Quantum Zero-Sum Games Overview

• Quantum zero-sum games are adversarial interactions where one player’s gain equals the other's loss, enhanced by quantum states, entanglement, or operations.
• They adopt various formulations—from classical matrix games accelerated by quantum Gibbs sampling to density-matrix strategies and quaternionic implementations—altering equilibrium computation.
• Researchers investigate nonclassical geometries on spectrahedra and S³, develop efficient quantum algorithms, and analyze learning dynamics and convergence in these settings.

Searching arXiv for recent and foundational papers on quantum zero-sum games. Quantum zero-sum games are adversarial games in which one player’s gain is exactly the other’s loss, but the strategy spaces, payoff rules, or equilibrium computation are modified by quantum information-theoretic structure. In the literature, the term covers at least three distinct but related settings. In one line of work, the underlying game remains a classical two-player zero-sum matrix game, while quantum algorithms accelerate the computation of approximate Nash equilibria (Apeldoorn et al., 2019, Bouland et al., 2023, Gao et al., 2023, Do et al., 8 Apr 2026). In a second line, the strategies themselves are quantum states or quantum operations, and payoffs are expectation values of Hermitian observables or measurement outcomes on joint quantum systems (Jain et al., 2022, Vasconcelos et al., 2023, Ickstadt et al., 2023, Erbanni et al., 2023). A third line studies quantum implementations of classically defined games, especially under Eisert–Wilkens–Lewenstein and related protocols, where entanglement and unitary actions induce nonclassical equilibrium structure; in the $2\times 2$ zero-sum case, a notable consequence is that any mixed quantum equilibrium yields the average of the four classical payoffs (Landsburg, 2011). Across these formulations, the central themes are minimax structure, equilibrium geometry on nonclassical domains such as $S^3$, density-matrix spaces, or spectraplexes, and the algorithmic and informational role of entanglement, discord, Gibbs sampling, and extra-gradient learning.

1. Formal models of quantum zero-sum interaction

The most classical formulation treats a zero-sum game as a payoff matrix $A \in [-1,1]^{n\times m}$, with mixed strategies $x \in \Delta^n$, $y \in \Delta^m$, and expected payoff $x^\top A y$. In this setting, the game itself is classical; the quantum aspect lies entirely in the algorithm used to compute an $\varepsilon$-approximate Nash equilibrium (Apeldoorn et al., 2019, Bouland et al., 2023, Gao et al., 2023). This formulation preserves the standard minimax equation

$$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$

and embeds equilibrium computation into LP- or no-regret-based frameworks.

A more intrinsically quantum formulation replaces probability vectors by density matrices. In the two-player zero-sum model analyzed via Matrix Multiplicative Weights Update and Quantum Replicator Dynamics, Alice’s strategy is a density operator $\rho \in D(\mathcal{A})$, Bob’s strategy is $\sigma \in D(\mathcal{B})$, and a referee measures a Hermitian payoff observable $S^3$0 on $S^3$1, yielding expected payoff

$S^3$2

Bob’s payoff is $S^3$3, so the game is exactly zero-sum (Jain et al., 2022). Closely related formulations define the same bilinear payoff through a positive super-operator $S^3$4, with

$S^3$5

and equilibrium defined as the quantum saddle point (Jain et al., 2022).

A third formalization uses quantum strategic implementations of classical games. In the Eisert–Wilkens–Lewenstein model for a classical $S^3$6 game $S^3$7, each player acts on one qubit of an entangled pair using a unitary in $S^3$8, identified with a unit quaternion on $S^3$9, and classical outcome probabilities are derived from the squared coordinates of the quaternion product $A \in [-1,1]^{n\times m}$0. The induced quantum game $A \in [-1,1]^{n\times m}$1 has payoff

$A \in [-1,1]^{n\times m}$2

where $A \in [-1,1]^{n\times m}$3 ranges over the four classical outcomes (Landsburg, 2011). In this model, pure quantum strategies are points on $A \in [-1,1]^{n\times m}$4, and mixed quantum strategies are Borel probability measures on $A \in [-1,1]^{n\times m}$5.

