acceptodds
Under review as a conference paper at ICLR 2027

Greedy MMD Without Replacement: Selecting and Stopping for Mixture Separation

Abstract

This paper studies how to split an unlabelled sample drawn from a mixture when a separate reference sample from is available. Greedy maximum mean discrepancy (MMD) selection without replacement chooses distinct observations to approximate , leaving a remainder intended to approximate . Success depends on order and stopping: early choices may remove observations needed later, while a wrong stopping size can leave a poor remainder even when the selected observations fit . An error bound tracks the effect of earlier selections. Under two geometric conditions and appropriate stopping, convergence rates are established for the MMD errors of the selected group relative to and the remainder relative to . Simulations and real-data experiments illustrate the effect of stopping and support the predicted convergence behaviour.

Primary area:
learning theory
Keywords:
mixture separation, maximum mean discrepancy, greedy subset selection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_eAkKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers the problem of separating an original dataset sampled from a two-component mixture using a reference dataset sampled from one component. It greedily selects observations without replacement from the two-component dataset (until a stopping criterion is met) aiming to minimize the MMD to the distribution of the reference dataset. Observations from the original dataset that have not been selected, constitute the remainder subset that is representative of the other component. The paper presents an extensive theoretical analysis of how the selection of candidate instances affects approximation error and how stopping size determines remainder quality, with convergence guarantees under geometric assumptions. Experiments on synthetic data and three real datasets examine selection quality and stopping rules.

Critical strengths

S1. The main strength of the paper is theoretical analysis. It provides an analysis of greedy selection without replacement, specifying how candidate selection affects approximation quality. It establishes an exact decomposition of remainder error and provides error bounds.

S2. Empirical evidence is provided that the approach performs better in approximating the target distribution.

Critical weaknesses

W1. The contribution and novelty is not clear from an aglorithmic point of view. The method is simple combining incremental greedy MMD minimization with existing stopping procedures. Its distinction from MMD-Critic and related approaches, and why it should perform better, remain insufficiently clear. The theoretical bounds do not guide selection or stopping, while the strongest guarantees require restrictive assumptions and are developed for the energy kernel.

W2. Empirical evidence is insufficient: experiments cover only three real datasets and comparison is made with only two competitors and using only the specific kernel. Improved results in positive-distribution matching do not consistently improve remainder recovery. Moreover, An application of the method to a real problem would illustrate practical significance.

Questions for authors

Q1. To clarify the algorithmic novelty, could you briefly describe the closest greedy MMD approaches and explicitly specify their objectives, selection rules, and stopping procedures illustrating the differences to the proposed method?

Q2. What practical guidance do the theoretical bounds provide? Are they exploited in the algorithm?

Q3. Which theoretical results extend beyond the energy kernel, and how sensitive is performance to kernel choice?

Q4. Given the small number of datasets and narrow range of baselines, could you expand the experimental evaluation to establish whether the improvements generalize, particularly for approximation of the remainder?

Q5. Are there real-world applications of the proposed method?

3 comments

We thank the reviewer for the positive assessment of our analysis and experiments. We respond to each concern below.

W1 & Q1: On Algorithmic Novelty and Differences from Existing Greedy MMD Methods

We agree that the update rule is closely related to existing methods and do not claim it as our innovation. We summarize the main differences from the closest related works in Section 5 and Table 24 in Appendix G.1, which we reproduce below

  1. MMD-Critic (Kim et al., 2016) focuses on the subset selection without replacement to represent the only distribution P, which depends on the candidate pool's empirical mean as the target (not an external target).
  2. Teymur et al. (2021) and Pronzato (2023) handle external targets, but over fixed domains with replacement, and so there is no candidate-set depletion.
  3. Shao et al. (2024) selects without replacement but target the candidate pool's own empirical distribution. Their guarantees rely on favourable candidates being available at each step, rather than addressing an independent external target and the resulting mixture-distribution remainder.

In contrast, Theorem 1 establishes a path-wise error bound for arbitrary external targets in a Hilbert space, separating the initial convex hull error, removal loss, and one-step greedy approximation cost. Theorem 2 shows that, given selected-set consistency, convergence of the stopping fraction to θ\theta is both necessary and sufficient for remainder consistency. This question does not arise in representative subset selection,. Thus, our novelty lies in the theoretical characterization of an existing greedyrule in a distinct statistical setting, not the update rule itself.

