Markets are competitive if and only if P != NP

Abstract: I prove that competitive market outcomes require computational intractability. If P = NP, firms can efficiently solve the collusion detection problem, identifying deviations from cooperative agreements in complex, noisy markets and thereby making collusion sustainable as an equilibrium. If P != NP, the collusion detection problem is computationally infeasible for markets satisfying a natural instance-hardness condition on their demand structure, rendering punishment threats non-credible and collusion unstable. Combined with Maymin (2011), who proved that market efficiency requires P = NP, this yields a fundamental impossibility: markets can be informationally efficient or competitive, but not both. Artificial intelligence, by expanding firms' computational capabilities, is pushing markets from the competitive regime toward the collusive regime, explaining the empirical emergence of algorithmic collusion without explicit coordination.

• The paper shows that sustaining collusion requires solving three NP-hard problems—joint pricing, deviation detection, and punishment design—while competitive best responses remain polynomial-time solvable.
• Under a stronger instance-hardness assumption alongside P ≠ NP, computational limits make collusion unsustainable and imply that the competitive stage-game equilibrium prevails in sufficiently complex markets.
• The paper derives an efficiency–competition impossibility and warns that greater transparency or AI adoption could ease monitoring, increase collusion, and require computationally informed antitrust policies.

Overview and central claim

This paper, by Philip Z. Maymin (2602.20415), establishes a formal equivalence between market structure and computational complexity. The central result is that collusion is sustainable as an equilibrium of a repeated oligopoly game if and only if $P = NP$; conversely, when $P \neq NP$ and the market's demand structure satisfies a generic instance-hardness condition, collusion unravels and the competitive outcome prevails. Combined with the author's earlier result that informational efficiency holds if and only if $P = NP$ [maymin2011], the paper derives an "Efficiency–Competition Impossibility": no market can be simultaneously informationally efficient and competitive. The paper's stated motivation is to provide a theoretical foundation for the empirical emergence of algorithmic collusion among AI pricing agents.

The argument inverts the standard institutional account of competition. Rather than attributing competitive outcomes to antitrust enforcement or entry conditions, the paper locates their source in computational limitations: firms cannot sustain collusion because they cannot solve the inference problems required to monitor and punish deviations. This formalizes Stigler's classic observation that secret price-cutting is the chief difficulty of collusion — but recasts informational scarcity as computational hardness.

Model and the three computational problems of collusion

The setting is a repeated game with $N \geq 2$ firms over $T$ periods (finite or infinite with discount factor $\delta$), each selling $K$ products, facing stochastic demand $$q_{ik}^t = D_{ik}(p^t, \theta^t) + \varepsilon_{ik}^t$$ where the demand state $\theta^t$ is unobserved by any firm. A "Richness" assumption requires $|\Theta| \geq NK + 1$, so observed price-quantity data never uniquely identifies the demand state. The game is specified in compact form (polynomial-time computable demand functions, sampling oracles), so the state space may be exponentially large relative to the description size — this gap is what generates hardness.

Sustaining collusion requires solving three decision problems:

• Collusion Strategy Problem (CSP): compute the joint profit-maximizing price vector across the combinatorial product space.
• Collusion Detection Problem (CDP): given observed $P \neq NP$0, determine whether some firm deviated or whether the data are consistent with compliant play under some unobserved demand state.
• Optimal Punishment Problem (OPP): compute the punishment profile minimizing the deviator's continuation payoff.

By contrast, the myopic Competitive Best-Response Problem (CBR) is solvable in polynomial time: it reduces to convex optimization over the firm's own prices using aggregate expected demand, solvable via interior-point methods, with equilibrium convergence under contraction conditions.

Complexity results

Each collusion problem is shown NP-hard by a textbook-style reduction:

Problem	Reduction from	Mechanism
CSP	Max-Weighted-SAT	Binary prices encode truth assignments; clause-weighted demand states make joint profit equal weighted satisfiability
CDP	3-SAT	Clause-firms' demands encode satisfaction indicators; verifying consistency of observed quantities requires finding a satisfying assignment
OPP	Minimum Vertex Cover	Aggressive-pricing choices must cover all complementarity edges at minimum cost

These reductions are direct and rely on constructing games whose description length is polynomial while the state space $P \neq NP$1 is exponential. The implication is stark: planning, monitoring, and enforcing collusion are each computationally hard, whereas competing requires none of these capabilities.

The main theorem

Direction 1 ($P \neq NP$2 implies collusion). If $P \neq NP$3, firms solve CSP, CDP, and OPP in polynomial time. The folk theorem for repeated games with imperfect public monitoring then applies: a perfect public equilibrium sustains the monopoly outcome $P \neq NP$4 provided $P \neq NP$5. Because optimal punishments are exactly computable, $P \neq NP$6 is minimized and $P \neq NP$7 is bounded away from 1 under standard demand structures. The key step is that credible punishment presupposes detection; here detection is exactly what $P \neq NP$8 restores.

