AI referee report
on “Sharp Minimax Rates for Average Treatment Effects with Discrete Confounding under Fixed Overlap”
The paper delivers a significant and well-positioned sharp minimax-rate result with verified statements, but several exposition and claim-fidelity issues should be cleaned up before publication.
Summary
The submission identifies the fixed-interior minimax MSE rate for ATE estimation with unrestricted finite-alphabet confounding under fixed overlap and constructs a computable split-sample ratio-polynomial estimator attaining the upper rate. The contribution is meaningful: it closes the logarithmic upper-bound gap left by the closest discrete-covariate causal benchmark and gives clear parametric and consistency regimes. The formal statements are largely faithful to the verification contract, but the manuscript needs tightening around implementability, theorem quantifiers, verification-boundary prose, and reader-facing clarity.
Strengths
- Sharp fixed-interior minimax rate is an important contribution for discrete-covariate causal inference.
- The estimator is explicit and tied to a clear approximation/factorial-moment mechanism rather than only an existence argument.
- The paper engages the closest predecessor, including the exact logarithmic gap in ZengBalakrishnanHanKennedy2024.
- The endpoint randomized bound and combined upper envelope add useful context beyond the fixed-interior theorem.
- The prose mostly states the theorem conditions accurately, including fixed epsilon, the calibrated range, and the source-backed lower-bound dependency.
Findings
The statement and surrounding prose call the selector a deterministic estimator, but the later paragraph says it is oracle-calibrated because C_epsilon is an existence constant. This creates a claim-fidelity tension for implementability: "For the deterministic estimator ... one also has" can be read as an implementable procedure, while the paper later says the rule is not computable from the sample alone.
Suggested fix
In the theorem item title and displayed definition, rename it to an oracle-calibrated selected estimator or certificate selector, and state immediately before the display that only the two branch estimators are implementable from data while the selector packages two risk certificates using an existence constant.
The sentence "The theorem statements and the proofs written in this paper are machine-checked in Lean 4" overstates the verification boundary for synthesized presentation material and for cited external inputs unless read with later qualifications.
Suggested fix
Rewrite the opening of the verification claim affirmatively and locally, e.g. "The formal development machine-checks the theorem statements listed in the verification boundary and the Lean proofs of the in-paper reductions; cited published lower-bound inputs enter through theorem-local interfaces identified in the footnotes."
The paragraph "The extrema in the minimax risk" says "an estimator takes finitely many values and is a bounded function of the sample." A measurable real-valued estimator on a finite sample space is bounded, but it need not take values in a bounded a priori range common across estimators; the prose then discusses totalized order extrema. The paragraph is hard to reconcile with the later bookkeeping convention.
Suggested fix
Clarify that each fixed estimator is bounded because the sample space is finite, while the bound can depend on the estimator; then separate this ordinary finite-sample observation from the Lean convention for empty or unbounded indexed suprema.
Theorem 1 quantifies over "every single-observation law P, and every sample law mu_n in the experiment class" even though the displayed minimax and worst-case bounds do not depend on that particular P or mu_n. The later explanatory paragraph acknowledges this, but the theorem statement remains distracting and nonstandard.
Suggested fix
Restate the theorem without the unused law-level quantifiers in the reader-facing version, or move the inhabited-class explanation into a brief parenthetical after the theorem. If the Lean statement requires these witnesses, present that as a formalization note rather than as the statistical theorem's main quantifier structure.
The estimator definition sets M(n) and B(n) only "For n>=2" but then defines the estimator "For every n" and later claims computability for every n,d. The formal layer totalizes these boundary cases, but the manuscript recipe leaves n=0 and n=1 underspecified for readers.
Suggested fix
Add a short boundary convention for n<2, or state the estimator definition is intended for the positive large-sample regimes and give the totalized convention used for the all-n computability theorem.
The comparison to ZengBalakrishnanHanKennedy2024 says "the consistency benchmark moves from the d/n -> 0 of the standard estimator class to d/(n log n) -> 0." This is substantively right only when clearly restricted to the calibrated fixed-interior range and to minimax risk rather than the standard estimator class.
Suggested fix
Revise to: "Within the fixed-interior range d <= rho_epsilon n log n, the minimax consistency criterion is d/(n log n)->0, improving on the d/n->0 criterion for the standard plug-in/IPW/DR class analyzed there."
Several formal labels use synthesized names such as synth_3, synth_10, synth_13, synth_19, and synth_20. These are reader-visible through labels and cross-references if errors arise, and they make the manuscript look less polished.
Suggested fix
Rename presentation-synthesized environments to semantic labels such as def:chebyshev-factorial-notation, def:centered-endpoint-estimator, def:product-sampling-laws, def:chebyshev-reciprocal-coefficients, and def:sparse-polynomial-means.
The title "Sharp minimax rates for average treatment effects with discrete confounding under fixed overlap" uses sentence case correctly, but theorem title "Overlap adaptive hybrid envelope" would read more naturally as "Overlap-adaptive hybrid envelope."
Suggested fix
Hyphenate compound modifiers in titles and prose where needed, while preserving sentence case.
The phrase "the minimax mean-squared risk ... is [displayed rate] up to constants" is clear to specialists but slightly informal for an abstract of a theory paper.
Suggested fix
Use "is bounded above and below by epsilon-dependent constant multiples of" before the display.
Questions for the authors
- Can the existence constants C_epsilon and rho_epsilon be made at least numerically traceable from the proof, even if conservative, so the selector/envelope discussion is more operational?
- Would the authors consider moving the long verification-boundary discussion to a concise appendix table and keeping the main text focused on the statistical result and estimator?
Review history
| Round | Date | Score | Findings |
|---|---|---|---|
| 1 | — | 8.2 | 6 |
| 2 | — | 8.2 | 5 |
| 3 | — | 8.0 | 4 |
| 4 | — | 8.2 | 5 |
| 5 | — | 7.4 | 6 |
| 6 | — | 7.2 | 12 |
| 7 · current | — | 8.0 | 9 |
Every referee draw is kept: scores are noisy, so the history is shown rather than only the last or the best pass.