acceptodds
Under review as a conference paper at ICLR 2027

Learning Conditional Distributions from Replicated Aggregates

Abstract

Learning an individual response distribution from group totals becomes difficult when each measurement carries unknown offsets and the underlying distribution is heavy-tailed. To tackle this, we introduce Pair-Triple Spectral Recovery (PTSR) where pair differences recover the conditional law up to reflection and triple contrasts use spectral phase to resolve the remaining orientation. PTSR provides moment-free recovery while also eliminating the shared nuisance offsets. We provide finite-sample guarantees for reflection selection and connect spectral signal-to-noise to intrinsic reflection information. Across two heavy-tailed benchmarks, PTSR improves conditional-density accuracy and held-out likelihood over winsorized method-of-moments and aggregate-likelihood estimator baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Distributional Disaggregation, Conditional Distribution Learning, Characteristic function, Spectral Methods, Inverse Problems, Alpha-Stable Distributions, Weak Supervision, Method of Moments

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ypv5OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes PTSR, a method based on paired tripled differences that recovers the conditional law of a response variable in an aggregate learning setting where each group shares an unknown offset. The poor writing and description of the setting and the missing motivation of the offset itself and thus of the main research problem, leave it open whether this work actually addresses an impactful research question. Together with the almost non-existent proper baselines and the evaluation focusing only on two seemingly cherry-picked synthetic data tasks selected by the authors, it is hard to judge whether the paper constitutes a contribution at all. Neither the writing nor the actual methodology and experimental setup of the paper hold up to the standards of this conference, which is why I recommend a clear rejection.

Critical strengths

The paper proposes to learn shape and orientation of a response distribution while eliminating shared offsets. The method itself is justified and leads to an estimation of the conditional distribution while simultaneously removing the offset.

Critical weaknesses

Baselines and experimental setup: The paper lacks a clear and thorough experimental evaluation, especially with respect to baselines and benchmarks. The authors claim "two heavy-tailed benchmarks" in the abstract. However, these are synthetic experiments, constructed by the authors with no justification why they should be a fair benchmark. In fact, the scenarios consider a specific heavy-tailed (α\alpha-stable) distribution which seems to be chosen specifically to accommodate the paper setting without addressing or discussing its suitability in practical applications. There are no actual benchmarks or real-world datasets used, not even from the referenced related work.

Furthermore, the paper introduces several baselines (mostly without any reference or justification) of which three are "privileged" baselines that use some notion of ground truth information and another two are trained on a reduced pair pool. While some of those are based on related work (the others are not referenced at all), they are implicitly flagged as "not comparable" and not adequately compared. The only remaining actual comparable baseline is then the Winsorized GMM for which neither an adequate justification nor any reference is given. Possible baselines to add (beyond properly evaluating the cited ones [1,2,3]) would include for example:

  • Delaigle et al. [7], who provide a density estimator based on pairwise differences (not triplets), which is also considered in this paper. They also already provide the estimator based on cosine/Fourier-features
  • Li et al. [8], who provide a nonparametric estimator for two or more available measurements also using characteristic functions

Overall, the experimental section seems constructed and cherry-picked specifically to highlight the performance of the proposed PTSR method. While Figure 4 shows a qualitative evaluation of the method regarding calibration, the quantitative comparison in the paper, especially with respect to the dataset and baseline selection, is not suitable for a venue like this. This would either require a detailed setup against benchmarks and baselines or alternatively good justification why this does not apply together with more thorough qualitative evaluation of the method and a placement within the broader literature of learning from aggregates.

Writing and style of the paper The paper is very poorly written and thus extremely difficult to follow and understand. Neither the motivation nor the research question are adequately introduced or explained. The paper uses seemingly complex scientific words and definitions without actually explaining the context or providing the thorough definition. While most of the definitions come from signal processing and might be known in that community they are neither cited nor sufficiently explained here for the reader coming more from the probabilistic methods community (which this paper is classified under). Also nothing of the definitions and the naming is similar to the cited related work [1,2,3]. The following constitutes a non-exhaustive list of poor wording and styling:

