Pith. sign in

REVIEW 2 major objections 5 minor 7 references

Equilibria in Large Position-Optimization Games

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Once the number of players is large, every Nash equilibrium in a position-optimization game concentrates on a finite set of pseudo-targets, and the distribution of play converges to the projected target distribution at rate O(1/n).

desk verdict Solid pure-equilibrium and convergence-rate results, but the symmetric mixed equilibrium existence theorem is false as stated — that half of the paper needs a real fix. read the letter →

arxiv 2602.15225 v2 pith:A7KXXXCL submitted 2026-02-16 cs.GT

Rafael Frongillo , Melody Hsu , Mary Monroe , Anish Thilagar This is my paper · ORCID
classification cs.GT MSC 91A1091A0691B72
keywords position-optimizationgamesNashequilibriumpseudo-targetsHotellingforecastingcompetitionsconvergenceratessymmetricmixedwisdomofcrowds
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper establishes that in any position-optimization game—a symmetric game in which n players pick positions to capture the mass of targets to which they are closest—once n is large enough, every Nash equilibrium (pure or symmetric mixed) is extreme: all players play on a finite set of pseudo-targets, the positions that are optimal for some target. Furthermore, the equilibrium distribution over pseudo-targets converges to the projected target distribution P at rate O(1/n). This gives the first convergence rates for Hotelling games and forecasting competitions, and the first general equilibrium characterization for forecasting competitions with finite n. The unifying insight is that large competitions produce a wisdom-of-crowds effect: individually strategic play collectively mimics the underlying distribution.

What carries the argument

The central object is the pseudo-target set X* = {x : x minimizes d(x,y) over X for some y}, together with the projected distribution P(x) = Q({y : x*(y) = x}). For pure equilibria, the argument rests on the observation that a strategy profile covering X* makes all profitable deviations lie inside X*, and on a greedy algorithm (Algorithm 1) that builds an equilibrium by initially placing about 2P(x)/p0 agents on each x and then iteratively adding agents to the position maximizing P(x)/(k+1). For mixed equilibria, the key identity is the utility decomposition into events E1, E2, E3; defining G(σ_x) = (σ_x − σ_x^n)/(1 − σ_x^n − (1−σ_x)^n), equilibrium forces σ_x ≈ G^{-1}(p_x), and a coupon-col

What would settle it

Take a game satisfying Condition 1 with n above the mixed-equilibrium threshold, compute a symmetric mixed equilibrium, and check whether any pseudo-target x has |σ_x − p_x| > 1/n; a single violation would refute Theorem 5. More decisively, construct a game with an infinite pseudo-target set (a continuous consumer distribution in Hotelling) and test whether some equilibrium's play distribution fails to converge to P at rate O(1/n).

Watch

Extended reading notes

Core claim

The paper's central claim is that, under Condition 1 (no ties with positive probability and a finite pseudo-target set X^*), the large-n behavior of the game G_n is fully tractable. For n >= 2/p0, a pure Nash equilibrium exists, is extreme, and has at least two agents on every pseudo-target; for n > 2/p0 this holds for every pure equilibrium. For symmetric mixed equilibria, once n > max{43, 8(4/p0)log(1/p0)}, every such equilibrium is extreme and its per-position probabilities satisfy |σ_x − p_x| ≤ 1/n. In both cases the play distribution converges to P, the projection of the target distribution Q onto X*, with an explicit O(1/n) rate; the paper also proves the pure-existence threshold is ti

Load-bearing premise

The whole proof depends on the set of pseudo-targets being finite and on ties in the closest-position mapping having zero probability; without finiteness, the covering and coupon-collector arguments collapse.

