acceptodds
Under review as a conference paper at ICLR 2027

Penalized-Weight Robust Optimal Transport: From Hard Clipping to Soft Compression

Abstract

Robust optimal transport (OT) methods reduce the influence of outliers, and several produce point-level outlier flags, but we know of no guarantee that these flags recover the contaminated observations. We introduce PW-ROT (Penalized-Weight Robust OT), which assigns each transport path a weight and penalizes linearly. Eliminating the weights in closed form yields ordinary OT with a bounded ground-cost family , indexed by a threshold and a power : is ROBOT's hard-clipped cost (up to its conventional factor of two in the threshold), and gives smooth soft-compression costs. In a general parametric model, we show that the bounded cost down-weights separated outliers automatically: although no data are discarded and the outlier fraction is not an input, the PW-ROT criterion matches, up to a constant, a criterion that removes that fraction of mass from the fitted model (not from the data), exactly for the hard clip () and up to an explicit error for . We bound the estimation error under arbitrary outliers, prove consistency of the hard clip with a fixed fraction of separated outliers in bounded-support models, and show that the flags recover the outlier set exactly with high probability whenever the estimation error is small relative to a separation margin. For observations with Gaussian inliers in of variance and a point target, exact recovery holds for every member, including ROBOT, with a fixed fraction of outliers separated at the same scale, when ; slower growth provably flags inliers for any estimator. The power governs a trade-off: soft powers down-weight rather than discard outliers beyond the threshold, so their weights vary continuously with the cost, but separated outliers then bias the estimate toward themselves, which the hard clip avoids. Empirically, hard clipping is best for one-sided, well-separated outliers; soft compression has lower bias near the threshold and more stable threshold selection on images; and is the most consistent default.

Primary area:
learning theory
Keywords:
Robust optimal transport, outlier detection, exact outlier recovery, penalized weights, bounded cost, minimum-distance estimation, contamination model, model-side trimming, robust statistics, ROBOT

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_5TahOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

By extending ROBOT, this paper proposes an OT-based approach to outlier detection. The outlier model is O∩I\mathcal O \cap \mathcal I-model, where the inlier points I\mathcal I follows an underlying law while the outlier points O\mathcal O are arbitrary. The basic idea is the joint optimization of minimum distance estimation and outlier flagging. The outlier flagging (ww-optimization) in PW-ROT can be written in closed form in Theorem 3.1, which leads to the standard OT with a transformed cost ρ∘c\rho\circ c. After solving this, the outlier set is estimated by enumerating the points xix_i that is not "reachable" from other points with high probability under the estimated model θ^n\hat\theta_n. The main theoretical contribution is the exact recovery in Theorem 4.2, where the estimated outlier set is exact under some margin conditions.

Critical strengths
  • The extension of ROBOT with the exponent rr is clear. The exponent rr originates from the intensity of the outlier flagging weights ww in Eq. (1). This translates into the cost clipping function ρ\rho in Eq. (2). In Figure 1, we can clearly see that rr softens the hard clipping r=1r=1. In Section 4.2, the authors discuss that soft r>1r>1 may be informed by inliers that are "false positive" (= falsely detected as outliers), while its theoretical benefit has yet to be formalized.
  • Fairly strong outlier guarantee since outlier points can be arbitrary.
  • Not quite a lot of hyperparameters. Given that the exponent rr is a fixed hyperparameter, the only hyperparameter that a user needs to choose is λ\lambda, which is the clipping parameter Eq. (1). Theoretically, this should be scaled with nn as shown in Theorem 4.2. Empirically, the authors use a trimmed validation strategy to choose λ\lambda minimizing the validation squared error after discarding 35% of outlier-like points.
    • This "35%" value may be a bit controversial because all the true outlier proportions tested in the experiments are less than 35%, which means that we somewhat have the knowledge that the outlier proportion is no larger than 35%. It might be okay to choose this 35% criterion large enough so that the validation error is not affected by outliers, while the statistical efficiency is lost. Anyway, I don't think there is a sufficient clean solution to this issue (other existing methods have a similar issue more or less), so I would not argue it so significantly.
Critical weaknesses
  • The paper is relatively hard to read. For example, there is a long paragraph after Theorem 4.2 to discuss the implication of this theorem but without a clear paragraph structure. Usually we prefer to have a well-separated paragraph structure. In this case, we may consider to split this paragraph into three pieces to discuss (a), (b), and (c), respectively, at least. The same issue persists in appendix too. For example, the proof of Lemma D.5 is compressed significantly. Since this lemma has five arguments, we should at least split into several paragraphs to make it easier to read.
    • The above is my suggestion to the authors but not the basis of paper assessment.
  • The theoretical contributions mainly consist of the two parts, the estimation error rate (Theorem 4.2(a)) and the exact recovery of the outlier set (Theorem 4.2(b)). I read the overall flow of the proofs, which looks valid while the assumptions looks a bit strong.
    • Theorem 4.2(a): Roughly speaking, the main piece of this claim is the rate estimation of M(θ)=J(θ)−J0M(\theta)=J(\theta)-J_0, which is a consequence of the Lipschitz continuity shown in Lemma D.4. Given that we consider the Lipschitz cost ρ∘c\rho\circ c and OT formulation, the Lipschitz continuity is natural. Also, the boundedness of ρ∘c\rho\circ c is crucial to apply McDiarmid's bound. Then the estimation error bound follows. Additionally, the bias rate on ∣∣θ^n−θ0∣∣||\hat\theta_n - \theta_0|| comes from the strong convexity assumption of MM (or by the modulus of continuity, in more general), which is a standard assumption. Overall, the flow is clear, and I think the claim is valid, but it is not very surprising (it is okay).
    • Theorem 4.2(b): The crucial assumption is that inlier points are included in the margin set Gλ(t)\mathcal G_\lambda(t) with high probability, while outlier points are included in Bλ(t)\mathcal B_\lambda(t). These margin sets are defined by whether νθ(Aλ(x))\nu_\theta(A_\lambda(x)) (model probability of the reachable points from xx) is large/small enough. Essentially, this can be read as a tautology to some extent. While the authors discuss the validity of the assumptions after Theorem 4.2, my concern on this point has not been cleared.

If the authors can adequately clarify the assumption validity on Theorem 4.2(b), I am ready to reconsider the evaluation.

Questions for authors
  • In Figure 1, why is rρr\rho is plotted instead of ρ\rho itself? In my opinion, we should plot ρ\rho so that its boundedness behavior becomes clear in the figure.
  • At some point, we should have a notation table since the paper has a lot of notations with slight variants.
  • (Very minor) In L.257, it looks better to write tn(δ):=...t_n(\delta) := ... to make it clear that this is the definition.
  • In Section 4.2, it is rather unclear what "pull" does mean; although I could guess the overall implication of this subsection.
  • In Lemma D.5, it looks θ†\theta^\dagger is not clearly defined. Probably it suggest the well-specified case as discussed in Remark D.1, which should be made clearer.
2 comments
AuthorsOpenReview ↗

Assumption in Theorem 4.2(b)

Thank you for the careful reading and for identifying the key condition underlying our theoretical guarantee. We agree that the margin-set condition needed a more direct explanation. We address this concern in three parts, with all changes included in the revised PDF (new Appendix D.6, highlighted in blue).

1. What the condition means and why it is not circular.

The margin sets Gλ(t)G_\lambda(t) and Bλ(t)B_\lambda(t) depend only on the model family, λ\lambda, the flag level α\alpha, the clean parameter θ0\theta_0, and the margin tt. They do not depend on the data, outliers, or estimator. For both targets, these sets can be bounded by explicit balls around θ0\theta_0 (Remark D.10 and Lemma D.16). Thus, requiring inliers to lie in Gλ(t)G_\lambda(t) is a tail condition on the inlier distribution, while requiring outliers to lie in Bλ(t)B_\lambda(t) is a separation condition.

Theorem D.11(i) is deterministic once these two events hold. The role of Theorem 4.2(b) is to establish them through (a) the estimation bound in Theorem 4.2(a), which places θ^n\hat{\theta}_n within distance tt of θ0\theta_0, uniformly over arbitrary outliers, including inlier-dependent ones; and (b) the requirement that none of the nn inliers is flagged, which forces λn≳σ2(d+log⁡n)2\lambda_n \gtrsim \sigma^2(\sqrt{d}+\sqrt{\log n})^2. This threshold growth is necessary, not an artifact of the proof: no estimator can avoid flagging an inlier below this scale (Theorem 4.2(c); Remark D.18 for the spread target).

2. Explicit conditions for the Gaussian spread model.

For the Gaussian spread target νθ=θ+N(0,Id)\nu_\theta=\theta+N(0,I_d) and the 95% rule in Section 5, Corollary D.17 replaces the abstract margin-set condition with three explicit conditions independent of the estimator:

(I) λ≥t+qα+d+2log⁡(n/δ)\sqrt{\lambda}\geq t+\sqrt{q_\alpha}+\sqrt{d}+\sqrt{2\log(n/\delta)}, controlling the inlier tail probability.

(S) Every outlier is farther than λ+z1−α+t\sqrt{\lambda}+z_{1-\alpha}+t from θ0\theta_0, ensuring separation.

(F) A bound on the outlier fraction and inlier fluctuation, which determines tt.

We also establish the curvature required by Assumption 3, which the original submission left to be verified for this target. Lemma D.15 gives κ=p12/r\kappa=p_1^2/r and t0=λ/2−1t_0=\sqrt{\lambda}/2-1, independently of dd.

Consequently, with

λn=4(2d+2log⁡(n/δ))2,\lambda_n=4(2\sqrt{d}+\sqrt{2\log(n/\delta)})^2,

outliers at distance at least 1.5λn+z1−α1.5\sqrt{\lambda_n}+z_{1-\alpha}, and

ε<p128r≈0.058r,\varepsilon<\frac{p_1^2}{8r}\approx\frac{0.058}{r},

the procedure recovers the outlier set exactly with probability at least 1−2δ1-2\delta, at the same threshold scale as the point target.

We acknowledge that these constants are asymptotic and conservative. This corollary does not cover the finite-sample experiments in Section 5 (ε\varepsilon up to 0.3), and its admissible contamination level is substantially below the point-target value of 1/21/2.

3. The conditions are not vacuous: they predict where recovery fails.

We provide empirical evidence in Appendix F.1 (Tables 8 and 9). In the mean-shift experiment (n=300,d=5,ε=0.2n=300,d=5,\varepsilon=0.2, one-sided), the flag radius R(λ∗)R(\lambda^*) at the selected threshold predicts both swamping and masking. Assuming a perfect parameter estimate, the predicted probabilities are

P(χd2>R2),P(χd2(nc,k2d)≤R2),P(\chi_d^2>R^2),\qquad P(\chi_d^2(nc\\,k^2d)\leq R^2),

respectively.

For r=2r=2, predicted versus observed swamping rates are 0.27 versus 0.31 at k=2k=2, and 0.10 versus 0.12 at k=4k=4. The corresponding masking rates are 0.63 versus 0.69 at k=1k=1, and 0.84 versus 0.85 at k=0.5k=0.5. Across-trial correlations range from 0.83 to 0.99 for swamping and for masking when k≤1k\leq1. ROBOT and r=1.5,3r=1.5,3 show similar agreement, with prediction errors within 0.07.

Swamping at k=2k=2 reflects a violation of (I): the selector optimizes estimation error and chooses R2≈6.5R^2\approx6.5, well below the approximately 17 needed to avoid flagging any inlier (the 1−1/N1-1/N quantile of χ52\chi_5^2, with N=240N=240). The necessary scale max⁡2log⁡N,d≈11\max\\{2\log N,d\\}\approx11 is only a lower bound; even at R2=11R^2=11, approximately 5% of inliers would still be flagged. Masking at k≤1k\leq1 reflects a violation of (S), since the outliers lie inside the flag radius.

Consistent with this explanation, exact recovery becomes more frequent as separation increases. At the data-selected threshold, ROBOT recovers the outlier set exactly in 0/50 runs for k≤2k\leq2, 5/50 at k=3k=3, 22/50 at k=4k=4, and 46/50 at k=6k=6.

We report these results as limitations of the experiments, not as empirical validation of the theoretical constants.

(1) Theorem 4.2(a): "not very surprising."

We agree and have clarified this point in the revision. Theorem 4.2(a) provides the supporting estimation bound. The transport-specific contributions are the inlier-sampling term, governed by the OT rate n−1/dn^{-1/d} through EW1\mathbb{E}W_1, and the fact that outliers enter the bound only through λε\lambda\varepsilon.

(2) Section 4.2 and the benefit of soft powers.

We agree that this benefit is not formally established. We have revised the abstract and contribution 3 to state only the proven trade-off: softer compression comes at the cost of a bias bound that increases with rr. Contribution 4 now clarifies that the experiments motivate, but do not theoretically establish, the use of a soft power in the range 1.5≤r≤21.5\leq r\leq2 (we use r=2r=2) when the contamination geometry is unknown and the threshold must be selected from data.

(3) The 35% trimming level.

We appreciate the reviewer raising this point. We now report a sensitivity analysis in Appendix F.1 (Table 7). When ε=0.2\varepsilon=0.2, the results are nearly identical (within numerical variation) for ζ∈0.25,0.35,0.45\zeta\in\\{0.25,0.35,0.45\\}. However, when the contamination level exceeds the trimming level (ε=0.4\varepsilon=0.4), the soft powers can degrade. For example, at r=2r=2 and k=4k=4, the error is 0.43 for ζ=0.25\zeta=0.25 and 0.22 for ζ=0.45\zeta=0.45, while ROBOT is insensitive to this choice. This confirms that an upper bound on the contamination level is needed, as the reviewer anticipated.

(4) Presentation (all changes completed and highlighted in blue).

  • The paragraph following Theorem 4.2 is now divided into (a) Estimation, (b) Exact Recovery, (c) Necessity, and the threshold window. The proof of Lemma D.5 is also reorganized into one paragraph per part, and θ†\theta^\dagger is now defined as the well-specified case of Remark D.1.
  • Figure 1 now plots ρ\rho itself, making its boundedness at λ\lambda visible.
  • tn(δ):=…t_n(\delta):=\ldots is now explicitly written as a definition.
  • The term "pull" is defined at its first occurrence as a nonzero contribution to the gradient of the criterion that moves the estimate toward the point.
  • A notation table has been added in Appendix I, along with a short algorithm overview (Algorithm 1) in the main text.
Reviewer_w56nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper considers robust optimal transport in the presence of outliers and proposes PW-ROT, which assigns a weight to each transport path. The authors show that optimizing these weights yields a family of bounded ground costs that includes the hard-clipping cost of ROBOT as a special case and allows soft compression through an additional parameter rr. They connect the proposed method to trimmed OT, derive estimation-error bounds under contamination, and provide sufficient conditions for exact outlier recovery. In particular, for Gaussian mean estimation, they characterize the threshold required for exact recovery and quantify the additional bias introduced by soft compression. Finally, the authors present simulation studies and experiments on contaminated MNIST data to illustrate the proposed method's performance under different contamination settings.

Critical strengths
  1. The paper is well written and easy to follow (although some elaboration in the theory section would be more helpful; see the weakness section for details).

  2. The idea is interesting, and I like the observation that ROBOT and related robust OT problems can be unified by optimizing the path-specific weight ww.

Critical weaknesses
  1. The choice of (λ,r)(\lambda,r) is not entirely clear to me. Although the authors provide theoretical guidance on the threshold λ\lambda for exact outlier recovery and recommend r=2r=2 based on numerical experiments, it is unclear how a practitioner should choose these parameters jointly. For example, suppose I know an upper bound on the contamination proportion εn\varepsilon_n. Can I use this information to choose (λ,r)(\lambda,r) in a principled way? It would be helpful if the authors could provide more guidance on this issue.

  2. As the authors have pointed out, when the inliers are generated from a parametric model, say f(x;θ∗)f(x;\theta_*), the population minimizer satisfies θ0=θ∗\theta_0=\theta_* for every fixed λ>0\lambda>0 and r≥1r\ge1. However, the estimation error and statistical efficiency (limiting variance) should still depend on the choice of (λ,r)(\lambda,r). Although Corollary 4.3 provides an explicit estimation-error bound for Gaussian mean estimation with a point target, it is not clear how this trade-off behaves for more general parametric models. In particular, can the authors characterize how the estimation rate and constants depend on (λ,r)(\lambda,r), and whether this characterization can be used to select these parameters?

  3. Essentially a follow-up of the previous point: I also wonder whether the authors can provide some results on statistical inference. In particular, how does the estimator's asymptotic variance depend on (λ,r)(\lambda,r) in the parametric model? Is the estimator asymptotically normal at a certain rate depending on (m,n)(m, n)?

Questions for authors

Kindly see the weakness section.

2 comments
AuthorsOpenReview ↗

choosing (λ, r) given an upper bound ε̄ on the contamination (W1)

Thank you. This is the practical question the paper should answer, and we now clarify what the theory supports and what it does not.

  • The trim level uses εˉ\bar{\varepsilon} directly. The only contamination information required by the selection rule is that the trimming level ζ\zeta of the validation error exceeds the contamination level (Section 5). Thus, setting ζ=εˉ\zeta = \bar{\varepsilon} is the principled use of the bound. Appendix F.1 (Table 7) shows that when ε<ζ\varepsilon < \zeta, the results remain unchanged within noise as ζ\zeta varies from 0.25 to 0.45, whereas the soft powers degrade when ζ<ε\zeta < \varepsilon.

  • The threshold λ\lambda lies in a window. Its lower bound is determined by the requirement that no inlier be flagged: λ≳σ(d+2log⁡(n/δ))\sqrt{\lambda} \gtrsim \sigma(\sqrt{d} + \sqrt{2\log(n/\delta)}), or equivalently, λ≳σ2(d+log⁡n)2\lambda \gtrsim \sigma^2(\sqrt{d} + \sqrt{\log n})^2. This lower bound is unavoidable, since below σ2max⁡{2log⁡n,d}\sigma^2\max\{2\log n,d\}, some inlier is flagged (Theorem 4.2(c); Remark D.18 for the spread target). Its upper bound is determined by bias: λnεn→0\lambda_n\varepsilon_n \to 0 for a fixed margin (Theorem 4.2(a)). For a point target with Gaussian inliers and a fixed contamination fraction εˉ<1/2\bar{\varepsilon} < 1/2, an explicit choice is λn=cλ2τn2\lambda_n = c_\lambda^2\tau_n^2, where cλ>2c_\lambda > 2 and cλ2>2(1−εˉ)/[r(1−2εˉ)]c_\lambda^2 > 2(1-\bar{\varepsilon})/[r(1-2\bar{\varepsilon})] (Corollary 4.3; Corollary E.8), provided the outliers are farther than CDτnC_D\tau_n. The condition involving εˉ\bar{\varepsilon} and rr is binding only when εˉ>1/3\bar{\varepsilon} > 1/3 for r=1r=1, εˉ>3/7\bar{\varepsilon} > 3/7 for r=2r=2, and εˉ>5/11\bar{\varepsilon} > 5/11 for r=3r=3; otherwise, cλ>2c_\lambda > 2 is the binding condition. We emphasize that this is an asymptotic and relatively large threshold scale: for n=300n=300 and d=5d=5, it gives λn>205\lambda_n > 205, whereas the thresholds selected in our experiments have medians between 3.4 and 12. In practice, we therefore select λ\lambda by trimmed validation with ζ=εˉ\zeta = \bar{\varepsilon} and use the theoretical window to explain what this selection can and cannot achieve (Appendix F.1, Table 9).

  • The power rr trades efficiency against bias, and we do not claim a joint selection rule. For efficiency at a given threshold, under clean Gaussian inliers and a point target, every soft power is at least as efficient as hard clipping at the same λ\lambda (Proposition E.14). In d=5d=5, the thresholds required for 95% efficiency are approximately λ=14.1\lambda=14.1 for r=1r=1, and 9.9, 8.8, and 7.85 for r=1.5r=1.5, 2, and 3, respectively (Table 10). This provides a Huber-style rule for choosing λ\lambda based on a target efficiency. For bias from separated outliers, when r>1r>1, the bias has an attained order, and consistency holds if and only if λn=o(Dn(r+1)/r)\lambda_n=o(D_n^{(r+1)/r}) (Corollary 4.3; Proposition E.11). Thus, a larger rr requires greater separation. Our practical guidance, now included in Section 5 and Algorithm 1, is to use r=1r=1 when outliers are expected to be one-sided and well separated, or when the dimension is large. When the geometry is unknown and the threshold must be selected from data, we recommend a soft power between 1.5 and 2 (we use r=2r=2). Neither choice is the worst in point error in more than 1 of 24 cells, and each requires the fallback selector in at most 1 of 20 image runs (Section 5.2; Table 18). We do not select rr from data (Limitations), and the contamination bound εˉ\bar{\varepsilon} alone does not determine rr.

  • Rates for a general parametric model (W2). Theorem 4.2(a) gives ∥θ^n−θ0∥≤(2λεn+2tn(δ))/κ\|\hat{\theta}_n-\theta_0\| \leq \sqrt{(2\lambda\varepsilon_n+2t_n(\delta))/\kappa}, where tn(δ)=E[Tn]+λlog⁡(1/δ)/(2(n−m))t_n(\delta)=\mathbb{E}[T_n]+\lambda\sqrt{\log(1/\delta)/(2(n-m))}, E[Tn]≤LE[W1]\mathbb{E}[T_n]\leq L\mathbb{E}[\mathcal{W}_1] with L=2λ/rL=2\sqrt{\lambda}/r, and κ=κ(λ,r)\kappa=\kappa(\lambda,r) is the curvature constant in Assumption 3. Thus, λ\lambda enters through the bias term λεn\lambda\varepsilon_n and the sampling term λ/r\sqrt{\lambda}/r, while rr enters through 1/r1/r in LL and through κ\kappa. The curvature constant is explicit in the two cases we study: for the Gaussian spread target, κ=p12/r\kappa=p_1^2/r, where p1=2Φ(1)−1p_1=2\Phi(1)-1 (Lemma D.15); for the point target, it is given by the curvature in Proposition E.3 and Lemma E.4. This result is an upper bound, and we do not use it to select (λ,r)(\lambda,r). We do not establish a limiting variance for general parametric models. A central limit theorem for minimum-distance estimators based on a robust transport cost in a general parametric model would require central limit theorems for (semi-discrete) optimal transport, which are beyond the scope of our results.

  • Variance for the point target (W2, W3). Proposition E.14 gives an exact expression. For clean Gaussian inliers, n(μ^n−μ0)→dN(0,v(λ,r,d)Id)\sqrt{n}(\hat{\mu}_n-\mu_0)\xrightarrow{d}N(0,v(\lambda,r,d)I_d), where v(λ,r,d)=dE[CWr(C)2]/(E[CWr(C)])2v(\lambda,r,d)=d\mathbb{E}[CW_r(C)^2]/(\mathbb{E}[CW_r(C)])^2 and C∼χd2C\sim\chi_d^2. This yields three consequences. First, v≥1v\geq 1. Second, for ROBOT, v=1/Fd+2(λ)v=1/F_{d+2}(\lambda), so the thresholds corresponding to 95% efficiency are the χd+22\chi_{d+2}^2 quantiles: 7.81, 14.07, and 33.92 for d=1,5,20d=1,5,20, respectively. Third, v(λ,r,d)≤v(λ,1,d)v(\lambda,r,d)\leq v(\lambda,1,d) for every r>1r>1. At the threshold that leaves 5% of clean points beyond it, ROBOT achieves efficiencies of 0.72, 0.86, and 0.91 for d=1,5,20d=1,5,20, respectively, compared with 0.88--0.997 for the soft powers. Monte Carlo results agree with the theoretical variance vv within 1.2% at 10 of 14 settings and within 4% at all settings. The largest deviations occur for hard clipping at low thresholds and decrease approximately at the rate n−1/2n^{-1/2}. For example, for ROBOT with λ=5\lambda=5 and d=5d=5, the relative deviations are −6.7-6.7\\%, −3.1-3.1\\%, and −2.4-2.4\\% at n=500n=500, 20002000, and 80008000, respectively. Thus, vv describes the limiting variance as n→∞n\to\infty. Asymptotic normality also holds for ROBOT; the difference across powers is in their asymptotic variance.

  • With outliers (W3, the role of mm and nn). Suppose dd, σ\sigma, rr, and ε<1/2\varepsilon<1/2 are fixed, m=εnm=\varepsilon n, and λn\lambda_n, τn\tau_n, and DnD_n satisfy the conditions of Corollary E.8, with Dn≥CDτnD_n\geq C_D\tau_n. Theorem E.7 gives an exact decomposition of the estimation error, on a high-probability event, into an inlier-noise term and an outlier term (Remark E.9). For r=1r=1, the outlier term is zero. Consequently, n−m(μ^−μ0)\sqrt{n-m}(\hat{\mu}-\mu_0) is within total variation distance 1/n1/n of N(0,σ2Id)N(0,\sigma^2 I_d). This is the oracle variance based on the n−mn-m inliers, corresponding to variance σ2/(1−ε)\sigma^2/(1-\varepsilon) on the n\sqrt{n} scale, uniformly over outlier configurations. For r>1r>1, the same limit holds if nε′λn(λn/(Dn−bn))(r+1)/(r−1)→0\sqrt{n}\varepsilon'\sqrt{\lambda_n}\left(\sqrt{\lambda_n}/(D_n-b_n)\right)^{(r+1)/(r-1)}\to 0. A sufficient separation condition is Dn≥nγD_n\geq n^\gamma with γ>(r−1)/(2(r+1))\gamma>(r-1)/(2(r+1)), which requires γ>1/6\gamma>1/6 for r=2r=2 and γ>1/4\gamma>1/4 for r=3r=3. This condition is stronger than Dn≫λnr/(r+1)D_n\gg\lambda_n^{r/(r+1)}, which is sufficient for the bias merely to vanish. For one-sided outliers, the stronger condition cannot generally be dropped because the bias bound is attained (Proposition E.11). These results assume that λ\lambda is fixed in advance; they do not cover the data-selected λ\lambda in Section 5 or the spread target used in the numerical experiments. For plug-in confidence intervals, one can evaluate v(λ^,r,d)v(\hat{\lambda},r,d) using a robust estimate of σ\sigma.

Reviewer_k9cuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies robust optimal transport (OT) with outliers. It aims to estimate a clean model and identify outliers without knowing their fraction. PW-ROT adds penalized weights to transport paths. Eliminating them gives bounded costs covering ROBOT's hard clipping (r=1) and soft compression (r>1). The paper proves estimation bounds and exact-recovery conditions. Tests on synthetic data and contaminated MNIST show gains for soft costs in some settings, but not consistent gains over baselines.

Critical strengths
  1. The paper provides theoretical bounds on parameter estimation error under outlier contamination and establishes sufficient conditions for exact recovery of the outlier set.

  2. The paper proves that all outliers can be identified with high probability under Gaussian inliers and a fixed fraction of sufficiently separated outliers. It also shows that hard clipping can recover the oracle inlier mean.

Critical weaknesses
  1. The novelty over ROBOT (ICML 2021) is unclear. The r=1 case recovers ROBOT, and eliminating latent weights is a known technique. The paper should state what is new beyond extending the cost family.

  2. The exact-recovery guarantee needs Gaussian inliers and outlier separation of order σ(√d+√log n). It does not cover nearby or overlapping outliers. The author should emphasize the scope of applicability and limitations for this assumption.

  3. The presentation is hard to follow. The links among path weights, model-side trimming, general theory, and point-target results are unclear. Algorithms 1 and 2 are only in Appendix H; the main text needs a short algorithm overview.

  4. Soft compression does not consistently beat prior methods. In Table 4, r=2 has F1=0.62 versus 0.69 for ROBOT and 0.79 for trimmed OT at k=2. The image tests use only two settings and three seeds.

  5. The theory covers an exact minimizer with a fixed threshold, while the experiments use a stochastic solver and data-selected thresholds. Moreover, the validation procedure uses a 35% trimming level, chosen above the maximum tested contamination of 30%. It is unclear how this level should be selected without prior knowledge of the contamination level.

Questions for authors
  1. For a point target, PW-ROT reduces to a robust M-estimator. Which results need OT rather than classical M-estimation?

  2. Why does soft compression help on MNIST? The final flag is still a threshold test. Can the authors separate the effect of the loss from threshold tuning and compare with other detectors?

  3. How close does Algorithm 1 get to the exact minimizer assumed in the proofs? Appendix H also seems to use different training/validation splits for PW-ROT and the baselines. Please clarify.

5 comments
AuthorsOpenReview ↗

Novelty (W1) and What needs OT (Q1)

Thank you. We agree that the novelty should be stated narrowly, and we have clarified this in the response below and added a new paragraph to Appendix A.

What is not new: eliminating a latent weight is the half-quadratic construction of Geman–Reynolds and is also used in self-paced learning. For r=1r=1, our family contains ROBOT. For a point target, PW-ROT is an M-estimator (Proposition E.2); when r=1r=1, it reduces to the hard-rejection estimator of Hinich and Talwar. The paper already acknowledges the first two connections in Section 2. Appendix A now also cites the Hinich–Talwar estimator and clarifies which ingredients are classical.

What we claim is new, in order of strength:

1. Exact finite-sample recovery of the outlier set for a fitted estimator in a general parametric model (Theorem 4.2(b)). ROBOT's authors identify outlier-detection guarantees as future work, while existing partial-OT results concern a fixed pair of measures. Theorem 4.2(c) further shows that the threshold must scale as σ2(d+log⁡n)2\sigma^2(\sqrt{d}+\sqrt{\log n})^2 for every estimator, establishing that our sufficient threshold is sharp for the point target. Remark D.18 provides the corresponding necessity result for the spread target. We have clarified the scope of this claim using the phrase "finite-sample, estimator-level" in the abstract and Appendix D.

2. Model-side trimming at the true, unknown contamination fraction (Proposition 4.1). For a general parametric model, our criterion is equivalent to trimming the fitted model at fraction ϵ\epsilon, exactly when r=1r=1 and with an explicit gap when r>1r>1. This mechanism is known for a fixed pair of measures (Nietert et al.). Our contribution is establishing this result for minimization over a parametric family.

3. Quantifying the price of softness. For r>1r>1, we establish an attained order for the bias caused by separated outliers and show that consistency holds if and only if λn=o(Dn(r+1)/r)\lambda_n=o(D_n^{(r+1)/r}) (Corollary 4.3 and Proposition E.11). This quantifies the trade-off between hard and soft rejection, rather than simply introducing another cost family.

Q1: What requires optimal transport? For a point target, Corollary 4.3 and the oracle-mean result for r=1r=1 are M-estimation and hard-rejection results, respectively, and we explicitly acknowledge this connection. Three aspects are specific to transport and become relevant when the model is not a point mass, such as the Gaussian spread target used in our experiments or a generative model:

(i) The criterion is not a sum of per-observation losses. Proposition 4.1 is therefore needed to characterize the contribution of outliers.

(ii) The inlier-sampling error in Theorem 4.2(a) is governed by the OT rate EW1≍n−1/d\mathbb{E}\mathcal{W}_1\asymp n^{-1/d}, which has no direct analogue in standard M-estimation.

(iii) The outlier-flagging rule uses the full-weight reachable mass of the transport plan, νθ(y:c(x,y)≤λ)\nu_\theta(\\{y:c(x,y)\leq\lambda\\}). This reduces to a residual-based test only for a point target.

Corollary D.17 illustrates the third aspect by establishing exact recovery for the spread target, with all required conditions explicitly stated in terms of the data distribution and the outliers.

AuthorsOpenReview ↗

Scope of exact recovery (W2)

We agree and have made the scope of our results and their limitations explicit in the revised Limitations section (highlighted in blue).

- Exact recovery requires outliers to be separated from the inliers by at least order σ(d+log⁡n)\sigma(\sqrt{d}+\sqrt{\log n}). It does not cover nearby or overlapping outliers. This separation is necessary for our estimator and flagging rule: an outlier inside the region that is never flagged remains unflagged whenever the estimated model is within the margin (Appendix D.6).

- The estimation bound (Theorem 4.2(a)) does not require separation and holds for arbitrary outliers, including those that overlap with the inlier distribution.

- Gaussian inliers are needed only for the explicit corollary, not for the general result. Theorem 4.2(b) requires only that the inliers lie within their margin set. For bounded-support inliers, a fixed threshold is sufficient (Theorem D.14).

- Our empirical results confirm that the exact-recovery condition is restrictive. In Appendix F.1, Table 8, at the data-selected threshold, ROBOT achieves exact outlier recovery in 0 of 50 runs for shift k≤2k\leq 2, compared with 46 of 50 runs at k=6k=6. Thus, most finite-sample settings in Section 5 fall outside the regime covered by the corollary, as now acknowledged in the Limitations.

- For overlapping outliers, we also study a setting with lognormal shift magnitudes, where some outliers lie within the inlier bulk. The threshold grid used in the paper, which is anchored on a non-robust fit, fails to reach the useful threshold scale, and PW-ROT performs substantially worse (error 0.86–1.26 at k=2k=2, compared with 0.20–0.24 for Nietert, ROBOT-χ2\chi^2, MCD, and Tukey). With an expanded 80-point grid in a post-hoc diagnostic, the errors decrease to 0.24–0.29, within 0.08 of Nietert (Appendix F.1, Tables 13–14). We now explicitly acknowledge in the Limitations that the threshold-selection procedure is not robust to outliers that inflate the initial anchor fit.

AuthorsOpenReview ↗

Soft compression versus prior methods (W4), and MNIST (Q2)

We agree with W4 and do not claim otherwise. Table 4 reports an F₁ score of 0.62 for r=2r=2, compared with 0.69 for ROBOT and 0.79 for trimmed OT at k=2k=2. In response to the reviews, we extended the comparison on shared data to include Huber, Tukey, MCD, and an OT-profile detector, with all tuning parameters selected using the same held-out split as PW-ROT (Appendix F.1, Table 5). At k=2k=2 (one-sided, ϵ=0.2\epsilon=0.2, 50 trials), the error and F₁ scores are as follows: PW-ROT (r=2r=2): 0.32 / 0.61; ROBOT: 0.19 / 0.72; Nietert: 0.19 / 0.80; MCD: 0.18 / 0.83; and OT-profile elbow: 0.16 / 0.90.

The AUROC based on the ranking ∣xi−μ^∣\\|x_i-\hat{\mu}\\| is nearly identical across methods (0.976–0.988, except for Huber and the coordinate median). Thus, the F₁ gap mainly comes from how each method selects its flagging threshold. We have revised the abstract, Contribution 4, Algorithm 1, and Section 5 to clarify that no single power performs best in all settings. We recommend r=1r=1 for one-sided, well-separated outliers, including in the higher dimensions we tested, and a soft power between 1.5 and 2 (we use r=2r=2) when the geometry is unknown and the threshold must be selected from data. Neither choice has the largest point error in more than 1 of the 24 experimental settings. Each also requires the fallback threshold selector in at most 1 of 20 MNIST runs, compared with 8 runs for the hard clip.

Huber, Tukey, and MCD are included to calibrate the point-target special case, where PW-ROT is itself an M-estimator, rather than as direct competitors. These methods do not apply to a generator and use different information. PW-ROT's model and flagging rule assume unit inlier variance, as do the threshold grids for Huber and Tukey. In contrast, Nietert, ROBOT-χ2\chi^2, the OT-profile detector, and the medians do not use a scale estimate, while MCD estimates its covariance. We have therefore grouped the robust-OT and classical estimators separately in the tables and made no overall ranking claim (Appendix F.1, "Scope of the comparison with classical estimators").

Q2: Loss versus threshold selection on MNIST. The final flagging rule is a threshold test for every power. To separate the effect of the cost from that of threshold selection, we compare, for each run, the F₁ score at the selected threshold with the best F₁ score over the entire λ\lambda path. The latter serves as an oracle benchmark that does not require threshold selection (Appendix G, Table 18). We report all 80 runs, with 10 runs per task and power.

ROBOT's cost can detect the outliers: its best-λ\lambda F₁ scores are 1.00 for all-white outliers and 0.95 for FashionMNIST. However, threshold selection reduces F₁ by 0.40 for all-white outliers and 0.18 for FashionMNIST. In comparison, the loss is only 0.00–0.03 for r=1.5,2,3r=1.5,2,3. Four of the ten r=3r=3 all-white runs collapse at every threshold, with F₁ equal to zero; these are also the four runs in which the selector falls back. The fallback is needed in none of the 20 runs for r=2r=2, in 1 run for r=1.5r=1.5, and in 8 runs for the hard clip. These results suggest that the advantage of soft powers on MNIST comes from the cost producing a threshold that is easier to select, rather than from a better threshold-selection procedure.

We did not compare other detectors on MNIST because trimming and profile-based detectors do not have a natural extension to a generator. Instead, comparisons with other detectors are provided for the mean-estimation task in Table 5. We acknowledge the absence of a broader MNIST comparison as a limitation. We have also increased the number of random seeds from 3 to 10.

Reviewer_QGALOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • Penalized-Weight Robust Optimal Transport (PW-ROT), assigns penalized weights on each transport path. Eliminating these weights in closed form yields a family of bounded ground costs unifying ROBOT-style hard clipping (r=1r=1) with smooth soft compression (r>1r>1)
  • This work is inspired by ideas from robust regression that are adapted to robust transport for mean estimation and automatic outlier identification without explicitly specifying the contamination fraction.
  • This paper provides theoretical guarantees for robust estimation, automatic down-weighting, and exact outlier recovery under separation assumptions.
  • Experiments on synthetic mean estimation and contaminated MNIST, indicate complementary strengths of hard and soft clipping.
  • Hard clipping performs better for well-separated, one-sided contamination, whereas soft compression can offer advantages near the threshold.
  • r=2r=2 is recommended as a practical compromise, although no power consistently dominates across all experiments.
Critical strengths

Mathematical formulation and theoretical contributions

  • This paper extends penalized-weight ideas from robust regression to optimal transport which is interesting and novel.
  • It provides a unified framework connecting hard clipping and soft compression.
  • Establishes nontrivial outlier-recovery guarantees, including fixed-fraction contamination and necessary threshold-growth rates.

Empirical analysis of robustness trade-offs

  • This paper provides extensive theoretical analysis, proofs, and supplementary experiments.
  • It investigates how the power parameter affects both estimation bias and outlier detection, complementing with theoretical and empirical evidence at different contamination levels, geometries, and soft-compression powers.
  • It demonstrates that soft compression is advantageous near the threshold while acknowledging scenarios where hard clipping performs better.
Critical weaknesses

Relevant literature on partial-OT and insufficient comparison with existing approaches

  • The related-work section misses relevant combinatorial and partial-OT methods, particularly Computing All Optimal Partial Transports (ICLR 2023), which studies the OT-profile across transported masses and provides separation-based results for outlier identification without requiring a predetermined transported mass.
  • Other relevant work includes PAWL (ICLR 2025) and recent robust/partial semi-discrete OT algorithms.
  • These approaches do not necessarily establish identical theoretical guarantees, but their relationship to PW-ROT should be discussed more carefully, particularly regarding automatic outlier identification, assumptions on contamination, and computational complexity.
  • More direct experimental comparisons with partial-OT outlier detectors and classical robust M-estimators as baselines would help establish the practical advantages of the proposed formulation.

Limited experimental diversity and insufficient validation of the practical advantages of soft compression

  • Most synthetic experiments use low-dimensional Gaussian data, while the image experiments consider only two contaminated-MNIST settings with three seeds. Additional experiments involving higher-dimensional, heavy-tailed, multimodal, and more realistic contamination scenarios would improve generalizability.
  • r=2r=2 is not consistently supported across contamination regimes.
  • Connection between smooth weights and stable binary outlier detection needs further justification.
  • Threshold selection uses a 35% validation-trimming budget, exceeding the tested contamination fractions.
  • Clearer proof organization would improve readability as the proofs are dense and dependencies are across appendices.
Questions for authors
  • How does PW-ROT differ from combinatorial partial-OT approaches, particularly Computing All Optimal Partial Transports which can identify outliers without a predefined trimming fraction? Also any other relevant literature like PAWL, Sliced-OT
  • Please provide an intuitive example explaining model-side trimming and its relationship to automatic outlier detection? How do model-based flags relate to the optimal transport plan?
  • Can you isolate the contribution of smooth weights from model fitting and threshold selection since binary outlier flags remain threshold-based for all powers?
  • Please compare with additional partial-OT methods, classical robust estimators, and efficient OT implementations.
  • Provide actual examples of correctly and incorrectly detected outliers across different thresholds, showing false positives and false negatives
3 comments
AuthorsOpenReview ↗

Partial-OT literature and baselines (W1, Q1, Q4)

Thank you for these pointers. The partial-OT works are now discussed in Section 2 and Appendix A, including Phatak et al. (ICLR 2023), PAWL (Chapel and Tavenard, ICLR 2025), sliced optimal partial transport (Bai et al., CVPR 2023), and the robust and partial semi-discrete algorithms of Agarwal et al. (NeurIPS 2025).

Q1: How PW-ROT differs. Along the three axes you identify:

  • Automatic identification. Phatak et al. compute optimal partial transport for each transported mass and identify outliers from the OT profile, without specifying the mass in advance, for a fixed pair of discrete measures. PW-ROT has no mass parameter. Instead, it takes a threshold λ in cost units, obtains the bounded cost from per-path weights, and flags a point when the paths it can use at full weight carry less than a fraction α of the fitted model's mass (Definition D.3). It also jointly fits a parametric or generative model, which is the focus of our theoretical guarantees; the partial-OT results concern the pair of measures themselves.

  • Assumptions on contamination. Phatak et al. assume deterministic separation: for a small ε and a constant C > 1, every outlier lies at least Cw/ε away from every inlier point, where w is the cost of transporting the inliers (Outlier Lemma 1.1). They show that the (α* − ε)-optimal partial transport moves no outlier mass and that the first derivative of the profile increases noticeably around α*, thereby identifying the inlier mass. In contrast, Theorem 4.2(b) requires the inliers to lie inside and the outliers outside a margin set around the fitted model, while allowing outliers at arbitrary positions beyond that margin. The result is probabilistic with respect to the inliers. Theorem 4.2(c) further shows that the threshold must grow at the scale σ²(√d + √log n)².

  • Computation. PW-ROT requires a single OT problem with an entrywise-transformed cost matrix (Theorem 3.1), so any OT solver can be used. The selector fits a warm-started path of approximately 25 thresholds. Phatak et al. provide an exact algorithm with Õ(n²K) complexity, where K is the number of profile segments (Chapel and Tavenard summarize it as O(n³)). PAWL computes the exact partial 1-Wasserstein solution on the line for all transported masses in O(n log n). In the semi-discrete setting, Agarwal et al. provide an exact algorithm with complexity n^{O(d)} log(1/ε) and an n^{1+o(1)} ε^{-O(d)} approximation for the α-partial and λ-robust (total-variation-regularized) problems. These methods address different problems from ours: they concern a pair of measures, whereas PW-ROT addresses parametric estimation. A direct runtime comparison is left for future work.

Q4: Baselines. Using shared data and selecting all tuned parameters on the same held-out split as PW-ROT, we added Huber and Tukey M-estimators, the MCD detector, and an OT-profile detector (Appendix F.1, Table 5). The latter adapts the profile idea to a point target, where partial transport to a Dirac measure reduces to the trimmed mean. It selects the knee of the trimmed cost as a function of the removed mass, without specifying the mass in advance. This is our adaptation, not the algorithm of Phatak et al.

At k = 2 (one-sided contamination, ε = 0.2, 50 trials), the error/F₁ results are: PW-ROT with r = 1 (ROBOT), 0.19/0.72; OT-profile, 0.16/0.90; MCD, 0.18/0.83; and Nietert, 0.19/0.80. We recommend r = 1 for one-sided contamination and do not recommend soft compression in this setting. Table 5 also reports the soft powers (r = 1.5, 2, 3), which are less accurate than r = 1 here. We do not claim that PW-ROT outperforms these detectors, and the revised paper now makes this explicit.

We did not run PAWL, sliced partial OT, or the semi-discrete solvers as detectors because they are formulated for pairs of measures; adapting them to the parametric setting is left for future work. We also did not compare OT implementations. Our theory concerns the exact minimizer, while Table 6 shows that our stochastic solver achieves results within 0.024–0.042 of a longer refit.

Finally, Huber, Tukey, and MCD serve to calibrate the point-target special case, where PW-ROT is itself an M-estimator, rather than to establish a direct comparison with competing methods. These classical estimators do not apply to a generator and use different information. Specifically, PW-ROT's model and flagging rule assume unit inlier variance, as does the threshold grid for Huber and Tukey. In contrast, Nietert, ROBOT-χ², the OT-profile detector, and the medians do not require a scale estimate, while MCD estimates its covariance. Accordingly, the tables separate robust-OT estimators from classical estimators, and we make no ranking claim (Appendix F.1, "Scope of the comparison with classical estimators").

AuthorsOpenReview ↗

Additional Experiments: Dimensionality, Contamination Robustness, and Hyperparameter Sensitivity (W2, W3, W5)

Dimension. We repeated the one-sided experiment at d = 20 and 50 (Table 12), initializing PW-ROT at the coordinate median, as for the baselines. The hard clip remains competitive with Nietert's trimmed OT in terms of error (k = 2: 0.28 vs. 0.32 at d = 20, and 0.47 vs. 0.49 at d = 50), although Nietert achieves higher F₁. The errors of the soft powers increase substantially faster with dimension than those of PW-ROT with r = 1 (ROBOT), which we recommend for this setting. At d = 50 and k = 1, the errors are 0.52 for r = 1 and 1.22, 1.36, and 1.39 for r = 1.5, 2, and 3, respectively, consistent with the price of softness (Proposition E.11). The selected thresholds are not at the grid boundaries in most runs, suggesting that this is not simply a grid-boundary effect. We have not separately isolated the contributions of the cost transformation and the selection rule. The number of MNIST seeds has also been increased from 3 to 10.

