AI referee report
on “Second-order Minimax Risk in Multi-arm Binary Randomization”
The paper delivers a substantial verified fixed-K minimax-rate result and useful finite-dimensional reductions, but its novelty positioning, organization, and presentation of several certificate and attainment claims require major repair before the contribution is journal-ready.
Summary
The paper establishes an exact reduction from the labeled binary potential-outcome game to a response-type orbit game, identifies the first-order minimax constant, and proves that improvement over the first-order envelope has order n^{-4/3} for every fixed nonzero zero-sum contrast. It also constructs rational grid certificates, transfers them to real contrasts, and gives a persuasive finite three-arm example showing that a scalar signed score can discard decision-relevant information. These are meaningful theoretical contributions, but the manuscript obscures them through extensive duplication, weak quantitative engagement with the sharp two-arm benchmark, and several misleading or ill-typed descriptions of what is attained and certified.
Strengths
- The exact labeled-to-orbit equality is a clean and useful reduction that preserves the unrestricted design-estimator minimax value.
- The first-order constant C_0(c)=L_c^2/4 and contrast-weighted allocation have a clear interpretation and are accompanied by an explicit finite-sample procedure.
- The two-sided bounds establish the universal n^{-4/3} order under precisely stated fixed-K and fixed-contrast conditions.
- The rational linear-program certificates and continuity transfer provide a principled finite-dimensional route to lower and upper risk bounds.
- The n=3 three-arm separation is concrete, exactly evaluated, and appropriately distinguishes the full assignment-outcome experiment from the scalar-score experiment.
- The limitations section accurately acknowledges that the sharp second-order constant, convergence, optimizers, and useful-scale computation remain open.
Findings
The discussion of the closest competitor is too cursory to establish the paper's incremental contribution. The sentence “Sudijono, Dobriban, and Tchetgen Tchetgen obtains a sharp two-arm binary expansion” omits the competitor's scalar reduction, exact Airy constant, nonlinear attaining rule, and least-favorable-prior characterization. Consequently, readers cannot see that this paper supplies an internal matching-order converse and a fixed-K extension while leaving the sharp multi-arm constant and optimizer open.
Suggested fix
Add a quantitative comparison stating the competitor's expansion, constant, design-estimator characterization, and conditions with theorem or section locators. Then distinguish explicitly among what the present paper proves internally for all fixed K, what follows exactly for support-two contrasts, and what remains unresolved for support at least three. Add a similarly specific comparison to the bounded-outcome and multi-treatment finite-population literature identified in the related-work brief.
The manuscript repeatedly repackages the same results in the exact-reduction theorem, the multi-arm specialization theorem, the universal-rate theorem, the certificate frontier theorem, and the final attainment theorem. Several packages add automatic consequences such as compactness of the cluster set or repeat complete certificate-transfer statements. This makes the logical contribution difficult to identify and gives bookkeeping results the same prominence as the core econometric theorems.
Suggested fix
Reorganize the main text around four nonoverlapping results: exact orbit reduction, first- and second-order risk, finite certificate construction, and the three-arm diagnostic. State corollaries such as support-two transfer, compactness, and mesh asymptotics briefly, and move large aggregate theorem packages, verification-oriented definitions, and full proofs to a supplement.
The “Hull risk-value inclusion” clause is tautological: it proves membership by choosing the same design and estimator. The subsequent “strict treatment-domain separation” merely observes that a real-valued two-arm schedule set is not a subset of a binary K-arm schedule set with a different ambient type. Neither clause provides a risk comparison or substantive extension, yet both appear inside a main theorem alongside the actual orbit and rate results.
Suggested fix
Remove these clauses from the main theorem or demote them to a short scope remark. Position the Hull paper through its actual bounded two-arm minimax result and explain affirmatively that the present converse proceeds through the binary sign-pair embedding, without presenting set-theoretic noninclusion as a contribution.
The certificate-frontier statement is type-confused: “there exist sequences G_n, P_n, and L_n such that G_n is an exact rational grid-certificate cluster” treats the scalar normalized grid value G_n as the certificate cluster defined earlier as a collection of primal weights, dual multipliers, an estimator, and derived bounds. The proof repeats the same categorical mismatch.