Direction 2 ($P \neq NP$9 implies competition). Under an Instance Hardness assumption — that no probabilistic polynomial-time algorithm solves the market's actual CDP instances with probability exceeding $P = NP$0 for negligible $P = NP$1 — the paper proves a lemma establishing profitable undetectable deviations: since $P = NP$2 generically is not a best response for any individual firm, and since richness plus noise ensures deviations lie within the ambiguity set of compliant explanations, each firm can profitably shade toward its best response in a way computationally indistinguishable from compliance. Punishment threats therefore lack credibility, and the unique stage-game Nash equilibrium $P = NP$3 becomes the only equilibrium of the repeated game. Notably, this unsustainability direction holds for the broader class of subgame-perfect equilibria, not merely PPE.

A critical caveat bears directly on the theorem's force: Assumption on instance hardness is strictly stronger than $P = NP$4. The paper argues it holds generically — the set of demand parameters admitting polynomial-time CDP solutions lies in a lower-dimensional algebraic variety of measure zero — but concedes that separable, low-rank, or sparse demand structures violate it, and explicitly does not prove that these exhaust the sources of tractability. The analogy drawn is to cryptographic practice, where specific instances are assumed hard without full characterization.

The Efficiency–Competition Impossibility

Combining the main theorem with Maymin's earlier efficiency result yields the impossibility: informational efficiency requires $P = NP$5; competition (under instance hardness) requires $P = NP$6. Since most cryptography rests on the widely believed conjecture $P = NP$7, the combined prediction is that markets are competitive but informationally inefficient — consistent, the author argues, with persistent mispricing alongside robust sectoral competition. The paper presents this as a strengthening of the Grossman–Stiglitz paradox: rather than a tension between incentives, the computational requirements for efficiency and competition are logically incompatible.

A corollary with immediate policy relevance is the Transparency Paradox: increasing transparency shrinks the ambiguity set of demand states consistent with observations, monotonically reducing the difficulty of CDP and thereby facilitating collusion. Mandated real-time price reporting and algorithmic transparency requirements, conventionally viewed as pro-competitive, may instead lower the computational barrier to monitoring tacit agreements.

Extensions: the AI transition, heterogeneity, and approximate collusion

Modeling firm computational capacity as a bound $P = NP$8 on solvable problem size, the paper characterizes three regimes as minimum capacity $P = NP$9 rises: a competitive regime ($N \geq 2$0), an unstable regime with intermittent collusion and detection-failure price wars ($N \geq 2$1), and a collusive regime ($N \geq 2$2). AI progress moves markets rightward through this curve, which the paper identifies as a "computational phase transition" underlying empirical findings that Q-learning agents learn collusive pricing autonomously [calvano2020], LLMs converge to supra-competitive prices without collusive instructions [fish2024], and algorithmic pricing raised German gasoline margins by 9% [assad2024].

With heterogeneous capacity, the paper derives asymmetric adoption results: AI firms collude among themselves while traditional firms compete, producing two-tier pricing, and AI firms gain a strategic incentive to increase market complexity beyond traditional firms' detection capacity. An approximate-collusion extension shows that a polynomial-time detector with accuracy $N \geq 2$3 suffices to sustain $N \geq 2$4-approximately collusive equilibria for $N \geq 2$5 near 1 — a practically significant concession, since modern AI systems are effective heuristic solvers even if exact CDP remains intractable. The competitive boundary is thus a gradient rather than a knife-edge.

Empirical predictions and policy implications

Five testable predictions follow: (P1) AI adoption correlates with higher markups and lower dispersion; (P2) conditional on AI adoption, more complex markets are more competitive — the novel claim that complexity protects competition; (P3) structural breaks in margins as capability thresholds are crossed; (P4) mandated transparency raises equilibrium prices; and (P5) collusion emerges without communication, shared algorithms, or third-party services. Prediction P5 challenges antitrust law's reliance on evidence of agreement or conspiracy, motivating the paper's proposal of "computational antitrust": treating market complexity as a design variable, increased through product differentiation incentives, demand opacity, asynchronous pricing, and algorithmic diversity mandates. The Efficiency–Competition Impossibility further implies a regulatory trilemma among efficiency, competition, and AI integration, of which at most two are jointly attainable.

Limitations and open questions

The paper is candid about several dependencies. The "only if" direction rests on the instance-hardness assumption, which fails for structured demand systems even if $N \geq 2$6; the conjecture that separability, low rank, and sparsity exhaust tractable cases is asserted without proof. Focal-point coordination through simple rules (e.g., matching an industry leader's price) can sustain partial collusion without solving CDP, particularly in simple markets; the theorem's predictive force is confined to sufficiently complex markets where focal rules fail. Rational profit maximization is assumed throughout, and small or enumerable state spaces permit brute-force detection, voiding the hardness results. Finally, the regime-shift thresholds $N \geq 2$7 and $N \geq 2$8 are characterized qualitatively but not computed for concrete market classes, leaving open how to operationalize the Market Complexity Index empirically.

Conclusion

The paper demonstrates that the sustainability of collusion in repeated oligopoly is equivalent to the tractability of three NP-hard problems — strategy computation, deviation detection, and optimal punishment — while competitive best-response dynamics remain polynomial-time solvable. Its principal contributions are the formal equivalence Competition $N \geq 2$9 (under generic instance hardness), the derived impossibility of simultaneous efficiency and competition, the transparency paradox, and the characterization of AI-driven regime shifts. Whether regulators can design institutions that preserve competition under expanding computational capability — and whether the instance-hardness condition can be verified for real markets rather than assumed generically — remain open questions raised by the analysis.