Other formulations broaden the notion further. In the Hilbert-space model of quantum superpositions, each player chooses a local state and a unitary on the joint system, and the outcome is the final joint state in $A \in [-1,1]^{n\times m}$6, evaluated through a preferred measurement basis (Khan et al., 2011). In semidefinite network games, each player’s strategy is a density matrix $A \in [-1,1]^{n\times m}$7, and payoffs are local expectation values $A \in [-1,1]^{n\times m}$8 on graph edges; the zero-sum condition is imposed globally by $A \in [-1,1]^{n\times m}$9 for all profiles (Ickstadt et al., 2023). In unitary many-body games, the shared object is a quantum register, each player applies a unitary constrained by entangling capability, and the payoff is the final energy expectation of a Hamiltonian, with one player maximizing and the other minimizing (Erbanni et al., 2023).

2. Equilibrium notions and minimax structure

In classical matrix games, a mixed Nash equilibrium is a saddle point of the bilinear objective. The quantum-algorithmic literature preserves this notion unchanged: an $x \in \Delta^n$0-optimal pair $x \in \Delta^n$1 is one satisfying an additive equilibrium gap bound such as

$x \in \Delta^n$2

or, equivalently, an $x \in \Delta^n$3-approximate Nash equilibrium in the minimax sense (Apeldoorn et al., 2019). The more recent online-learning formulation expresses the same condition through regret: if the row and column players incur total regrets $x \in \Delta^n$4 and $x \in \Delta^n$5, then the average strategies $x \in \Delta^n$6 satisfy

$x \in \Delta^n$7

linking low-regret dynamics directly to approximate equilibrium (Gao et al., 2023).

In density-matrix quantum zero-sum games, a quantum Nash equilibrium $x \in \Delta^n$8 is defined by

$x \in \Delta^n$9

so no player can unilaterally improve by changing their quantum state (Jain et al., 2022). Because $y \in \Delta^m$0 and $y \in \Delta^m$1 are convex and compact and $y \in \Delta^m$2 is bilinear, a quantum analogue of the minimax theorem applies in this setting (Jain et al., 2022).

In the Jain–Watrous–Vasconcelos line of non-interactive refereed quantum zero-sum games, the strategy sets are spectraplexes

$y \in \Delta^m$3

the payoff is $y \in \Delta^m$4, and equilibrium is characterized through a variational inequality associated with the monotone gradient operator

$y \in \Delta^m$5

where $y \in \Delta^m$6 and $y \in \Delta^m$7 are super-operators induced by the payoff observable (Vasconcelos et al., 2023, Su et al., 25 Sep 2025). This formulation supports first-order methods such as matrix mirror descent, matrix extragradient, and OGDA.

The EWL quaternionic formulation introduces a distinct but still minimax-compatible equilibrium notion. Mixed strategies are measures $y \in \Delta^m$8, with expected payoff

$y \in \Delta^m$9

A mixed-strategy Nash equilibrium in $x^\top A y$0 is a pair $x^\top A y$1 such that each measure maximizes expected payoff against the other (Landsburg, 2011). Although the strategy space is continuous and high-dimensional, the equilibrium structure is unexpectedly rigid.

3. Geometry of quantum strategy spaces

A central difference between classical and quantum zero-sum games is the geometry of feasible strategies. Classical mixed strategies live on simplices. Quantum state strategies live on spectrahedra or spectraplexes, while EWL strategies live on $x^\top A y$2 or distributions over $x^\top A y$3.

In the EWL $x^\top A y$4 model, each unitary is identified with a unit quaternion

$x^\top A y$5

The outcome probabilities of the four classical profiles are proportional to $x^\top A y$6, so the quantum strategy geometry is literally encoded by quaternion multiplication on the 3-sphere (Landsburg, 2011). Theorem 2.1 shows that every mixed strategy is equivalent to one supported on at most four points, which can be taken to form an orthonormal basis of $x^\top A y$7 (Landsburg, 2011). Best responses are intersections of $x^\top A y$8 with linear subspaces of $x^\top A y$9, because optimizing against a fixed mixed strategy reduces to maximizing a quadratic form on the sphere (Landsburg, 2011). For generic games, equilibria fall into a short list of support geometries, including four-point uniform supports, three-point subsets of $\varepsilon$0, two-point orthogonal supports in parallel planes, special fully intertwined configurations, and pure strategies (Landsburg, 2011).

In density-matrix models, the geometry is semidefinite rather than spherical. The strategy set is

$\varepsilon$1

whose extreme points are rank-one projectors, so unlike simplices it has uncountably many extreme points (Jain et al., 2022, Su et al., 25 Sep 2025). This nonpolyhedral geometry originally suggested that the classical linear last-iterate convergence phenomena might fail in quantum zero-sum games. The 2025 result on metric subregularity over spectraplexes shows otherwise: despite curved boundaries and infinitely many extreme points, an error bound of the form

$\varepsilon$2

holds for bilinear quantum zero-sum games, where $\varepsilon$3 is the duality gap and $\varepsilon$4 is a joint density-matrix profile (Su et al., 25 Sep 2025). This geometric fact underpins logarithmic last-iterate convergence.

In semidefinite network games, each player’s strategy space is likewise

$\varepsilon$5

and the network equilibrium set in the zero-sum case is characterized as the projection of a spectrahedron, namely the feasible region of an SDP (Ickstadt et al., 2023). This establishes a precise geometric bridge between multiplayer quantum-like zero-sum interaction and semidefinite optimization.

4. Zero-sum specializations and payoff identities

Several papers isolate zero-sum structure as producing unusually clean equilibrium payoff statements. The most explicit is the EWL $\varepsilon$6 result. If the underlying classical payoffs are $\varepsilon$7 with $\varepsilon$8, then Theorem 4.1 implies that in any mixed quantum equilibrium, Player 1 earns at least $\varepsilon$9, and the zero-sum specialization forces equality (Landsburg, 2011). Thus, in any mixed quantum equilibrium of a $$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$0 zero-sum game,

$$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$1

Equivalently, the value of the quantum game is the average of the four classical payoffs, independent of which equilibrium geometry occurs (Landsburg, 2011). This contrasts sharply with classical $$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$2 zero-sum games, where the value generally depends on the full matrix structure rather than the simple arithmetic mean.

A different zero-sum specialization arises in the study of shared correlations and quantum advantage. For the $$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$3-strategy “quantum dice” family with payoff matrices

$$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$4

the paper proves that the unique classical correlated equilibrium is the uniform distribution $$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$5, which is fair and gives both players expected payoff $$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$6 (Wei et al., 2015). This classical equilibrium remains stable under classical deviations, but it can be exploitable by a quantum player depending on the structure of the shared state. If the shared state is a symmetric zero-discord state

$$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$7

then the possibility of quantum advantage is governed by the basis-overlap matrix

$$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$8

When $$\min_{y \in \Delta^m} \max_{x \in \Delta^n} x^\top A y = \max_{x \in \Delta^n} \min_{y \in \Delta^m} x^\top A y,$$9 is full rank, no quantum advantage exists; for $\rho \in D(\mathcal{A})$0, zero discord always rules out advantage; but for $\rho \in D(\mathcal{A})$1, rank deficiency of $\rho \in D(\mathcal{A})$2 enables explicit symmetric zero-discord, separable states yielding positive quantum advantage in a zero-sum game (Wei et al., 2015). The paper’s $\rho \in D(\mathcal{A})$3 example attains quantum advantage $\rho \in D(\mathcal{A})$4 while the state remains separable and discord-free in the $\rho \in D(\mathcal{A})$5 basis (Wei et al., 2015). This directly challenges the common identification of discord as the necessary resource for quantum advantage in adversarial settings.

In unitary many-body zero-sum games, the payoff is the energy of

$\rho \in D(\mathcal{A})$6

with Alice maximizing and Bob minimizing the final expectation value (Erbanni et al., 2023). If both players can implement arbitrary unitaries on the full register, the second mover can undo the first player’s move and then apply their own optimal unitary, producing a strong second-mover advantage (Erbanni et al., 2023). If, however, the first player can entangle more qubits than the second, the second mover cannot generally reverse the entanglement structure. In the minimal-advantage case $\rho \in D(\mathcal{A})$7, $\rho \in D(\mathcal{A})$8, the minimum energy per qubit the second player can enforce is

$\rho \in D(\mathcal{A})$9

which approaches the ground-state energy density as $\sigma \in D(\mathcal{B})$0 grows but equals $\sigma \in D(\mathcal{B})$1 for the $\sigma \in D(\mathcal{B})$2 Bell-state case (Erbanni et al., 2023). This suggests that entangling capability acts as a quantum strategic resource even in a pure zero-sum objective.

5. Quantum advantage, correlation, and strategic resources

A recurrent question in quantum zero-sum games is what resource makes quantum play outperform classical play. The answer varies across models.

In the correlation-based zero-sum matrix-game setting, entanglement is sufficient but not necessary. The paper on advantage without discord first recalls an entangled Penny Matching example in which a player can convert a fair correlated equilibrium into a certain win via a Hadamard operation (Wei et al., 2015). It then gives a separable positive-discord example where a Hadamard transform changes the outcome probabilities from uniform to $\sigma \in D(\mathcal{B})$3, raising the winning probability from $\sigma \in D(\mathcal{B})$4 to $\sigma \in D(\mathcal{B})$5 and expected payoff to $\sigma \in D(\mathcal{B})$6 (Wei et al., 2015). Most significantly, for $\sigma \in D(\mathcal{B})$7 it constructs zero-discord, separable, symmetric states that still enable quantum advantage, so neither entanglement nor discord is a universal operational characterization of strategic quantum advantage in zero-sum games (Wei et al., 2015).

In the EPR-based formulation of $\sigma \in D(\mathcal{B})$8-player quantum games, players’ strategic choices remain classical—they choose between two measurement directions—but the shared state is an entangled $\sigma \in D(\mathcal{B})$9-qubit GHZ or W state (Abu-Zayyad et al., 2012). The probability of a joint outcome under a GHZ-type state contains terms involving $S^3$00 and $S^3$01, where $S^3$02 is the entanglement angle and $S^3$03 encodes multipartite correlations (Abu-Zayyad et al., 2012). The game reduces to the classical one at $S^3$04, because the distribution factorizes into a product of local terms. Although that work focuses on Prisoners’ Dilemma and related games rather than an explicit zero-sum family, the formalism directly accommodates zero-sum payoff tensors satisfying $S^3$05 for each outcome string (Abu-Zayyad et al., 2012). This suggests that in EPR-style zero-sum games, entanglement alters equilibrium structure without enlarging the players’ local strategic menus, a useful distinction from EWL-style quantization.

In unitary sequential games, the relevant resource is not merely entanglement in the shared state but control over entangling unitaries. The paper studies absolutely maximally entangled states, Haar-random unitaries, and mixed-state ergotropy as means of defending against the second mover (Erbanni et al., 2023). When absolutely maximally entangled states exist, the first player can render any subsystem up to half the register maximally mixed, neutralizing certain local counterstrategies; when they do not exist, random unitaries still provide high average subsystem entropy and partially suppress the second-mover advantage (Erbanni et al., 2023). For mixed initial states, the player with larger entangling capability can access more ergotropy than any local strategy, linking zero-sum game advantage to the theory of quantum batteries (Erbanni et al., 2023).