Heavy-tailed, bimodal, and overlapping contamination. We conducted three additional experiments, with their settings and predicted outcomes specified before running them (Appendix F.1, Tables 13–14): Student-t(3) noise on all observations, contamination with lognormal shift magnitudes (allowing some outliers to overlap with the inlier bulk), and bimodal inliers. Under Student-t noise, the errors of all methods range from 0.157 to 0.197, except for Tukey (0.225). PW-ROT achieves errors of 0.16–0.17, comparable to the coordinate median (0.16). With bimodal inliers, PW-ROT with r = 1 (ROBOT) remains competitive, achieving errors of 0.32 and 0.19 at k = 2 and 4, respectively, compared with 0.30 and 0.25 for Nietert. Under lognormal shifts, the pre-specified experiment shows that PW-ROT underperforms Nietert, ROBOT-χ², MCD, and Tukey, with errors of 0.86–1.26 at k = 2. This failure is largely attributable to the threshold grid, which is anchored to the largest residual from a non-robust fit and is therefore inflated by these outliers. The smallest grid value was 75–270, compared with a useful scale of 3–12. As a post-hoc diagnostic using the same data and selector, increasing the grid from 25 to 80 points reduces the errors to 0.24 for PW-ROT with r = 1 (ROBOT) and 0.29 for r = 2 at k = 2, within 0.08 of Nietert. We report the pre-specified experiment as the main result and the additional experiment as a diagnostic explanation. The Limitations section now explicitly states that the selection procedure is not robust to outliers that inflate the anchor fit. Since λ is a threshold rather than a scale parameter, PW-ROT targets separated contamination with a specified inlier bulk, rather than gap-free heavy-tailed contamination.

W3, r = 2. The paper now recommends PW-ROT with r = 1 (ROBOT) for one-sided, well-separated outliers, including in the higher-dimensional settings we tested, where we do not recommend soft compression. For unknown contamination geometry, where the threshold must be selected from data, we recommend a soft power between 1.5 and 2 (we use r = 2). Neither choice produces the worst point error in more than 1 of the 24 experimental cells, and each requires the fallback selector in at most 1 of 20 image runs, compared with 8 for the hard clip. This replaces our earlier recommendation of r = 2 as the default, since Tables 5 and 12 show that the hard clip performs better under moderate separation and in higher dimensions. The abstract, Contribution 4, Algorithm 1, and Section 5 have been updated accordingly.

W5, the 35% trim. Table 7 examines sensitivity to the trimming parameter. For ε = 0.2, the results are essentially identical, up to experimental noise, for ζ ∈ {0.25, 0.35, 0.45}. For ε = 0.4, when ζ exceeds 0.25 and 0.35, PW-ROT with r = 1 (ROBOT) remains insensitive to the choice of ζ, whereas the soft powers deteriorate when ζ < ε. For example, with r = 2 and k = 4, the errors are 0.43, 0.26, and 0.22 for ζ = 0.25, 0.35, and 0.45, respectively. Thus, the procedure requires an upper bound on the contamination proportion, which we now explicitly state. The other baselines have the same requirement.

AuthorsOpenReview ↗

Intuition, smooth weights, examples, proofs (Q2, Q3, Q5, W4, W6)

Q2: An example. Appendix D now opens with a one-dimensional illustration. Consider the model N(θ, 1), with inliers drawn from N(0, 1), a fraction ε of outliers located at 20, and λ = 16. Every path from an outlier to a model point near the inliers has cost greater than λ and therefore receives zero weight. Its contribution to the objective is consequently the constant λ per unit of mass, regardless of θ. Thus, the outliers contribute ελ to the criterion independently of θ, while the remaining transport matches the inliers to a subset of the model with mass 1 − ε. In other words, PW-ROT trims the model rather than the data, without requiring ε as an input (Proposition 4.1). The flagging rule examines the same paths: an outlier can reach at full weight only the model mass within distance √λ = 4, which is approximately zero, so it is flagged. An inlier at 0, by contrast, can reach almost all the model mass and is therefore not flagged. For r > 1, the outlier paths retain a small positive weight, illustrating the explicit gap identified in Proposition 4.1.

Q3 and W4: Smooth weights versus binary flags. The binary flag is a threshold test for every power, so the smooth weights do not themselves generate the flags. The difference lies in how the flag set behaves as λ varies and how reliably the threshold can be selected. We disentangle these effects from model fitting and threshold selection in two ways (Appendix G, Table 18; 80 runs in total, with 10 runs per task and power). First, we compare the F₁ score at the selected threshold with the best F₁ score across the entire λ path, without threshold selection. PW-ROT with r = 1 (ROBOT) can achieve strong detection performance (best-λ F₁ of 1.00 on all-white and 0.95 on FashionMNIST), but threshold selection reduces F₁ by 0.40 and 0.18, respectively. For r = 1.5, 2, and 3, the corresponding losses are only 0.00–0.03. Second, Table 17 directly reports the flagged-image counts across thresholds. We clarify that the relationship between smooth weights and flag-set stability is empirical: Section 4.2 explicitly states that our theory does not quantify this relationship.

Q5: Examples across thresholds. Table 17 (Appendix G; seed 0, 1,200 outliers among 6,000 images, using the paper's 95% rule) reports the numbers of flagged images, true positives, false positives, and false negatives at five thresholds around the selected value. On FashionMNIST, the hard clip produces nonmonotone flagged-image counts as λ decreases (0, 22, 88, 0, and 313). This occurs because each λ corresponds to a separate training run, and the selected threshold is a fallback. In contrast, for r = 2, the counts increase smoothly (1,037, 1,122, 1,150, and 1,196 flagged images), with at most 14 false positives, until the threshold becomes sufficiently small that every inlier is flagged. We report aggregate counts rather than individual images because the per-image flag sets under the 95% rule were not stored. Re-running the experiments to save these sets would require approximately two GPU-hours per run.

W6: Proof organization. All corresponding revisions are highlighted in blue. A dependency map at the beginning of Appendix C now lists the lemmas and theorems supporting each main-text result. A roadmap sentence at the beginning of Section 4 connects path weights, flagging, trimming, estimation, and recovery. We also reorganized the discussion following Theorem 4.2 into separate paragraphs, structured the proof of Lemma D.5 with one paragraph for each part, and added a notation table in Appendix I.

In summary. We have expanded the discussion of partial-OT literature and narrowed our novelty claims to what the theoretical guarantees support. The experiments now cover higher dimensions (d = 20 and 50), heavy-tailed noise, lognormal-shift contamination, and bimodal inliers, with additional baselines evaluated on shared data. The paper also clarifies the practical recommendation: use r = 1 for one-sided contamination and a soft power between 1.5 and 2 otherwise.

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

This paper proposes Penalized-Weight Robust Optimal Transport (PW-ROT), a family of robust OT methods that introduces penalized weights on individual transport paths. Eliminating these weights yields a bounded ground-cost function that generalizes ROBOT's hard clipping to smooth soft compression. The paper establishes theoretical guarantees for robust parameter estimation and exact outlier recovery under separation conditions, including sufficient and necessary threshold scaling for Gaussian inliers. Experiments on synthetic mean estimation and contaminated MNIST demonstrate the trade-off between hard clipping and soft compression, with the authors recommending r=2 as a practical default.

Critical strengths
  1. Theoretical guarantees for exact outlier recovery. The paper provides a systematic analysis of robust OT estimators, including estimation error bounds, exact outlier recovery guarantees, and necessary threshold scaling. These results address a meaningful gap between robust estimation and reliable outlier identification.
  2. A unified characterization of hard clipping and soft compression. The proposed formulation connects ROBOT's hard clipping with smoothly compressed costs through a closed-form weight optimization. The theoretical analysis and experiments help clarify the trade-off between robustness to well-separated outliers and stability near the clipping threshold.
Critical weaknesses
  1. Limited methodological novelty beyond existing robust estimators. The penalized-weight construction closely follows classical half-quadratic optimization and robust M-estimation. For point-target estimation, PW-ROT reduces directly to an M-estimator with a bounded loss. Moreover, the exact recovery guarantees also apply to the existing ROBOT method (r=1). Therefore, the main contribution appears to be the theoretical characterization of an existing class of robust estimators rather than a fundamentally new robust OT mechanism. The paper could better distinguish its novel theoretical results from established robust statistics.
  2. Insufficient empirical evidence for the practical advantages of soft compression. The experiments are largely limited to low-dimensional Gaussian mean estimation and MNIST with only three seeds. In several settings, hard clipping or existing baselines outperform the proposed soft variants. In particular, the constructed trimming and reweighting detectors achieve higher F1 scores in several comparisons (Table 4). The strong MNIST results also depend on a data-driven threshold-selection procedure, making it difficult to isolate the benefit of soft compression itself. Consequently, the recommendation of (r=2) as a general default is not yet sufficiently supported.
Questions for authors
  1. Since the point-target formulation reduces to classical robust M-estimation, which theoretical results are fundamentally specific to the OT setting and cannot be obtained through existing robust M-estimation analyses?
  2. The exact recovery guarantees apply equally to ROBOT (r=1). What additional theoretical or practical advantages does PW-ROT provide over ROBOT beyond the smoothness of its weights?
  3. Could the authors provide a more controlled comparison of threshold sensitivity and detection performance, ideally separating the effects of the robust cost function from those of the threshold-selection procedure? This would help determine whether the improvements on MNIST are primarily due to soft compression or threshold selection.
2 comments
AuthorsOpenReview ↗

Novelty (W1) and what PW-ROT offers over ROBOT (Q1, Q2)

Thank you for stating the objection so precisely. We agree with its premises and answer within them.

For what is not new: weight elimination is the half-quadratic construction; r=1r=1 is ROBOT; and, for a point target, PW-ROT is an M-estimator with a bounded loss (Proposition E.2). The paper already states the first two points in Section 2, and Appendix A now clarifies which ingredients are classical.

Q1: What is specific to OT?

For a point target, Corollary 4.3 and the oracle-mean statement for r=1r=1 are M-estimation results, and we now say so explicitly. Three results concern a non-point-mass model (the Gaussian spread target used in the experiments, or a generator), where the criterion is not a sum of per-point losses:

(i) Proposition 4.1: model-side trimming at the true, unknown outlier fraction, which is exact for r=1r=1 and has an explicit gap for r>1r>1.

