__ Summary and Contributions__: The paper proves multiple coupling based invertible nets can universally approximate smooth diffeomorphisms, and as a byproduct establishes the equivalence of single-coordinate transforms, triangular transforms and C^2 diffeomorphisms through function composition, which is a somewhat surprising and stronger result than distributional universality commonly seen in the literature.

__ Strengths__: The results proven in this work do not seem trivial, and can potentially help practitioners gain a better intuition to improve the parameterization of invertible neural networks, especially because coupling-based INN is the most commonly adopted architecture.

__ Weaknesses__: The need of having a diffeomorphic universality for discrete time NF is not well motivated, asides from the fact that it implies distributional universality. On what other occasion would this result be useful?

__ Correctness__: I didn’t go through the appendix to read the proofs in detail, but the sketch of proof presented in the main text seems convincing and reasonable.

__ Clarity__: The paper is well written and easy to follow on the whole.

__ Relation to Prior Work__: The following refs are missing. [1] attempts to prove Real-NVP type of coupling flow are distributionally universal. [2] also introduces universal non-linear 1D transformation via spline. [3] discusses the expressivity power of planar and sylvester (perhaps not as relevant since it’s not the same type of flow).
[1] Solving ODE with Universal Flows: Approximation Theory for Flow-Based Models
[2] Neural Spline Flows
[3] The Expressive Power of a Class of Normalizing Flow Models

__ Reproducibility__: Yes

__ Additional Feedback__:
[POST REBUTTAL]
---------
I thank the authors for the detailed response. I guess I see how one can parameterize some (Real NVP-style) linear couplings combined with permutation and sign flipping to represent any regular matrices (regular here denotes invertible I suppose?). Perhaps this could be explicitly constructed in the paper to complement the results. Does it also imply the the general linear group in the main result can be replaced with permutation group + sign flipping? Furthermore, if someone is only concerned with diffeomorphisms with a jacobian having strictly positive eigenvalues, then can the sign flipping be dropped?
I raised my overall rating to 8 due to the non-triviality of the results presented in the paper, and the clear exposition. I hope the authors can take into account the feedback/additional questions to update the paper, and I look forward to reading the updated version.
---------
Can you comment on the argument of [4], and whether it contradicts the results presented in this work. One key difference that I observe is that INN_H-ACF is intertwined with linear maps whereas regular Real-NVP type of coupling is followed by permutation. This perhaps emphasizes the importance in linear maps (aside from the fact that it allows one to flip the direction). Can you point out the use of this linear map in establishing diffeomorphic universality?
Perhaps this explains the empirical advantage of Glow over RealNVP as it generalizes the permutation to linear maps (albeit 1x1). Also, I think it’s not accurate enough to directly say (in line 191) Glow is universal since the family of linear maps it uses is limited (so is the ACF class, since it’s convolutional).
[4] You say Normalizing Flows I see Bayesian Networks

__ Summary and Contributions__: The work provides several theoretical results concerning the universality of coupling-based invertible neural networks (CF-INNs).
The central contribution is the rigorous proof that affine coupling flows (ACFs), the most widely used architecture in current research, is distributional- and Lp-universal.

__ Strengths__: To the INN- and NF-community, I see this work having a tremendous significance, and being an important missing ingredient in bringing the field of NFs forward.
Firstly, it addresses an important and long-standing question:
Specialized architectures such as NAF and SoS flows have been shown to be distributional universal approximators in the past. But in reality, the majority of practical works use affine coupling flows (ACFs), where distributional universality had not been shown. But this work goes above and beyond, and even shows that they are Lp-universal Diffeomorphism approximators, which is a much stronger property, and relevant for a broader range of works.
Secondly, the results are formulated and presented in such a way that make them easy for others in the community to use:
The work shows various connections and implications between different notions of universality and architectures, e.g. Theorem 1, Lemma 1, lines 186-192, etc. This is extremely useful for the future, in proving the reliability of newly proposed methods in different contexts.
These extra results are stated clearly and provided in a helpful and accessible manner.

__ Weaknesses__: I do not see a weakness concerning the content itself or the methodology.

__ Correctness__: The proofs are carefully and rigorously conduced, having read most of the appendix. However, I am no expert in differential geometry, and I did not myself verify that the conditions for all cited theorems that are used were met (therefore my confidence score of 3).

