Information Games" (Ganzfried, 30 Sep 2025) instead encodes the sequence-form KKT system as a nonconvex QCQP and solves it with a modern nonconvex quadratic solver, yielding exact equilibria on reduced three-player Kuhn poker.

Dynamic programming on beliefs remains central in asymmetric-information stochastic games. In (Kartik et al., 2019), the value recursion operates on common-information beliefs and prescriptions; in (Saha, 2014), discounted Shapley equations on the belief space are combined with a vanishing-discount argument for the average-payoff case. For partially and asymmetrically informed LQG zero-sum games, (Guan et al., 18 Mar 2026) proposes a forward-backward dynamic programming algorithm that iterates between coupled filtering recursions and a backward Riccati-like saddle-point recursion. For DAG-structured stochastic games with persistent imperfect information, (Ganzfried, 2020) introduces Sequential Topological PI-FP and its type-dependent-value variant, combining fictitious play at each DAG state with Bayes-consistent belief propagation across the graph.

A different algorithmic paradigm appears in finite local-information design. "Potential Games Design Using Local Information" (Li et al., 2018) asks whether one can choose utility functions depending only on local neighborhoods so that a given system-level objective is the exact potential. The existence question reduces to linear equations

$I_t^i=C_t \cup P_t^i$2

and, when these are solvable, the resulting local-information game can be optimized by logit learning or binary restrictive logit learning using only neighborhood observations.

4. Existence, convergence, value, and complexity

Several works establish equilibrium existence and uniqueness under strong regularity. In (Salehisadaghiani et al., 2017), strong monotonicity and Lipschitz continuity of the pseudo-gradient imply a unique Nash equilibrium, and the inexact-ADMM iterates converge to $I_t^i=C_t \cup P_t^i$3 under explicit spectral inequalities on $I_t^i=C_t \cup P_t^i$4 and the proximal penalties $I_t^i=C_t \cup P_t^i$5. In (A et al., 2016), existence of stationary pure-strategy Nash equilibria in complete, partial, and incomplete information power-control games follows from Debreu–Glicksberg–Fan because action sets are compact and convex and utilities are continuous in all powers and concave in own power. In (Saha, 2014), the average-payoff partially observable zero-sum stochastic game has a value, independent of the initial distribution, and admits stationary saddle-point strategies under Lyapunov drift and vanishing normalized growth assumptions. In the one-sided complete-information subclass of (Kartik et al., 2019), the value exists and equilibrium strategies depend only on private information and the common-information belief.

Convergence guarantees vary substantially by method. The distributed ADMM method of (Salehisadaghiani et al., 2017) establishes non-ergodic $I_t^i=C_t \cup P_t^i$6 decay of both the consensus residual and an optimality residual, while empirical comparisons show faster convergence than a consensus-based diminishing-step-size gradient method. The partial-information Q-learning algorithm of (Medhin et al., 2023) converges almost surely for general two-player stochastic games with finitely many states and actions. By contrast, PED in (Ganzfried, 28 Jun 2026) is a first-order method for a nonconvex and nonsmooth objective and has no global convergence guarantees to Nash equilibria in general multiplayer, non-zero-sum settings, although $I_t^i=C_t \cup P_t^i$7 and $I_t^i=C_t \cup P_t^i$8 iff $I_t^i=C_t \cup P_t^i$9 is a Nash equilibrium.

Partial information can also destroy classical value properties. The poorman bidding analysis of (Avni et al., 2022) proves that pure-strategy determinacy can fail: with $C_t$0 and a uniform budget distribution over $C_t$1 in the bowtie game, the lower value is $C_t$2 while the upper value is $C_t$3. This is not a pathology of approximation but a structural statement that the value under pure strategies need not exist. Conversely, for first-price parity and reachability objectives under partial information, and for first-price Richman mean-payoff, the value exists and can be expressed through threshold ratios or random-turn equivalences (Avni et al., 2022).

Formal-language, graph-game, and bounded-memory perspectives reveal sharp decidability boundaries. "Games with a Weak Adversary" (Chatterjee et al., 2014) proves that when player 1 is less informed than player 2, the three-player decision problem for parity objectives is 2-EXPTIME-complete, and EXPSPACE-complete when player 1 is blind. "Bounded-Memory Strategies in Partial-Information Games" (Bose et al., 2024) shows NP-hardness for approximating zero-sum values already with memoryless strategies and 1-player reachability games, while for any fixed number of players $C_t$4 it is decidable in polynomial space whether there exists an $C_t$5-Nash equilibrium in which all strategies use at most $C_t$6 memory modes. "Nash Equilibria in Symmetric Games with Partial Observation" (Bouyer et al., 2014) proves that the bounded-memory symmetric Nash equilibrium existence problem for LTL objectives is EXPSPACE-complete, while the unrestricted imperfect-information problem is undecidable for $C_t$7 players. "Lossy Channel Games under Incomplete Information" (Dimitrova et al., 2013) shows that safety and reachability games are decidable and finite-state observation-based strategies can be effectively computed, but parity and weak parity objectives are undecidable. In hyperproperty verification, (Beutner et al., 4 Oct 2025) exploits hierarchical information to obtain a decidable multiplayer partial-information game for HyperLTL with arbitrary quantifier alternations, while retaining completeness only for the fragment $C_t$8.

