AI referee report
on “A Lower-Bound Calibration for Joint Margin–Overlap Decay in Offline Policy Learning”
The core lower-bound calibration appears correct and potentially useful, but the paper is still too narrow and too conditionally framed, with several claim-fidelity and presentation problems that prevent publication as written.
Summary
The paper proves an observed-law minimax lower bound for policy regret under a joint margin-overlap decay condition, and separately analyzes a clipped cross-fitted AIPW ERM under supplied nuisance and empirical-process side conditions. The lower-bound exponent is the main genuine contribution and is presented with commendable care about scope. However, the feasible upper result remains highly conditional, the comparison to the closest weak-overlap literature is not yet precise enough, and several prose and verification-scope claims still need correction.
Strengths
- The paper clearly separates observed-law welfare regret from causal identification and avoids claiming a new identification result.
- The lower-bound construction is transparent: active block mass, contrast size, and weak-arm probability are balanced cleanly into the displayed exponent.
- The manuscript is unusually explicit about the conditional nature of the clipped-AIPW upper bound and about the open feasible-tightness question.
- The formal statement layer, as summarized in the verification contract, largely supports the main mathematical claims.
Findings
The paper's delivered contribution is primarily a two-point observed-law lower-bound calibration plus a conditional upper bound for one supplied-nuisance procedure. As written, this is not yet a complete minimax rate characterization and may be too narrow for a leading econometrics journal unless the paper either proves more or more sharply explains why this lower-bound calibration itself changes practice or theory.
Suggested fix
Either add a substantive theorem closing an important part of the feasible/minimax gap, or substantially reframe the paper as a lower-bound note and add concrete corollaries/examples showing how the exponent changes conclusions relative to strict-overlap policy learning.
The comparison to the closest weak-overlap and clipping literature remains too high level. Table 1 says Liu2026 is closest to the feasible upper-bound side, but it does not compare target, overlap condition, clipping schedule, nuisance assumptions, or rate exponent closely enough to establish novelty.
Suggested fix
Add a focused subsection comparing the present law class, target, clipping choice, nuisance assumptions, and polynomial exponents against Liu2026, Zhao2023, and the closest limited-overlap papers; state exactly which assumptions/results are weaker, stronger, or incomparable.
The upper bound is still too easy to read as an achievability result. The sentence "Theorem \ref{obj:oeq:feasible-upper} is the promised conditional achievability statement" overstates what is delivered because the theorem assumes supplied nuisance estimators, boundedness, fixed folds, oracle-in-class, and primitive localized envelope and offset-envelope conditions.
Suggested fix
Rename this result throughout as a conditional risk bound or oracle-side-condition bound, not an achievability theorem, unless the paper proves that the side conditions are satisfied by an implementable nuisance-learning method over a meaningful primitive class.
The sentence "The theorem statements and proofs described above are machine-checked in Lean 4" is too broad relative to the verification contract: the Le Cam chi-square testing lemma is listed with no Lean declaration attached.
Suggested fix
Either add the missing machine-checked declaration for the Le Cam testing lemma or qualify the verification note to list precisely which statements are machine-checked and which testing step is used as an external mathematical lemma.
The title "Bounded potential outcomes" and the sentence "Potential-outcome notation Y(a) is used for the boundedness assumption below" do not match the verified assumption, which bounds the observed outcome and observed-law treatment-arm conditional means, not potential outcomes.
Suggested fix
Retitle the assumption as "Bounded observed outcome and arm regressions" and remove the potential-outcome wording unless a separate potential-outcome assumption is actually introduced and used.
The localized VC and offset-envelope assumptions are still named and introduced in a way that may suggest they are standard consequences of finite VC dimension. The formal conditions are primitive high-level bounds for broad policy-compatible increment classes and are not derived from finite VC alone.
Suggested fix
Rename them as "Primitive localized empirical-process envelope" and "Primitive localized offset envelope," and add one sentence after each assumption saying finite VC motivates but does not imply the displayed condition without additional envelope, measurability, and moment arguments.
The notation E_P[R_P(hat pi_n)] is imprecise when hat pi_n is sample-dependent. The formal statements integrate over the product experiment P^{otimes n}, and the theorem statement for the lower bound later reverts to E_P notation.
Suggested fix
Use E_{P^{otimes n}} consistently in the minimax definition, upper-risk definition, theorem statements, and explanatory prose whenever the estimator is sample-dependent.
The notation imposes unnecessary cognitive load by alternating among r_up, r_feas, g_joint, r_star, A_alpha, beta, and D, while r_up and r_feas are identical by definition.
Suggested fix
Use a single symbol for the feasible upper exponent, keep r_star for the lower-bound exponent, and move secondary symbols such as D and beta into a short notation table or a single definition block.
The proof appendices read like a Lean transcript rather than an econometrics article, with phrases such as "Lean derives," inline proof tags, and internal bookkeeping comments.
Suggested fix
Move Lean trace comments and proof-tag markers to a separate verification supplement, and rewrite the appendix proofs in conventional mathematical prose emphasizing the statistical argument rather than proof-assistant bookkeeping.
Several cross-references use theorem or assumption labels as section labels, for example "Section \ref{obj:thm:minimax-lower}" and "Section \ref{obj:ass:iid}."
Suggested fix
Create section labels for the setup, lower-bound, upper-bound, and appendix sections, and reserve theorem/assumption labels for references to those formal environments.
Questions for the authors
- Can the localized envelope and offset-envelope conditions be verified for a concrete class of nuisance estimators and policy classes, even in one representative nonparametric example?
- How exactly do the rate exponents and clipping schedules compare to Liu2026 under the closest common assumptions?
- Is the intended contribution mainly the lower-bound exponent, or do the authors want readers to view the clipped-AIPW result as an implementable procedure? The manuscript should make this choice explicit.