Kernel quadrature with randomly pivoted Cholesky

This paper presents new quadrature rules for functions in a reproducing kernel Hilbert space using nodes drawn by a sampling algorithm known as randomly pivoted Cholesky. The resulting computational procedure compares favorably to previous kernel quadrature methods, which either achieve low accuracy or require solving a computationally challenging sampling problem. Theoretical and numerical results show that randomly pivoted Cholesky is fast and achieves comparable quadrature error rates to more computationally expensive quadrature schemes based on continuous volume sampling, thinning, and recombination. Randomly pivoted Cholesky is easily adapted to complicated geometries with arbitrary kernels, unlocking new potential for kernel quadrature.

Paper

Similar papers

Peer review

Reviewer PPmk6/10 · confidence 3/52023-06-29

Summary

The kernel quadrature problem computes quadrature weights and nodes that minimize the maximum quadrature error over all functions in RKHS with the norm bounded by 1. This paper proposes new quadrature rules for functions in RKHS using nodes drawn by randomly pivoted Cholesky algorithm. The randomly pivoted Cholesky used in the paper is an extension of the finite variant introduced in [10]. The authors provide the unoptimized implementation followed by the variant with efficient rejection sampling.

Strengths

1. Quadrature problem is one of the fundamental problems in computational mathematics and the paper proposes an efficient method for it. 2. The experiments offer comprehensive comparisons against other state-of-the-art methods. 3. Theorem 1 provides error bounds which achieve spectral accuracy with just logarithmic number of nodes. 4. The proposed method (both the unoptimized and the optimized variants) is easy to implement from the description.

Weaknesses

1. There are no empirical results on real datasets despite earlier work such as [19] having results on real datasets. 2. Discussion on the application of the proposed method in machine learning is largely missing.

Questions

1. For the right side of the Figure 2 (time vs number of nodes): what is the error level considered for all methods? 2. Could the authors comment on how the results could change for other Mercer kernels? 3. How about the effect of dimensionality on the empirical results? The reference [19] does offer more parameter settings used for evaluation. 4. Seeing that Algorithm 2 and Algorithm 3 stores Cholesky factors, could the authors comment on scalability of the method as the number of quadrature points is increased?

Rating

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

There are no results on real datasets and even on synthetic datasets, the results are limited to one type of kernel. Although it is clear the proposed method has wide applications in machine learning, the authors should demonstrate a usecase in which the proposed method can be applied (e.g. scaling kernel machines).

Reviewer D4rs8/10 · confidence 4/52023-07-03

Summary

This paper introduces a new method for kernel quadrature which relies on a randomly pivoted Cholesky decomposition. The sampling of the quadrature points is done thanks to a rejection sampling algorithm. To improve performance, an optimized rejection sampling procedure is also introduced, so that the acceptance probability is updated if the number of rejections becomes too large. Numerical simulations show that this method is fast and often as accurate as other performant methods. An error bound for the expectation of the quadrature method is also derived.

Strengths

- This paper is very well written. It mentions other important methods for kernel quadrature. - The main idea of this research work is simple and elegant. - The theoretical results are well introduced and the proofs are clear although non-trivial. - From the practical perspective, the numerical simulations are convincing. There is a clear gain in sampling time, in particular with respect to continuous volume sampling which is a closely related method.

Weaknesses

There are very few weaknesses in this paper. - It seems that the proof techniques are mainly inspired from [10] to which the authors refer. This is not an issue and does not limit the significance of this contribution. - The end of the proof of proposition 3 is not very explicit. However, the authors provide the main elements to finish the proof so that the reader can complete the argument on a sheet of paper.

Questions

- Concerning the simulations reported in Fig 2 (a) (left-hand side), I wonder if there is a numerical issue in the computation of the curve corresponding to continuous volume sampling (CVS). It seems to me that CVS should have at least a similar performance as RPCholesky. However, for the last two data points, CVS suddenly has a worse performance (following the IID case). This phenomenon is not happening in Fig 2 (b) (left-hand side). Hence, I wonder if there could be an issue in the implementation of CVS such as a numerical instability. Could you comment about that ?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

4 excellent

Contribution

3 good

Limitations

The authors mention possible limitations.

