Optimal deep learning of holomorphic operators between Banach spaces

Operator learning problems arise in many key areas of scientific computing where Partial Differential Equations (PDEs) are used to model physical systems. In such scenarios, the operators map between Banach or Hilbert spaces. In this work, we tackle the problem of learning operators between Banach spaces, in contrast to the vast majority of past works considering only Hilbert spaces. We focus on learning holomorphic operators - an important class of problems with many applications. We combine arbitrary approximate encoders and decoders with standard feedforward Deep Neural Network (DNN) architectures - specifically, those with constant width exceeding the depth - under standard $\ell^2$-loss minimization. We first identify a family of DNNs such that the resulting Deep Learning (DL) procedure achieves optimal generalization bounds for such operators. For standard fully-connected architectures, we then show that there are uncountably many minimizers of the training problem that yield equivalent optimal performance. The DNN architectures we consider are `problem agnostic', with width and depth only depending on the amount of training data $m$ and not on regularity assumptions of the target operator. Next, we show that DL is optimal for this problem: no recovery procedure can surpass these generalization bounds up to log terms. Finally, we present numerical results demonstrating the practical performance on challenging problems including the parametric diffusion, Navier-Stokes-Brinkman and Boussinesq PDEs.

Paper

Similar papers

Peer review

Reviewer 4yPo8/10 · confidence 4/52024-07-08

Summary

This work examines learning holomorphic operators between Banach spaces. It shows that these learning tasks can be tackled using encoder-decoder approaches with deep neural networks (DNNs), achieving optimal approximation rates up to log factors. The architectures used are problem-agnostic and depend only on the amount of available training data. The study also delves into theoretical aspects of the associated training problem for standard feed-forward networks and presents numerical findings to support its claims.

Strengths

* The paper is well-written and includes a thorough review of current research on learning holomorphic operators. * The results are new and interesting, expanding on previous work by addressing holomorphic mappings between arbitrary Banach spaces and allowing standard training methods for the DNNs utilized. * Although some assumptions and results are technical, I appreciate how the authors provide a clear overview of their meaning and interpretation. The proof techniques are explained well and summarized in the appendix, making them accessible without diving into deep details. * The theoretical findings are also backed up by practical experiments.

Weaknesses

Overall I do not have any serious concerns about this work and consider it to be a valuable contribution. Some questions are listed below.

Questions

* Perhaps I missed it, but I could not locate a precise reference discussing the holomorphicity of the Navier-Stokes-Brinkman and Boussinesq equations, similar to Section B.3.2. for the Diffusion problem. It would be helpful if such a discussion could be included in the manuscript. * Could you briefly summarize to what extent Theorems 4.1 and 4.2 require different techniques compared to those used in [7]? How does the use of a general Banach space $\mathcal{X}$ complicate the analysis? * Regarding the experimental results, do you have any insights into why different activation functions perform differently across individual tasks? Can this behavior be explained using the results from Theorems 3.1 and 3.2? * Do the authors think the results from Theorem 3.2 could be extended to include information on optimization procedures used for solving (2.5) (e.g., stochastic gradient descent and variants thereof)? Such an extension would enhance the practical relevance of the result. I am only seeking an educated guess or a high-level opinion, without expecting detailed analysis. * I would appreciate it if the authors could briefly summarize the differences between the Banach space and Hilbert space cases for $\mathcal{Y}$. Specifically, could you indicate precisely where the inner product structure is utilized in the proofs to achieve better rates? Also please double check formulas (3.3) and (3.5) in this context, currently it seems that they are the same.

Rating

8

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

Limitations and social impact have been discussed.

Reviewer XKAz8/10 · confidence 3/52024-07-12

Summary

Authors provide theoretical results on sample efficiency for learning holomorphic neural operators $\mathcal{X}\rightarrow\mathcal{Y}$ (results are available for both Hilbert and Banach spaces) with a neural network of a particular structure. The structure of the network is encoder $\mathcal{X}\rightarrow\mathbb{R}^{d_1}$, decoder $\mathcal{Y}\rightarrow\mathbb{R}^{d_2}$ and DNN $\mathbb{R}^{d_1}\rightarrow \mathbb{R}^{d_2}$, with both encoder and decoder considered to be given and DNN trained on $l_2$ loss in a usual supervised manner. Besides the upper bound on approximation error, authors also provide lower bound and conclude that active learning (deliberate collection of data used to train neural operator) is not justified, since non-adaptive training is already optimal up to log terms. Authors illustrate their theoretical results (relative test error vs number of samples for selected DNN configurations) with several numerical experiments, including diffusion equation in $D=2$, Navier-Stokes-Brinkman equation in $D=2$ and Boussinesq equation in $D=3$.

Strengths

Operator learning is a relatively recent DL setup, so novel theoretical results are of considerable interest. Authors managed to significantly improve previously known bounds and show that DNN can obtain near-optimal performance (in terms of the number of samples needed) in non-adaptive iid settings. The contribution is primarily theoretical, but formal statements are well explained, supplemented with discussions and examples which significantly simplifies understanding of the results.

Weaknesses

Overall, in my opinion, the research is of excellent quality, so I can not point to any major weaknesses. The two main issues that can be relevant are: (i) the requirements for the encoder and decoder are not sufficiently discussed, and I find restricting the assumption that they are already available, and (ii) the significance of the numerical confirmation is not apparent.

Questions

1. After line 102 there is a definition of Bochner and Pettis $L^p$ norms. Presumably, the later expression is a definition of $|||F|||$ which is not apparent. 2. I want to clarify several statements on the encoder and decoder. 1. In (1.3) authors introduce encoder $\mathcal{E}$ and decoder $\mathcal{D}$, and later approximate encoder and decoder $\widetilde{\mathcal{E}}$, $\widetilde{\mathcal{D}}$ after line 114. What is the difference between those? 2. In particular, in the discussion (lines 145, 146) of assumption A.IV (line 131) that requires both $\mathcal{E}$ and $\mathcal{D}$ to be linear and bounded, authors claim that $\widetilde{\mathcal{E}}$ is not required to be linear. It is not clear to me how this agrees with the fact that $\mathcal{E}_{\mathcal{X}}$ is defined in A.III as a composition of approximate encoder and decoder. Can the authors please clarify that? 3. According to (1.3) encoder and decoder are defined for the approximate operator $\hat{F}$. This means, as I understand, they are not rigidly fixed by $F$ since presumably $F$ can be approximated with many different encoders and decoders. Approximate encoders and decoders are not formally defined, so it is not clear what they approximate. Does it all mean that authors consider not a continuous operator learning problem but suppose that discretization (of the space of parameters and $\mathcal{Y}$) is already performed ($\mathcal{E}$ and $\mathcal{D}$ are selected) and $\widetilde{\mathcal{E}}$ and $\widetilde{\mathcal{D}}$ approximate $\mathcal{E}$ and $\mathcal{D}$? 4. Why there are $\widetilde{\mathcal{E}}\_{\mathcal{X}}$ and $\widetilde{\mathcal{D}}\_{\mathcal{X}}$ but not the approximate encoder and decoder for $\mathcal{Y}$? 3. Several questions on Theorem 3.1: 1. Why $a_{\mathcal{Y}}$ is called approximation error? What is the meaning of this term (it looks like some kind of volume)? In the ideal setting $\mathcal{D}\_{\mathcal{Y}}\circ\mathcal{E}\_{\mathcal{Y}} = I$, so we have a perfect approximation that has nonzero approximation error. 2. Since $\lim_{m\rightarrow\infty}m \big/ L(m) = \infty$ many terms in the upper bound will diverge unless the encoder and decoder are of sufficient quality. More specifically, the encoder and decoder should improve their accuracy as $\sim m^{-1/2}$ for u.b. to be useful in case of $2$ norm and as $\sim m^{-1}$ in case of Chebyshev norm. Can the authors please discuss this consequence of their result in more detail? 3. Can the authors also explain the consequence of the optimization error term? It seems that $\tau$ that appears in it is to some extent beyond control and completely independent of $m$. 4. Questions about experiments: 1. In experiments authors strive to confirm theoretical scaling $m^{-1}$. Since the encoder and decoder are perfect, and noise is absent, the only two sources of error are approximation error and optimization error. In my opinion, one can clearly see the signs of saturation for larger $m$. This is also clear from operator learning literature that relative error rarely fell below $10^{-3}$. Can the authors please comment on that? Is it because optimization error prevails for large $m$? 2. Dimensions $d=4,8$ in Figures 1, 2, 3 and in the text (e.g., lines 296, 297: "PDE in $d=4$ and $d=8$ dimensions") are slightly misleading because they refer to the dimension of the parameter space and can be confused with the number of spatial dimensions. 3. All results are given for perfect encoder and decoder. Can the authors share some thoughts on whether it is possible to test their upper bound with an imperfect decoder and encoder? What are those experiments might be? Is it possible to create a controllable setting to test terms in ub that contain encoder and decoder errors?

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

yes

Reviewer ZpED7/10 · confidence 4/52024-07-12

Summary

The paper addresses the challenge of learning operators between Banach spaces, focusing on holomorphic operators which have many applications. The authors employ standard (DNN) architectures with constant width greater than depth and $\ell_2$-loss minimization. They identify a family of DNNs that achieve optimal generalization bounds for such operators. For fully-connected architectures, they show there are infinitely many solutions that offer optimal performance. The results demonstrate that deep learning achieves the best possible generalization bounds, and no method can surpass these bounds, up to logarithmic terms.

Strengths

This work presents powerful and well-formulated theorems. The techniques used on holomorphic operators are interesting, and thanks to them, a generalization bound can be demonstrated. The work is novel, highly original, and I believe it could have a fruitful impact in the context of deep learning.

Weaknesses

I think some of the figures of the experiments results are too small, making it a bit difficult to recognize the results. It would be good if the authors could review this issue.

Questions

No questions.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

Theorem 3.2 only asserts that some minimizers are ‘good’, not all.

Reviewer 4yPo2024-08-07

I am pleased with the explanations provided by the authors and am willing to increase my score to 8. I am also looking forward to follow-up work that may address my first and fourth points (as mentioned, I did not expect detailed responses here). Here are a few additional comments: * Regarding the tanh network: One possible reason for its behavior could be the vanishing gradient property of the tanh function, which may lead to suboptimal performance of the optimization algorithm in larger (deeper) networks. However, this is merely a rough guess, and further investigation is needed since there could be many other contributing factors, as the authors have noted. * Regarding the Pettis norm: It might be helpful for the reader to introduce the associated notation ($|||.|||$) in the equation between 102 and 103. I believe this would address the source of my confusion.

Authorsrebuttal2024-08-08

Thanks for your second positive assessment and further comments. Your comment about tanh networks is a good one. We agree this is a good starting point for further investigations and will take it on board in our future work. Thanks also for flagging the point about the Pettis norm. As another also referee pointed out, this was a typo in the equation between lines 102 and 103. We will add this notation to that equation in the revision.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC