Efficient Sampling of Stochastic Differential Equations with Positive Semi-Definite Models

This paper deals with the problem of efficient sampling from a stochastic differential equation, given the drift function and the diffusion matrix. The proposed approach leverages a recent model for probabilities \cite{rudi2021psd} (the positive semi-definite -- PSD model) from which it is possible to obtain independent and identically distributed (i.i.d.) samples at precision $\varepsilon$ with a cost that is $m^2 d \log(1/\varepsilon)$ where $m$ is the dimension of the model, $d$ the dimension of the space. The proposed approach consists in: first, computing the PSD model that satisfies the Fokker-Planck equation (or its fractional variant) associated with the SDE, up to error $\varepsilon$, and then sampling from the resulting PSD model. Assuming some regularity of the Fokker-Planck solution (i.e. $\beta$-times differentiability plus some geometric condition on its zeros) We obtain an algorithm that: (a) in the preparatory phase obtains a PSD model with L2 distance $\varepsilon$ from the solution of the equation, with a model of dimension $m = \varepsilon^{-(d+1)/(\beta-2s)} (\log(1/\varepsilon))^{d+1}$ where $1/2\leq s\leq1$ is the fractional power to the Laplacian, and total computational complexity of $O(m^{3.5} \log(1/\varepsilon))$ and then (b) for Fokker-Planck equation, it is able to produce i.i.d.\ samples with error $\varepsilon$ in Wasserstein-1 distance, with a cost that is $O(d \varepsilon^{-2(d+1)/\beta-2} \log(1/\varepsilon)^{2d+3})$ per sample. This means that, if the probability associated with the SDE is somewhat regular, i.e. $\beta \geq 4d+2$, then the algorithm requires $O(\varepsilon^{-0.88} \log(1/\varepsilon)^{4.5d})$ in the preparatory phase, and $O(\varepsilon^{-1/2}\log(1/\varepsilon)^{2d+2})$ for each sample. Our results suggest that as the true solution gets smoother, we can circumvent the curse of dimensionality without requiring any sort of convexity.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer jfmV6/10 · confidence 2/52023-06-24

Summary

This article investigates the approximation of the Fokker-Planck equation and the fractional Fokker-Planck equation using positive semi-definite (PSD) models. The authors demonstrate that for a given accuracy $\epsilon$, there exists a PSD model with a dimension not exceeding $\epsilon^{-(d+1)/(\beta - 2s)}(\log 1/\epsilon)^{d+1}$ that approximates the corresponding partial differential equation (PDE). By utilizing this result in conjunction with the PSD model-based sampling method proposed in [1], the authors are able to provide samples of the stochastic differential equation (SDE) corresponding to the Fokker-Planck equation or fractional Fokker-Planck equation at any given time $t$. [1]Marteau-Ferey U, Bach F, Rudi A. Sampling from arbitrary functions via psd models[C]//International Conference on Artificial Intelligence and Statistics. PMLR, 2022: 2823-2861.

Strengths

1. This article demonstrates clear and logical writing, making it easy to understand the problem formulation considered by the authors as well as the fundamental concepts and properties of the PSD model adopted. 2. As a newly proposed model, this paper presents a novel application scenario for the PSD model, which has not been previously explored in related works. This contributes to its significant novelty. 3. From the perspective of sampling based on SDEs, the authors provide a new design approach for sampling by combining existing works, offering a fresh perspective on sampling techniques.

Weaknesses

1. Although the author mentions it in the paper, the influence of initial conditions on the solution of the PDE can sometimes be significant. It would enhance the completeness of the theoretical analysis in this article if the author could provide an analysis of whether the PSD model can satisfy the initial conditions or, in cases where exact satisfaction is not possible, characterize the approximation of the initial conditions' impact on the final solution's approximation. 2. Personally, I have limited knowledge of non-log-concave sampling and numerical modeling of PDEs. However, it would be beneficial if the author could compare the effectiveness of the PSD model-based sampling method with other traditional methods (such as Langevin Monte Carlo) in terms of sampling performance. Additionally, comparing the approximation methods of the PSD model for PDE solutions with other existing approaches, either from the perspective of complexity analysis or purely experimental results, would provide valuable insights.