Editorial extensions

If this is right

  • In forecasting competitions with enough forecasters, equilibrium reports are extreme (deterministic outcome predictions) and the empirical distribution of reports converges to the outcome distribution Q, so collecting many reports recovers Q even though each forecaster is strategic.
  • In Hotelling games with finite or discrete consumer distributions, retailers locate on first-preference locations and the empirical distribution of locations approaches the consumer distribution with an explicit O(1/n) KL-divergence bound, extending prior results to non-differentiable consumer distributions.
  • The symmetric mixed equilibrium converges to P componentwise within 1/n, meaning aggregate random play is nearly exact even though each player randomizes.
  • The pure-equilibrium existence threshold n ≥ 2/p0 is tight; below it equilibria may fail, exactly identifying when coordination guarantees a stable outcome.
  • The convergence-rate bound applies to every equilibrium (not just the constructed one), so the wisdom-of-crowds guarantee is a property of all large equilibria.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors conjecture in Section 6 that convergence holds without the finiteness condition; if true, the rate likely depends on the geometry of X*, and a natural test is whether continuous Hotelling games also exhibit O(1/n) convergence in total variation.
  • The mixed-equilibrium result implies a falsifiable prediction for real forecasting contests: winners' submissions should be less calibrated (more extreme) than the average forecaster, and the gap should shrink like 1/n; this could be tested on archived Kaggle-style data.
  • The wide gap between the pure threshold (2/p0) and the mixed threshold (exponential in 1/p0) suggests that randomizing players need far larger populations to concentrate than deterministic ones; contest designers might exploit this by making participation sizes public.
  • The componentwise bound |σ_x − p_x| ≤ 1/n gives a simple model-checking test: repeated plays of a symmetric mixed equilibrium should produce empirical frequencies within ~1/n of P for every pseudo-target; systematic deviations indicate misspecification or non-equilibrium behavior.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces position-optimization games, in which n symmetric agents choose positions to maximize the expected mass of targets for which they are closest, with ties split evenly. Under Condition 1 (a.s. unique pseudo-target and finite pseudo-target set X*), it proves: for n ≥ 2/p0, an extreme pure Nash equilibrium exists and can be constructed in O(|X*|n) time; for n > 1/p0 all pure equilibria are extreme and cover X*; and the empirical equilibrium distribution converges to the projected target distribution P in KL divergence with the explicit bound log((⌊cn⌋+1)/⌊cn⌋). It further claims existence for all n > 1 of symmetric mixed equilibria (Theorem 4 via the general Theorem 6) and, for large n, extremeness and |σ_x − p_x| ≤ 1/n (Theorem 5). Applications to forecasting competitions, finite and classic Hotelling games, spatial voting, and discrete Voronoi games are then drawn.

Significance. If the theorems were correct, the paper would be a valuable unifying treatment: it gives finite-n existence, structural predictions (all equilibria extreme), and the first O(1/n) convergence rates for a class that includes forecasting competitions and Hotelling variants. The pure-strategy side is constructive, with an explicit algorithm and a tight threshold; these are concrete contributions. However, the mixed-strategy existence claim rests on a false general theorem and an invalid Reny-security proof, so the mixed-strategy half of the advertised contribution is not established. The pseudo-target framework and the explicit rates are genuine strengths, but the mixed-equilibrium results need new proofs before they can be accepted.

major comments (2)
  1. [Appendix C.1, Theorem 6] Theorem 6 is false as stated, and its proof does not verify Reny payoff security. The proof only checks symmetric diagonal profiles: if every opponent plays σ′, the deviator plays σ′ and secures c/n. Payoff security requires, for every profile and ε > 0, a single deviation that remains ε-optimal against every strategy tuple in a neighborhood of the given opponents' tuple; the diagonal is not a neighborhood in the product topology. The statement is not merely unproved: on X=[0,1], define u(a,b)=1 if (b=1 and a<1) or (a>b and a,b<1); u(a,b)=−1 if (a=1 and b<1) or (b>a and a,b<1); u(a,a)=0. This is a compact symmetric constant-sum game with no symmetric mixed equilibrium. Hence Theorem 4 and the mixed-existence claims in the abstract and Section 5 are unsupported.
  2. [Appendix C.2, Lemma 12 / Proposition 4] The proof of the upper bound in Theorem 5 is not complete. Lemma 12 claims monotonicity of the function Gbar by arguing that N′>0 and 0<D′<1 imply N/D is increasing; this is not a valid criterion for a quotient. Proposition 4 then applies Gbar^{-1} outside the domain on which monotonicity was stated and contains mismatched references ('Lemma 3' where Observation 3 is meant). Since the conclusion σ_x ≤ p_x + 1/n depends on this step, the upper-bound half of Theorem 5 is not established as written.
minor comments (5)
  1. [Theorem 3 proof, after Eq. (4)] The displayed bound after Eq. (4) should be |log(k/(nP))| ≤ log((⌊cn⌋+1)/⌊cn⌋), not the ratio (⌊cn⌋+1)/⌊cn⌋; as printed, the subsequent KL inequality does not follow. This appears to be a local typo.
  2. [Lemma 4] The hypercube vertices x1 and x3 are both labeled {1,0}; presumably one is {0,1}. In the same proof, the text refers to ui(x1,x_−i) when discussing an agent on x2; the notation is confusing and should be cleaned up.
  3. [Appendix B.4, Lemma 5 proof] The inequality 'P(x)>2/c≥2/n' should presumably be 'P(x)>2c≥2/n'; as written the claimed chain is false. The following display has a similar typo, using c/(⌊cn⌋) instead of the intended 2c/(⌊cn⌋) argument.
  4. [Lemma 11 proof] The line 'we have |X*|≤p0' should be |X*|≤1/p0; the subsequent bound uses the corrected inequality. This is a typographical issue, but it makes the displayed derivation hard to follow.
  5. [Abstract] The abstract contains the typo 'psuedo-targets'; the term should be 'pseudo-targets'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: extremeness and O(1/n) convergence are proved from the game definition; self-citations are contextual only.

