An explicit solution of the five-expert prediction PDE and the exact optimality set of COMB
math.AP · cs.GT · cs.LG · math.OC
Quick verdict
Explicit five-expert prediction PDE solution disproves COMB optimality; exact Bernstein certificates and Lean 4 formalization give rigorous verification.
Summary
The paper derives, for the first time, a closed-form solution of the stationary prediction-with-expert-advice PDE with five experts under geometric stopping. Working in the ordered sector, the authors split the domain into three regions: in Regions I and II the solution is the four-expert formula of Bayraktar–Ekren–Zhang plus a single Laplace-type integral with an explicit positive density; in Region III it is a finite combination of hyperbolic products whose four coefficients come from one scalar quadrature defining the trace function e. The formula yields the game value u(0)=45π²/(512√2) and a second-order expansion at the origin. It proves the rank direction (1,0,1,0,0) is optimal throughout the sector, while the COMB strategy (1,0,1,0,1) attains the Hamiltonian maximum only where x1=x2 and x3=x4 — a lower-dimensional set — sharply disproving the Gravin–Peres–Sivan conjecture. Verification of the sixteen Hamiltonian inequalities is reduced to 21 scalar inequalities certified by 147 exact rational Bernstein polynomials, rechecked by independent fraction-arithmetic checkers in a pinned, offline Zenodo supplement; a Lean 4 formalization on Mathlib machine-checks the reductions, regularity, and both theorems, excluding only the viscosity characterization. Main limitations: the unformalized uniqueness step, and a presentation whose key checks live in the supplement rather than the paper.
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 1
● low 2
4 strengths · 4 weaknesses · 4 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The stationary five-expert prediction PDE has the explicit three-region solution of Theorem 2.1, and it is the unique viscosity solution with u−maxᵢxᵢ bounded.
- The game value at the origin is u(0)=45π²/(512√2), with the second-order expansion (2.18) at the origin.
- The rank direction v*=(1,0,1,0,0) attains the Hamiltonian maximum throughout the ordered sector, while COMB (1,0,1,0,1) is optimal exactly where x1=x2 and x3=x4, disproving the Gravin–Peres–Sivan conjecture for n=5.
- The Hamiltonian verification reduces to 21 scalar inequalities, each proven by exact rational Bernstein polynomial certificates (147 total) with independent arithmetic rechecks.
- A Lean 4 formalization machine-checks the reductions, regularity, and both main theorems, excluding only the viscosity characterization.
Strengths
Definitive resolution of a decade-old conjecture
Theorem 2.2 gives not merely non-optimality but the exact optimality set of COMB — equality in the curvature gap Δ (2.19) holds iff x1=x2 and x3=x4 — a sharp form of disproof of the Gravin–Peres–Sivan conjecture predicted numerically by Chase [16] and [13].
Verification with no numerical slack
All 21 scalar inequalities are proven by 147 exact rational Bernstein polynomial certificates (56 for Region III, 91 for Regions I/II, largest degree 79), over seven dyadic intervals using a degree-12 arctanh bound (4.35); independent checkers recompute with stdlib fractions only, and no floating point enters anywhere.
Machine-checked formalization
Conceptual payoff beyond the result
Weaknesses
One unformalized logical link
The paper alone is not checkable
Cross-subsection dependency for boundedness
Presentation density
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper settles the five-expert case of prediction with expert advice under geometric stopping, whose continuum limit is the fully nonlinear elliptic equation (2.1): u − ½ max_v vᵀ∇²u v = maxᵢ xᵢ. Three claims are made. First, Theorem 2.1 gives an explicit formula for the unique viscosity solution with bounded correction, in three regions of the ordered sector x1≥…≥x5: Regions I–II as the four-expert solution F4 (2.14) of Bayraktar–Ekren–Zhang plus a single integral against the explicit positive density p (2.11)–(2.12), and Region III as a finite hyperbolic-product sum (2.16) whose coefficients b0–b3 (2.17) derive from the quadrature-defined trace e (2.8). It also gives the value u(0)=45π²/(512√2) (Lemma 3.1) and the expansion (2.18). Second, Theorem 2.2 pins the exact COMB optimality set: Δ(x)=0 iff x1=x2 and x3=x4 (2.20). Third, a verification architecture: 16 inequivalent binary controls reduced to 21 scalar inequalities, each certified by exact rational Bernstein polynomials.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 4 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the COMB strategy in prediction with expert advice?
After sorting the experts by current regret, the adversary advances all odd-ranked experts with probability 1/2 and all even-ranked experts with probability 1/2. Gravin, Peres and Sivan proved it optimal for three experts and conjectured asymptotic optimality for any number of experts; this paper shows the conjecture fails at five experts.
What exactly does this paper disprove?
It disproves the conjecture that COMB is optimal for the five-expert geometric-stopping game: COMB attains the optimality condition (zero curvature gap against the best direction) only on the lower-dimensional set where x1=x2 and x3=x4, not throughout the domain, while the direction (1,0,1,0,0) is optimal everywhere.
Is the proof computer-assisted, and can I verify it myself?
Yes, partially. Twenty-one scalar inequalities are certified by 147 exact rational Bernstein polynomials checked in exact integer/rational arithmetic — the archived supplement (five-expert-proof-1.0.6 on Zenodo, pinned dependencies, offline, a few minutes on CPU) reruns all checks. A Lean 4 formalization additionally machine-checks the reductions and both main theorems, except the viscosity-solution characterization.
What is the value of the five-expert prediction game?
The value at the origin is u(0)=45π²/(512√2), the leading-order minimax regret constant under geometric stopping. For comparison, the values are √2/4 for two experts, √2/3 for three, and π/(4√2) for four experts.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extraction marks two omitted sections ([...section omitted...]); some details of Section 4.3's region reduction and the Region III vertex table context were reconstructed from surrounding text, but all cited numbers appear verbatim in the extracted text.
- The Zenodo certificates and the Lean build were not executed in this review; reproducibility assessment is based on the described checksummed, pinned release procedure.
- The companion preprint [4] and reference [7] were not examined; claims about them rest on this paper's description.
- No figures exist in the extracted text; the paper appears to be purely formal mathematics with no figures required.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.