Questions

1. In line 180, should it be $E e^{j\langle u, L_{t-s}^\alpha\rangle} = \exp\{-(t-s)||u||_2^\alpha\}$? 2. In line 216, should it be $Y \in \mathbb{R}^{n\times d^\prime}$?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Despite the theoretical guarantees provided by the author, I believe the biggest limitation of this article lies in the lack of clear demonstration of the advantages of the PSD model. It would be beneficial if the author could explain how the PSD model stands out compared to other similar models in terms of approximating FPE and subsequent sampling tasks. This would significantly enhance the persuasiveness of the article.

Reviewer T6Bn7/10 · confidence 3/52023-07-04

Summary

The authors introduced a new method of sampling solutions of the Fokker--Planck equation (FPE) using a positive semidefinite (PSD) model. While the algorithm used for the PSD model is not new, the authors created a framework to sample from the FPE, which is quite original based on my knowledge. While I believe the results are very neat, I would like to clarify several questions with the authors before providing a confident score. For now I will recommend borderline accept, and will raise the score once my questions are adequately addressed.

Strengths

The approach to drawing samples for the Fokker--Planck is very unique and original. I believe alternative methods that hold promise should be welcomed and more carefully studied, regardless of how practical it is at the moment.

Weaknesses

There are a few clarifications I would like to have regarding the implementation of the algorithm.

Questions

1. To start, while I understand the algorithm of [MFBR22] is not part of the contribution for this work, can the authors provide a high level explanation of how the algorithm works, and comment on its computational complexity? I believe all readers can benefit from understanding the core algorithm, and plus I believe readers should be able to implement the algorithm presented in this paper without having to reference other works. 2. Can authors clarify the role that $q$ plays in this problem? In particular, is $q$ given or chosen by the user, and how does the value of $q$ affect the computational complexity of this algorithm? 3. Under the discussion of "Efficient sampling method" the authors mentioned a step of finding the best model $\hat{A}_m$ by minimizing an error function. Suppose I am an applied scientist who wants to sample from a given FPE, how would I compute this model $\hat{A}_m$ from the PDE? On a high level, I would like to understand how to implement this sampler end to end, and I don't think I fully understand this part. Once again, I would be happy to raise my score once my questions are adequately answered.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer oceV7/10 · confidence 4/52023-07-06

Summary

This paper considers numerical methods for the Fokker-Plank equation and its "fractional Laplacian" version. These partial differential equations describe the evolution of densities of certain random processes: in the case of standard FP, the process is a regular diffusion (an SDE driven by Brownian motion), whereas the fractional version corresponds to a process driven by $\alpha$-stable noise. Accordingly, the authors focus on initial conditions given by probability densities. The idea behind the paper is to adapt the fairly recent methodology of positive semidefinite models (PSD) to this PDE problem. PSD models estimate nonnegative functions via feature maps $\phi:X\to H$, where $X$ is the domain of the function and $H$ is a Hilbert space. Given $\phi$, the nonnegative function of interest is estimated as $\langle \phi, M\phi\rangle$, where the PSD linear operator $M$ is chosen from the data. The whole method is based on the same idea as the well-known kernel trick: there is no need to explicitly deal with the Hilbert space $H$; only the kernel matrix over data points will matter. In a nutshell, the paper gives a method for approximately solving the two aforementioned PDEs via PSD models. The salient features of this approach are as follows. 1) The PSD estimator can be obtained via a semidefinite program (as shown in previous work). 2) The method will output probability densities, and one can sample from them via methods from previous work. 3) The complexity of the problem grows like $\approx \epsilon^{-d/\beta}$, where $\epsilon$ is the target accuracy, $d$ is the dimension and $\beta$ measures the degree of smoothness of the solution. The exponent is not quite right, but the point is that sufficient smoothness will avoid the curse of dimensionality. One may note that this method has rigorous error guarantees. This sets it apart from other PDE solvers based on machine learning ideas, such as Physics Inspired Neural Nets (PINNs).

