Optimal approximation using complex-valued neural networks

Complex-valued neural networks (CVNNs) have recently shown promising empirical success, for instance for increasing the stability of recurrent neural networks and for improving the performance in tasks with complex-valued inputs, such as in MRI fingerprinting. While the overwhelming success of Deep Learning in the real-valued case is supported by a growing mathematical foundation, such a foundation is still largely lacking in the complex-valued case. We thus analyze the expressivity of CVNNs by studying their approximation properties. Our results yield the first quantitative approximation bounds for CVNNs that apply to a wide class of activation functions including the popular modReLU and complex cardioid activation functions. Precisely, our results apply to any activation function that is smooth but not polyharmonic on some non-empty open set; this is the natural generalization of the class of smooth and non-polynomial activation functions to the complex setting. Our main result shows that the error for the approximation of $C^k$-functions scales as $m^{-k/(2n)}$ for $m \to \infty$ where $m$ is the number of neurons, $k$ the smoothness of the target function and $n$ is the (complex) input dimension. Under a natural continuity assumption, we show that this rate is optimal; we further discuss the optimality when dropping this assumption. Moreover, we prove that the problem of approximating $C^k$-functions using continuous approximation methods unavoidably suffers from the curse of dimensionality.

Paper

Similar papers

Peer review

Reviewer 6LvE7/10 · confidence 4/52023-07-06

Summary

The paper studies approximation rates of shallow complex-valued neural networks (CVNN) with general non-polyharmonic activation functions. First, the paper establishes upper bounds for the error of approximation of polynomial and general smooth functions by CVNNs. Then, various aspects of the optimality of these bounds are discussed. First, it is shown that these bounds are tight assuming continuous weight selection. Next, a special activation function is presented for which the bound can be improved. Finally, it is shown that for a standard sigmoid-type activation function the rate bound is tight.

Strengths

**Contribution.** The paper seems to be the first paper that establishes optimal approximation rates for CVNNs with general non-polyharmonic activation functions. Previous results were either limited to specific activations or only established the universal approximation property, without convergence rates. Moreover, the paper comprehensively analyzes the optimality of its convergence rates. It proves their tightness assuming a continuous weight selection and shows that faster rates can be achieved if this assumption is dropped and a specially designed activation is used. It should be noted, however, that the CVNN model considered in the paper does not seem to be very important, and that most results of the paper are analogs of existing results for usual real-valued neural networks - see Weaknesses below. **Quality and clarity.** The paper is very well written. The theorems are clearly stated, sketches of key ideas of the proofs are provided in the main text. Proof details are provided in a large appendix and also seem to be carefully written, though I did not study all of them closely.

Weaknesses

**Questionable significance of the CVNN model.** The usual real-valued neural networks are important both mathematically and practically. They represent simple and natural non-linear models whose significance is well-established. In contrast, CVNNs do not seem to be significant from either perspective. Mathematically, they mix holomorphic linear operations with generic non-linear operations. This combination does not seem to have interesting analytic properties (except for the original observation by Voigtlaender that the condition of non-polynomiality for RVNN gets replaced by non-polyharmonicity for CVNN) or obvious practical or computational meaning. The paper does not explain why CVNNs are important, instead referring to a small number of papers from 2016-2018 where a similar structure was applied. **Limited conceptual novelty.** The paper generally adapts existing methods and results from the real-valued to the complex-valued setting - admittedly with many extra new twists. The established convergence rate is the same as for the respective real network with doubled real dimension. **High technicality for a conference.** This is quite a technical paper with 40 pages of proofs in the appendix. It is unlikely that any NeurIPS reviewer properly checks all these, so the paper might be more suitable for a journal with a more comprehensive review process. However, this is only a minor point, since the paper makes a good effort to present key ideas and sketches of proofs already in the main text.

Questions

Why specifically is the particular computational structure used in CVNNs (a layer of holomorphic linear operations + a layer of pointwise non-holomorphic non-linear operations + another layer of holomorphic linear operations) important?

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

See above

Reviewer ve6Y7/10 · confidence 3/52023-07-07

Summary

The paper studies approximation error bounds for complex-valued neural networks based. The authors relied on several techniques from the work by Mhaskar (1996) and proved several theorems in the paper. I will summarize two important results here. 1. Given a function from $C^k$, an optimal error bound is proved under some conditions on the activation function and the hypothesis of continuous weight selection. Furthermore, the complex-valued neural network achieving this bound has a universal first layer. 2. When the hypothesis of continuous weight selection is dropped, the authors proved that there exists an activation function allowing the complex-valued neural network to achieve a better bound. No optimality is claimed under this case. On the other hand, there is an activation function such that the bound cannot be improved up to logarithmic factors.

Strengths

The novelty of this paper is clear. The paper provides error bounds for complex-valued neural networks using more general activation functions and shows that the optimal bound can be achieved under the hypothesis of continuous weight selection. The results in this paper generalize the results in the previous work [9]. The problem is well-motivated, and the theorems proved are sound. Overall, the paper is well-written, and I enjoy reading the paper.

Weaknesses

The hypothesis of continuous weight selection is not a practical assumption. The optimal choice of the network weights is discontinuous in general (see the reference below). It would be better if the authors could make this clear in the paper. Kainen, Paul C., Věra Kůrková, and Andrew Vogt. "Approximation by neural networks is not continuous." Neurocomputing 29, no. 1-3 (1999): 47-56. Under this assumption, the error bound can be proved to be optimal. However, no optimal results are provided if this assumption is dropped. Regarding this, it would be more convincing for the paper if the authors could provide insights into difficulties that arose in proving the statements and potential workarounds. The results provided are similar to [9] so it would be more interesting if the authors could describe what difficulties they have overcome to prove these results. Given the high similarity of the results, I would expect more discussions in the related work section or in the descriptions of the proof sketch.

Questions

Line 28: Would it be more precise if referencing the work by Cybenko in 1989? Line 140: What is your definition of a smooth function? Please make it precise. From Theorem 3.1, the complexity of an approximating complex-value network is established for a polynomial. Would it be possible to use the Stone-Weierstrass theorem to establish an approximation bound? This seems to be a natural step to apply.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The authors have sufficiently addressed the limitations of the paper in Section 5. In my view, the greatest limitation of this work is the use of continuous weight selection for guaranteeing the optimality of the error bound. It would be more convincing if the authors can give some insights into this in Section 5.

Reviewer yApE5/10 · confidence 3/52023-07-07

Summary

This paper studies the approximation power of complex-valued neural networks (CVNNs). They derive that the approximation error is with the order $m^{-k/2n}$, where m is the number of neurons, k is the smoothness of the target function, and n is the input dimension of the neural network. They also show that this approximation error is optimal under some mild assumptions.

Strengths

1. This paper derives the approximation bounds of CVNNs for any continuous activation functions. 2. Furthermore, they show that the approximation bounds they obtain are optimal with a natural continuity assumption. 3. For several specific activation functions, the authors derive the upper bounds without the assumption of the continuity of weight selection.

Weaknesses

1. For general activation functions, the authors need to assume the continuity of weight selection to derive the optimal approximation rates. It will be more convincing if they can show results for general activation functions without the assumption of the continuity of weight selection.

Questions

Please see the weaknesses.

Rating

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

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

No.

Reviewer GqHo3/10 · confidence 4/52023-07-10

Summary

The paper studies the expressive power of complex-valued neural networks. It is shown that depth 2 networks with non-polyharmonic (the complex equivalent of non-polynomial) activations can approximate any continuous function on a compact domain to arbitrary accuracy which decays polynomially with the width and exponentially with the smoothness of the target function, but nevertheless suffers from the curse of dimensionality. This upper bound is also shown to be optimal in general, and that furthermore, there are more tailored examples where one can slightly improve the approximation rate. Post-rebuttal: I still find the technical contribution of the paper rather fair (as my original score indicates). My main concern with the paper is that its impact feels somewhat limited as it provides a guarantee which suffers from the curse of dimensionality. While I understand that this cannot be evaded in general, obtaining a result which is tight in the worst case where the worst case is intractable is a clear limitation. While I appreciate the fact that there are other older results in the literature which are highly cited and obtain such results for the real setting, this paper's merits should be considered compared to what is already known, and from this perspective, obtaining results in the complex setting that are analogous to the real setting is of interest but not groundbreaking. I have therefore decided to keep my original score.

Strengths

- The problem feels overall well-motivated, and CVNNs seem like an interesting model to study. - The analogies and differences between real and complex neural networks highlighted by the results in this paper are interesting.

Weaknesses

- On a technical perspective, it seems like the main technical contributions here are to adapt existing technique to the CVNN setting. Most results seem to take existing proofs and just apply them to the complex setting, which doesn't feel novel enough. - The exact dependence on the rate of approximation is unclear. There's a `constant' $c(n,k)$ which hides some dependence on the input dimension and smoothness parameters, which could be potentially huge (see questions below). - The separation established in Section 4 by Theorem 4.2 and Theorem 4.3 is not particularly strong, and is only significant in cases where the input dimension $n$ is very small. - I didn't find the main approximation result in Theorem 3.2 very surprising. This is an analogous result for something which is already known for real-valued networks, so obtaining this result for complex-valued functions feels incremental. In particular, it would be of greater interest, in my opinion, to study cases where complex-valued networks attain an \emph{efficient} approximation in the input dimension $n$, since these have stronger practical implications.

Questions

Questions: - "complex-valued networks behave significantly different from real-valued networks": I don't see how this is `significantly different'. The second property does indicate a difference, but the first characteristic is analogous (non-polynomial and non-polyhedral activations). - There's no clear motivation for the very technical choice of the $C^k$-norm in the paper. Can you provide a more intuitive explanation for it and why it is chosen? In particular, why is this choice not stylized for obtaining the main result? - "It is crucial that the size of the networks considered in Theorem 3.1 is independent of the approximation quality $\varepsilon$": Why is this important? - What is the quantity $c(n,k)$ which appears in many of the theorem? The bounds in the paper are interesting since they allow us to study the dependence of the accuracy attained as a function of, say, $n$ for some fixed $k$; but the dependence of $c(n,k)$ on $n$ isn't clear. I tried to better understand it by looking at the proofs in the appendix but it's not made explicit there as well. $c(n,k)$ could, potentially, have magnitude $m^{k/(2n)}$ which would render the upper bound vacuous. Why is this not the case? Is this the same quantity in both the upper bound in Theorem 3.2 and the lower bound in Theorem 4.3? If not, it could make the either bound very loose despite appearing to provide a tight result. Comments: - Abstract: "the real-valued case is supported by a firm mathematical foundation" -- I wouldn't say our mathematical understanding of deep learning is firm. - The abstract doesn't explicitly state what are the functions the main result applies to. - Line 27: The universal approximation theorem dates back to 1989. See "Approximation by Superpositions of a Sigmoidal Function" by G. Cybenko - "CVNNs have the same excellent approximation properties as real-valued networks.": Why is this conclusion reached? The results in the paper hold for a very broad class of smooth function and therefore suffer from the curse of dimensionality in the worst case.

Rating

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

- I think that the lack of clarity in the theorem statements regarding $c(n,k)$ is problematic, yet this is not discussed in the paper. I urge the authors to clarify this and discuss this more clearly. - The limitations of the provided results which suffer from the curse of dimensionality should be clearly discussed.

Reviewer 6LvE2023-08-12

Thank you for your answers. Your arguments look convincing. I'm increasing my score. I still have an impression, though, that your setting is a special case of a more general setting that would be more natural. The usual shallow real networks are constructed from layers that obey two kinds of constraints: linear layers obey the linearity constraint, while the pointwise nonlinear layers obey the constraint that the action is pointwise. My understanding is that your setting extends this setup in the following way. First, you group the variables in blocks of two (real/imaginary parts), and then you impose Cauchy-Riemann conditions on the linear layers. In the linear case the Cauchy-Riemann conditions reduce to a pair of linear algebraic conditions in each $2\times 2$ block of the respective matrix. You also replace pointwise nonlinearity by blockwise nonlinearity. Now, my understanding is that your results show that these modifications do not break the classical approximation rate. This suggests that none of the specific modifications that you do (blocking, Cauchy-Riemann, blockwise activation) actually has any effect on the approximation rate. So if you divided the variables into blocks of say three variables rather than two, and imposed some blockwise linear algebraic constraints other than Cauchy-Riemann, and used a suitably non-degenerate blockwise activation, would this also preserve the classical rate?

Authorsrebuttal2023-08-15

Thanks for your positive feedback and for increasing the score! We greatly appreciate it. It is indeed an interesting question whether our results can be embedded into an even more general setting, where arbitrarily big groups of neurons are considered. And indeed, some older works (such as "Quaternionic Neural Networks for Associative Memories" by Teijiro Isokawa, Haruhiko Nishimura, and Nobuyuki Matsui) have for instance considered quaternion-valued neural networks or even so-called Clifford-algebra-valued neural networks (although we are not aware of significant real-world applications of these). However, in general going beyond the setting of complex numbers and quaternions, the considered structures get much less canonical, i.e., it is not quite clear which algebraic constraints one should impose in the case of, e.g., blocks of three neurons. This is mostly since it is not possible to endow $\mathbb{R}^3$ with an algebraic structure that represents a sensible multiplication (allowing for division). In fact, it is not possible to endow $\mathbb{R}^s$ with such a structure whenever $s \notin \{1,2,4\}$ (which is exactly the case of the usual real numbers, the complex numbers and quaternions). This is known as Frobenius' Theorem for real division algebras. Finally, the set of "good" blockwise activation functions would heavily depend on the chosen algebraic constraint. Thus, one should probably first settle the question of universality in these general settings, and then study the question of approximation rates. We would expect that the reachable approximation rate (under suitable assumptions such as continuity) again agrees with the "expected" rate, but this is not clear at all and would have to be carefully investigated.

Reviewer GqHo2023-08-18

Post-rebuttal response

Dear authors, Thank you for your detailed response which addresses the questions I raised. I appreciate your intensions to incorporate some of my suggestions into the paper. I want to point out that a setting where $n$ is taken as fixed is limiting. While such results are mainly of interest when $n$ is moderate, the resulting approximations quickly become inefficient as $n$ grows. I understand that this is the best that can be done in the worst case, but this ignores potential "average case" results and might be very loose in many instances. Such a limitation should be clearly discussed in the paper. Moreover, even though $c(n,k)$ is independent of $m$, in cases where it behaves as $k^n$ for example still make the upper bound weak.

Authorsrebuttal2023-08-19

Thanks for your reply. We understand that you are dissatisfied with the dependence of our approximation bounds on the input dimension, which makes them subject to the curse of dimensionality, and we agree that it is important to study alternative function classes for which neural networks can avoid the curse. We strongly disagree, however, with the conclusion that this makes our results uninteresting, or not worth publishing. In the following, we try one final time to make our point. 1. As we already pointed out in an earlier response, for the $C^k$ function class that we consider, the curse of dimensionality is an inescapable fact of life and cannot be avoided. This holds in a worst-case sense, but quite likely also in an "average case" sense. Indeed, the paper "Phase Transitions in Rate Distortion Theory and Deep Learning" by Grohs, Klotz, and Voigtlaender shows this optimality in an average sense in a slightly, but closely related setting. We are strongly convinced that this result also extends to our setting. Verifying this, however, is outside the scope of this paper. 2. In the real-valued setting, there are several highly influential (well published and highly cited) works that are subject to the same limitations as our result. As selected examples we mention the following: - "Neural networks for optimal approximation of smooth and analytic functions" by Mhaskar - "Error bounds for approximations with deep ReLU networks" by Yarotsky - "Optimal approximation of piecewise smooth functions using deep ReLU neural networks" by Petersen and Voigtlaender - "Optimal approximation of continuous functions by very deep ReLU networks" by Yarotsky - "The phase diagram of approximation rates for deep neural networks" by Yarotsky and Zhevnerchuk - "Deep network approximation characterized by number of neurons" by Shen, Yang, and Zhang. This underlines that such results are of high interest in the community. 3. Our results are not strictly limited to the class of $C^k$ functions. As an important auxiliary result (which might be of independent interest), we show that CVNNs can well approximate algebraic polynomials. There are natural and widely studied classes of functions that can be very well approximated by polynomials; for instance this holds for certain classes of holomorphic functions; see e.g. Example 2 in the paper "Approximation of smooth functionals using deep ReLU networks" by Song, Liu, Fan, and Zhou. For instance, for this class, our results on the approximation of polynomials using CVNNs would imply that using CVNNs with $N$ neurons, one can obtain an approximation error bound of $C \cdot \rho^{- N^{1/(2n)} / 5}$, where $C > 0$ and $\rho > 1$ only depend on the size of the polyellipse on which the considered functions are holomorphic, but not on the dimension. This bound is still subject to the curse of dimensionality, but much less than our bound for $C^k$ functions. This shows that our results and proof techniques can be useful to tackle the question of alternative function classes for which the curse can be avoided. We emphasize that there are many other function classes that can be well approximated by polynomials; for these, our results will thus be helpful. We will be happy to add a brief discussion of these points to the final version of the paper.

Reviewer ve6Y2023-08-21

I would like to thank the authors for their detailed responses. My concerns are fully addressed, and in light of this, I have increased my rating by 1. I hope the authors keep their promises and deliver extra analyses and clarifications in their final version. One of the concerns raised by Reviewer GqHo is about the constant $c(n,k)$. I think this is not that problematic given the independence of $\epsilon$. However, I believe it is helpful for the reader to know the upper bound of the constant and know why the constant is huge and perhaps some potential approaches to improve it. I also agree with the authors that the curse of dimensionality cannot be avoided without making strong assumptions about the function family.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC