acceptodds
Under review as a conference paper at ICLR 2027

Private Evolution for IPM Optimization: Limits and Alternatives

Abstract

Private Evolution (PE) is a framework for differentially private (DP) synthetic data generation. Given a sensitive dataset and a public dataset , PE evolves by generating public variations of and privately selecting from them to reduce a target discrepancy to . Prior work proves that PE converges to in -Wasserstein () distance, but generating DP synthetic data with accuracy suffers from the curse of dimensionality, even outside the evolution framework. The distance is an integral probability metric (IPM), defined by the largest discrepancy over a class of queries. It is known that more structured query classes can produce DP synthetic data that converge faster than existing lower bounds on convergence. We ask whether PE can exploit this structure by optimizing IPMs other than . Our first result is that even simple IPMs induced by linear or convex-quadratic queries can be difficult to optimize directly via evolution. Even without privacy, exponentially many variations may be needed to generate a candidate dataset with strictly smaller target IPM. For linear queries, we recover statistically- and computationally-efficient PE by replacing the target IPM with a surrogate distance that bounds it, and for which polynomial-size Gaussian variation pools decrease the surrogate in expectation. We then use a DP Frank–Wolfe subroutine within PE to efficiently optimize a distribution over the variations and construct to match the mean of it. The resulting PE achieves IPM error and avoids the curse of dimensionality. This works in part because a dataset with a prescribed mean is easy to construct; for richer queries, reconstruction can itself be hard: we prove that matching only two bounded convex quadratic queries with a fixed-size dataset is already NP-hard. Overall, our results disentangle statistical and computational aspects of the three pieces of PE: the IPM, the evolution, and privacy. Our lower bounds show how non-private evolution fails (computationally), even for simple IPMs, while our positive result shows that replacing the target IPM by a compatible surrogate can recover efficient PE in some special cases.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, Synthetic Data, Private Evolution, Private Optimization, Integral Probability Metrics

Reviews

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

Disclaimer. My expertise lies in generative modeling where I had experience in DP and IPM optimization, but I have no a priori knowledge of Private Evolution and the related literature. I therefore have low confidence for my assessment below. I will let the other reviewers and the AC judge the relevance of my review.


This paper studies an alternative to Private Evolution (a training-free method to derive a DP synthetic dataset from a real one) that the authors name direct evolution, consisting in optimizing an IPM with an evolutionary algorithm. The convergence of such an algorithm is proven to require exponentially many variations even for simple function classes in the IPM. The paper proposes an alternative based on mean matching for linear functions but finds that even data reconstruction is NP-hard for quadratic functions.

Critical strengths
  1. The paper derives a new paradigm for private evolution consisting in directly optimizing an IPM, together with theoretical results about the difficulty of following such a paradigm, which is a negative result that may be conceptually interesting.
Critical weaknesses

As an outsider of the research area of Private Evolution, but having studied the problem for the present review, I find that this paper suffers from two important weaknesses severely limiting the significance of its contributions.

Motivation and clarity

  1. The paper clarity could be improved. It explains its framework and Private Evolution in a more convoluted manner than prior work (Lin et al., 2023; González Lara et al., 2026). The motivation for direct evolution is seldom discussed. Furthermore, the derived insights are scattered throughout the paper, leaving their recollection up to the reader.
  2. Regarding direct evolution, it is especially unclear to me why, when optimizing the IPM via direct evolution, the authors choose to do so via an evolutionary process when the IPM could be optimized even more directly via, e.g., a gradient flow.

Impractical and unrealistic results

  1. While I appreciate negative results, the paper does not convey valuable insights that could be leveraged elsewhere in the literature -- at least to my knowledge. The significance of the presented results seem to be bounded to the present paper.
  2. Moreover, the presented results are presented in a very restricted setting with unrealistic assumptions of linear or quadratic queries. Their discriminative power being limited, bare mean matching can be used to solve the studied problem when realistic data require more elaborate care, and it is unclear whether the optimization and difficulty results can transfer to more complex queries.
Questions for authors

Given the above weaknesses, and in my current limited understanding of the tackled research problem, I recommend rejecting this paper. However, given my lack of expertise, I remain open for discussion with the authors, reviewers and AC.

