acceptodds
Under review as a conference paper at ICLR 2027

Four-Certificate Minimax Rates over Public Class Pairs for Bounded Partially Linear Models

Abstract

We study estimation of a bounded partially linear coefficient from two public nuisance classes with separate approximation and stochastic complexity certificates. In the strict envelope interior, the worst-public-pair expected absolute risk for has rate . Matching directional lower bounds close the gap between the Gu–Yin–Cai–Fan lower envelope and TAME's calibration upper. We also construct one rule attaining this rate without any of the four budget inputs. Double calibration removes approximation-dependent orientation, and empirical local Rademacher fixed points calibrate the stochastic radii. The general rate persists over full finite dictionary balls and finite unions of rank-one balls. Full interval boxes instead have sharp rate , and full affine Hilbert balls have rate ; both are attained without any budget input. The model requires only positive average residual variance and bounded envelopes. Results include all sample sizes and zero budgets; a uniform upper connects to saturation. The guarantees are statistical; effective optimization requires additional interfaces.

Primary area:
learning theory
Keywords:
partially linear models, minimax estimation, nuisance classes, local Rademacher complexity, adaptive estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KNhyOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers the problem of estimating partially linear model coefficients within four error budgets for both function class approximation error and statistical error. This paper claims to close the gap between the upper and lower bounds for expected absolute risk. The paper achieves a bound of n−1/2+(a+s2)(b+t2)+(s∧t)2+t(a∧t)+s(b∧s)n^{-1/2} + (a+ s^2)(b+ t^2) + (s\wedge t)^2 + t(a\wedge t) + s(b\wedge s), where a,s,b,ta,s,b,t are four error budges

Critical strengths

The paper closes the gap of upper and lower bounds for bounded public classes carefully established by two prior works. The tight lower bound established in this paper could be very interesting.

Critical weaknesses

It seems the paper largely follows the works of Gu et al. 2026 and Gu 2026 (TAME). However, the presentation can be improved such that this paper is more self-contained.

  • The partially linear model was not introduced throughout the introduced, nor the association of four budget parameters to the model. This makes the discussion of the contribution hard to read for the general audience.
  • Notation ∧\wedge and ∨\vee are not defined. We should not assume that the readers are familiar with the notations from prior works.
  • Line 43-44: "Since W≍AB+M+O+PW \asymp AB+ M + O+ P, the two-branch rate is minimax optimal in this experiment", while the paper has not introduced any experiment yet.
Questions for authors

It seems Gu 2026 already achieved matching upper and lower bound. How do you compare your results to theirs?

Reviewer_CUZtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies estimation of the scalar coefficient in a bounded partially linear model when the outcome and treatment nuisance functions are approximated by two publicly supplied function classes. Each class has a separate approximation-error certificate and stochastic-complexity certificate, giving four budgets: (a,s)(a,s) for the outcome nuisance and (b,t)(b,t) for the treatment nuisance.

Critical strengths
  1. The additional terms are consequential because they can change the polynomial order of the lower bound. For example, on the ray (a,s,b,t)=(r,0,0,r)(a,s,b,t)=(r,0,0,r), the previous nuisance lower envelope is of order r3r^3, while the new directional obstruction is of order r2r^2. This establishes that the existing calibration upper bound captures a real statistical difficulty. The contribution is appropriately positioned relative to TAME: the leading two-branch upper rate was already available, while the present paper supplies its matching lower-bound justification in the stated public-class experiment.
  2. Theorem 4 establishes that one estimator attains the sharp two-class rate without receiving the approximation or stochastic budgets as inputs. This is stronger than selecting a calibration branch using known nuisance-error bounds. The proof combines two analyses of the same double-calibrated estimator with empirical local Rademacher fixed points computed on an independent covariate block. It handles arbitrary represented classes without imposing convexity or star-shapedness on the original classes. The precise advance is simultaneous adaptation to the sharp four-budget rate: earlier SADE results already provided budget-free adaptation in a single-outcome-class setting.
Critical weaknesses
  1. The M1 interface guarantees measurable evaluations and sequential approximation, but it does not provide terminating optimization or separation procedures. Theorem 4 requires suprema over represented classes, empirical local-complexity calculations, and approximately optimal feasible selections from countable candidate sets. Consequently, the theorem does not yet provide a generally usable adaptive learning procedure. The paper acknowledges this limitation, but it materially restricts the algorithmic significance of removing all four budget inputs: the statistical guarantee is established, while the computational requirements for realizing it remain unresolved.

  2. The box estimator exploits the availability of every admissible coordinate combination, while the Hilbert estimator exploits an entire ball with a uniformly bounded pointwise envelope. These results therefore require more information and structure than merely having a fitted nuisance estimate represented in a box or an RKHS. Lemma 26 correctly quantifies the envelope cost of enclosing the hard dictionaries in a Hilbert ball, but the manuscript would benefit from a worked example showing how the assumptions and constants behave for a familiar learning model. Without such an example, the practical reach of the geometry comparison remains difficult to judge.

Questions for authors

See Weakness

Reviewer_JGghOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the problem of parameter estimation in a partially linear model. Concretely, the paper studies the model T=π0(X)+u,Y=β0T+μ0(X)+ε,T = \pi_0(X) + u, \qquad Y = \beta_0 T + \mu_0(X) + \varepsilon, where TT is the treatment variable, XX is the covariate, u,εu, \varepsilon are zero-mean noise variables, and π0,μ0\pi_0, \mu_0 are nuisance functions.

The paper studies the problem of estimating β0\beta_0 in the setting where we do not posit any assumptions on μ0\mu_0 and π0\pi_0 apart from the fact that they can be approximated by auxiliary function classes (Gμ,Gπ\mathcal G_{\mu}, \mathcal G_{\pi}, resp.). The key contribution is a characterization of the minimax rates using four parameters: approximation errors of each of π0,μ0\pi_0, \mu_0 w.r.t. its respective function class, and two terms that capture the complexities of the respective function classes through localized Rademacher complexity.

Critical strengths

As far as I can tell, the authors give a sharp characterization of the minimax rate in their problem as a function of the four parameters, and provide an algorithm that adaptively achieves the upper bound without knowledge of parameters. However, due to the poor quality of presentation, it is difficult to understand what the precise contributions are relative to prior work.

Critical weaknesses

Presentation. The presentation of the paper is poor. The motivation for the problem is essentially missing from the introduction. The precise statement of the problem under consideration is spread over Section 2, and the reader has to track many pieces of auxiliary notation to assemble the problem definition. The exposition is extremely dense, with many pieces of terminology introduced quickly and with little explanation ("ordered countable cores", "stochastic/approximation certificates", "directional interactions", "M1").

The quality of presentation is the main reason behind my score.

Questions for authors
  • Could authors clarify the contributions of the paper relative to prior work?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.