W1 & Q2: On the Practical Implications of the Theoretical Error Bounds

We agree that our error bounds explain greedy selection; they are not used to choose the next point or the stopping time. Their practical value is to identify the mechanisms behind the error (depletion v. stopping), give conditions for consistency, and justify evaluating selection and stopping separately, as our experiments do.

W1 & Q3: Which Theoretical Results Extend to Other Kernels?

  1. Theorem 1 holds in an arbitrary real Hilbert space.
  2. The identity (12) and Theorem 2 use only the linearity of mean embeddings, so they hold in any RKHSs.
  3. The geometric conditions of Theorem 4 use properties of the Energy kernel feature map. We do not claim that they extend to arbitrary kernels. Empirically, we have added new experiments based on another kernel, Gaussian RBF kernel, to test the sensitivity of the kernel choice. The added content (blue text) is in the section 4 (Experiments, Page 8) and Appendix F.4 (Page 30). The experiments show the stability of our greedy strategy beyond the energy kernel.

W2 & Q4: On Baselines, Experimental Scope, Kernel, and Remainder Recovery

We have expanded the evaluation to five real-world datasets, adding Dry Bean and Spambase, and have added the Gaussian RBF kernel. The added content (blue text) is in the section 4 (Experiments, Page 8) and Appendix F.4 (Page 30). We compared Min-MMD Greedy (the same greedy strategy as EG), Kernel Herding, R-GKHR, and LabelOracle. All methods were evaluated at the same selection budget k=θN2k=\theta N_2, allowing us to isolate selection-order effects from stopping errors. The new results show that the improvement can be generalized to another kernel and the advantage holds in remainder recovery.

We do not claim that EG always gives the best remainder. Our claim is that EG selects accurately and, with accurate stopping, also recovers the remainder well. Let $ e_P =MMD ( \hat{P}_S,P) $ and $ e_Q = MMD(\hat{P}_R,Q) $. When the stopping fraction equals θ, Equation (12) implies ∣eQ−θ1−θeP∣≤MMD(M^,M)1−θ,|e_Q-\frac{\theta }{1-\theta } e_P| \leq \frac{MMD(\hat{M},M)}{1-\theta }, where $ \hat{M} $ is the empirical distribution of the candidate pool. Thus, the two errors are tied, but a smaller $ e_P $ does not mean a smaller $ e_Q $ in finite samples due to mixture-pool sampling error. This is why we report both errors rather than $ e_P $ alone.

W2 & Q5: On Real-World Applications.

Our framework fits data curation where a reliable reference sample exists, but the candidate data mix sources. For example, a factory has two production lines A and B, whose sensor measurements follow distributions PP and QQ, respectively. The factory maintains a set of verified reference samples from line A, but source labels were lost during historical data integration, making it impossible to determine which production line generated each observation. Nevertheless, production logs still provide the overall mixture proportion θ. The pooled data can thus be modeled as M=θP+(1−θ)QM=\theta P+(1-\theta )Q. Our method splits the poolinto two groups approximate PP and QQ, enabling production-line-level quality monitoring and distributional analysis without recovering individual labels.

I thank the authors for their response and the additional experiments. I remain concerned about the limited novelty of the proposed methodology relative to existing methods, its real-world impact and the practical value of the theoretical results.

We thank the reviewer for the follow-up and address the concerns as follows.

Novelty. We agree that the greedy update rule is not new and do not claim it as our contribution. Our contribution is the analysis of this rule in a setting that, to our knowledge, prior work has not combined: an external target, selection without replacement, and recovery of the unselected remainder. Teymur et al. (2021) and Pronzato (2023) consider external targets, but the set of available points stays the same at every step, so depletion does not arise. MMD-Critic and Shao et al. (2024) select without replacement, but their target is the candidate pool itself, so it always lies within the initial candidate hull, and no second distribution is recovered from the remainder. Our bounds explicitly account for depletion, i.e., the remaining candidate hull moving away from the target as candidates are removed.

Practical value. The results lead to concrete practical conclusions.

  • Theorem 2 shows that, given an accurate selected set, the remainder is consistent if and only if the stopping fraction converges to θ\theta. Selected-set error alone therefore cannot certify separation, which is why we report both ePe_P and eQe_Q in the experiments, and a stopping rule should be judged by how accurately it estimates θ\theta.
  • Equation (12) splits the remainder error into selection and stopping terms. This explains, for example, why PathMin gives a smaller remainder error than KM2 on Page Blocks with θ=0.1\theta = 0.1: it stops closer to θ\theta.