6. Algorithmic computation of equilibria

One major branch of the literature studies quantum speedups for computing equilibria of classical zero-sum matrix games. The 2019 algorithm based on efficient Gibbs sampling methods reformulates equilibrium computation as LP solving and obtains $S^3$06-approximate Nash equilibria in complexity

$S^3$07

in the dense access model and

$S^3$08

in the sparse model, where $S^3$09 is sparsity (Apeldoorn et al., 2019). This work also reduces general LP solving to zero-sum games, producing quantum LP solvers with corresponding complexities in terms of a scale-invariant precision parameter $S^3$10 (Apeldoorn et al., 2019).

The 2023 improvement via dynamic Gibbs sampling reduces the dense-model runtime to

$S^3$11

improving on the prior $S^3$12 dependence by designing a data structure for sampling from slowly changing Gibbs distributions (Bouland et al., 2023). The core insight is that in the Grigoriadis–Khachiyan style mirror-descent process, each Gibbs distribution changes only multiplicatively by a constant factor over a phase of $S^3$13 iterations, allowing reusable overestimating “hints” and amplitude-amplified rejection sampling (Bouland et al., 2023).

A complementary 2023 line develops an online quantum algorithm based on optimistic multiplicative weights. For an $S^3$14 matrix game, it achieves total regret $S^3$15 and computes an $S^3$16-approximate Nash equilibrium in quantum time

$S^3$17

using a fast quantum multi-sampling procedure for Gibbs sampling (Gao et al., 2023). Unlike earlier quantum MWU algorithms with $S^3$18 regret, this combines the optimistic FTRL/RVU framework with quantum Gibbs multi-sampling, matching the best known dimension dependence while improving regret to the logarithmic regime in the number of actions (Gao et al., 2023).

For quantum zero-sum games in which strategies are density matrices rather than classical mixed strategies, the classical baseline is the Jain–Watrous Matrix Multiplicative Weights Update method, which achieves $S^3$19 iterations to $S^3$20-equilibria in the $S^3$21-dimensional spectraplex (Vasconcelos et al., 2023). The 2023 extra-gradient work introduces a hierarchy of matrix optimization algorithms, culminating in Optimistic Matrix Multiplicative Weights Update (OMMWU), with average-iterate complexity $S^3$22 (Vasconcelos et al., 2023). The gradient operator

$S^3$23

is monotone and Lipschitz, enabling a matrix mirror-prox analysis and a quadratic speedup in $S^3$24 relative to Jain–Watrous (Vasconcelos et al., 2023).

The 2025 result breaks the apparent $S^3$25 barrier. By proving a semidefinite metric-subregularity condition over spectraplexes, it shows that matrix variants of Nesterov’s iterative smoothing and OGDA achieve linear last-iterate convergence,

$S^3$26

matching the classical polyhedral case despite the nonpolyhedral quantum feasible set (Su et al., 25 Sep 2025). This removes the previously conjectured separation between simplex geometry and spectraplex geometry for bilinear zero-sum games.

A different algorithmic branch uses variational quantum circuits to solve classical zero-sum matrix games. The projected variational quantum extragradient framework represents mixed strategies as Born distributions of PQCs, embeds arbitrary $S^3$27 games into power-of-two dimensions via a dominated embedding that preserves equilibria, and optimizes the parametric saddle objective

$S^3$28

with stochastic extragradient updates (Do et al., 8 Apr 2026). The parameter-shift rule gives unbiased gradient estimators with variance $S^3$29 in the number of shots $S^3$30, and the method converges to approximate first-order stationarity under bounded-domain and smoothness assumptions (Do et al., 8 Apr 2026). Since stationarity in parameter space does not imply game-space equilibrium, the paper evaluates performance using the Nash gap and reports high-precision solutions on structured instances up to $S^3$31, while emphasizing the difficulties of unstructured games (Do et al., 8 Apr 2026).

