acceptodds
Under review as a conference paper at ICLR 2027

Learning under Mechanism-Induced Covariate Shift: Finite-Scale Coverage and Source Allocation

Abstract

Simulators and physical acquisition systems can generate abundant data yet still leave deployment-relevant neighborhoods poorly observed, even when those neighborhoods are reachable. We study this as mechanism-induced covariate shift: the conditional law is shared, but a generator's acquisition law and forward map determine where training examples appear in task space. The key quantity is the finite-scale local-mass field—the probability that one generated example falls within a neighborhood of each deployment point. This quantity connects the generator directly to finite-budget learning. For Nadaraya–Watson (NW) regression, it yields a finite-sample risk certificate with constants uniform over the generator family, and on a controlled one-dimensional common-support family the optimized certificate matches minimax risk up to family-uniform constants. Keeping the full spatial field matters: generators can have identical normalized-Jacobian distributions and identical distributions of source-density values at deployment draws while having different finite-scale coverage. The same field also turns upstream data collection into a pre-training decision problem: task matching removes generator-dependent risk differences at fixed labels, exact rejection sampling exposes the corresponding proposal cost, and source mixing gives a convex allocation problem that can exploit spatial complementarity. On a controlled 23-design family, the coverage score tracks exact finite- NW risk with median target-wise Spearman correlation 0.97–1.00 across six settings. Estimated from unlabeled forward samples, finite-scale coverage can therefore guide source selection and allocation before training, without learner-side feedback.

Primary area:
learning theory
Keywords:
covariate shift, nonparametric regression, data selection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CckHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This manuscript introduces a new approach for overcoming covariate shift due training models on synthetically generated data instead of real data. Specifically, the authors assume access to a collection of synthetic generators, each of which samples different regions of the task space. The goal is to selectively combine their outputs so as to produce a useful finite sample training set. Each generator is assessed by the probability that it produces samples near the observed training data and the authors show these can be combined so as to reduce prediction error. This approach is theoretically justified on Nadarya-Watson kernel predictors via a risk bound that connects these probabilities to its prediction error. They then empirically demonstrate this approach on, among others experiments, a 23-design common-support family task and show that task matching removes the fixed-label gap. They then apply their approach to 8 synthetic seven-joint arm-like variants using kernel ridge regression to demonstrate its performance beyond NW kernel estimators.

Critical strengths
  1. This paper studies an important problem which is overcoming the distribution shift incurred by the sim-to-real gap. Their approach to optimally allocate generators to improve training coverage appears to be a novel contribution.

  2. The authors derive many theoretical results to support their arguments.

Critical weaknesses
  1. This paper is very difficult to read and its exposition has made it close to impossible to determine its contributions. For example, many relatively simple concepts are given long and complex names. Neighborhood probabilities are termed "Finite-scale local-mass fields" and “Rigidity of local summaries without finite-scale rigidity” means the statistical summary stays the same when the neighborhood probabilities change. There are also a mass of theoretical results and its not clear how they all relate to each other or the problem. There is a risk bound on NW estimators, a counter example about histograms, another bound on rejection sampling, and so on. Many of these theoretical results also do not fully define their notation, which makes them difficult to assess.

  2. This paper needs to more clearly connect its theoretical contributions to past work on kernel regression. For example, the upper bound in theorem 1 appears to be an corollary of Lemma 2 in Pathak et. al. (2022), which the authors cite. If you take the bound Eqn. 10 in the proof of Lemma 2 and substitute it into the last equation of the proof, I believe you would nearly recover the upper bound in Theorem 1 just using different notation. The only difference seems to be that Theorem 1 (Eqn. 4) assumes ∣∣f∣∣∞≤B||f||_\infty \leq B (which isn't noted until the table at the end). However, in that case the bound in Lemma 2 would already be uniform over ff, so adding on the supfsup_f would not change the bound. I was not able to verify whether the other theoretical results are also similarly connected to prior work.

Questions for authors
  1. Could the authors please revise this work so that the problem and logical progression through the theoretical results are more clear? Please also make sure that all notation is defined where it used, rather than a large table at the end.

  2. What is the practical benefit of theorem 1? I'm not sure how exactly the risk certificate should be used for either improving the source mixing or understanding this problem more abstractly.

  3. This approach appears to be intended as a tool for guiding synthetic data acquisition more generally than for just NW estimators. How well does this approach apply for other predictors beyond kernel methods? Could the authors provide more empirical evidence to demonstrate this that

  4. Do the experiments show that this approach improves predictive performance? Figure 4 seems to be showing that the distribution can be corrected, but does this result in meaningful predictive gains?

Reviewer_B3TROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies covariate shift induced by a data generator's acquisition law and forward map. It uses the local mass field, m_theta(z,h) = P_theta(B_h(z)), to connect finite sample availability to Nadaraya–Watson regression risk. The contributions include an explicit finite-sample certificate, restricted minimax sharpness, a construction separating density-value summaries from spatial coverage, and convex source allocation objectives. Controlled kinematic examples validate risk tracking and source complementarity; additional learners illustrate the scope and limitations of coverage-based allocation.

Critical strengths
  1. The finite-budget analysis is precise and carefully scoped. Theorem 1 separates smoothing bias, the reciprocal nonempty-neighborhood count, and empty-neighborhood loss, with constants uniform across generators. Theorem 2 gives an interesting explicit family with identical normalized-Jacobian and density-value distributions but different local coverage. This construction explains exactly which spatial information global summaries discard.

  2. The allocation analysis makes the coverage perspective actionable. Proposition 2 provides optimality conditions and a Frank–Wolfe gap, while Corollary 7 identifies how spatial complementarity determines mixture gains. The paper also distinguishes label budgets from proposal budgets and separates population guarantees from numerical checks. These distinctions make the theoretical and empirical claims easier to assess.

Critical weaknesses
  1. The learning-theoretic advance over the closest prior work is limited. The profile Psi_theta(h) is the source-only functional of Pathak et al. [1], whose Theorem 1 and proof already connect NW risk to local counts and empty neighborhoods. Retaining the binomial terms exactly is a useful refinement, but much of the generator dependence follows by substituting P_theta = (T_theta)#mu. The sharpness result assumes uniformly bounded source densities, yielding the common classical rate across the family; it supplies little resolution for ranking individual generators. The rigidity construction and allocation results add value, but the overall theoretical increment remains modest for a learning-theory submission.

  2. The guarantee for allocation from finitely estimated coverage fields remains incomplete. Proposition 9 assumes a uniform mass error delta_m and, for the profile objective, requires delta_m < eta_h. It does not derive this error from an unlabeled sampling budget or control estimation of the deployment integral. The profile error scales as delta_m/[eta_h(eta_h-delta_m)], making accuracy in poorly covered regions consequential. Section C.1 uses 2^20 forward samples per design for each mass-field estimate; all experimental task spaces are one- or two-dimensional. The numerical stability checks do not establish the population error assumed by the proposition. This leaves the central practical claim of reliable allocation before training without a quantitative sampling requirement.

[1] A new similarity measure for covariate shift with applications to nonparametric regression, ICML 2022.

Questions for authors
  1. Beyond the exact binomial refinement, which result gives a generator-dependent learning or allocation conclusion unavailable from the local-coverage analysis of Pathak et al. [1]? A concrete comparison under a common task metric and deployment law would help establish the theoretical increment, particularly the role of Theorem 1(b) in source selection.

  2. With a finite number of unlabeled forward samples per source and deployment samples from Q, what conditions give the uniform error required by Proposition 9, and what allocation-objective regret follows? Please clarify how the requirement depends on the neighborhood scale, minimum local mass, and task-space complexity.

Reviewer_J3AkOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies regression when the training-input distribution can be chosen from a family Pθ{P_\theta}, while the test-input distribution QQ remains fixed. Labels satisfy Y=f(Z)+εY=f(Z)+\varepsilon, with the same regression function and conditional noise distribution for every choice of PθP_\theta. The question is whether one can choose a useful training distribution before fitting a predictor.

The motivating example is a collection of robot arms with different link lengths. Randomly sampling their joint angles produces different distributions over endpoint locations or distances. These locations or distances are the regression inputs ZZ. The experiments assign synthetic regression functions to these inputs; they do not involve learning from physical robot measurements.

The main quantity is mθ(z,h)=Pθ(Bh(z))m_\theta(z,h)=P_\theta(B_h(z)), the probability of drawing a training input within distance hh of zz. The paper derives an expected prediction-error bound for Nadaraya–Watson regression, which predicts by averaging nearby training labels. It also proves a minimax comparison on a restricted one-dimensional family and constructs an example showing that distributions can have the same histogram of density values while placing probability mass in different locations. Finally, it proposes choosing proportions of samples from different training distributions by minimizing $\Psi(\alpha)=\int\frac{1}{P_\alpha(B_h(z))},dQ(z), \qquad P_\alpha=\sum_i\alpha_iP_i.$

Critical strengths

Choosing how to collect training inputs before spending resources on labels or model fitting is potentially interesting. In particular, two sampling procedures may be useful together because they produce examples in different parts of the test-input domain. This could motivate a useful discussion of when local sampling probabilities provide actionable information about data collection.

Critical weaknesses
  1. The presentation obscures a relatively straightforward statistical setup. The abstract introduces “acquisition laws,” “task space,” “spatial fields,” and “risk certificates” without first explaining their roles in the learning problem. Here, these mean a sampling distribution, the regression-input domain, a function of neighborhood probabilities, and an upper bound on expected prediction error. Using these terms before establishing the underlying objects makes the paper unnecessarily difficult to follow.

The central example should appear at the beginning: sample joint angles, map them to an endpoint location, and learn a function of that location. This would explain what ZZ represents, why changing a robot changes its distribution, and why two robots reaching the same locations can nevertheless provide different numbers of training examples near them. Instead, robots and “design families” appear around line 110 without enough preparation.

The abstract’s reference to “Jacobian values” is particularly uninformative without specifying the differentiated map and the distribution under which those values are sampled. The paper also needs to distinguish its terminology from its contributions: introducing a new name for a familiar statistical object does not explain what new problem or result is being introduced.

  1. The additional theoretical contribution is not sufficiently compelling. The reciprocal-neighborhood quantity ∫P(Bh(z))−1,dQ(z)\int P(B_h(z))^{-1},dQ(z) already appears in Pathak et al. (2022), together with nonasymptotic Nadaraya–Watson bounds and minimax analysis under covariate shift. The present paper refines the upper bound by retaining the random number of nearby observations and the probability that there are none. However, the argument remains a local-average error decomposition, and the paper needs to explain more clearly why this refinement constitutes a substantial advance.

Describing PθP_\theta through a procedure that generates inputs does not itself change the statistical problem. Indeed, the paper explicitly states that the procedure affects learning only through PθP_\theta. Similarly, making every training distribution equal to QQ necessarily removes differences between the resulting learning problems, because the conditional label distribution is already shared.

The minimax result is also too restricted to justify the proposed use for comparing training distributions. It concerns a one-dimensional family with densities bounded above and below, where all members have the same standard nonparametric rate up to constants. This does not establish that the proposed score correctly ranks their prediction risks or that optimizing it selects a better training distribution. The paper needs to separate these claims.

  1. The experiments do not establish that optimizing the proposed quantity reliably improves learning. The strongest reported correlations come from a single synthetic family with a one-dimensional regression input and analytically evaluated Nadaraya–Watson risk. Appendix C.1 additionally reports that elementary functions of robot link lengths correlate with the proposed score at 0.978–0.991. Consequently, the headline correlations do not establish an advantage from estimating neighborhood probabilities over simpler ways of ranking the candidate procedures.

Evidence beyond this setting is mixed. The seven-joint robot experiment still uses only a two-dimensional regression input and synthetic target functions. Only one of six kernel ridge regression comparisons between the proposed allocation and equal sampling has a confidence interval excluding zero. None of the reported downstream decision comparisons does. The separate experiment varying the sampling distribution gives essentially no aggregate improvement for random-feature regression or the multilayer perceptron.

There is also a gap between the quantity optimized and the outcome of interest. Appendix C.5 acknowledges that minimizing Ψ\Psi does not generally minimize the finite-sample error bound, let alone the prediction risk of a particular target function and learner. Furthermore, the bound uses population neighborhood probabilities. The approximation result assumes uniformly accurate estimates of those probabilities without establishing the sample budget needed to obtain them. The practical recommendation therefore remains insufficiently supported.

Questions for authors
  1. Can you introduce the problem using an explicit regression example, defining the sampled variables, regression inputs, labels, and test distribution before introducing the application-specific terminology?

  2. What statistical restriction or result is added by representing a training distribution through a data-generating procedure? Which conclusions depend essentially on that representation?

  3. What is the main theoretical advance over Pathak et al. (2022), and what does the minimax result imply about comparing two candidate training distributions?

  4. Under what conditions does minimizing the proposed reciprocal-neighborhood objective improve actual prediction risk? How should the neighborhood radius and distance metric be chosen for the learner being trained?

  5. How many unlabeled samples are needed to estimate the objective accurately enough to make reliable choices, particularly where neighborhood probabilities are small?

  6. Can the authors demonstrate a clear advantage over simple ranking rules and equal sampling in a setting beyond low-dimensional synthetic regression, with improvements that persist for learners other than local averaging methods?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.