- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 3
Read review
Paper summary
This paper studies the problem of nonconvex optimization over Wasserstein space. In particular, it develops a notion of perturbation based on Hessian information to prevent gradient-flow-based methods from getting stuck at saddle points. This mirrors the developments in perturbation methods for nonconvex optimization in Euclidean space. The work is purely theoretical. The authors establish a finite-sample theorem for tracking the curvature of the objective, along with a result on a matrix factorization model showing, in more detail, the curvature locally around a stationary point.
Critical strengths
The work tackles an important problem of nonconvex optimization over the space of measures. Translating ideas from Euclidean optimization to optimization over measures is natural and can lead to many interesting applications. The theory developed naturally considers the geometry of Wasserstein space.
The substantial theory may contribute an important foundation to the literature. In particular, the study and application of Wasserstein Hessians is important.
Critical weaknesses
The paper is almost impossible for a human to read and understand. The main theorems don't seem to say anything directly about the actual method of interest - Hessian Perturbed Wasserstein Gradient Flow - in a direct way, and instead talk about tracking finite-sample bounds and local geometry in a population setting. But there is no signposting or roadmap that I can grab onto. No clear algorithm that can be referenced and explained. Furthermore, at a full 128 pages of material, it is impossible to check these results for a conference.
While I agree that optimization over measures is important, it is unclear how practical the results of this paper could be. Especially in the context of ICLR, I think that motivating practical applications is important. Does the developed theory change at all how we would approach any of the mentioned nonconvex optimization problems over Wasserstein space? Do we observe advantages?
Questions for authors
What applications is this work important for?
Can you comment on the main ideas of the proof?
Can you give a better, more human readable overview of the algorithm, theoretical framework, and important results in the paper?
3 comments
I think that the readability issue persists - for example you state that Theorem 3.1 discusses Algorithm 1, but Algorithm 1 is referenced nowhere in text. Instead, the reader is left to parse the assumptions and earlier discussion to try to parse what the theorem is about and can tell us. More discussion of the result itself would also help, especially at a higher/pedagogical level.
Highlighting your response to Q2 in text would be particularly useful.
Your response to Q3 is helpful, where you give a simple and intuitive explanation of the algorithm. But again I fail to see how the text clearly explains and motivates this.
Response to the follow-up comment by Reviewer LWEU: second revision
Thank you for reading the revision and for the suggestions. Based on them, we rewrote the explanations in Sections 2 and 3 and uploaded a second revision. The assumptions, the theorem and the proofs are the same as before. In the new PDF, the text added this time is in green, and the changes of the first revision are still in blue. The changes are listed below.
Section 2 (the method). We now explain what the method does before the formal definitions.
- The section now starts with a short description of the particle method. It is similar to our response to Q3.
- Algorithm 1 is now in Section 2.2 (it was in Appendix C.1). We refer to it at the start of Section 2, in Section 2.2 and before Theorem 3.1.
- The subsections have new titles. Each subsection starts with one sentence that says what it defines. Section 2.3 explains why we use reference particles.
Section 3 (the theorem). We now give the main idea of the bound first, and explain each formal statement briefly.
- The start of the section explains why the bound uses the curvature along the path and not a Lipschitz constant.
- Section 3.1: we added a short explanation after each assumption.
- Section 3.2: we added an explanation of the two factors after Definition 3.1. It is based on our response to Q2. It explains how the error changes between jumps and at a jump.
- Section 3.3: the paragraph before Theorem 3.1 explains what the theorem shows. It is based on our response to Q3. After the theorem, we discuss the bound in two simple cases and explain what it suggests in practice.
To Reviewer LWEU
We thank the reviewer for the comments and questions, which led us to reorganize the main text.
W1. We are sorry that the paper was long and hard to follow. In the revision, the assumptions are stated before Theorem 3.1, and the proof overview comes right after the theorem and its example. The Introduction now ends with an outline of the paper. The method is written out as Algorithm 1 in Appendix C.1.
The method in Theorem 3.1 is PWGF run with N particles (Algorithm 1), and it uses only the particles. Under Assumptions 3.1–3.3 and a jump schedule fixed in advance, the theorem bounds the distance rt between these particles and reference particles drawn from the population PWGF of Yamamoto et al. (2025), and the gap in objective value ∣F(ρXt)−F(μt)∣. Section 5 studies the population-level geometry of a matrix-factorization model. It motivates the curvature scenario of Figure 1 and is separate from the theorem.
W2 and Q1. Nonconvex optimization over probability measures appears in several learning problems. Kim & Suzuki (2024) study in-context learning with an attention layer and a three-layer teacher–student network. Yamamoto et al. (2025) give matrix decomposition and a three-layer network as examples, and our Section 5 studies a product-measure version of the matrix problem. Yamamoto et al. (2025) proposed PWGF to escape saddle points in such problems and proved guarantees for the flow of measures. A run uses finitely many particles, and our paper bounds the gap between the particle run and the flow of measures when the jump schedule is fixed in advance.
Advantages over other methods were observed by Yamamoto et al. (2025). In their experiments on the in-context learning objective of Kim & Suzuki (2024) and on matrix decomposition, run with their Algorithm 2, PWGF reduced the loss faster than gradient flow without noise, and as fast as or faster than isotropic noise. Our experiments (Appendix C) measure how closely a particle run follows the population flow. With a common prescribed schedule (Appendix C.4), the median of rt over trials has a negative log–log slope in N at every recorded time.
Our bound weights sampling errors by the curvature accumulated along the path. This suggests that a run needs more particles when the path spends a long time in negative curvature, and that earlier errors matter less after the curvature turns positive.
Q2. We compare the particles Xi with reference particles Yi. The Yi start from the same points and move with the population drift and the population jumps. Given the population fields, they are i.i.d. samples of μt, so concentration inequalities apply to them. At each jump the two systems are driven by one common Gaussian vector (Lemma A.14), and the two jumps then differ only through their covariance factors.
Between jumps, linearizing the two ODEs gives, to leading order, r˙t≤−κtrt+Γt. Here κt is the smallest eigenvalue of the Hessian sampled along the reference particles, and Γt is a sampling error of order N−1/2. By this inequality, an error created at time s arrives at time t multiplied by exp(−∫stκu,du). It grows while κu<0 and shrinks while κu>0. We keep this weight in the bound, and its integral over s is the factor CtY of Theorem 3.1.
At a jump, the past error is multiplied by a bounded factor and a new error is added. The new error is bounded by the square root of a covariance mismatch of order N−1/2 (Lemma A.13), which gives the N−1/4 term. Jump errors are carried forward with the same weight, and their sum is the factor JtY.
Q3. The algorithm alternates two steps. Between jump times, each of the N particles follows the gradient flow of F evaluated at the particles' empirical measure. At each jump time, one Gaussian field is drawn, with covariance given by the squared Hessian of F at the empirical measure, and each particle is shifted by ηℓ times the value of the field at its position.
We compare this particle system with the population PWGF μt through rt, the root-mean-square distance between the particles and coupled samples of μt, and through the gap in objective value.
Theorem 3.1 bounds both quantities by N−1/2(1+CtY)+N−1/4JtY, up to small powers of N. The first term comes from errors made between jumps and the second from errors made at jumps, each weighted by the curvature along the path as in Q2. The theorem holds under Assumptions 3.1–3.3, on an event of probability at least 1−p, for T≤N1/4−υ−ζ, and for reference paths on which the two factors satisfy the stated bounds. Example 3.1 checks all conditions in a variance-plus-cosine model. There they hold with probability at least 1−N−1/16, and the error is at most N−3/16+ε up to time N1/8.