  • The abstract mentions a conditional law, but so far no covariates are introduced
  • Many words are thrown in the abstract and introduction that are not properly defined (or maybe defined later) and not really clear from the context, e.g., "triple contrasts", "remaining orientation" (orientation of what), "totals" (not defined in introduction and not a consistent wording across literature), "identified up to reflection by pair spectra" (before defining pair spectra), "orbit" (used in contributions but never explained until later), "spatial coverage" (how is the scenario now spatial),...
  • "Baseline", "offset", and "location" are three words used for the same thing in the introduction
  • Section 4.2: Several metrics are introduced but only one is actually defined in the main paper
  • 2.1 uses the notation Zbn(r)Z_{bn}^{(r)}, while Proposition 1 suddenly changes it. Also the proof of Prop. 1 is very informal.
  • l. 162 "[...] share an anchor context xjx_{j}, as in the recovery benchmark". Since the recovery benchmark is not defined (and also is not actually a benchmark), it is not clear what an anchor context is
  • Algorithm 1: The input requires fixed frequencies and candidate bandwidths, however these are only defined in the appendix (without an actual selection heuristic) and should be in the main part
  • Algorithm 1: three out of five equations are actually in the appendix. Since the algorithms constitutes the contribution the equations should be in the main text. Also the algorithm is written very informal
  • l. 267 "Covariates are scrambled Sobol points": Why choose a quasi-random Sobol sequence, as opposed to random sampling. The choice should be justified and the Sobol sequence equipped with a reference.
  • Figure 2: "Coordinates [...] and ticks are shared with x2x_{2} labelled only in the Truth column". This is obvious and does not need to go in the Figure description as it serves now purpose. At the same time raises the question why x1x_{1} is not also shared, since both panels have the same xx-axis
  • The contributions section names sections 3,5,6, thus in the wrong order
  • l. 169 The space of θ\theta is undefined
  • Wording errors such as "rspctively" in l. 148
  • ...

While individual mistakes are not a problem, the writing of the paper overall does not hold up to scientific standards and makes it very hard for any reader to understand what the contribution and proposed method actually is.

Missing references and literature The references are very sparse and there are many missing references, as well as important related work that is not cited. Examples of missing references:

  • l. 272 Winsorized GMM not referenced (or adequately explained). Same for the other benchmarks, i.e., Nuisance-profiled MLE, Shift-known MLE, Raw ECF, ...
  • l. 29 "Recovering [...] is an inverse problem" not referenced of justified
  • l. 267 Sobol points (sequence) not explained or referenced
  • l. 40 "For a family identified up to reflection by pair spectra, this leaves two candidate orientations." Claim not supported by a reference
  • While the main contribution is concerned with offsets across replicate groups, no reference is given to support the relevance of the setting in practice

Exemplary additional relevant related work:

  • Tanaka et al. [4], while focusing on the spatial setting, propose a probabilistic model to model different levels of granularity (aggregates) based on Gaussian processes
  • Javanmard et al. [5] consider an aggregate learning setting and study the impact of designing a loss function either on the instance-level or the bagged level
  • Yousefi et al. [6] consider an aggregated learning setting and propose to a linear combination of latent Gaussian processes as a model.
  • ...

References

[1] Zhang, Y. et al. Learning from Aggregate Observations, NeurIPS 2020.

[2] Bhowmik, A. et al. Frequency domain predictive modelling with aggregated data. UAI 2017.

[3] Law, H. et al. Variational Learning on aggregate outputs with gaussian processes. NeurIPS 2018.

[4] Tanaka, Y. et al. Refining Coarse-Grained Spatial Data Using Auxiliary Spatial Data Sets with Various Granularities. AAAI 2019.

[5] Javanmard, A. et al. Learning from aggregate responses: Instance level versus bag level loss functions. ICLR 2024.

[6] Yousefi, F. et al. Multi-task learning for aggregated data using gaussian processes, NeurIPS 2019.

[7] Delaigle A. et al. On deconvolution with repeated measurements. An. Statist. 36(2).

[8] Li, T. et al. Nonparametric estimation of the measurement error model using multiple indicators. Journal of Multivariate Analysis 1998.

Questions for authors
  1. While the examples in l.25-l.30 are examples of aggregates, it is not clear at all why a shared offset makes sense here. It would be good to further motivate this to highlight the relevance of the actual research problem.

  2. You claim that estimating the response is difficult under heavy tails (l.13) and that "in general the mean [...] may not exist" (l.79). Since estimation under heavy tails is a proposed advantage of PTSR: Why does the heavy-tailed assumption matter in your specific setting beyond the constructed synthetic dataset. Are there any examples of real-world data from aggregates where heavy-tailed data are present or the data has undefined mean (i.e., Cauchy data)?