Strengths

The paper gives strong convergence guarantees for a new PDE solving method. Although several ingredients come from previous work, they are combined with new ideas into an impressive package. As mentioned above, this is a ML-based PDE solver with rigorous guarantees. Importantly, these guarantees do not require coercivity or hypercontractivity properties that are usually required from Langevin samples for solutions.

Weaknesses

* No experiments are performed to compare the method with other PDE solvers (though this is understandable in a heavily mathematical paper). * The writing is a bit sloppy in places: I found quite a few typos and have several minor comments on writing (I could probably find some more typos if I kept looking). See the list below. NB: were it not for the sloppiness, I would have given this paper a higher score. --- _List of typos and comments on writing_ (Line numbers refer to the full paper in the SM.) Lines 146/7: "It was" and "It has" should be "They were" and "They have". Line 180: why define $j=\sqrt{-1}$ if you could have just written $\sqrt{-1}$ in the exponent? (You never use this notation again, do you?) Line 185: Shouldn't the condition for infinite $p$-th moment be $p\geq \alpha$? Line 201 and elsewhere: I find it a bit odd that the transpose notation is used for vectors in Hilbert space. If you are going to use it, at least explain here what is going on. Line 251: why is this sensence needed here? Line 205: $\eta$ was treated as a vector before, and will be treated as a vector later in the text. Here, however, it seems to be a scalar. Line 223: "The cost of the algorithm is..." -- per sample point, right? Line 254: "Hence, $p^\star(x,t)$" Line 255 should end with ":". Line 270: the definition of $\psi$ should be in a numbered display for easier refererence. Display (8): if $p^*$ is the true solution, why do we have it in here? (Its contribtion vanishes.) Indeed, the definition of the loss later does not involve $p^*$; see eg. (9). (On the other hand, $p^\star$ shows up again in the display right after line 989.) Line 294 and elsewhere: when you are mentioned a numbered theorem or assumption, start the corresponding word with a capital letter. Line 295: $\mathcal{H}_X\otimes\mathcal{H}_T$ Line 318: "by Gaussian kernel approximation" -- should this be here? Line 365: "Similarly to the previous section, two steps are required to obtain" Line 327 and elsewhere: I think $\mathbb{E}_t$ was not defined anywhere. Line 352: "If the kernel satisfies the Bochner..." Line 355 and elsewhere: this is a matter of taste, but couldn't you replace "utilize" with "use". Line 376: "The major difficulty" Line 391 (taste): "does require" could be "require". Line 395: I think "approximated" should be "approximating". Proposition 4 is also in Evans's book, which was cited previously (I am saying this from memory, so please check. $W_2$ is replaced by $H$ in that book.) Many spots of the SM: the inequality $\|f\star g\|_p\leq \|f\|_p\|g\|_1$ should be mentioned, as it is used in several steps. Lemma 2: it seems that $\tilde{p}$ is replaced by $f$ in the statement of the Lemma and at several steps of the proof. (Eg. the display right after 665.) This also seems to take place in the next Lemma. Line 680: Since you are repeating a definition from the preceding proof, please say so. Line 682: "We use the result of Lemma 1" Line 746: Should "The following result holds" be here? Display (34): this "fill distance" is the smallest value of $h$ such that $\tilde{X}$ is an $h$-net of $[-R,R]^d$. Display below line 777: how should I interpret the norm in the RHS? Lines 873/4: why are you spaeking of conditional probabilities here? Line 891: what is $f(\delta)$? Line 910: "redefine" should be "define again" (they are not the same). Line 928: capitalize "Holder".

Questions

* Have you attempted experiments against other ML approaches, or other methods from the literature? * Methods have been proposed recently to speed up SDPs, including randomized sketches (https://math.paperswithcode.com/paper/scalable-semidefinite-programming). Could these methods be applicable in the present setting?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

4 excellent

Limitations

They have not addressed limitations explicitly, but I do not think that that would have been necessary.