In particular, answers to the following questions could make me reconsider my assessment.

  1. Could the authors explain why they choose not to optimize the IPM via gradient methods?
  2. Could the authors justify their assumptions and the relevance they carry w.r.t. a realistic setting and the relevant literature?

Post-rebuttal update

I would like to thank the authors for their response, and to apologize for not having had the time to acknowledge it sooner. Unfortunately, I am not convinced by the provided answers, and I tend to agree with Reviewer M1ek. Regarding my specific questions:

  1. The reason why direct optimization cannot be used in this context is still unclear: I do not understand how it is less relevant than changing paradigm from private evolution to direct evolution. However, this may be due to my lack of expertise on the topic.
  2. Negativity results are not shown to transfer to more complex queries. I understand the FID argument, but it is generally a bad idea to optimize for FID directly (unless for some fine-tuning) precisely for the reasons provided by the authors; MMD is a better choice but requires more complex queries (Deng et al., 2026).

To put it more broadly, I think the paper lacks a general insight about how the presented results on direct evolution transfers to the actual private evolution algorithms. Therefore, I am inclined to maintain my score while acknowledging my low confidence and waiting for the reviewer-AC discussion period.

1 comment

Thanks for your feedback. We answer below.

The paper clarity could be improved. It explains its framework and Private Evolution in a more convoluted manner than prior work (Lin et al., 2023; González Lara et al., 2026).

We have now slightly repurposed how we introduce PE and added a flowchart for better intuition in Section 2, and we hope it is gentler to the uninitiated reader.

The motivation for direct evolution is seldom discussed.

Lin et al., 2023 proposed PE and González Lara et al., 2026 proved that the algorithm is effectively minimizing a W1 distance. We therefore start directly from the interpretation of PE as an algorithm optimizing W1, which motivates studying other IPMs. Direct evolution is a natural candidate for this task: make as much progress on the target IPM as the variation set allows at each iteration.

Furthermore, the derived insights are scattered throughout the paper, leaving their recollection up to the reader.

We have rewritten a good amount of the main body to explain these insights better (in our opinion). As concrete examples, in Section 3 we now include some of the intuition behind why direct evolution gets stuck with high probability in the quadratic case and we have rewritten Theorem 3.1 in a more clear way.

Regarding direct evolution, it is especially unclear to me why, when optimizing the IPM via direct evolution, the authors choose to do so via an evolutionary process when the IPM could be optimized even more directly via, e.g., a gradient flow.

We want to emphasize that our goal in this work was not to propose an algorithm for better solving a given class of IPMs, but to explore what is feasible (and not) using private evolution and related algorithms. Indeed, you are correct that one could optimize the IPM more directly with other methods, but our goal in this work is to deepen our understanding of evolution-based algorithms. We now discuss related methods for optimizing distances between datasets without evolution in the related work.

While I appreciate negative results, the paper does not convey valuable insights that could be leveraged elsewhere in the literature -- at least to my knowledge. The significance of the presented results seem to be bounded to the present paper.

Understanding PE for IPM optimization is relevant in other places in the literature. We give two examples where this appears. First, there are already works adapting PE to query-based utility. For example, “Is API Access to LLMs Useful for Generating Private Synthetic Tabular Data?” by Swanberg et al. (2025) studies PE for tabular data using an IPM distance as accuracy measure. Second, Lin et al., (2023) develops PE for images, where preserving quadratic queries is important; we elaborate on this second example below. We have added these to the introduction.

Moreover, the presented results are presented in a very restricted setting with unrealistic assumptions of linear or quadratic queries. Their discriminative power being limited, bare mean matching can be used to solve the studied problem when realistic data require more elaborate care, and it is unclear whether the optimization and difficulty results can transfer to more complex queries.

We remark that the simplicity of the queries is actually a good thing for our lower bounds: even for simple IPMs, we show that evolution may fail to make progress. Our lower bounds also establish worst-case obstructions for broader problem classes that allow higher-degree queries, since these classes include our quadratic instances.

That said, although their discriminative power is limited, there are also practical motivations for understanding linear and quadratic IPMs. We give two examples. First, in image generation, a widely used metric for evaluating synthetic images is Fréchet Inception Distance (FID), which depends only on the first and second moments of image features. Thus, matching linear and quadratic queries in feature space is sufficient to preserve the quantities entering FID. We mention this example in the introduction. Second, recent work on evaluating LLM-based user simulators (Meshi et al., 2026) compares synthetic and real conversations using simple statistics, such as how often users accept recommendations. These are applications where the goal is to preserve simple statistics.

Reviewer_M1ekOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether Private Evolution (PE) can optimize integral probability metrics (IPMs) other than the 1-Wasserstein distance and thereby avoid the curse of dimensionality. It first shows that direct evolution, which selects the next dataset from projected variations to minimize the target IPM, can fail even without privacy: for linear and convex-quadratic queries, there are instances where improvement requires exponentially many variations (Theorem 3.1), and a fixed variation pool can improve W1 but not a simple affine IPM (Proposition 3.1). For linear queries, the paper replaces the IPM with a mean-matching objective, privately selects a distribution over Gaussian variations by stochastic Frank-Wolfe with the exponential mechanism, and constructs any next dataset with the resulting mean, obtaining error O~((d2/(nε))1/3)\tilde O((d^2/(n\varepsilon))^{1/3}) (Theorems 4.1-4.2). A simulation in a 10-dimensional ball compares the method with W1-based PE in terms of mean error.

Critical strengths
  1. The question is relevant, and the decomposition into the IPM, the evolution step, and privacy is clear. Theorem 3.1 shows that the difficulty can come from whether the variation mechanism generates improving candidates, independently of privacy, and Proposition 3.1 uses a small example to show that the same variation pool can improve W1 but not an affine IPM. This compatibility viewpoint helps explain the limits of PE (Section 3).
  2. The proofs I checked are correct, and the positive result is a complete private algorithm. The one-step contraction constants in Theorem 4.1, the exponential-mechanism sensitivity 2Bt2/n2B_t^2/n and the choice of ε0\varepsilon_0 in Theorem 4.2 (consistent with zCDP composition of pure-DP steps), the rate derivation, and the constants of the Max-Cut reduction in Proposition 5.1 are all consistent. The resulting rate is polynomial in dd, in contrast to the (nε)−1/d(n\varepsilon)^{-1/d} rate under W1.
Critical weaknesses
  1. The exponential lower bound applies only to a single synthetic point. Theorem 3.1 sets Φ(x)=DF({y},{x})\Phi(x)=D_{\mathcal F}(\{y\},\{x\}) and starts at a cube vertex x0=θx_0=\theta, where projection cancels every outward move, so improvement requires a draw in one orthant, which has probability 2−d2^{-d}. PE, however, maintains many points, and for linear queries the IPM depends only on the mean. In a simple simulation on the same linear instance, I placed mm points at θ\theta and gave each one Gaussian variation; these mm variations form an admissible candidate in ΠX(Variations(St))m\Pi_{\mathcal X}(\mathrm{Variations}(S_t))^m. With d=200d=200 and m=100m=100, i.e., only 100 proposals, this candidate strictly reduced Φ\Phi in about 99% of trials, whereas the single-point bound implies roughly 22002^{200} proposals. The abstract's claim that exponentially many variations may be needed to generate a candidate dataset with a smaller IPM therefore overgeneralizes the result. A lower bound for m≥2m\ge2, or an explicitly restricted statement, is needed.
  2. The positive result says little about PE exploiting structure. For linear queries the IPM depends only on the mean, so HH is an ℓ2\ell_2 version of the same statistic rather than a surrogate for a harder objective. Algorithms 1 and 2 also allow St+1S_{t+1} to be any dataset with the prescribed mean, e.g., mm copies of it, so the method amounts to private mean estimation by Frank-Wolfe over random Gaussian atoms followed by reconstruction. This is the estimate-then-reconstruct paradigm that Sections 1.1 and 5 contrast with PE. The rate O~(d2/3/(nε)1/3)\tilde O(d^{2/3}/(n\varepsilon)^{1/3}) is also polynomially worse in nn than the Gaussian mechanism, roughly Rd/(nε)R\sqrt d/(n\varepsilon). In the paper's simulation setting with n=40,000n=40{,}000 and ε=1\varepsilon=1, the latter gives an expected error of about 0.0013, versus about 0.037 for FW-PE in Figure 1. The claim of statistically efficient PE is therefore too strong, and Figure 1 should include this reference.
  3. The reconstruction hardness result is not positioned against classic work. Proposition 5.1 is a non-private worst-case result in which both the queries and the targets are chosen adversarially, proved by a standard gap Max-Cut reduction. Computational hardness of DP synthetic data is well established: [1] characterizes when synthetic-data release is computationally feasible, and [2] shows that, assuming one-way functions exist, no efficient DP algorithm can generate synthetic data that preserves even all 2-way marginals. Neither work is cited. Since in PE the targets come from real data and the query class is fixed, the paper should clarify what Proposition 5.1 adds beyond these results.