5. Canonical domains and applications

Wireless communication and networked control provide some of the most developed application classes. In (A et al., 2016), Gaussian interference channels yield stochastic games in which each transmitter-receiver pair maximizes long-term average rate under an average power constraint, with information regimes ranging from full channel knowledge to direct-link-only knowledge. The same paper derives lower-bound water-filling policies and a distributed method for Pareto points. In (Salehisadaghiani et al., 2017), the partial-decision information framework is illustrated on a networked Nash-Cournot competition with $C_t$9 firms and $P_t^i$0 markets and on a rate-control game in a WANET with 16 links and 15 users. In the WANET example, the ADMM-based approach converges about two orders of magnitude faster than the gossip-based diminishing-step-size method.

Imperfect-information poker remains a canonical benchmark for multiplayer equilibrium computation. In (Ganzfried, 28 Jun 2026), generalized three-player Kuhn poker with deck size $P_t^i$1 is used to compare PED, fictitious play, counterfactual regret minimization, and the FP-PED hybrid. For the 5-card game, optimizing $P_t^i$2 rather than $P_t^i$3 yields final exploitability $P_t^i$4 versus $P_t^i$5, and in the 6-card game the best FP-PED hybrid achieves exploitability and sum gap significantly below $P_t^i$6. Exact computation is still challenging: (Ganzfried, 30 Sep 2025) reports that its QCQP solves reduced three-player Kuhn poker in 2.47 seconds with max gain from deviation at most $P_t^i$7, while the full game was not solved within 24 hours.

Public-resource and competitive-search models show how partial information changes equilibrium timing. In (Li, 2023), a multiplayer war of attrition with asymmetric private information has a unique equilibrium in which one player exits instantly with positive probability and the remaining players exhibit player-specific strict waiting times. In (Kavitha et al., 2019), two-agent two-lock acquisition games with partial-asymmetric information admit Nash equilibria in partial state-dependent time-threshold policies. In (Avni et al., 2022), partially observable budgets in bidding graph games generate wallet-splitting strategies and can eliminate pure-strategy values. These works collectively show that partial information can induce stratified waiting, probing, or revelation dynamics rather than only estimation error.

Differential and stochastic dynamic games provide a control-theoretic counterpart. In (Zhao et al., 23 Jun 2026), defender-side information delay in reach-avoid games yields an explicit delayed attack region,

$P_t^i$8

a convex capture-point optimization problem, and subgame-perfect Nash equilibrium strategies in the one-on-one setting, together with pairwise winning relations and maximum matching for the general multiplayer case. In (Guan et al., 18 Mar 2026), partial and asymmetric information in two-player zero-sum LQG games leads to coupled filtering and control recursions because signaling prevents separation. In (Ganzfried, 2020), a 4-player naval strategic planning scenario with persistent imperfect information serves as a testbed for approximate Nash computation; the type-dependent-value variant attains exploitability about $P_t^i$9, roughly $\Phi$0 of the 100-point terminal payoff scale.

Distributed systems and formal methods provide a different applied frontier. In (Li et al., 2018), local-information-based utilities are designed so that a given system objective becomes the exact potential, enabling local-information learning for consensus and gathering. In (Beutner et al., 4 Oct 2025), multiplayer games under partial information are used to verify HyperLTL formulas with arbitrary quantifier alternations, and prophecy variables are introduced to provide controlled “peeks” into the future. This suggests that partial-information game models are now serving not only as descriptions of strategic interaction but also as algorithmic backends for synthesis and verification.

6. Limitations and open directions

A recurring feature of the literature is that positive results depend on strong structure. Distributed NE seeking in (Salehisadaghiani et al., 2017) assumes convex compact local constraints, differentiable costs, strong monotonicity, Lipschitz continuity, synchronous updates, and an undirected connected communication graph. The stochastic interference-channel models of (A et al., 2016) assume i.i.d. channels, known distributions, stationary policies, and finite action reductions for Bayesian learning. Sequence-form methods such as PED and QCQP assume perfect recall and explicit sequence-form storage, which becomes problematic in larger games (Ganzfried, 28 Jun 2026, Ganzfried, 30 Sep 2025). Belief-state dynamic programming in (Kartik et al., 2019) and (Saha, 2014) relies on finite spaces or compactness, common knowledge of the model, and public observation structures that support a sufficient common belief.

Many papers identify explicit open problems. The ADMM framework of (Salehisadaghiani et al., 2017) leaves directed, time-varying, and asynchronous graphs for future work. The bidding analysis of (Avni et al., 2022) leaves mixed-strategy determinacy, general strongly-connected upper-value characterizations, and extensions beyond mean-payoff poorman games open. PED does not handle imperfect recall and would need sampling or decomposition for games such as three-player Leduc Hold’em (Ganzfried, 28 Jun 2026). The HyperLTL framework of (Beutner et al., 4 Oct 2025) leaves automated synthesis of prophecy families and completeness beyond $\Phi$1 unresolved. The LQG dynamic-game program of (Guan et al., 18 Mar 2026) points to nonzero-sum and $\Phi$2-player extensions, nonlinear dynamics, and convergence analysis of the forward-backward operators. The DAG-structured stochastic-game work of (Ganzfried, 2020) identifies cyclic graphs, stronger refinement notions, and more scalable POMDP-based ex-post evaluation as future directions. The local-information potential-design framework of (Li et al., 2018) suggests near-potential design when the exact linear equations are infeasible.

Taken together, these results suggest a unifying picture. Partial information does not merely make equilibrium computation harder; it changes the appropriate state variables, the equilibrium objects that are well-defined, the geometry of best responses, and the algorithms that remain tractable. In some settings it can be regularized by consensus penalties, common-information beliefs, hierarchical observation, or bounded-memory restrictions; in others it produces nonexistence of pure values, undecidability, or severe scalability barriers. The modern theory of multiplayer games under partial information is therefore best understood as a collection of structurally specialized frameworks, each balancing expressiveness, equilibrium fidelity, and computational feasibility.