Real-world impact. The setting arises whenever a trusted reference sample exists but pooled data mix sources whose labels are lost, as in the production line example in our earlier response. In such applications, individual labels often cannot be recovered reliably where components overlap, but group-level distributions, which our paper targets, can still be recovered and suffice for downstream tasks. Following your suggestion, we added more real-world datasets for experiments; the results agree with our theoretical analysis.

Reviewer_YA5mOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies how to split a mixture sample into two groups using reference data from one component. It analyzes greedy MMD selection without replacement, which selects observations to match the reference distribution while leaving the remainder to approximate the other component. The authors derive an error bound accounting for how earlier selections deplete the candidate pool. They also show why matching the reference distribution is not enough: successful separation additionally requires selecting the right proportion of observations. Under two geometric conditions and appropriate stopping, they establish convergence rates for both groups. Synthetic and real-data experiments illustrate how different stopping points can produce substantially different remainder quality despite similar target fit.

Critical strengths
  • The paper provides a useful and novel analysis of greedy MMD selection without replacement. Beyond the general error bound, it identifies two geometric conditions that yield convergence guarantees for both groups when selection is stopped appropriately.

  • The paper clearly distinguishes matching the reference distribution from successfully separating the mixture. Its stopping analysis explains why a well-matched selected subset can still leave a poor remainder, and the synthetic and real-data experiments illustrate this point through comparisons across stopping sizes.

Critical weaknesses
  • I believe the main limitation is the gap between the theoretical guarantees and the procedure used in the experiments. Theorem 4 requires an accurate stopping fraction and either exact RKHS orthogonality or a finite-support pairing condition. However, the paper does not show that PathMin or the implemented KM1/KM2 rules consistently choose the right stopping fraction, and neither geometric condition covers the continuous overlapping mixtures used in the main experiments. While the authors acknowledge these limitations, I am not fully convinced that the theory explains why the practical procedure should reliably separate the mixture components in these settings.

  • I also find the empirical evidence for the benefit of estimated stopping somewhat mixed. For example, Table 7 shows that KM2 overestimates the mixture proportion in the Page Blocks experiments with θ=0.1\theta=0.1, and EG/KM2 has higher mean remainder error than EG/PathMin at all three sample sizes. I would have liked a clearer discussion of when estimated stopping helps and when it does not, particularly given its central role in the practical procedure.

Questions for authors
  1. In Appendix F.1, I noticed a mismatch between the sample ordering in the Gram matrix and the coefficient vector a(λ)a(\lambda). The Gram matrix lists the reference samples first, while a(λ)a(\lambda) puts the mixture coefficients first. It seems that the coefficient blocks should be swapped to match the stated ordering. Could you clarify whether this is only a notation error or whether the same mismatch is present in the implementation used to produce the experimental results?

  2. I am having trouble understanding the claimed equivalence between EG and GKHR at αH=1\alpha_H=1 in Section 5 with the displayed updates. Expanding Equation (7) and Equations (59)–(60), I obtain a target-similarity coefficient of j+1j+1 for EG after jj selections, compared with jj for GKHR. EG also includes a kernel-diagonal term that need not be constant across candidates. I do not see why choosing the candidate-pool mean as the target removes these differences. Is a different update or indexing convention intended, or should the equivalence statement be revised?

  3. The introduction claims that the dependence on departures from the projection condition is sharp, but I could not find a matching lower bound for the ζlog⁡(e/ζ)\zeta\log(e/\zeta) term in Theorem 4. The constructions in Appendix E seem to establish sharpness for the removal-loss coefficient, the inverse-square-root approximation rate, and the remainder-bound coefficients instead. Appendix G also states that the necessity of the logarithmic factor has not been established. Is this sharpness claim intended to refer only to those quantities?

3 comments

We thank the reviewer for the careful reading. All issues below have been corrected in the revision.

Q1: Ordering in a(λ)a( \lambda )

Appendix F.1 orders samples reference-first, while the displayed coefficient vector is mixture-first. The correct vector places $ (1-\lambda ) N_1^{-1} \mathbf 1_{N_1} $ before $ \lambda N_{2}^{-1}\mathbf 1_{N_2} $. The code uses consistent block ordering, so this is a notational error and has no effect on the results. Appendix F.1 has been corrected.