(ii) Theorem 4.2(a): the inlier-sampling term is governed by the OT rate EW1≍n−1/d\mathbb{E}\mathcal{W}_1\asymp n^{-1/d}.

(iii) Theorem 4.2(b): exact recovery of the outlier set, with flags defined through the plan's full-weight reachable mass. Corollary D.17 establishes this result for the spread target, with all conditions stated in terms of the data law and the outliers.

We scope the novelty claim accordingly: this is the first finite-sample, estimator-level exact-recovery guarantee for robust-OT flags in a general parametric model. The authors of ROBOT list such guarantees as future work, while partial-OT results concern a fixed pair of measures.

Q2: PW-ROT versus ROBOT.

Exact recovery applies to every r≥1r\geq 1, including ROBOT. We do not claim it as an advantage of soft powers. Instead, the family provides a quantified trade-off, and we state its cost.

1. Efficiency at the same threshold (Proposition E.14 in Appendix E.7; Tables 10–11 and Figure 7 in Appendix F.1).

For a point target and clean Gaussian inliers, n(μ^−μ0)\sqrt{n}(\hat{\mu}-\mu_0) is asymptotically normal, with an explicit variance v(λ,r,d)v(\lambda,r,d). For ROBOT, v=1/Fd+2(λ)v=1/F_{d+2}(\lambda), where Fd+2F_{d+2} is the χd+22\chi^2_{d+2} distribution function. We prove that, at every λ\lambda, each soft power has vr≤v1v_r\leq v_1. Under this parametrization, every member flags exactly the points with cost above λ\lambda, so the same λ\lambda gives the same flag radius.

For d=5d=5, achieving 95% efficiency requires λ≈14.1\lambda\approx14.1 for ROBOT, compared with 9.99.9, 8.88.8, and 7.857.85 for r=1.5r=1.5, 22, and 33, respectively. At the threshold that leaves 5% of clean points beyond it, ROBOT's efficiency is 0.720.72, 0.860.86, and 0.910.91 for d=1d=1, 55, and 2020, respectively, compared with 0.880.88–0.9970.997 for the soft powers. Asymptotic normality also holds for ROBOT; the difference is in the variance. This gain concerns clean data and must be weighed against the cost of softness when outliers are well separated.

2. Bias for outliers near the threshold (Section 5.1).

When outliers are placed at squared distance approximately λ\lambda, the error at ε=0.4\varepsilon=0.4 is 1.771.77, 1.391.39, 1.321.32, and 1.291.29 for r=1r=1, 1.51.5, 22, and 33, respectively. Only the soft powers retain part of the inliers' pull beyond the threshold. We explicitly state that our theory does not cover this effect.

3. Selectability of the threshold on MNIST (Q3 below).

The cost is the price of softness: for one-sided, well-separated outliers, hard clipping performs better (Corollary 4.3; Table 5). At k=4k=4, the error is 0.1340.134 for ROBOT, compared with 0.1620.162 for r=2r=2 and 0.2840.284 for r=3r=3. No power dominates. Our claim is a trade-off, not that soft powers outperform hard clipping.

AuthorsOpenReview ↗

Empirical evidence (W2) and cost versus selection (Q3)

(1) Constructed detectors and other baselines

We extended the comparison to Huber, Tukey, MCD, and an OT-profile detector, all tuned on the same held-out split (Appendix F.1, Table 5). At k=2k=2, error/F₁ scores are: r=2r=2 (0.32/0.61), ROBOT (0.19/0.72), Nietert (0.19/0.80), MCD (0.18/0.83), and OT-profile (0.16/0.90). AUROC is similar across most methods (0.976–0.988). Thus, PW-ROT is not the most accurate detector in this setting; much of the difference comes from threshold selection rather than ranking quality.

Huber, Tukey, and MCD are calibration methods for the point-target special case, not direct competitors for generator-based detection. Methods also differ in scale assumptions and estimation procedures. We therefore report robust-OT and classical estimators separately and make no overall ranking claim (Appendix F.1).

(2) Q3 and MNIST: cost versus selector

All powers use the same selector. We compare selected-threshold F₁ with the oracle best F₁ over the entire λ\lambda path (Appendix G, Table 18; 80 runs). ROBOT achieves best-λ\lambda F₁ of 1.00 on all-white outliers and 0.95 on FashionMNIST, but threshold selection reduces these by 0.40 and 0.18, respectively. For r=1.5,2,3r=1.5,2,3, reductions are only 0.00–0.03. The fallback selector is needed in 0/20 runs for r=2r=2, versus 8/20 for the hard clip. Thus, soft powers perform better mainly because their costs produce thresholds that are easier to select, not because they use a better selector. We also increased the seeds from 3 to 10.

  • Dimension and default r=2r=2

At d=20,50d=20,50, the hard clip remains competitive with Nietert in error, but Nietert achieves higher F₁ (Appendix F.1, Table 12). At k=2k=2, error/F₁ scores are 0.28/0.79 versus 0.32/0.89 at d=20d=20, and 0.47/0.89 versus 0.49/0.93 at d=50d=50, respectively.

Soft-power errors increase faster with dimension: at d=50,k=1d=50,k=1, errors are 1.22, 1.36, and 1.39 for r=1.5,2,3r=1.5,2,3, versus 0.52 for ROBOT. These results do not support r=2r=2 as universally best. We have narrowed our claims: no power is best in every setting. We recommend r=1r=1 for one-sided, well-separated outliers and a soft power between 1.5 and 2 (using r=2r=2) when contamination geometry is unknown. Neither choice has the worst estimation error in more than 1 of 24 settings or requires fallback selection in more than 1 of 20 image runs, versus 8/20 for the hard clip. The abstract, Contribution 4, and Section 5 are revised accordingly.

  • Heavy-tailed and bimodal settings

We added Student-t(3)t(3) noise, lognormal-shift contamination, and bimodal inliers (Appendix F.1, Tables 13–14). PW-ROT is comparable to the coordinate median under Student-tt noise (errors 0.16–0.17 versus 0.16), while ROBOT is competitive with Nietert under bimodal inliers (0.32 versus 0.30 at k=2k=2). Under lognormal shifts, however, PW-ROT performs substantially worse (0.86–1.26 versus 0.20–0.24 for competing methods). Its data-derived grid, anchored on a non-robust fit, fails to reach the useful scale. Extending the grid to 80 points reduces errors to 0.24–0.29. We now discuss this limitation explicitly.

  • Grid sensitivity

We compared 25 and 80 thresholds in a paired experiment (ROBOT and r=2r=2, k=2,4k=2,4, 30 trials). Mean estimation errors change by at most 0.011, indicating stable point estimation. However, the pre-specified criterion is not met: ROBOT at k=4k=4 is 2.2 standard errors from zero. Detection is more sensitive: for r=2,k=4r=2,k=4, swamping increases from 0.11 to 0.19 and F₁ decreases from 0.84 to 0.75 because the extended grid permits smaller selected thresholds. Both results are reported in Appendix F.1.

Comments on the paper

3 comments

We thank the five reviewers for their detailed and constructive feedback. Across the reviews, three central strengths of our work were recognized: (1) a closed-form family of bounded costs that includes ROBOT as its hard-clipping member and introduces smooth soft compression (5Tah, QDJx, QGAL, w56n); (2) exact recovery of the outlier set, together with a threshold-rate condition that we establish to be necessary (5Tah, QDJx, QGAL, k9cu); and (3) an analysis of how the power parameter rr controls the trade-off between bias and weight continuity (QDJx, QGAL, w56n). Our new MNIST experiments further demonstrate the practical advantage of soft powers on image data: when the threshold is selected from data, the F₁-score decreases by only 0.00–0.03 for r=1.5,2,r=1.5, 2, and 33, compared with decreases of 0.40 and 0.18 for hard clipping.

The reviewers also raised five important concerns. We have addressed them through additional theoretical results, extensive experiments, and a more precise articulation of the scope and limitations of our contributions.

1. Novelty relative to M-estimation and ROBOT (QDJx, k9cu). We now explicitly distinguish the classical connections from our contributions: weight elimination is a half-quadratic representation, r=1r=1 recovers ROBOT, and, for a point target, PW-ROT is an M-estimator (Proposition E.2). We have accordingly narrowed our novelty claim to a finite-sample, estimator-level exact-recovery guarantee for identifying outliers using a fitted robust-OT estimator in a general parametric model, together with a necessary threshold-rate condition (Theorem 4.2(c) for the point target and Remark D.18 for the spread target). Our transport-specific theoretical contributions concern target models beyond point masses.

2. Validity of the assumption in Theorem 4.2(b) (5Tah, k9cu). We clarify that the margin sets depend only on the model, λ\lambda, α\alpha, and θ0\theta_0, rather than on the data or the estimator. Moreover, for the Gaussian spread-target setting and the 95% rule used in our experiments, we replace the abstract margin-set condition with three interpretable conditions on the inlier distribution, outlier distribution, and outlier fraction. We then establish the required curvature under these conditions (Lemma D.15, Corollary D.17, and Remark D.18).

3. Evidence for soft compression and the default choice of r=2r=2 (QDJx, QGAL, k9cu, w56n). We have added comparisons with additional baselines, a dimensionality study, heavy-tailed and bimodal settings, a separation-parameter sweep, and an analysis of the cost function versus the threshold selector on MNIST. These results also lead us to refine our practical recommendation: r=1r=1 is preferable for one-sided, well-separated outliers, whereas a soft power in the range of 1.5–2 (with r=2r=2 as our default) is more appropriate when the outlier geometry is unknown and the threshold must be selected from data.

4. Theory versus practice (k9cu, QGAL, 5Tah, w56n). We have added empirical analyses of the optimization solver gap, the effect of the 35% trimming level, sensitivity to the threshold grid, and the extent to which the theoretically derived flag radius predicts observed swamping and masking. These analyses help clarify the relationship between our theoretical guarantees and their practical behavior.

5. Presentation and related literature (5Tah, k9cu, QGAL). We have improved the organization and accessibility of the manuscript by adding an algorithm overview and a roadmap to the main text, a notation table, a proof-dependency map, and a worked example. We have also expanded our discussion of the partial optimal transport literature to better position our work relative to existing approaches.

Finally, all new experiments were specified, including their evaluation criteria and decision rules, before execution. We report all results transparently, including outcomes that are unfavorable to PW-ROT, to provide a balanced assessment of its strengths and limitations.

AuthorsOpenReview ↗

Detailed Summary of Revisions: Additional Theory, Experiments, and Presentation Improvement

Theory (Appendices D and E; changes highlighted in blue).

  • Exact recovery for the Gaussian spread target: We state all assumptions directly in terms of the data-generating distributions. Lemma D.15 establishes the required curvature, with κ=p12/r\kappa=p_1^2/r, independent of the dimension dd. Corollary D.17 establishes the exact recovery result, and Remark D.18 shows that the threshold scale is necessary for this target.
  • Asymptotic normality and efficiency of the point-target estimator: Proposition E.14 studies the asymptotic variance as a function of (λ,r)(\lambda,r). For ROBOT, the variance is 1/Fd+2(λ)1/F_{d+2}(\lambda). We prove that, for every λ\lambda, each soft power is at least as efficient as the hard clip. For d=5d=5, achieving 95% efficiency requires λ≈14.1\lambda\approx14.1 for ROBOT, compared with 7.857.85–9.99.9 for r=1.5,2,3r=1.5,2,3 (Table 10).
  • Asymptotic distribution with a fixed outlier fraction: We show that n−m(μ^−μ0)\sqrt{n-m}(\hat{\mu}-\mu_0) converges to N(0,σ2I)N(0,\sigma^2 I) for the hard clip and, under an explicit separation condition, also for r>1r>1 (Remark E.9). Asymptotic normality also holds for ROBOT, with the difference lying in the asymptotic variance.

Experiments (Appendix F.1; Tables 5–14).

Unless otherwise specified, the experiments use a one-sided mean shift with n=300n=300 and an outlier fraction of ϵ=0.2\epsilon=0.2.

  • Additional baselines (Table 5): We compare Huber, Tukey, MCD, and an OT-profile detector (our adaptation of the profile-based approach to a point target). All methods use the same data and held-out split for tuning as PW-ROT. These experiments provide a reference for the point-target setting, where PW-ROT is itself an M-estimator, rather than a direct competition between methods. Accordingly, the paper discusses robust-OT estimators and classical estimators separately and makes no ranking claims.
  • Solver accuracy (Table 6): A longer refit changes the estimate by only 0.024–0.042, compared with errors of 0.12–0.43. The resulting outlier flag sets remain unchanged.
  • Trimming level and exact recovery (Tables 7–8): Results remain stable when the trimming level ζ\zeta is above the contamination fraction. With the data-selected threshold, the hard clip recovers the outlier set exactly in 46 of 50 runs at shift k=6k=6, but in none of the runs when k≤2k\leq2.
  • Flag radius (Table 9): The radius determined by the selected threshold predicts the observed swamping and masking within 0.07.
  • Additional settings and sensitivity analyses: We consider higher dimensions (d=20,50d=20,50; Table 12), heavy-tailed, lognormal-shift, and bimodal distributions (Tables 13–14), and the sensitivity of the main results to the threshold grid (last paragraph of Appendix F.1).
  • MNIST experiments (Appendix G): We compare the cost function and threshold selector (Table 18), report the numbers of correct and incorrect outlier flags across thresholds (Table 17), and include results from 10 random seeds.

Presentation and Related Literature (changes highlighted in blue).

  • Manuscript organization: We add Algorithm 1 and a roadmap sentence to the main text, a notation table (Appendix I), a proof-dependency map (beginning of Appendix C), and a worked example illustrating model-side trimming (beginning of Appendix D).
  • Theoretical presentation: We reorganize the discussion following Theorem 4.2 into parts (a), (b), and (c), and revise Figure 1 to show ρ\rho itself.
  • Related literature: We expand the discussion of partial optimal transport, including Phatak et al., PAWL, sliced partial OT, and robust and partial semi-discrete OT, in Section 2 and Appendix A. We also revise our novelty claims to better reflect the relationship with these existing works.
AuthorsOpenReview ↗

Summary of New Results

We summarize the main findings from the new experiments and analyses.

1. Where soft compression helps: symmetric and near-threshold contamination.

When outliers are symmetrically distributed around the mean, their effects largely cancel. In this setting, robustness is less important, and the main question is how much efficiency a robust estimator sacrifices. At ε=0.2\varepsilon = 0.2 and k=0.5k = 0.5 (100 trials; Appendix, Table "one-sided/symmetric, all cells"), the errors are 0.131 for ROBOT and 0.126–0.127 for the soft powers r=1.5,2,3r = 1.5, 2, 3. These results are close to the sample mean (0.125), and better than Nietert's estimator (0.169) and ROBOT-χ2\chi^2 (0.177). Thus, the soft powers remain close to the sample mean, while hard clipping and trimming sacrifice some efficiency. They also outperform ROBOT in most symmetric settings, with error reductions of 0.001–0.011. When outliers are placed near the threshold, the soft powers also achieve lower bias (Section 5.1: errors of 1.29–1.39, compared with 1.77 for ROBOT at ε=0.4\varepsilon = 0.4). We do not claim that the soft powers outperform the sample mean under symmetric contamination, where the sample mean is already adequate.

2. Threshold selection on MNIST: the cost of hard clipping.

Across 10 random seeds and 80 runs, ROBOT's cost-based threshold selection can detect outliers, achieving a best-threshold F1F_1 score of 1.00 on the all-white task and 0.95 on FashionMNIST. However, its selected threshold leads to F1F_1 losses of 0.40 and 0.18, respectively, relative to the best threshold. In contrast, the corresponding losses for r=1.5,2,3r = 1.5, 2, 3 are only 0.00–0.03. Moreover, r=3r = 3 collapses in 4 of the 10 all-white runs. At the selected threshold on the all-white task, both masking and swamping rates are 0.00 for r=1.5r = 1.5 and r=2r = 2. In comparison, ROBOT has a masking rate of 0.40 (four of ten runs fail to flag any outlier), while r=3r = 3 also exhibits masking in its collapsed runs (Table 18).

3. Recommendation: choose the power based on the contamination structure.

Since no single power performs best in every setting (see below), we have removed the previous claim that r=2r = 2 is the "most consistent default." Instead, we recommend r=1r = 1 for one-sided, well-separated outliers, and a soft power in the range 1.5≤r≤21.5 \leq r \leq 2 (we use r=2r = 2) when the contamination structure is unknown and the threshold must be selected from data. The two soft powers perform similarly in our experiments: neither has the largest point-estimation error in more than 1 of the 24 contaminated settings, and each requires the selector's fallback in at most 1 of the 20 image runs. For comparison, the hard clip requires the fallback in 8 runs, while r=3r = 3 requires it in 4 runs.

4. Exact recovery requires separation.

Our exact-recovery guarantee requires the outliers to be separated from the inliers by a distance of order σ(d+log⁡n)\sigma(\sqrt{d} + \sqrt{\log n}). It therefore does not cover overlapping outliers. In contrast, our estimation error bound does not require such a separation condition. We have clarified this distinction in the Limitations section.

5. No single power is best in every setting.

At k=2k = 2 under one-sided contamination (ε=0.2\varepsilon = 0.2, d=5d = 5), trimming-based and χ2\chi^2-based estimators, MCD, and an OT-profile detector achieve errors comparable to or lower than PW-ROT, while also achieving higher F1F_1 scores: 0.61 for r=2r = 2, 0.72 for ROBOT, 0.80 for Nietert's estimator, and 0.90 for the OT-profile detector.

Among the powers, hard clipping performs best for one-sided, well-separated outliers, including in higher dimensions (d=20d = 20 and 5050). For example, at d=50d = 50 and k=2k = 2, its error is 0.47, compared with 0.49 for Nietert's estimator. The main advantages of the soft powers are their more reliable threshold selection on MNIST (as discussed above) and their higher efficiency than hard clipping, at the same threshold, for point estimation under Gaussian inliers.

6. Heavy-tailed shifts.

Under lognormal shift magnitudes, the data-derived threshold grid used in the paper, which is anchored on a non-robust fit, does not cover the useful threshold scale. Expanding the grid to 80 points reduces the errors to 0.24 for ROBOT and 0.29 for r=2r = 2 at k=2k = 2, bringing both within 0.08 of Nietert's estimator (Appendix F.1).

We hope these revisions make the contribution and limitations of the paper clearer. PW-ROT provides a unified family of estimators that includes ROBOT as a special case. It offers exact-recovery guarantees under explicit separation conditions and quantifies the trade-off between hard and soft compression. Importantly, the revised paper also identifies the settings in which each approach works well and acknowledges where competing methods perform better.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.