Reviewer biEz3/10 · confidence 2/52023-07-07

Summary

This paper studies the problem of efficiently sampling from a SDE given a drift function and diffusion tensor. It proposes a solution based on computing a PSD model that satisfies the Fokker-Planck solution associated with the SDE, and then sampling from the resulting PSD model.

Strengths

The paper seems to have a lot of technical content, and it is clear that the authors have put in a lot of effort.

Weaknesses

This paper is extremely difficult to read. The theorem statements (and the paper overall) need to be drastically simplified to be understandable. The problem doesn't seem to be well motivated, and it's not clear from reading the paper how the solution compares to others in the literature. For instance, the [LCL+21] paper is mentioned in related work, but it's not at all clear how the solution proposed of learning a PSD model compares to their work. There are no experiments to support this idea of fitting a PSD model. Overall, I think this paper needs to be rewritten to be more easily understandable.

Questions

- Can you provide experimental evidence for the theoretical claims? - Can you provide simplified theorem statements? - Can you provide a more clear comparison with related work?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

Yes

Authorsrebuttal2023-08-11

Start of the author reviewer discussion

We hope that our response answers the questions the reviewer asked. Please let us know the further questions or confusion about this work so that we can address it. If we have answered all the questions from the reviewer, then in light of all the novel contributions we have made in this paper, we would like to request the reviewer to reevaluate the score assigned to this paper.

Authorsrebuttal2023-08-14

Reminder: Start of the author reviewer discussion

Please let us know if you have further questions or confusion about this work that we can address. If we have answered all the questions from the reviewer, then in light of all the novel contributions we have made in this paper, we would like to request the reviewer to reevaluate the score assigned to this paper.

Reviewer T6Bn2023-08-16

Response

I apologize for my delayed response, I have been traveling and wrestling with a heavy review load this year. I'm sorry if this created additional stress for the authors. Indeed, I believe my concerns have been addressed, and I will raise my score to accept. I find this work quite interesting, and I hope the authors continue to develop this topic further.

Authorsrebuttal2023-08-11

Start of the author reviewer discussion

We hope that our response answers the questions the reviewer asked. Please let us know the further questions or confusion about this work so that we can address it. If we have answered all the questions from the reviewer, then in light of all the novel contributions we have made in this paper, we would like to request the reviewer to reevaluate the score assigned to this paper.

Authorsrebuttal2023-08-14

Reminder: Start of the author reviewer discussion

Please let us know if you have further questions or confusion about this work that we can address. If we have answered all the questions from the reviewer, then in light of all the novel contributions we have made in this paper, we would like to request the reviewer to reevaluate the score assigned to this paper.

Authorsrebuttal2023-08-11

Start of the author reviewer discussion

We hope that our response answers the questions the reviewer asked. Please let us know the further questions or confusion about this work so that we can address it.

Reviewer oceV2023-08-11

Thank you

Thank you for the rebuttal. I do want to insist that it would have been nice to see experiments, though I understand that math is more important here.

Authorsrebuttal2023-08-14

RE: Thank you.

Thanks a lot for the acknowledgment.

Authorsrebuttal2023-08-11

Start of the author reviewer discussion

We hope that our response to the reviews (all the reviews) provided more clarification on the contributions we made to this work. Please let us know the further questions or confusion about this work so that we can address it. Please also look at other reviews and our responses to the questions asked by other reviewers. In light of all the novel contributions we have made to this paper, we would like to request the reviewer to reevaluate the score assigned to this paper.

Authorsrebuttal2023-08-14

Reminder: Start of the author reviewer discussion

Please let us know if you have further questions or confusion about this work that we can address. If we have answered all the questions from the reviewer, then in light of all the novel contributions we have made in this paper, we would like to request the reviewer to reevaluate the score assigned to this paper.