Suggested fix
Define scalar sequences explicitly, for example as normalized improvements formed from the grid value, upper endpoint, and lower endpoint, and state separately that for each n these numbers are generated by an exact certificate cluster at mesh M=n. Then express asymptotic coincidence among the three scalar sequences.
The sentence claiming that support-at-least-three contrasts are cases “where the finite orbit geometry governs the certificate frontier” suggests a sharper support-specific asymptotic characterization than the theorems deliver. The verified results give a universal order sandwich and finite orbit certificates, but they do not characterize the sharp constant, subsequential-limit set, optimizer, or a distinct multi-arm limiting geometry.
Suggested fix
Replace this with the precise positive claim that support-at-least-three contrasts are evaluated through the full finite orbit program and inherit the universal n^{-4/3} bounds and certificate transfer. Keep the unresolved sharp asymptotic geometry in the labeled limitations and future-work subsection.
The prose says, “The lower side of the displayed envelope is attained by the explicit clipped-shrinkage procedure.” The procedure witnesses a guaranteed improvement with constant C_0(c)κ_c; it is not shown to attain the minimax second-order constant or an equality in the liminf bound. The theorem title “Attainment and certificates” creates the same ambiguity.
Suggested fix
Use “witnessed” or “achieved as a uniform lower bound on the minimax improvement,” and rename the aggregate result to “Rate-achieving shrinkage and finite certificates.” Reserve “attainment” for the finite orbit saddle point or an exact equality.
The paper repeatedly calls the certificate bounds computable and recommends them as a design-practice check, but it presents no solved grid-program instance, certificate artifact, runtime, or scaling evidence. The n=3 lookup-table witness is an exact diagnostic calculation, not a reported primal-dual solution of the proposed grid program.
Suggested fix
Provide at least one reproducible nontrivial LP instance with primal and dual outputs, certified gap, dimensions, runtime, and code or machine-readable certificate. Explain how the reported instance relates to the three-arm diagnostic and where computational growth becomes prohibitive.
The manuscript violates its affirmative-framing rule outside the limitations subsection. Examples include “what is deliberately outside the paper's formal claims” in the introduction and “it does not by itself isolate the identity of the individual negative-weight arms” in the three-arm discussion. The ACTG paragraph also places two noncoverage statements in the design-implications subsection before the labeled limitations subsection.
Suggested fix
Rewrite the introduction and diagnostic passage affirmatively, for example by saying that the exhibited rule aggregates the two negative-weight arms. Move the ACTG design and endpoint discrepancies into the labeled limitations subsection, while describing the binary-endpoint, prespecified-contrast application positively in the design-implications subsection.
The claim that a matched environment is a “verbatim rendering of a Lean declaration” is inaccurate terminology: the displayed mathematics is a reader-facing semantic rendering of structures and composite predicates, not literal Lean source. The appendix's own declaration list makes this distinction visible.
Suggested fix
Replace “verbatim rendering” with “faithful reader-facing rendering whose mathematical content matches the extracted declaration,” while retaining the statement- and proof-audit scope.
The first appearance of K in “fixed-K randomized experiments” lacks the requested plain-word gloss.
Suggested fix
Write “experiments with a fixed number K of treatment arms” at first use.
Questions for the authors
- Is the intended central novelty the universal multi-arm rate exponent, or do the authors intend the certificate program to support a future sharp multi-arm expansion? The title and theorem hierarchy should make this choice explicit.
- Can the authors provide an actual primal-dual certificate produced by the proposed grid program, including a machine-readable rational witness and computational cost?
- What substantive comparison to the bounded two-arm problem is intended beyond the sign-pair embedding, and why should the current tautological Hull clauses remain in a main theorem?
Review history
| Round | Date | Score | Findings |
|---|---|---|---|
| 1 | — | 6.2 | 10 |
| 2 | — | 6.1 | 11 |
| 3 · current | — | 6.1 | 10 |
Every referee draw is kept: scores are noisy, so the history is shown rather than only the last or the best pass.