7. Learning dynamics and recurrence

The learning-theoretic view of quantum zero-sum games asks not only whether equilibria can be computed, but what trajectories arise under adaptive play. In the finite-dimensional density-matrix model, Matrix Multiplicative Weights Update takes the form

$S^3$32

$S^3$33

which generalizes multiplicative weights from vectors to density operators (Jain et al., 2022). The continuous-time limit yields quantum replicator dynamics.

The crucial dynamical result is that if $S^3$34 is a fully mixed Nash equilibrium, then the total quantum relative entropy to equilibrium is conserved: $S^3$35 Thus, the dynamics preserve an information-theoretic integral of motion (Jain et al., 2022). Combined with a volume-preservation argument on a canonically transformed matrix space, this yields a Poincaré recurrence theorem: for almost all interior initial conditions, trajectories return arbitrarily close to their starting point infinitely often (Jain et al., 2022). The implication is that day-to-day learning trajectories in zero-sum quantum games need not converge pointwise to equilibrium, even though time averages can converge and MMWU can compute approximate equilibria. This mirrors classical cycling phenomena in zero-sum replicator dynamics, but in a noncommutative state space.

A plausible implication is that adversarial quantum learning systems such as quantum GANs inherit the same tension already known classically: equilibrium computation via averaging may be tractable, while naive online dynamics remain oscillatory or recurrent unless modified by optimism, extra-gradient correction, or additional regularization (Jain et al., 2022, Vasconcelos et al., 2023, Gao et al., 2023).

8. Multiplayer and network generalizations

Although much of the literature focuses on two-player games, several works extend zero-sum structure to larger systems.

The semidefinite network-game framework generalizes two-player SDP games to a graph $S^3$36, where each player $S^3$37 chooses a density matrix $S^3$38, and edge payoffs are

$S^3$39

The game is zero-sum if

$S^3$40

In this case, the set of Nash equilibria is exactly the projection of the optimal solution set of an SDP

$S^3$41

so equilibrium computation reduces to semidefinite optimization (Ickstadt et al., 2023). Beyond zero-sum, the same paper shows that equilibria solve a semidefinite linear complementarity problem, extending classical LCP characterizations of bimatrix and polymatrix games (Ickstadt et al., 2023).

In combinatorial game theory, “quantum zero-sum games” can also refer to quantum algorithms for solving classical impartial constant-sum games such as one-heap Nim and subtraction games. For balanced subtraction games, classical query complexity is $S^3$42, while a quantum dynamic-programming algorithm based on Grover search achieves $S^3$43; for restricted games, an exact $S^3$44 quantum algorithm computes all positional values, again improving on $S^3$45 classical complexity (Kravchenko et al., 2020). The paper extends these results to $S^3$46-player variants without changing asymptotic complexities, thereby supplying an explicitly defined class of zero-sum or constant-sum combinatorial games with provable quantum-over-classical separation in the complexity of solving them (Kravchenko et al., 2020).

The EPR-based $S^3$47-player framework likewise allows zero-sum extensions by specifying outcome-dependent payoffs whose sum is zero for each outcome string. Because the entanglement parameter $S^3$48 tunes the departure from the classical product distribution, this setting offers a controlled way to study how multipartite entanglement alters the minimax landscape while leaving local strategy sets classical (Abu-Zayyad et al., 2012).

9. Conceptual distinctions and recurrent misconceptions

One recurring source of confusion is the phrase “quantum zero-sum game” itself. In some papers, the game is classical and only the solver is quantum (Apeldoorn et al., 2019, Bouland et al., 2023, Gao et al., 2023, Do et al., 8 Apr 2026). In others, the strategy spaces are quantum states or operations and the payoff rule is intrinsically quantum (Jain et al., 2022, Vasconcelos et al., 2023, Ickstadt et al., 2023, Erbanni et al., 2023). These are related but not interchangeable notions. Conflating them obscures which statements are about computational complexity, which are about equilibrium geometry, and which are about physical resources such as entanglement or discord.