  3. l. 40 "For a family identified up to reflection by pair spectra". Since this seems to be the main assumption on the distributional family, how restrictive is this assumption and what families are included here beyond the α\alpha-stable family in the experiments. Do any other works consider similar restrictions or similar distributional families?

36 comments
AuthorsOpenReview ↗

Concern Regarding Review Accuracy and Integrity

We wish to flag a specific factual error that raises concerns about the accuracy and reliability of this review to the ACs, SACs and anyone else reading this:

The reviewer cites [2] (Bhowmik et al., Frequency Domain Predictive Modelling with Aggregated Data), as appearing at UAI 2017. This is incorrect. The paper was published at AISTATS 2017 (or PMLR, Vol. 54, pp. 971–980 as cited in our submission). This is directly verifiable from the official proceedings (https://proceedings.mlr.press/v54/bhowmik17a). And can be double checked against UAI's accepted papers (https://www.auai.org/uai2017/accepted.php).

This is an objectively false bibliographic attribution and it is particularly concerning that a reviewer who explicitly questions our referencing and scientific standards presents factually false publication information in their own assessment. We regard this as a potential review-integrity concern and request that the ACs and SACs examine whether the review meets ICLR's standards of responsible reviewing. Factually incorrect citations are a serious academic integrity violation, and we believe the same standards expected from submissions should apply equally to reviews.

We have already discussed and consolidated our responses to most of the reviewer's scientific, methodological, and presentation related criticisms. However, we have not had sufficient time yet to organize these into a clear and easy to read coherent point-by-point rebuttal due to the sheer volume of criticisms. We will post our response by the coming weekend, when all authors can once again reconvene to properly finalize it.

(after further discussion with the reviewer, we would like to retract parts of this statement such as the reliability concerns because we believe the reviewer was reviewing in good faith and are very grateful for the reviewer for providing us detailed comments that is helpful to understand their perspective)

I acknowledge the author's response and raised concerns and provide answers to several points below:

  1. You are correct that Bhowmik et al. [2] was published in AISTATS 2017 not in UAI 2017 as I have mistakenly stated and it is absolutely fair to point that out. However, trying to discredit the whole review on this basis seems to push it quite far, although the motivation behind discrediting seems quite clear given the ratings.
  2. I did not (and did not mean to) question the authors' scientific integrity and standards in my comment. I state that the paper does not hold up to the (scientific) standards, explicitly mentioning that this goes beyond small mistakes such as a false attribution of a reference. In my opinion, the overall standard of ICLR is not met, neither in writing nor in the baselines and experiments, as also pointed out by reviewer u6v4. This does not imply that the authors purposefully committed misconduct or violated scientific integrity; it just means that the paper does not meet the standards.
  3. While the review could be seen as detailed, it mentions overall only three main weaknesses. The "[...] sheer volume of criticisms" you mention seems overstated, especially since W1 aligns with reviewer u6v4, which you already seemed to have answered.

Of course, there is no timeline for when to reply, but I am happy to participate in additional rebuttal and am looking forward to further constructive comments when the authors "once again reconvene".

Hi, one of the authors here. Thanks for writing back :)

We appreciate the message and wanted to briefly clarify your three points to avoid any further misunderstandings before we post our consolidated rebuttal:

Regarding (1): To be clear, we did not suggest that this single error discredits your entire review, nor were we attempting to dismiss your remaining concerns. On the contrary, we had already responded to the other reviewers and did not want to leave your review unanswered in the meantime. This is precisely why we explained at the end of our preliminary comment that we had already discussed and consolidated our responses, but needed additional time for the coauthors to meet again and organise a clear rebuttal. Our intention is to give your review the same careful attention as the others, not to dismiss or discredit it.

Regarding your remark that "the motivation behind discrediting seems quite clear given the ratings," , it's best to and let's try to avoid such assumptions about each other's motivations where possible and keep the discussion focused on the scientific arguments and evidence.

Regarding (2): This was clear and not misunderstood on our end, we do understand that in your assessment, you feel our paper falls short in your view on what ICLR's (scientific) writing/exposition standards require. Either way, this is why we hope our consolidated rebuttal will give us the space to explain our reasoning and evidence more clearly, discuss the specific points of disagreement, find common ground, and hopefully address your concerns through constructive discussion.