__ Clarity__: While the paper is quite technical, the language and notation is very precise and well thought-out, avoiding most misunderstandings.
Furthermore, I think dividing the main paper into "Main Results" (Sec 3) and "Proof Outline" (Sec 4) is a good idea. Section 3 is compact and simple enough that more practically oriented researchers should be able to understand and cite the results correctly.
Section 4 explains the main ideas of the proofs and illustrates them with figures. The figures are helpful, and most explanations are easy to follow.
However, I do think lines 212-235 could be more straightforward; I had to read this twice to understand all the reasoning and look at parts of the appendix in between.
Perhaps the arguments and explanations could be sorted in a different order.
E.g. flow endpoints are introduced, but it is not clear why this is useful until l.227. Instead, a sentence could be spent around line 216, saying that the "additivity" of flow endpoints will later be useful to decompose a non-nearly-id map into a composition of identical nearly-id maps.

__ Relation to Prior Work__: The relevant related work is discussed. Some related work is also extended upon, see line 162ff.
A remark that does not influence my score:
For the related work, I think some works from Bayesian inference could be relevant, for instance Bigoni et al, "Greedy inference with layers of lazy maps". While the nomenclature is different, the 'Lazy maps' are essentially also CF-INNs, and they show some guarantees in terms of KL-divergence.

__ Reproducibility__: Yes

__ Additional Feedback__: ================
Update after rebuttal
================
Some additional criticisms were brought up by the other reviewers.
The ones I found most relevant were also addressed properly in the rebuttal.
I leave my score unchanged.

__ Summary and Contributions__: The paper shows universality results for affine coupling flows for invertible function approximation.
******Post Rebuttal**********
Thank you for the response. My comment about the work not being quantitative was indeed not meant as a major criticism and this being one of the first works on the problem it addresses, I agree with the authors that this is a good first step.

__ Strengths__: The results are clean and the proof techniques based on some basic theorems from differential topology/geometry seem new. I found the results somewhat surprising

__ Weaknesses__: My main concern is that the results here are not "quantitative": e.g. they does not show what depth r (number of composition) is required to achieve error \epsilon. It is known that in certain cases this can be an issue:
[The Expressive Power of a Class of Normalizing Flow Models, Zhifeng Kong, Kamalika Chaudhuri, https://arxiv.org/abs/2006.00392]. Thus, even if the class of functions is universal in the sense of the present paper, it may not be so in practice where the depth r may be a relatively small number (say 100).
This sort of limitation is also shared by various classical universality results such as Cybenko's.

__ Correctness__: Proofs seem correct though I did not check all the details.

__ Clarity__: Well written.

__ Relation to Prior Work__: Well done (but see the reference above).

__ Reproducibility__: Yes

__ Additional Feedback__: Marther --> Mather (Line 221)

__ Summary and Contributions__: In this paper, the author provides rigorous proof of a proposed theorem to show the universal approximation properties for certain types of functions. This will provide the guarantee of universality of a CF-INN. This paper also provides a solid answer t othe distributional universality of ACF-based CF-INNs.

__ Strengths__: The claims in this paper are rigorously proved in view of differential geometry. The results provide the theoretical understanding of a CF-INNs by illustrating the universal approximation. Furthermore, it provides some deeper insight on the functionality of different flow layer designs in the model class.

__ Weaknesses__: As a theoretical paper, it focuses on illustrating the universal approximation on a special type of INN, thus the contribution is limited. No experimental results are presented to demonstrate how the theory works on existing algorithms.

__ Correctness__: The claims in this paper are correct and the proofs are rigorous and elegant.

__ Clarity__: The paper is well structured and well written.

__ Relation to Prior Work__: Due to the novelty of this topic, the related literature is limited. This paper advances the theoretical understanding of previous work.

__ Reproducibility__: Yes

__ Additional Feedback__: [POST REBUTTAL]
I thank the authors for the detailed response, which address all of my concerns.
I carefully check the proofs in the supplementary materials, and am convinced. Hence I change to my score to 7 accept.
It will be helpful to further discuss the following problem: current method aims at using a sequence of diffeomorphisms to map one probability measure to the other one, this can be achieved more directly using rigorous optimal tansportation method. The comparison between these two approaches will be helpful. The second issue is the choice of function famility as the approximator, why the current one is the most appropriate one to reflect the reality?