A second misconception is that discord or entanglement alone should serve as a universal resource measure for quantum strategic advantage. The zero-discord advantage construction shows that, at least in adversarial two-player settings with different objectives, a player can exploit basis geometry and conditional structure in separable, discord-free states once the strategy dimension exceeds two (Wei et al., 2015). This does not imply that discord is irrelevant, but it does show that its operational meaning in game theory differs from tasks such as state discrimination or communication.

A third misconception is that quantum enlargement of strategy spaces must always favor one player or destroy zero-sum symmetry. The EWL zero-sum result indicates the opposite: the enlarged strategy space can regularize the value to a fixed spectral average while preserving exact zero-sum payoffs (Landsburg, 2011). Likewise, in density-matrix models the zero-sum property is encoded directly at the level of payoff observables or super-operators, so quantum mechanics changes the feasible set and dynamics without changing the adversarial accounting (Jain et al., 2022, Vasconcelos et al., 2023).

A fourth misconception is that nonpolyhedral semidefinite geometry inevitably prevents the linear last-iterate rates seen in classical simplex-constrained bilinear games. The 2025 metric-subregularity result refutes this by establishing an error bound over spectraplexes despite their infinitely many extreme points, thereby supporting $S^3$49 convergence (Su et al., 25 Sep 2025). This suggests that the right analogue of Hoffman-type error bounds survives in semidefinite game geometry.

10. Research directions

Several directions emerge repeatedly. In the algorithmic line, precision dependence remains central: the progression from $S^3$50 to $S^3$51 and then to logarithmic last-iterate convergence in density-matrix quantum games suggests that the main open question is no longer whether quantum zero-sum equilibrium computation can match classical first-order rates, but under what access models, regularity conditions, and output requirements such rates remain practical (Apeldoorn et al., 2019, Bouland et al., 2023, Gao et al., 2023, Su et al., 25 Sep 2025).

In the variational line, the key challenge is the mismatch between parametric stationarity and game-space optimality. Projected VQEG proves convergence to approximate first-order stationarity with shot variance $S^3$52, but this does not guarantee a small Nash gap unless the PQC ansatz is sufficiently expressive and the nonconvex landscape is benign (Do et al., 8 Apr 2026). This suggests that future work must integrate equilibrium-aware certificates more directly into the variational objective, perhaps by optimizing duality-gap surrogates rather than only the saddle objective in parameter space.

In the dynamical line, the coexistence of equilibrium computability and recurrent learning trajectories implies that stability-enhancing methods—optimistic updates, extra-gradient steps, or other regularizations—are likely essential in practical quantum adversarial learning (Jain et al., 2022, Vasconcelos et al., 2023, Gao et al., 2023). This is particularly relevant for quantum GANs and other quantum min–max training problems.

In the multiplayer line, semidefinite network games establish that zero-sum quantum-like equilibria can remain SDP-representable well beyond two players, but only under product-state assumptions (Ickstadt et al., 2023). A plausible implication is that allowing controlled forms of entanglement across players or edges may be the next frontier: it could connect networked quantum games to nonlocal games and multipartite interactive proof systems, while sharply changing tractability.

Overall, quantum zero-sum games constitute a meeting point of quantum information, game theory, optimization, and complexity theory. They range from quaternionic reformulations of $S^3$53 games and density-matrix saddle problems to Gibbs-sampling-based quantum algorithms and semidefinite multiplayer network models. What unifies them is the zero-sum principle itself: equilibrium is a minimax object, and quantum structure—whether in the feasible set, the shared state, the dynamics, or the solver—changes the geometry and computation of that object in ways that are now both mathematically explicit and algorithmically consequential (Landsburg, 2011, Jain et al., 2022, Apeldoorn et al., 2019, Vasconcelos et al., 2023, Su et al., 25 Sep 2025).