Regarding (3): Yes, we recognise that you have organised your criticisms into three main categories, which is why we referred to the "scientific, methodological, and presentation related criticisms". To clarify "sheer volume of criticisms," we meant the numerous individual points raised within some categories, which take time to address carefully. We want to give your review the proper attention and address your comments thoughtfully, rather than post a rushed or fragmented reply.


Finally, we want to reiterate that we genuinely do appreciate the time you have taken to read and review our paper, as well as your willingness to continue the discussion. We want to make it clear our intention is to engage with all your concerns seriously and to explain the reasoning and choices underlying the work.

With the consolidated rebuttal, we hope our subsequent discussions will help clarify the precise points of disagreement and establish common ground where possible. We believe constructive scientific discussion is valuable even when assessments and views differ. So we look forward to engaging with your feedback in detail when we post our full response.

Thanks again for your time, and wishing you a good weekend!

Reviewer_u6v4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies learning individual conditional distributions from replicated aggregate observations with unknown shared offsets. The authors propose Pair–Triple Spectral Recovery (PTSR), which uses pair contrasts to remove nuisance offsets and recover distributions up to reflection, and triple contrasts to resolve the remaining orientation ambiguity through spectral phase information. The paper provides finite-sample guarantees for reflection recovery and connects spectral signal-to-noise with intrinsic information. Experiments on synthetic heavy-tailed benchmarks show that PTSR improves conditional density recovery and likelihood compared with aggregate likelihood and winsorized moment-based baselines.

Critical strengths
  1. The paper identifies an interesting structure in replicated aggregates: pair contrasts eliminate unknown offsets but lose orientation, while triple contrasts recover phase information. This leads to a principled decomposition of shape estimation and orientation recovery, rather than a heuristic algorithm.
  2. The paper provides meaningful theoretical guarantees for reflection selection and connects the proposed spectral score to statistical information quantities. The use of bounded Fourier features is particularly suitable for heavy-tailed distributions where moments may not exist.
Critical weaknesses
  1. Experiments are entirely conducted on synthetic stable-mixture benchmarks. Although these experiments demonstrate effectiveness under controlled heavy-tailed settings, they do not show whether PTSR works on realistic aggregate learning problems or under distribution mismatch. This weakens the evidence for the claimed general applicability of the framework.
  2. The method appears to rely on a correctly specified conditional distribution family with tractable characteristic functions. It remains unclear how PTSR performs under model misspecification or with more flexible/non-parametric distribution models.
  3. The proposed method uses local evidence pooling over covariate locations with Gaussian kernel smoothing. The experiments are restricted to a two-dimensional covariate space, where spatial pooling is relatively straightforward. It remains unclear whether the approach can scale to high-dimensional covariates, where kernel-based aggregation may suffer from the curse of dimensionality.
  4. The paper reports average recovery metrics but provides limited discussion of when PTSR fails. For example, the method depends on sufficient triple-phase information, but the conditions under which reflection information becomes too weak are not systematically characterized in the experiments. More comprehensive failure analysis would improve understanding of the practical limitations.
  5. Although the paper provides theoretical justification for different components, experimental ablations separating the contributions of:
  • pair-stage distribution estimation;
  • triple spectral evidence;
  • spatial pooling;
  • frequency selection; are limited. It is difficult to quantify which components contribute most to the final improvements.
Questions for authors
  1. The experiments are limited to synthetic stable-mixture benchmarks. Has PTSR been evaluated on real aggregate datasets, and how does it handle distribution mismatch?
  2. How robust is PTSR when the assumed conditional distribution family is incorrect or replaced with more flexible/non-parametric models?
  3. The experiments use only 2D covariates with kernel-based pooling. How does PTSR scale to high-dimensional covariates?
  4. Under what conditions does reflection information become insufficient for recovery? Can the authors provide a systematic failure analysis?
  5. Can the authors provide ablations isolating the effects of pair estimation, triple spectral evidence, spatial pooling, and frequency selection?
10 comments

Thank you for identifying where the scope of this work, in particular the boundaries and limitations, needed clarification. We have completed additional studies per your suggestions using real-world data, increasing the covariates' dimension and isolating failure conditions contributions of different stages of the method. These studies distinguish where the method is effective from where its present implementation remains limited. We think this will greatly strengthened the clarity of the manuscript and the method's scope and clarify the conditions under it is effective.