Minor issues:

  1. The section structure is incorrect. Related Work appears as Section 1.1, the only subsection of the Introduction, rather than as a standalone section, so the Introduction mixes motivation, an overview of the results, and a literature review. It should be a separate section.
  2. The simulation is only illustrative. It uses d=10d=10, uniform data in a ball, and two values of ε\varepsilon, and its hyperparameters deviate from the theory (T=⌈Tth/64⌉T=\lceil T_{\mathrm{th}}/64\rceil, Q=32QbalQ=32Q_{\mathrm{bal}}, ηq=2/(q+3)\eta_q=2/(q+3); Appendix B). The error of W1-PE at ε=1\varepsilon=1 increases from n=20,000n=20{,}000 to n=40,000n=40{,}000 without explanation.
  3. In Proposition 3.1, no affine improvement exists only because the best attainable value ties with D(S,S0)=1/40D(S,S_0)=1/40, which makes the example weak evidence of incompatibility.
  4. Typos: "scaly" (after Theorem 3.1), "and the and the" (Section 5), and "erros" and "We consider and ε\varepsilon" (Section 6).
Questions for authors

Questions that could affect my assessment:

  1. Can Theorem 3.1 be extended to m≥2m\ge2 synthetic points? In the linear instance, averaging one Gaussian variation per point seems to improve the IPM with poly(dd) proposals. Does the quadratic case still yield an exponential lower bound for populations?
  2. Since St+1S_{t+1} may be any dataset with the prescribed mean, what distinguishes Algorithm 2 from privately estimating the mean and then reconstructing a dataset? Which part of the guarantee relies on the evolution structure?
  3. What does Proposition 5.1 add beyond the hardness results in [1] and [2]?

Suggestion:

  1. Could Figure 1 include a Gaussian-mechanism mean estimator as a reference?
1 comment

Thanks for the feedback. We address your concerns below.

Does the quadratic case still yield an exponential lower bound for populations?

It does. We have extended the quadratic lower bound to m≥1m\geq1, using the same three quadratic queries as the original theorem. The key observation is that any population different from 0m0^m whose error does not exceed the initial error must contain a nonzero record zz satisfying ∣⟨θ,z⟩∣≥d/2∥z∥2|\langle\theta,z\rangle| \geq \sqrt{d/2}\lVert z\rVert_2, where θ∈−1,1d\theta\in\\{-1,1\\}^d is the vector used to construct the quadratic queries. For any prescribed menu of LL additive variation laws, we choose θ\theta so that generating such a record is exponentially unlikely in dd, uniformly over all positive proposal scales.

As a consequence, starting from S0=0mS_0=0^m, after TT rounds of exact direct evolution with at most kk individual variations per record per round, we have P[direct evolution leaves 0m]≤TkmLde−d/18\mathbb{P}[\text{direct evolution leaves }0^m] \leq TkmLd e^{-d/18}. Hence, when T,k,mT,k,m, and LL are polynomial in dd, the method retains its initial error 3/673/67 with high probability, although the optimal error is zero. Thus, the obstruction persists for polynomially large synthetic populations. We have included this result and clarified that the original linear lower bound remains restricted to the singleton setting.

What distinguishes Algorithm 2 from privately estimating the mean and then reconstructing a dataset?

Our goal in this work is not to propose an algorithm that outperforms existing baselines, but to understand the limits of what (private) evolution can and cannot do. You are right, DP mean estimation followed by reconstruction works in this setting. However, we are not trying to solve mean estimation, per se. Our contribution is to show that an evolution-based procedure can overcome the obstruction faced by direct IPM selection and achieve a polynomial convergence guarantee.