full rationale

The central derivation is self-contained from the model in Section 2. The pseudo-target set X* and projection P are defined from Q and the proximity minimizer x*(y), but the claims that equilibria concentrate on X* and that the empirical or mixed distribution approaches P are proved through utility bounds and concentration arguments (Lemmas 1, 2, 5, 6, 7, 11, 13; Theorems 2, 3, 5), not assumed by construction. The O(1/n) rate in Theorem 3 follows from the equilibrium utility interval in Lemma 6 and the KL expansion; Theorem 5's |sigma_x - p_x| <= 1/n is derived from the inverse-relation G and a coupon-collector bound, not from a fitted parameter. Algorithm 1 uses P to construct an equilibrium, but the convergence claims apply to equilibria generally and do not depend on the construction being identified with the prediction. Self-citations (Frongillo et al. 2021; Monroe et al. 2025) appear only in related-work and motivation discussion and are not used as black boxes for the main theorems. The proof of symmetric mixed-equilibrium existence via Theorem 6 and Reny payoff security is a correctness concern—indeed the stated theorem is false under a compact-Hausdorff antisymmetric counterexample—but that is a soundness gap, not circularity: no result is assumed equivalent to its conclusion. Section 6 openly restricts results to Condition 1 and conjectures broader convergence, and that stated limitation likewise does not indicate circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted; the constants 43 and 8 are universal. 'Pseudo-targets' are derived from the model, not newly postulated entities. The main assumptions are Condition 1, implicit argmin existence, and the background theorem of Reny.

assumptions (3)
  • domain assumption Condition 1: Q({y:|x*(y)|>1}) = 0 and |X*| < ∞
    Assumed in Definition 1. Controls tie-breaking and gives a finite pseudo-target set; load-bearing in all theorems. The paper admits (§6) that relaxing finiteness is challenging and only conjectures the results extend.
  • domain assumption For Q-almost every y, the argmin x*(y) exists
    Implicit in the definition of pseudo-targets and Condition 1. In general arbitrary X (no compactness assumed in the main model) the minimizer may not exist; the paper does not discuss this topological assumption.
  • standard math Reny (1999) Corollary 5.3 as the existence theorem for discontinuous symmetric games
    Used in Theorem 6 to claim symmetric mixed equilibrium existence. The paper's verification of payoff security is incomplete (only diagonal deviations are treated), so reliance on this background theorem is part of the correctness risk.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Equilibria in Large Position-Optimization Games." pith.science (2026). https://pith.science/paper/A7KXXXCL

@misc{pith2026260215225,
  author       = {Pith},
  title        = {Pith review of: Equilibria in Large Position-Optimization Games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A7KXXXCL}},
  note         = {Machine review of arXiv:2602.15225}
}
abstract

We propose a general class of symmetric games called position-optimization games. Given a probability distribution $Q$ over a set of targets $\mathcal{Y}$, the $n$ players each choose a position in a space $\mathcal{X}$. A player's utility is the $Q$-mass of targets they are closest to under some proximity measure, with ties broken evenly. Our model captures Hotelling games and forecasting competitions, among other applications. We show that for sufficiently large $n$, both pure and symmetric mixed Nash equilibria exist, and moreover are extreme: all players play on a finite set of pseudo-targets $\mathcal{X}^* \subseteq \mathcal{X}$. We further show that both pure and symmetric mixed equilibria converge to the distribution $P$ on $\mathcal{X}^*$ induced by $Q$, and bound the convergence rate in $n$. The generality of our model allows us to extend and strengthen previous work in Hotelling games, and prove entirely new results in forecasting competitions and other applications.

Figures

Figures reproduced from arXiv: 2602.15225 by the authors.