Indeed, the central claim is conditional-distribution recovery within a family identified up to reflection by the pair designs, together with a moment-free guarantee for local reflection selection. Not that this is a nonparametric or dimension free distribution estimation methodology. The additional experiments below assess the robustness and operating range of that construction.

We hope these additions address the main concerns around the manuscript's soundness and would appreciate reconsideration in light of the additional ablations and evidence (once included). Thank you for your perspectives and suggestions.

with regards to question 1:

This method originated from an application within the energy industry. Due to compliance reasons, we cannot distribute the original dataset that inspired and validated the method. We thought the synthetic 2D experiments was a nice controlled and visual way to understand what is going on in this method but indeed the criticism you gave is fair, so we have added measured-response evidence using power-line communication (PLC) voltage recordings, while distinguishing the measured data from the constructed aggregate protocol.

The recordings come from the INESC-ID/IST time-domain PLC measurement archive which should be easily accessible. Original samples were divided chronologically into disjoint calibration, training, validation, and test segments, with gaps at the boundaries.

Of course, in order to assess the method, we need to know the ground truth (unaggregated data). As such, we aggregate individual observations from this dataset and assess the conditional distribution recovered by PTSR from the totals against the underlying empirical distribution generated by whatever real-world data generating process underlies the observations.

We evaluate two settings. Recorded-voltage estimation learns each recording’s marginal distribution from repeated sums of measured responses. Measured-noise conditional recovery replaces the simulated stable noise in the conditional-mixture task with samples from the voltage recordings, retaining the predefined bounded, covariate-dependent response component.

Both settings construct repeated totals and shared offsets by resampling within the appropriate source split. All methods receive the same separate, unshifted aggregate-only location/scale calibration. Thus these are real-response aggregate studies. The recorded-voltage setting uses constant covariates within each recording; the conditional-mixture setting uses generated covariates. This makes explicit what each experiment tests.

Table R1. Predictive performance on measured PLC responses. Entries are binned negative log scores, averaged equally over the seven recordings and then over six seeds; lower is better. All methods are evaluated on the same held-out responses using fixed bins, including tail bins.

MethodRecorded voltage ↓Measured-noise conditional mixture ↓
PTSR3.81663.9994
Matched Winsorized GMM3.83744.0124
Profiled MLE3.87564.1553
Aggregate MLE4.11884.1569
Gaussian pair-spectrum reference4.36994.4021
Individual supervision, privileged3.67203.8861

The profiled likelihood accounts for a shared offset, while the raw aggregate likelihood does not. Matched GMM shares the pair estimate and pooling with PTSR, isolating the orientation statistic.

PTSR has the lowest average log score among the feasible methods in both settings. The clearest matched result is the measured-noise conditional mixture: the PTSR-minus-GMM difference is −0.0131, with interval [−0.0190, −0.0071]. Differences relative to profiled and aggregate MLE are −0.1559 [−0.1922, −0.1197] and −0.1575 [−0.1877, −0.1273], respectively.

We do not assume that the measured voltages follow an exact stable law. The controlled experiments remain important because they provide the ground-truth conditional laws needed to isolate the pair–triple mechanism and measure recovery error directly. Nonetheless, this is a very useful improvement to the manuscript and we will include the real-world validation of the method.

With regards to question 2:

Thank you for this feedback. In general this method uses the empirical characteristic function as its base statistic to estimate upon. Thus wherever the empirical characteristic function (eCF) is a valid estimator of the true charcteristic function, this method is valid. The eCF may even be computable where the true characteristic function is intractible, but we do not to consider this particularly pathological setting.

Indeed, more restrictive to the exact methodology demonstrated in this manuscript is that the noise in the system needs to be symmetric such that the pair contrasts can pinpoint the distribution up to orientation (reflection). The triple contrasts would then provide a choice of this orientation. More generally, one could extend the method to higher order contrasts (e.g. use triple contrasts to help pinpoint the distribution up to 4th moment orientation...). Thank you for highlighting this, we will make clear of this requirement in the manuscript and also the potential extension, which is outside the scope here.

The reason experiments were run on the α\alpha-stable class of distributions is because the heavy tails and absence of raw moments are features which often inhibit other classical statistical methods that do not use the eCF. As such, able to handle heavy tailed symmetric distributions is an advantage of the method, rather than a cherry-picked distribution class (inasmuch as they are symmetric distributions).