Q2: The Problem Between GKHR and EG

We agree that the equivalence statement in Section 5 is not correct in general. Expanding the EG update gives a kernel-diagonal term and different coefficients for target and selected-set similarity, which are absent from the GKHR update at αH=1\alpha_{H}=1. These differences do not vanish even when the target is the empirical mean of the candidate pool. We have corrected Section 5 (Page 9). Since other reviewers raised related fairness concerns, we have also added some new experiments in Section 4 (Experiments, Page 8) and Appendix F.4 (Page 30).

Q3: Sharpness

We agree with the reviewer. The introduction overstates this. Our constructions establish sharpness of the removal-loss coefficient, the worst-case $ O (1/k ) $ squared rate, and the coefficients in the remainder-error bound, but not the $\zeta \log(e/ \zeta ) $ dependence in Theorem 4(a). We have corrected the statement.

W1: Gap Between Theory and Practice

We agree that Theorem 4 does not cover the continuous overlapping mixtures used in our experiments, and that consistency is not established for KM1/KM2, whose heuristic thresholds lack the guarantees od Ramaswamy et al. (2016). The experiment of the overlapping mixture is meant to illustrate Theorems 1 and 2.

W2: When Does Estimated Stopping Help?

The observation is consistent with Theorem 2. For Page Blocks with θ=0.1\theta=0.1 and N2=400N_2=400, PathMin already stops closer to θ\theta than KM2 (0.097 v.s. 0.149) and its remainder error is correspondingly lower (0.1804 v.s. 0.1890). Theorem 2 shows that remainder accuracy depends on both selected-set error and stopping-fraction error. Therefore, an external estimator only helps if it is more accurate than PathMin.

Thanks for addressing my questions. Your clarifications resolve my technical concerns, so I’m raising my soundness score to 3. I’m keeping my overall rating at weak acceptance because the theory still doesn’t explain when the practical stopping rules will reliably separate overlapping mixtures.

We would like to clarify that for overlapping mixtures. Even when the mixture proportion is identifiable, the practical KM2 threshold rule of Ramaswamy et al. (2016), applied with the energy kernel, can have a nonzero population estimation error. To illustrate this distinction, consider $ P=\mathrm{Unif} (0,1) $, $ Q=\mathrm{Beta}(2,1) $, and $ M=0.3P+0.7Q $. Although $ P $ and $ Q $ have identical supports, $ Q $ is irreducible with respect to $ P $, since the density ratio $ q(x)p(x)=2x $ has essential infimum zero. Hence, the true mixture proportion θ=0.3 is identifiable by taking $ \theta =\lim_{h\rightarrow 0^{+}} \frac{M([0,h])}{P([0,h])} $.

For the energy kernel, the population kernel-distance function satisfies $ d(\lambda )=0 $ for $1\leq \lambda \leq \lambda ^{\ast }=10/7 $, while for $\lambda \gt \lambda ^{\ast } $, d(λ)=s5/230(1+s)3/2,s=0.7λ−1.d\left(\lambda \right)=\frac{s^{5/2}}{\sqrt{30}{\left(1+s\right)}^{3/2}},\quad \quad s=0.7\lambda -1. Although the mixture proportion is identifiable, $ d_{+}' (\lambda ^\ast )=0 $. But the practical KM2 rule uses a fixed combination of the initial and limiting slopes, sinit=dE′(1)=0,s_{\mathrm{init}}=d_{E}' (1)=0, and s∞=lim⁡λ→∞dE′(λ)=∣∣μM−μP∣∣=0.730,s_{\infty }=\lim_{\lambda \rightarrow \infty } d_{E}' (\lambda ) ={||\mu_{M}-\mu_{P}||}=\frac{0.7}{\sqrt{30}}, giving the population threshold $ \nu_{KM2}=0.14/\sqrt{30}\gt 0$. This threshold is reached at λ≈1.791, corresponding to $\theta_{KM2}\approx 0.442 $, rather than the true value 0.3. Thus, even without sampling noise, the population-level KM2 thresholding rule can exhibit a nonvanishing bias.

Reviewer_XVfjOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies greedy Maximum Mean Discrepancy (MMD) selection without replacement for mixture separation, where an independent reference sample from one component is available. It derives a pathwise error bound that accounts for candidate depletion, establishes a relationship between selected-set accuracy, stopping-fraction error, and remainder quality, and provides convergence rates under structured geometric assumptions. Experiments on synthetic and real datasets investigate different selection and stopping strategies.

Critical strengths
  • The paper provides a meaningful theoretical analysis of greedy MMD selection without replacement, explicitly characterizing how earlier selections affect the remaining candidate set.

  • The theoretical analysis distinguishes selected-set approximation from stopping accuracy and remainder quality, providing useful insights into mixture separation.

  • The paper presents detailed proofs and clearly discusses the limitations of its geometric assumptions and stopping guarantees.

Critical weaknesses
  • The proposed PathMin stopping rule underestimates the mixture fraction in several experiments, while the stronger results rely on the external KM2 estimator. The main convergence guarantees also assume accurate stopping, leaving the practical stopping problem unresolved.

  • The mixture fraction is not identifiable under the general model without additional assumptions. The paper does not establish joint conditions ensuring both consistent mixture-proportion estimation and greedy-selection convergence, limiting its end-to-end theoretical guarantees.

  • The convergence guarantees in Theorem 4 rely on restrictive geometric assumptions that do not cover the continuous overlapping mixtures used in the practical experiments. Although Fig. 2 examines a structured theoretical setting, the guarantees do not establish convergence under broader experimental conditions.

  • The energy-kernel GKHR adaptation is not translation-invariant in the same way as the proposed energy-distance greedy rule. Its kernel-based initialization and selection scores can change under common translations, making it unclear whether the reported advantages arise from the greedy strategy or differences in the underlying objectives.

Questions for authors
  • Can the authors provide an end-to-end convergence guarantee for EG combined with a practical stopping estimator such as KM1 or KM2?

  • What additional assumptions would jointly ensure identifiability of the mixture fraction and consistency of the proposed separation procedure?

  • Can the theoretical analysis be extended to a meaningful class of continuous overlapping mixtures representative of the practical experiments?

  • Do the reported advantages over GKHR persist under common translations or a translation-invariant baseline formulation?

  • Can the authors provide an intuitive illustration connecting greedy selection, convex-hull depletion, and stopping decisions?

3 comments

We thank the reviewer for the careful reading. One central theoretical contribution is to separate the error arising from candidate-set depletion from the error caused by the stopping proportion and to establish guarantees for each. We address the concerns below.

W1/Q1: On PathMin, KM1/KM2, and End-to-End Convergence.

We do not propose a new mixture proportion estimator, and we do not claim end-to-end consistency for EG with KM1/KM2. Ramaswamy et al. (2016) prove rates for theoretically calibrated thresholds, whereas KM1/KM2 use heuristic thresholds. Our guarantees are modular: any proportion estimator with a proven convergence rate can be plugged into Theorem 4 to give an end-to-end rate under the stated geometric assumptions.

W2 & Q2: On Mixture Proportion Identifiability and Joint Sufficient Conditions.

We agree that θ\theta is generally unidentifiable without additional assumptions. This is inherent to mixture proportion estimation rather than EG itself. A classical condition is the irreducibility of QQ with respect to PP, under which θ\theta is uniquely determined by PP and MM [1].

We provide a fully data-driven finite-support example. Suppose PP has finite support AA with pmin⁡>0p_{\min}> 0, Q(A)=0Q\left(A\right)=0, and Assumptions 1 and 3 hold. Let TN1T_{N_{1}} denote the distinct values observed in the reference sample, and define θ~=1N2∑j=1N21[zj∈TN1].\tilde{\theta}=\frac{1}{N_{2}} \sum_{j=1}^{N_{2}} \mathbf{1}[z_j \in T_{N_1}]. Since Pr⁡(TN1≠A)≤∣A∣e−N1pmin⁡,\Pr{\left(T_{N_{1}}\neq A\right)}\leq \left|A\right|e^{-N_{1}p_{\min}}, $ \tilde{\theta} $ equals the empirical fraction of PP-candidates with probability tending to 1. Consequently, ∣θ~−θ∣=OP(N2−1/2).\left|\tilde{\theta }-\theta \right|=O_{P}\left(N_{2}^{-1/2}\right). After rounding and truncating N2θ~N_{2}\tilde{\theta } as in Eq. (35), Theorem 4(b), together with the exact greedy rule (7), yields MMD(P^S,P)+MMD(P^R,Q)=OP(N1−1/2+N2−1/2).MMD(\hat{P}_{S},P)+MMD(\hat{P}_R,Q)=O_P (N_1^{-1/2}+N_2^{-1/2}). This is an explicit end-to-end consistency guarantee, and covers the four-point model in Appendix D.3.

[1] Scott C. A rate of convergence for mixture proportion estimation, with application to learning from noisy labels. Artificial Intelligence and Statistics. PMLR, 2015: 838-846.

W3 & Q3: On Geometric Assumptions and Continuous Overlapping Mixtures.

We agree that the conditions in Theorem 4 are sufficient, not necessary. Our results sit at three levels of generality. Theorem 1 needs no geometric assumption. Theorem 2 shows that given selected-set consistency, remainder consistency is equivalent to stopping-fraction consistency. Theorem 4 gives rates under two explicit geometric conditions. Our overlapping-mixture experiments are intended to illustrate the general error decomposition in Theorems 1 and 2, rather than verify the assumptions of Theorem 4.

Importantly, overlap does not preclude identifiability. In our Gaussian experiments, P=N(mP,I2)P=\mathcal{N}\left(m_{P},I_{2}\right) and Q=N(0,I2)Q=\mathcal{N}\left(0,I_{2}\right), with mP≠0m_{P}\neq 0. The supports coincide, yet QQ is irreducible with respect to PP, so θ\theta is identifiable. Since EM[Z]=θmPE_M [Z]=\theta m_P the moment estimator θ^=y‾⊤z‾∥y‾∥2,θ^−θ=OP(N1−1/2+N2−1/2).\hat{\theta }=\frac{{\overline{y}}^{\top }\overline{z}}{{\left\lVert \overline{y}\right\lVert }^{2}},\quad \quad \hat{\theta }-\theta =O_{P}\left(N_{1}^{-1/2}+N_{2}^{-1/2}\right). where $ \bar{y}=N_{1}^{-1}\sum_{i=1}^{N_{1}} y_{i} $ and zˉ=N2−1∑j=1N2zj\bar{z}=N_{2}^{-1}\sum_{j=1}^{N_{2}} z_{j} are the sample means of the reference and mixture samples, respectively. The rate of θ^−θ\hat{\theta }-\theta can be used in place of rN1,N2r_{N_{1},N_{2}} into Theorem 2.

W4 & Q4: Translation Invariance and Baseline Fairness.

We agree that the energy-kernel GKHR may depend on the coordinate origin. To address this, we add new experiments of common translations to test the energy-kernel GKHR and conduct additional experiments using the translation-invariant Gaussian RBF kernel on five real-world datasets (Page Blocks, Shuttle, Wine Quality, Dry Bean, and Spambase). The added content (blue text) is in the section 4 (Experiments, Page 8) and Appendix F.4 (Page 30).

Q5: On the Intuitive Connection Between Greedy Selection, Convex-Hull Depletion, and Stopping.

We thank the reviewer for this suggestion. Proposition 12 in Appendix E illustrates depletion through a simple example: let u=0u=0 and x1=0, x2=⋯=xn=1x_{1}=0,\ x_{2}=\cdots =x_{n}=1. Initially, d0=0d_{0}=0, and greedy selection achieves zero error by selecting x1x_{1}. However, the remaining convex hull immediately shrinks to 1{1}, so err⁡k=k−1k.\operatorname{err}_{k}{=}\frac{k-1}{k}. Thus, a perfect start does not prevent depletion loss.

Stopping is a separate issue. Even an accurate selection path must stop near θ\theta to recover QQ and Equation (12) splits the remainder error into mixture sampling, selected-set and stopping-fraction errors. This is why we analyse selection and stopping jointly rather than minimizing selected-set MMD alone.

I thank the authors for their detailed and thoughtful responses. The clarifications on the theoretical assumptions, the finite-support consistency example, and the additional experiments have addressed several of my concerns. Additionally, I suggest including an overview figure to better illustrate the proposed framework and improve the clarity of the presentation. Although some limitations remain, I appreciate the authors' efforts and have decided to raise my rating.

We thank the reviewer for your appreciation. And we have added an illustration Figure 3 in appendix on Page 14 for the clarity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.