Figure 1
Figure 1. An example of spaces X and Y = {y1, y2, y3, y4}, here both subsets of some Euclidean space with proximity function d. Note multiple targets (y1, y2) can map to the same pseudo-target, and targets in X are themselves pseudo-targets (see y3). These games and their many variations have been studied extensively, but mainly under limited conditions—for specific choices of X , Y, and Q, or under strong structural constrai… view at source ↗
Figure 3
Figure 3. A forecasting competition with m = 3 binary events. Targets correspond to vertices of the cube {0, 1} 3 . Forecasters submit predictions (red) in the filled cube [0, 1]3 . The outcome (blue) is drawn from distribution Q over ver￾tices, with the winner minimizing ℓ2 distance. X . Their results show that pure equilibria exist, and the empirical distribution in pure equilibria converges to the true underlying mass of f… view at source ↗
Figure 4
Figure 4. All best-response dynamics for each strategy profile [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: explicitly illustrates the extremizing phenomenon that is observed in real-life forecasting competitions: for any px ̸= 0.5, σx places more weight than px on the outcome of x that is more likely under P. The striking emergence of non-truthful behavior even under our st…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references

  1. [1]

    (1,1, n−2) : agentionx 1 has a profitable deviation tox 2. Specifically,u i(xi,x −i) = 2−2ϵ n , while deviating tox 2 leads to utility 2−ϵ 2n + 2−2ϵ 2n = 4−3ϵ 2n > 2−2ϵ n .(That is, deviating tox 2 allowsito share the utility of bothx 1 andx 2.)

  2. [2]

    Specifically,u i(xi,x −i) = 2−ϵ 2n , and deviating tox 3 leads to utility n−4+3ϵ n(n−2) > 2−ϵ 2n forϵ > 4 n+4

    (1,2, n−3) : an agentionx 2 has a profitable deviation tox 3. Specifically,u i(xi,x −i) = 2−ϵ 2n , and deviating tox 3 leads to utility n−4+3ϵ n(n−2) > 2−ϵ 2n forϵ > 4 n+4

  3. [3]

    Specifically,u i(xi,x −i) = 2−2ϵ 2n , while deviating tox 2 leads to utility 2−ϵ 2n > 2−2ϵ 2n

    (2,1, n−3) : agentionx 1 has a profitable deviation tox 2. Specifically,u i(xi,x −i) = 2−2ϵ 2n , while deviating tox 2 leads to utility 2−ϵ 2n > 2−2ϵ 2n

  4. [4]

    Specifically,u i(xi,x −i) = 1−ϵ n , and deviating tox 3 leads to utility n−4+3ϵ n(n−3) > 1−ϵ n forϵ > 1 n

    (2,2, n−4) : an agentionx 1 has a profitable deviation tox 3. Specifically,u i(xi,x −i) = 1−ϵ n , and deviating tox 3 leads to utility n−4+3ϵ n(n−3) > 1−ϵ n forϵ > 1 n . 20 B.3 Proof of Theorem 2 Proof.We aim to prove that the strategy profile returned by Algorithm 1 is an equilibrium. Note that for anyn≥ 2 p0 , the value ofn 0 in Algorithm 1 satisfies n0...

    1999

  5. [5]

    Then1 x∈Xmin(x,y) = 1, and whenkother agents are onx, sinceQ({y: x∗(y)|>1}) = 0,|X min(x, y)|=k+ 1

    Consider eventE 1. Then1 x∈Xmin(x,y) = 1, and whenkother agents are onx, sinceQ({y: x∗(y)|>1}) = 0,|X min(x, y)|=k+ 1. LetB(n;k, p) denote the binomial probability mass function. It follows that Pr[E1] E y∼Q x−i∼σ−i 1 x∈Xmin(x,y) |Xmin(x, y)| E1 =P(x) n−2X k=0 B(k;n−1, σ x) k+ 1 ,(13) where we use that Pr[x ∗(Y) =x] =P(x) and E[ 1 |Xmin(x,y)| |E 1] = Pn−2...

  6. [6]

    For the second event, we have Pr[E2] E [πi(xi,x −i, y))|E 2] = σn−1 x n ,(14) because the probabilityn−1 agents playxisσ n x , and the utility of agenticonditioned on everyone else playingxis 1 n

  7. [7]

    ideal position

    We simply lower bound the last term: Pr[E3] E[πi(xi,x −i, y)|E 3]≥0.(15) We now simplify Equation 13 to explicitly construct the expression for a lower bound on ui(x,σ −i), which excludes the third term. We have px n−2X k=0 B(k;n−1, σ x) k =p x n−2X k=0 n−1 k σk x(1−σ x)n−1−k k+ 1 ! = px n n−2X k=0 n k+ 1 σk x(1−σ x)n−(k+1) (since 1 k+1 n−1 k = 1 n n k+1 ...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.