Accordingly, our supported claim is family-conditional recovery and the tested finite-family adaptation under symmetric noise. The present results do not establish unrestricted nonparametric identification. Replacing the family with an arbitrary flexible distribution model would require an identification and estimation analysis beyond the local reflection-selection guarantee.

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

The paper estimates a conditional response distribution from repeated group totals when each totals contains an unknown shared offset. They used the differences of two totals cancel the offset and yield a characteristic function. The PTSR fits a conditional model to these pair spectra, then uses the imaginary part of triple characteristic functions to choose between the learned reflected candidates. The sine and cosine features are bounded, so the estimator does not require raw response moments. The paper gives a local reflection bound and a spectral interpretation, then reports two synthetic heavy tailed recovery tasks and supporting controls.

Critical strengths

The construction handles a meaningful nuisance structure. Both pair and triple contrasts cancel an offset that can change between replicates.

Critical weaknesses

The paper seems not compare PTSR with a direct Fourier-based learner trained on the same offset free pair and triple observations. Such a comparison is needed to isolate the benefit of PTSR’s two-stage design.

Questions for authors

In the experiment, the baseline Winsorized GMM uses the offset free triples, but replaces Fourier evidence with a clipped cubic statistic. Could you provide a more fair comparisons where the baselines are trained on the same offset free pair and triple observations?

3 comments

Thank you for reading the paper carefully and highlighting the choice of cubic statistic. We agree that including its unclipped counterpart allows for fair comparisons (and also makes the rationale of winsorization clearer).

(Our rationale for winsorization: We initially used winsorization because we wanted to provide a fair robust cubic comparator (following Equation (25), the statistic is gλ(Cjn)g_{\lambda}(C_{jn})) in our heavy-tailed setting. Since at the benchmark's α=1.5\alpha=1.5, the triple contrast still contains heavy-tailed noise, whose absolute third moment is infinite so consequently, the empirical cubic average can be dominated by extreme observations. In contrast, we used winsorization to ensure ∣gλ(Cjn)∣≤λ|g_{\lambda}(C_{jn})|\leq\lambda, giving the statistic finite variance.)

As you requested, for fairness we additionally evaluated the unclipped counterpart , λ=∞  ⟹  g∞(Cjn)\lambda=\infty \implies g_{\infty}(C_{jn}) :

Here clipping is removed from both the observed statistic and its candidate calibration. All 3 methods share the same learned pair model, offset-free triple observations, and evaluation panel. Both cubic variants use the same candidate-calibration draws, variance normalization, bandwidth grid etc.

As per the manuscript's Tasks A and B, with N = 8192 triple groups per anchor, the results across twelve independent runs are:

MethodTask A conditional TV, mean ± SD ↓Task B conditional TV, mean ± SD ↓
PTSR0.0674 ± 0.00450.0694 ± 0.0088
Winsorized GMM0.1225 ± 0.01500.1112 ± 0.0266
Unclipped cubic GMM0.1138 ± 0.01630.1281 ± 0.0245

PTSR achieves approx 41% and 46% lower mean conditional TV than unclipped cubic GMM on Tasks A and B, respectively. PTSR also achieves lower mean conditional TV error than both cubic GMM variants on Tasks A and B at all three sample budgets (512, 2,048, and 8,192 triple groups per anchor).


Outlier diagnostics further support our motivation for winsorization: At the largest budget, a single observation contributes a median 60.5% and 61.5% of the total absolute cubic magnitude across training anchors in Tasks A and B. With the selected winsorization thresholds, these median shares fall to approximately 0.023% and 0.022%. Thus, clipping substantially limits the influence of extreme observations, although it does not uniformly improve conditional-TV recovery; the relative ordering of the two cubic variants varies across settings.

Overall, PTSR's advantage persists with or without winsorization. We will include the unclipped results and clarify the mathematical rationale for the original robustification using winsorization in the manuscript.

We thank the reviewer again for suggesting this additional comparison, which indeed helps clarify the role of clipping and strengthens the empirical evaluation :)

Thank you for providing the new comparison and outlier diagnostics. My concern about the cubic comparison is largely resolved.

Thank you for your suggestions and feedback. We are glad to hear this is resolved!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.