AI referee report
on “Pointwise Causal Estimation with Weak Dose Support and Gaussian Measurement Error”
The verified uniform frontier is a substantial theoretical contribution, while publication requires a more focused exposition, numerical assessment of the explicit procedure, and sharper attribution of methodological antecedents.
Summary
The paper establishes matching minimax absolute-error and honest connected-interval-length orders for pointwise causal evaluation under globally calibrated polynomial density envelopes and known Gaussian measurement error. Its strongest contribution is the uniform connection between direct estimation, Fourier inversion, and compact-support polynomial inversion. The prose generally preserves the verified parameter ranges, coverage quantifiers, and external-source boundaries. A substantial revision should improve accessibility and establish the numerical behavior of the advertised observable construction.
Strengths
- The same-class upper and observed-experiment lower bounds give a coherent three-regime characterization, including transition neighborhoods and fixed positive noise.
- The manuscript carefully distinguishes finite-sample coverage for all noncoverage levels in (0,1) from the matched length frontier for levels below one half.
- Identification accommodates schedule-level exchangeability and vanishing density at the target dose through explicit measurable structural primitives.
- Related work engages several close competitors through their actual rates, losses, experiments, and operator conditions.
Findings
The numerical assessment of the explicit construction is assigned to future work, although the manuscript itself establishes that its 95% interval equals [0,1] for every sample size through 4800. This makes the distinction between asymptotic rate attainment and finite-sample informativeness central to evaluating the procedure. Exact score minimization also involves Fourier quadrature and potentially unstable high-degree polynomial inverses.
Suggested fix
Provide a reproducible implementation and deterministic certificate diagnostics over representative calibrated K, beta, kappa, sample sizes, and noise scales. Report the selected family and index, score components, radius, and sample sizes at which the score permits informative intervals. Document numerical precision and quadrature controls, distinguishing numerical approximations from the exact coverage guarantee.
The exposition repeatedly restates the calibration inequality, parameter ranges, frontier interpretation, and upper/lower-bound roadmap. Setup presents elementary restrictions and witness-specific restrictions as a long sequence of separate assumptions; the main observed-law lower-bound section largely repeats narrative developed again in the appendix. This organization obscures the principal statistical advance and substantially increases the reading burden.
Suggested fix
Establish the maintained calibrated specification once in each relevant section and use concise references thereafter. Group the setup into structural identification, statistical regularity, and observation conditions; move rank-generated witness restrictions beside the witness construction. Consolidate repeated roadmaps and regime interpretations, and give the main lower-bound section a concise statement of target separation and observed-sample closeness. Preserve every verified condition and theorem-local formalization disclosure.
The compact-support comparisons are detailed, but the methodological positioning of polynomial Gaussian inversion and honest scalar inference would benefit from engagement with their direct antecedents. The supplied literature context identifies Masry and Stefanski for differentiation-based inversion and unbiased normal-mean estimation, and Donoho and Low for functional estimation and honest inference.
Suggested fix
Research and add precise citations for the analytic inversion construction and the estimation-to-inference connection. Explain how the paper's uniform weak-design frontier and weighted observed-law alternatives extend those antecedents, while describing the interval calibration as the certificate-based construction actually delivered.
Many assumption explanations cite HuangZhang2023 without locators, including the bounded-potential-outcome explanation and several witness-subclass explanations. Elsewhere the manuscript distinguishes its numerical boundedness and schedule-level exchangeability from that source's structural class. Readers need a clear distinction between a related causal setting and a source maintaining the displayed restriction.
Suggested fix
Check each assumption-level attribution against the source. Supply exact assumption or section locators for restrictions actually maintained there; describe the unit outcome bound, schedule-level formulation, and Bernoulli witness restrictions as this paper's specification where appropriate. Consolidate citations that merely establish the broader application context.
Ordinary prose violates the affirmative-framing contract in “we do not claim that the frontier fails when ” and “The measurable polynomial envelopes and Hölder mean restriction used here do not impose that increasing-dimensional operator condition uniformly over the model class.”
Suggested fix
State affirmatively that the matched length characterization uses alpha in (0,1/2), while the coverage construction applies throughout (0,1). Describe the spline comparison through its maintained uniform operator conditions and the present class through its measurable density envelopes and Hölder means. Place genuine unresolved coverage questions in Limitations and future work.
The title “Local design density bounds” suggests a neighborhood restriction, whereas the displayed envelopes govern the entire centered latent support [-1/2,1/2]. This distinction matters for interpreting the compact-support characterization and its scope relative to local degenerate-design benchmarks.
Suggested fix
Rename the assumption “Polynomial density envelopes on the latent support” and emphasize the full-support domain at its introduction. Preserve the global bounds and clarify that local dose scarcity is the phenomenon characterized within this globally restricted class.
The references to “steps 3 and 4 of the proof” use manually numbered internal targets. These references bypass the exclusive cleveref contract and become fragile when proofs are reorganized.
Suggested fix
Refer to the proof using Proposition 1 and describe the positive-mass and shrinking-average arguments in words, removing the manual step numbers.
Questions for the authors
- For representative calibrated specifications, at what sample sizes does the selected public score permit intervals narrower than the full outcome range?
- Which analytic ingredients distinguish the uniform frontier from existing polynomial Gaussian inversion and compact-support regression constructions?
Review history
| Round | Date | Score | Findings |
|---|---|---|---|
| 1 | 2 Oct 2026 | 7.2 | 7 |
| 2 | 2 Oct 2026 | 7.2 | 7 |
| 3 | 2 Oct 2026 | 8.0 | 5 |
| 4 | 4 Oct 2026 | 6.5 | 9 |
| 5 | 4 Oct 2026 | 6.8 | 6 |
| 6 | 4 Oct 2026 | 7.2 | 7 |
| 7 · current | 5 Oct 2026 | 7.1 | 7 |
Every referee draw is kept: scores are noisy, so the history is shown rather than only the last or the best pass.
Read against paper.tex sha256 92a4d34215b3.