AI referee report
on “Outcome Annotation and Minimax Risk for Discrete Average Treatment Effects”
The paper delivers a significant sharp minimax result with faithful formal statements, but its publication case is weakened by an unstated load-bearing lower-bound input, proof-led organization, and several misleading presentation choices.
Summary
The paper characterizes the finite-sample minimax risk for estimating a discrete-covariate ATE from n outcome-labeled observations and m additional treatment-covariate observations, obtaining a sharp rate with separate outcome and marginal-information components. The verified main statements support the central contribution, including the supervised endpoint, known-marginal limit, and sequence-regime characterizations. The result is potentially important for econometric theory, but the manuscript requires substantial restructuring and a self-contained presentation of its key lower-bound analytic input before publication.
Strengths
- The central minimax frontier is sharp up to constants depending only on the overlap parameter and cleanly separates labeled-outcome information from pooled treatment-covariate information.
- The common-marginal lower-bound construction is well aligned with the two-sample experiment and addresses the auxiliary-data design rather than merely importing a supervised converse.
- The paper supplies an explicit upper-bound witness, a simpler inverse-count benchmark, a known-marginal comparison, and exact sequence criteria for consistency, parametric risk, and strict rate improvement.
- The observed-data model, causal interpretation, overlap condition, sampling experiment, target, and minimax risk are defined carefully.
- The related-work section identifies the closest discrete minimax benchmark and accurately explains the logarithmic gap addressed by the new polynomial construction.
Findings
The paragraph titled “One-coordinate marked-Poisson input” identifies a “load-bearing analytic estimate,” but the manuscript supplies only an informal likelihood-expansion intuition rather than a named statement with exact hypotheses and a proof. The proof of the uniform-intensity construction invokes this estimate to obtain the geometric total-variation decay, so the reader-facing lower-bound argument is not self-contained even though the final theorem is verified.
Suggested fix
Promote the existing verified one-coordinate estimate to a numbered lemma, state its normalization, support, moment-matching, bandwidth, marking, and total-variation conventions exactly, and include a readable proof or a precise cross-reference to a complete supplementary proof. Then cite that lemma explicitly in the uniform-intensity proof.
The manuscript is organized too closely to a generated proof ledger for a journal article. Examples include the full dependency inventory inside the “Binary length” definition, repeated calibration bookkeeping, and many long proofs whose main statistical ideas are obscured by measure-theoretic and constant-tracking details. This makes the contribution difficult to evaluate despite the strong main theorem.
Suggested fix
Rebuild the main paper around the model, frontier theorem, estimator intuition, lower-bound construction, regimes, and closest comparisons. Move the dependency inventory, explicit numerical calibration, Lean-rendered proof details, and low-level transport lemmas to a clearly indexed technical supplement; retain concise proof arguments in the paper that expose the statistical logic and identify every supplementary input.
The theorem calls the estimator “total computable,” followed by the explanation “total computable means that the displayed finite sums define an everywhere-defined statistic.” Everywhere-defined and computable are different properties, and the formal artifact presented here establishes an explicit measurable statistic rather than an executable computability theorem.
Suggested fix
Replace “total computable” throughout with “explicit, everywhere-defined” or “fully specified measurable,” unless the authors add and document a genuine computability result with a representation and evaluation model.
The sentence “The upper bound is attained by a fully specified hybrid estimator” is formally defensible, but it invites an operational reading that sits uneasily with the later disclosure that the displayed calibration can require n on the order of 1.6×10^17 even at degree two and otherwise returns zero. The estimator is best understood as an explicit proof-calibrated minimax witness.
Suggested fix
Describe it affirmatively as an “explicit proof-calibrated hybrid witness attaining the minimax order up to overlap-dependent constants,” and distinguish the statistical frontier from implementable tuning when the estimator is first introduced in the main text.
The dedicated comparison does not engage sufficiently with Cheng, Ananthakrishnan, and Cai, although their chart-review design is the closest empirical motivation identified in the surrounding literature. Their surrogate-rich, doubly robust, locally efficient setting and this paper's baseline-only unrestricted finite-cell minimax benchmark answer related but distinct questions.
Suggested fix
Add Cheng, Ananthakrishnan, and Cai to the closest-comparisons table and give a focused paragraph comparing observed auxiliary variables, nuisance structure, inferential target, and guarantee. State positively that the present result supplies the baseline-only discrete minimax benchmark, without attributing surrogate-efficiency gains or model fit to the present theorem.
The label “Unannotated endpoint” for m=0 is liable to reverse the experiment's meaning: this endpoint still contains n outcome-annotated records and has no auxiliary outcome-unlabeled records. The manuscript later has to explain the generated phrase.
Suggested fix
Rename the bullet and all surrounding references to “Labeled-only endpoint” or “No-auxiliary endpoint.”
The sentence “its conservative calibrated witness is not an operational allocation rule” appears outside the explicitly titled limitations paragraph and violates the affirmative contribution-framing contract.
Suggested fix
Rewrite the sentence affirmatively as “The theorem provides a theoretical sample-size frontier, while the calibrated witness serves as its explicit proof construction,” and keep discussion of practical tuning in “Limitations and future work.”
The verification note gives filenames, a commit hash, and tool versions but no stable repository or archival location and no concise reproduction command. The formal verification is central to the submission's rigor claim, so readers need an independently retrievable artifact.
Suggested fix
Provide a persistent repository or archive identifier, bundle the crosswalk and audit files with the supplement, and state the exact command and expected result for checking the mapped declarations at the cited commit.
The displayed “Inverse-count arm mean” formula begins with q_ax/(u/u), which is algebraically just q_ax under u>0 and obscures the statistical quantity being averaged.
Suggested fix
Display either q_ax E[1+K_{1-a,x}/(K_{ax}+1)] or the equivalent expectation involving S_ax/u, and remove the artificial u/u factor.
Questions for the authors
- Is the one-coordinate marked-Poisson total-variation estimate already available as a separate verified declaration that can be promoted into the manuscript, or is it currently embedded only inside the uniform-intensity construction?
- Will the Lean sources, crosswalk, audit cache, macros, and build instructions be deposited in a persistent public archive tied to the cited commit?
- Do the authors intend “computable” in an executable formal sense, or only to mean that the estimator is explicitly and everywhere defined?
Review history
| Round | Date | Score | Findings |
|---|---|---|---|
| 1 | — | 6.3 | 11 |
| 2 | — | 6.4 | 10 |
| 3 · current | — | 7.2 | 9 |
Every referee draw is kept: scores are noisy, so the history is shown rather than only the last or the best pass.