Reviewer yiWH6/10 · confidence 3/52023-07-19

Summary

The authors provide theoretical analysis and show that RPCholesky has near optimal error rates. Meanwhile, the authors propose to combine RPCholesky with rejection sampling in order to use it in the continuous setting. Experiments show that the proposed method is compared favorably to continuous volume sampling which is much more difficult to sample.

Strengths

The paper is clearly written. Both assumptions and theorems are well presented and discussed. The experiments demonstrate the good performance of the proposed method.

Weaknesses

For ''Err'' under line 140 and ''Err'' under line 141, it’s better to include the derivation in the appendix to make the paper self-contained. The description in Line 226 and Line 227 is unclear. For instance, “opt” is imprecise. The proof of the theorem in Section 3.4 can be moved to the appendix. It seems not necessary to include it in the main paper. I think the biggest weakness of the paper is: the experiments are very limited. I would like to see more experiments as the authors claim that their method is very general.

Questions

Can the authors include additional experiments and show that the proposed method can also perform well in practice, e.g. real-world experiments? Is this something difficult to implement?

Rating

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

2 fair

Limitations

The limitations are discussed in Section 6.

Reviewer 4xjP7/10 · confidence 2/52023-07-28

Summary

The recently proposed Randomly Pivoted Cholesky procedure is an efficient and accurate method for approximating symmetric PSD (e.g. kernel) matrices. It builds a Nystrom approximation via randomizing the choice of pivot in the classical Cholesky decomposition procedure. The authors of the present submission transport the same sampling idea to the continuous case and develop new kernel quadrature rules. The proposed algorithm is based on rejection sampling, it's simple to implement and it's much faster than many other kernel quadrature procedures. Rigorous theoretical guarantees and experimental analysis with synthetic data accompany the algorithm.

Strengths

1) Solid theory. The error of the quadrature depends on the eigenvalue distribution of the kernel (it's spectrally accurate). 2) The rejection sampling algorithm is practical to implement and it's efficient. 3) Empirical evaluation supports the claims.

Weaknesses

1) As the authors themselves point out in Section 6 all spectrally accurate quadratures, including this, currently require the exact computation of the Hilbert–Schmidt integral of function $g$, which may limit the applicability of the method. E.g. the experiments are all based on $g\equiv 1$. 2) Although the first sentence of the introduction in lines 12-13 refers to several applications all the experiments in Figures 1 and 2 rely on computing the worst case error of kernel integrals in low dimensions on a given domain with a particular (complicated) kernel. Albeit this style of synthetic evaluation seems common in the cited prior work I believe it would increase the impact of the paper if the authors demonstrated end to end improvements on an actual machine learning tasks that requires computing these integrals. Including a concrete application would boost the accessibility and motivation for non experts reading the paper as well in my opinion.

Questions

1) The analysis and experiment focus on and minimize the worst case error (line 140). What, if anything, would change if you considered some expected error (loss) when $f$ is drawn from some distribution? 2) Since the experiments in Figure 2) are rather low dimensional ($d=1$ and $d=3$) could you also include a Quasi-Monte Carlo baseline with equal and optimal (eq (5)) weights similarly to current Monte Carlo and IID? Since QMC is not spectrally accurate one would still expect RPCholesky to have much lower error, the gap might shrink though. Could you also consider running experiments with much higher $d$?

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

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Theory paper, societal impact is not applicable. The authors do a thorough job in disclosing and discussing technical limitations.

Reviewer dsgV4/10 · confidence 5/52023-07-28

Summary

This article studies a family of quadrature rules suited for functions that belongs to an RKHS. The proposed construction is based on an algorithm called randomly pivoted Cholesky (RPCholesky). The latter was studied for the task of approximating positive-semidefinite matrices. In particular, the authors extended RPC to continuous domain, and investigated its effectiveness for kernel quadrature. More specifically, the authors studied the quadrature rule obtained by taking the nodes to be the configuration of points that RPCholesky and by taking the weights that minimises the worst-case integration error on the unit ball of the RKHS (WCE). In particular, they proved that the expected value of the squared WCE of this quadrature rule converges to 0, with a rate of convergence that depends on the eigenvalues of the integration operator. Moreover, they studied the expected runtime of the proposed algorithmes. The contributions of the article may be summarised as follows:  * An extension of the RPCholesky algorithm to a continuous domain (Algorithm 1) along with a detailed implementation of it using rejection sampling (Algorithm 2). * An implementation of RPCholesky with optimized rejection sampling (Algorithm 4). * A generic result (Theorem 1) gives the upper bound of the expected squared WCE for the proposed quadrature rule. * The study of the expected runtime of Algorithm 2 (Theorem 4).

Strengths

- The article is very well written. - The article expands the application of the randomly pivoted Cholesky algorithm into the continuous domain, establishing another bridge between the numerical linear algebra community and the kernel quadrature community. - The proof of the main result (Theorem 1) is original and may be beneficial for the community of kernel quadrature in the future.

Weaknesses

At first sight, the proposed algorithm seems to bring a significant algorithmic improvement upon continuous volume sampling. This is at least what we may understand reading the comments of Figure 2 (line 235-line 238). While, the relative improvement is undeniable, RPCholesky suffers from the same limitation of continuous volume sampling that is poor scalability for large values of $N$, especially for smooth kernels. Figure 2 of the article is corroborating this claim: the dependence of the empirical expected time as a function of $N$ seems to be polynomial. This is particularly the case in dimension 1, where we have $\lambda_m = \mathcal{O}(m^{-2s})$ so that $1/\sum_{m=N+1}^{+\infty}\lambda_{m} = \mathcal{O}(N^{2s+1})$. In other words, the smoother is the kernel the longer it will takes to sample from the RPCholesky, which is exactly the same problem with Gibbs sampling proposed in [1], which provably approximate continuous volume sampling. It turns out that Algorithm 1 is nothing but the algorithm proposed in [1] to initialise an MCMC algorithm that approximate continuous volume sampling; see Algorithm 1 in [1]. The difference between the two works is in the subroutine that consist in sampling from the probability measure $x \mapsto k(x,x) \mathrm{d}\mu(x)/\int_{\mathcal{X}}k(u,u)\mathrm{d}\mu(u)$: [1] proposed to use a ‘stochastic greedy’ algorithm using rejection sampling, while this article proposed to use RPCholesky with rejection sampling. Interestingly, the two algorithms have an expected time that depends on $1/\sum_{m=N+1}^{+\infty}\lambda_{m}$, where is the number of nodes in the quadrature rule, see Lemma 5.1 in [1]. In other words, this work does not break the ‘curse of smoothness’ that hinders the scalability of continuous volume sampling.
 [1] Rezaei, A., & Gharan, S. O. (2019, May). A polynomial time MCMC method for sampling from continuous determinantal point processes. In International Conference on Machine Learning (pp. 5438-5447). PMLR. [2] Anari, N., Gharan, S. O., & Rezaei, A. (2016, June). Monte Carlo Markov chain algorithms for sampling strongly Rayleigh distributions and determinantal point processes. In Conference on Learning Theory (pp. 103-115). PMLR.

Questions

- The empirical comparison in section 5 was limited to Sobolev spaces of order s=3. It may be interesting to compare RPCholesky to other algorithms on other RKHSs. - The empirical experiments make use of a continuous analog of a ’basis-exchange’ MCMC algorithm proposed in [2] instead of the MCMC algorithm based on Gibbs sampling proposed in [1]. Up to my knowledge, only the latter was studied theoretically, and it is possible that the convergence of ’basis-exchange’ MCMC algorithm is slower than the Gibbs sampling.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

None

Reviewer PPmk2023-08-14

Thanks to the authors for the feedback

The feedback I received from the authors is very helpful in clarifying some parts of the paper. In particular, the QM dataset application (chemical engineering) is an interesting application which could strengthen the paper. I think in order to gain more tractability for general machine learning applications, the authors should pursue how to scale to larger values of $n$. I will maintain my current score as the authors addressed all of my concerns.

Reviewer 4xjP2023-08-21

Thanks so much for the detailed reply and for adding the experiment with the chemistry task. I believe all these, especially the latter, strengthen the submission further.

Reviewer yiWH2023-08-21

Thank you for the response. I will increase my score.

Reviewer dsgV2023-08-21

Thank you for all the clarifications that you have provided.

© 2026 NYSGPT2525 LLC