Which part of the guarantee relies on the evolution structure?

The convergence guarantee does rely on Gaussian variations generating points that can be combined to make progress toward the target mean.

The rate O~(d2/3/(nε)1/3)\widetilde O(d^{2/3}/(n\varepsilon)^{1/3}) is also polynomially worse in nn than the Gaussian mechanism, roughly Rd/(nε)R\sqrt d/(n\varepsilon). The claim of statistically efficient PE is therefore too strong, and Figure 1 should include this reference.

Good point, we now include the Gaussian mechanism and avoid the term statistical efficiency and use "polynomial dependence on dimension" instead.

The reconstruction hardness result is not positioned against classic work. What does Proposition 5.1 add beyond the hardness results in [1] and [2]?

We are not sure which papers [1] and [2] refer to. We have added a discussion of known hardness results for DP synthetic data, including “PCPs and the Hardness of Generating Synthetic Data” by Ullman and Vadhan. Their lower bound concerns discrete data while we work with continuous data. Moreover, our result does not assume the existence of one-way functions. Finally, our result establishes NP-hardness of reconstruction even for two bounded convex quadratic queries. This rules out a generic polynomial-time reconstruction algorithm for quadratic IPMs but it does not imply hardness for every IPM. We have clarified this scope.

The simulation is only illustrative.

We have expanded the number of hyperparameter settings in the simulations. Specifically, we now include d=10,20,30d=10,20,30, and besides the original uniform distribution on a ball centered at zero, we consider three new sensitive-data distributions with nonzero means. The results are mixed, but generally show three main findings: (1) FW-PE has lower average mean error in every setting tested with the original distribution, while W1-PE performs better in many nonzero-mean settings, particularly at smaller sample sizes. (2) FW-PE performs better in some settings with larger sample sizes and larger privacy budgets. For example, at d=10d=10, n=80,000n=80{,}000, and ε=4\varepsilon=4, it has lower average error on all three nonzero-mean distributions. (3) As predicted by our theory, across all distributions, our method shows an overall decrease in average mean error as nn increases and degrades in larger dimensions.

The error of W1-PE at ε=1\varepsilon=1 increases from n=20,000n=20{,}000 to n=40,000n=40{,}000 without explanation.

The increase in Figure 1 is small relative to the Monte Carlo uncertainty shown by the error bars. We increased the number of runs to 50, and the mean error at n=40,000n=40{,}000 (0.09910.0991) is now lower than at n=20,000n=20{,}000 (0.10120.1012), although the difference remains small relative to the Monte Carlo uncertainty.

Proposition 3.1 is weak evidence of incompatibility.

The purpose of this proposition is to show that the same set of variations can improve one IPM while allowing no strict improvement for another; hence, a single example is enough to show that.

Typos.

Corrected

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

This paper investigates whether Private Evolution (PE) can leverage the structure of specific Integral Probability Metrics (IPMs) in order to obtain faster synthetic data convergence. A first result, Theorem 3.1, shows that for linear and quadratic queries, even non-private PE may need exponentially many variations over the records dimension to leave the initial state. Then, the paper shows that faster convergence can be obtained in the linear setting by optimizing a surrogate instead of the initial objective. A DP Franck-Wolfe algorithm allows to avoid the curse of dimensionality. Experiments demonstrate improved mean estimation compared with a Wasserstein PE baseline.

Critical strengths
  • The paper offers a useful distinction between the simplicity of an IPM and its compatibility with a variation mechanism. Theorem 3.1 yields an obstruction for non private PE about direct evolution. For linear queries, the paper constructs an alternative with a surrogate. Franck-Wolfe is well motivated here as a tractable selection procedure. The experiments also show loer mean error than the Wasserstein PE baseline.
  • The theory seems sound. I checked most of the proofs and did not find a substantial mathematical error in the paper.
Critical weaknesses
  • The relevance of the obstruction of Theorem 3.1 to larger syntethic datasets remains unclear. Theorem 3.1 establishes a valid lower bound when the synthetic dataset is a singleton, but the exponential dependence on dd remains to be clarified when the synthetic dataset has a larger size. In fact, in the linear setting, for every sign vector θ\theta, initializing with mm copies of θ\theta and generating one independent projected Gaussian variation per record yields, by concentration, a strict improvement with probability 1−2de−am1-2de^{-am} for some absolute constant a>0a > 0. Therefore, taking m≥O(log⁡(d/b)m \geq O(\log (d/b) may remove the described exponential dependence by allowing a first improvement with probability 1−β1-\beta. This does not rule out obstructions for larger synthetic dataset sizes, but this is an example supporting that the paper should clarify this limitation clearly.
  • The experimental setting is limited. While the dependence on dd is heavily discussed in the paper, the evaluation is only performed at d=10d=10. Moreover, X\mathcal X is in a centered ball of small radius. For nn independent variables uniformly distributed over a centered ball of radius RR, its empirical mean mm satisfies E[∥m(S)∥22]=dR2(d+2)n\mathbb E [\|m(S)\|_2^2] = \frac{d R^2}{(d+2) n}. At n=5000n=5000, the constant 00 estimator, which does not depend on data and therefore is (0,0)(0,0)-DP, has an expected mean error at most ≈0.02\approx 0.02. This analytical upper bound is substantially below the reported 0.120.12 value reported for FW-PE. I understand that this paper does not aim to compete against DP mean estimation algorithms. Nevertheless, experiments across dimensions and on private distributions with unknown nonzero means would provide stronger evidence of the proposed method’s usefulness.
Questions for authors
  • Can an analogous result to Theorem 3.1 be proven for larger synthetic datasets? How does the dimension dependence change with the size?
  • Does FW-PE still outperform Wasserstein PE over a more complete set of experiments, including varying the dimension?
1 comment

Thanks for your review. We address your concerns below.

The relevance of the obstruction of Theorem 3.1 to larger synthetic datasets remains unclear.

We have extended the quadratic part of Theorem 3.1. We use the same three quadratic queries as the original theorem. The key observation is that any population of size mm different from 0m0^m whose error does not exceed the initial error must contain a nonzero record zz satisfying ∣⟨θ,z⟩∣≥d/2∥z∥2|\langle\theta,z\rangle| \geq \sqrt{d/2}\lVert z\rVert_2, where θ∈−1,1d\theta\in\\{-1,1\\}^d is the vector used to construct the quadratic queries. Thus, even when selection can combine many proposed records, at least one proposal must be aligned with θ\theta or −θ-\theta. We choose θ\theta so that generating such a record is exponentially unlikely in dd, uniformly over all positive proposal scales. This necessary condition holds regardless of mm.

As a consequence, starting from S0=0mS_0=0^m, after TT rounds of exact direct evolution with at most kk individual variations per record per round drawn from a fixed menu of LL noise laws, we have P[direct evolution leaves 0m]≤TkmLde−d/18\mathbb{P}[\text{direct evolution leaves }0^m] \leq TkmLd e^{-d/18}. Hence, when T,k,mT,k,m, and LL are polynomial in dd, the method retains its initial error 3/673/67 with high probability, although the optimal error is zero. Thus, the obstruction persists for polynomially large synthetic populations, even if m≫nm\gg n. We have included this stronger result in the revised paper.

The experimental setting is limited.

We have expanded the number of hyperparameter settings in the simulations. Specifically, we now include d=10,20,30d=10,20,30, and besides the original uniform distribution on a ball centered at zero, we consider three new sensitive-data distributions with nonzero means. The results are mixed, but generally show three main findings: (1) FW-PE has lower average mean error in every setting tested with the original distribution, while W1-PE performs better in many nonzero-mean settings, particularly at smaller sample sizes. (2) FW-PE performs better in some settings with larger sample sizes and larger privacy budgets. For example, at d=10d=10, n=80,000n=80{,}000, and ε=4\varepsilon=4, it has lower average error on all three nonzero-mean distributions. (3) As predicted by our theory, across all distributions, our method shows an overall decrease in average mean error as nn increases and degrades in larger dimensions.

These results show that neither method uniformly dominates the other empirically. Our theoretical contribution is a guarantee tailored to the target IPM. Existing guarantees for the W1-PE baseline do not provide comparable rates for linear IPMs without incurring the dimensional dependence of full W1W_1 approximation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.