Reviewer biEz2023-08-16

I reread the paper after reading your comments. Currently I am not at all convinced by the story provided in this paper. The PSD model doesn't seem to be well-motivated, beyond statements like "PSD models are an effective way to model non-negative functions and enjoy the nice properties of linear models and more generally RKHS" which are provided without justification. The theorems and effectiveness of the solution are difficult to interpret. I would be much more convinced if you could: 1. Provide experiments that demonstrate distributions where this algorithm performs much better than simpler algorithms like Langevin Monte Carlo 2. Provide theoretical examples where Langevin Monte Carlo provably fails to sample correctly, while your proposed algorithm succeeds. In particular, even though you claim: "Remark. It is evident from the aforementioned result that our method excels in sampling from the Stochastic Differential Equation (SDE) while ensuring the particle density adheres to Assumption 1 within a bounded domain. In contrast, traditional sampling methods such as Langevin struggle to sample from this extensive class due to the absence of dissipativity, log-sobolev, and Poincaré-like conditions.", this seems like it could possibly just be a failing of the existing analyses of Langevin, rather than a failing of the actual algorithm. It would be much clearer if you could provide a specific example, along with a proof that your algorithm succeeds on this specific example, but Langevin fails. 3. Provided more easily interpretable theorem statements with direct comparisons to simpler algorithms (langevin). Currently the theorem statements are very difficult to interpret. For the reasons above, I maintain my score.

Authorsrebuttal2023-08-16

Dear Reviewer, Thanks a lot for your response. We believe there has been a misunderstanding that we would like to clarify. 1. PSD Models for Probability Representation: We dedicated an entire section 3.2 to the effectiveness of using PSD Models for Probability Representation. It is clear from the formulation in Equation 5 that f is a non-negative function. Proposition 1 in the main paper says that marginalization is possible in closed form for the PSD model with a Gaussian kernel. Please also look at Proposition 3. We are not aware of another representation that comes with these guarantees. This has also been discussed in detail in [RC21] and [MFBR20]. Please also look at the paper [RMFB20]. 2. Theorem statements: We clearly explained in our previous rebuttal that what are the implications of each of the theorem statements. Our proposed approach works in two parts. In the first approach, we approximate the (fractional)-Fokker Planck equation and we provide the guarantee for that in our theorems. If you are familiar with fokker planck and fractional Fokker Planck equations (section 3.1 in the paper) then the theorem statements are really saying that PSD-based approximation can approximate the solution of the fokker planck equation well. The theorem statements are currently in their simplest form stating the approximation error guarantee. The sampling result is given in Corollary 1. 3. Theoretically, we are not aware of any prior work that doesn't utilize either log-Sobolev or Poincare-type inequality to provide convergence for Langevin monte carlo. Even the latest results in this area (COLT 2023) rely on some kind of Poincare or log-Sobolev inequality to be satisfied. (i) Towards a Complete Analysis of Langevin Monte Carlo: Beyond Poincar\'e Inequality: Mousavi-Hosseini et al. (ii) Utilising the CLT Structure in Stochastic Gradient-based Sampling: Improved Analysis and Faster Algorithms: Das et al. There is no one-on-one comparison between Poincare or log-Sobolev type inequality and our condition (that is beta times differentiable). Proposition 3 in our paper describes the extent to which the PSD model can approximate. Even if it is the artifact of analysis for Langevin that sampling guarantee does not hold when Poincare or log-Sobolev inequality fails to hold, our result theoretically improves upon existing sampling results. In that sense, our work is a valuable contribution. Log-Sobolev/Poincare inequality does not hold for the class of density discussed in Proposition 3. 4. We have already provided clarification regarding the experiments that we would reiterate. There is a significant lack of literature on solving the (fractional) Fokker-Planck equation and sampling from the solution. Our framework not only addresses this problem but also allows for the approximation of the fractional Laplacian acting on density, which is a distinct research problem. We believe that our work represents a valuable contribution to the field and has the potential to inspire further research in this area. The theoretical results presented in this paper are non-trivial and come with worst-case guarantees, which makes them stand out from methods that rely on stronger assumptions. Our findings are particularly significant in situations where traditional assumptions, such as log-Sobolev and Poincare fail. Despite the absence of real-world experiments, we hope that the reviewer could re-evaluate our theoretical contributions from the perspective of not requiring strict assumptions and the novel contributions we make in multiple research problems. Regards, Authors.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC