acceptodds
Under review as a conference paper at ICLR 2027

Bounded Domains for Symplectic Groups and Applications to Representation Learning

Abstract

Neural networks on the Siegel disk domain have recently been introduced as an effective tool for addressing a number of machine learning tasks. These networks are built upon the complex structure of the Siegel disk domain and its connection with the real symplectic group. In this paper, we study generalizations of the real symplectic group referred to as symplectic groups over involutive algebras, and propose a class of upper half-space and precompact models associated with these groups. We derive closed-form formulas for some mappings required to implement neural networks on the precompact model, i.e., automorphisms and their inverses. We give explicit expressions for these mappings on a class of bounded domains, and investigate their representation power on the tasks of radar clutter classification, action recognition, graph reconstruction, and node classification. Our experimental results demonstrate the efficacy of the proposed approach.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Symplectic geometry, bounded domains, Riemannian neural networks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7FMqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper generalises Siegel-space neural networks to a larger family of complex matrix spaces and builds neural network layers on one such space, termed HB_n. It evaluates these networks on radar clutter classification, action recognition, graph reconstruction, and node classification.

Critical strengths
  1. The proposed models outperform Siegel and SPD baselines on most benchmarks across several different tasks (Tables 1-4).
  2. The ablations (Table 5, Fig 1) suggest that the new geometry contributes to the gains, which is the paper’s main claim.
Critical weaknesses
  1. The theoretical sections are very dense and give little intuition for a general ML reader, which makes the contribution hard to assess.
  2. The method is slower to train than the Siegel baseline (about 1.5× per batch - see Table 8), which the authors already acknowledge as a limitation in Appendix D. A brief discussion of possible ways to reduce this cost would be helpful.
Questions for authors
  1. Could the authors give some intuition as to why HB_n is a good fit for these tasks? This would help readers who are less familiar with the underlying mathematics.
  2. HBNetBN has more parameters than SiegelNetBN (see Table 7). Have the authors compared the two at a similar parameter count? This would help me judge how much of the gain comes from the HB_n geometry itself.
3 comments

We thank the reviewer for their insightful questions, which provides us an opportunity to clarify and strengthen our work.

Q1: Could the authors give some intuition as to why HB_n is a good fit for these tasks? This would help readers who are less familiar with the underlying mathematics.

For the radar clutter classification task, by assumption, the time series are stationary centered autoregressive (AR) Gaussian time series. Consequently, they can be represented in a product space of HBn\mathbb{HB}_n and Sym⁡n+\operatorname{Sym}_n^+ (Cabanes, 2022). Specifically, each time series can be represented in product space Sym⁡n+×SDnr−1\operatorname{Sym}_n^+ \times \mathbb{SD}_n^{r-1}, where rr (r>1r > 1) is the order of the AR model. Since SDn\mathbb{SD}_n is a subdomain of HBn\mathbb{HB}_n, both these spaces are appropriate representation spaces for this task.

For the human action recognition task, we have shown that meaningful mappings between the proposed bounded domain and spaces of matrices with positive-definiteness constraints (i.e., SPD and HPD manifolds) can be built (see Proposition 3.6 and Definition 5.2). Since SPD manifolds have been extensively used as representation spaces for human action recognition, HBn\mathbb{HB}_n could potentially be effective for this task. This was the main motivation behind the use of HBn\mathbb{HB}_n for human action recognition experiments.

Regarding the graph reconstruction and node classification tasks, our work was mainly motivated by the results in López et al. (2021) where the authors have shown that Siegel models when equipped with vector-valued distances outperform SPD models on these tasks. Since SDn\mathbb{SD}_n is a subdomain of HBn\mathbb{HB}_n, the latter can provide geometric inductive biases that the former may fail to capture. This has been demonstrated by our experimental results provided in Tab. 3.

Q2: HBNetBN has more parameters than SiegelNetBN (see Table 7). Have the authors compared the two at a similar parameter count? This would help me judge how much of the gain comes from the HB_n geometry itself.

We thank the reviewer for raising this question. We agree that this is important to evaluate the real geometric benefits of the proposed domain.

For the radar clutter classification and human action recognition experiments, our model consists of a BN layer and a classification layer, which is a minimal architecture and cannot be further simplified. This is also the case of the Siegel models (i.e., SiegelNetFC and SiegelNetBN). For the SPD models (i.e., SPDNetBN and LieBN), we used the best architectures in the original papers. Thus, matching the number of trainable parameters is not necessarily a meaningful way to ensure a fair comparison because it would require artificially constraining our model or the baselines without obtaining a better interpretation of the final results. Note that for these experiments, we cannot change the dimensions of input matrices for the models (so that the models are matched in terms of representation dimension as in the graph reconstruction and node classification experiments) because these dimensions are determined by the dimensions of input time series and of video sequences.

For the node classification experiments, our model has fewer learnable parameters than the SPD model (see Appendix A.4.3). We performed additional experiments where our model and the SPD model are matched in terms of representation dimension. Specifically, we learn embeddings on HB8\mathbb{HB}_8 and Sym⁡16+\operatorname{Sym}^+_{16}. The real dimensions of HB8\mathbb{HB}_8 and Sym⁡16+\operatorname{Sym}^+_{16} are 128 and 136, respectively. The results are provided in Table E. We can observe that increasing the dimension of the SPD embeddings has little impact on mean accuracy of the SPD model.

ModelAirportPubmedCoraDisease
SPD-GCN82.8$\pm$0.978.5$\pm$0.482.1$\pm$0.696.6$\pm$2.5
HB-GCN91.1$\pm$0.779.2$\pm$0.482.6$\pm$0.597.2$\pm$1.4

Table E: Node classification results (mean accuracy ±\pm standard deviation) across ten runs.

We now compare our model, the Siegel models, and the SPD model on the graph reconstruction task where they are matched in terms of representation dimension. Specifically, we learn embeddings on HB6\mathbb{HB}_6, SH8\mathbb{SH}_8, SD8\mathbb{SD}_8, and Sym⁡12+\operatorname{Sym}^+_{12}. The real dimensions of HB6\mathbb{HB}_6, SH8\mathbb{SH}_8, SD8\mathbb{SD}_8, and Sym⁡12+\operatorname{Sym}^+_{12} are 72, 72, 72, and 78, respectively. The results are provided in Table F.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_{12}$14.48$\pm$0.035.81$\pm$0.048.43$\pm$0.028.56$\pm$0.021.84$\pm$0.031.59$\pm$0.03
S$^{F_1}_{8}$0.01$\pm$0.01.02$\pm$0.080.44$\pm$0.050.43$\pm$0.040.52$\pm$0.060.48$\pm$0.05
B$^{F_1}_{8}$0.01$\pm$0.00.82$\pm$0.040.41$\pm$0.030.42$\pm$0.030.40$\pm$0.030.43$\pm$0.03
HB6F1\mathbb{HB}^{F_1}_60.01$\pm$0.00.80$\pm$0.040.50$\pm$0.030.48$\pm$0.020.37$\pm$0.020.40$\pm$0.04

Table F: Graph reconstruction results (mean distortion ±\pm standard deviation) across five runs.

In terms of mean distortion, our model and the Siegel models significantly outperform the SPD model, and our model achieves mean accuracy comparable to or better than the best Siegel model (i.e., B$^{F_1}_{8}$) on 4 of the 6 datasets.

Finally, the training times of the competing models on the graph reconstruction task are reported in Table G. For comparable representation dimensions, our model is computationally more efficient during the training phase than the competing models. It is important to note that the training time of the best Siegel model is about 4$\times$ that of our model.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_{12}$1.080.0325.5325.478.2820.96
S$^{F_1}_{8}$1.950.0548.7349.3615.8043.15
B$^{F_1}_{8}$2.920.0783.3283.4424.5669.94
HB6F1\mathbb{HB}^{F_1}_60.940.0220.4720.366.9417.68

Table G: Training times (seconds) per epoch on the graph reconstruction task.

Thank you for the detailed response and additional experiments. The intuition that you have given for each task is helpful, and the dimension-matched results in Tables E and F address my concern for the graph tasks. I also appreciate the training-time comparison in Table G. I will maintain my score of 3 (weak acceptance), noting that I am not an expert in the underlying mathematics.

Reviewer_rTgrOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents new mathematical model extending functions between the upper half plane and bounded disk to a broader set of involutive algebras. Among them are definitions of automorphisms that preserve symplectic forms, invertible centering operations, and a constructed connection between the upper half plane and "positive" matrices. The Hermitian positive definite group is chosen as the primary algebra, and using the defintions, a distance function and batch normalization and classification layers are presented. The formulations are applied to graph convolution networks on semi-positive definite matrices and benchmarked on four different tasks. The results show that the presented model consistently outperforms competing methods, often by a reasonable margin.

Critical strengths

The strong mathematical presentation in the main paper body and derivations in the appendix ground the model in a principled way. The proposed layers and definitions are logical and interesting. Experimentally, the formulations appear to be dominant on a diverse set of tasks. The cited literature is reasonably comprehensive.

Critical weaknesses

The mathematical exposition is overly compilcated and could be simplified, especially since only one algebra is utilized later in the paper. Many of the formulations follow quite naturally and intuitively, mirroring previously existing definitions in literature (particularly the Hermitian modular group that is actually utilized). Consequently, there was limited theoretical discussion in the main paper body of the actual network layers and how/why they are applied to GCNs which seem to be the main paper contributions. Additionally, explicit connections to familiar concepts would make it simpler to discern which steps are actually novel (e.g. the map φ\varphi in Sec. 3 is an extension of the well-known Cayley map). Most formulas are also presented without motivation, explanation, or appropriate links to derivations in the appendix.

Additional motivation (beyond citation) on why these mathematical structures are used for the applied task would make the work more self-contained.

Questions for authors
  1. ϕX\phi_{\mathbf{X}} in 3.3 appears to be non-unique. Is/are the degree(s) of freedom in choosing such a transformation material?
  2. Is there a purpose (other than mathematical novelty) for deriving the results in Sections 2 and 3 without specifying a specific algebra as done in Section 4?
  3. What is the motivation/intuition behind the specific form of ρ(Z)\rho(\mathbf{Z}) in 3.6? I would imagine there are many ways to draw that connection.
  4. Does the distance function relate to ϕX\phi_{\mathbf{X}}?
  5. Is Eq. 5 an "almost geodesic" or minimizing geodesic?
3 comments

We thank the reviewer for raising insightful questions which helped us improve the clarity and correctness of the manuscript.

Q1: ϕX\phi_{\mathbf{X}} in 3.3 appears to be non-unique. Is/are the degree(s) of freedom in choosing such a transformation material?

We thank the reviewer for raising this question. We also take this opportunity to clarify that the assumption I−σˉC(X)X∈(ACσˉC)+\mathbf{I} - \bar{\sigma}_{\mathbb{C}}(\mathbf{X})\mathbf{X} \in ( A_{\mathbb{C}}^{\bar{\sigma}_{\mathbb{C}}} )_+, analogous to the existing assumptions I−XσˉC(X)∈(ACσˉC)+\mathbf{I} - \mathbf{X}\bar{\sigma}_{\mathbb{C}}(\mathbf{X}) \in ( A_{\mathbb{C}}^{\bar{\sigma}_{\mathbb{C}}} )_+ and I−σˉC(X)Y∈AC×\mathbf{I} - \bar{\sigma}_{\mathbb{C}}(\mathbf{X})\mathbf{Y} \in A_{\mathbb{C}}^{\times}, should be explicitly included in Proposition 3.3. This was unintentionally omitted from the original statement, and we will correct the statements in Propositions 3.3 and 3.5 accordingly in the revised paper.

Under these assumptions, all the terms (I−XσˉC(X))−12\left ( \mathbf{I} - \mathbf{X}\bar{\sigma}_{\mathbb{C}}(\mathbf{X}) \right )^{-\frac{1}{2}}, (I−σˉC(X)Y)−1(\mathbf{I} - \bar{\sigma}_{\mathbb{C}}(\mathbf{X})\mathbf{Y})^{-1}, and (I−σˉC(X)X)12\left ( \mathbf{I} - \bar{\sigma}_{\mathbb{C}}(\mathbf{X})\mathbf{X} \right )^{\frac{1}{2}} are unique, which means that ϕX(Y)\phi_{\mathbf{X}}(\mathbf{Y}) is unique. The uniqueness of (I−σˉC(X)Y)−1(\mathbf{I} - \bar{\sigma}_{\mathbb{C}}(\mathbf{X})\mathbf{Y})^{-1} follows from the fact that the inverse of an invertible element in a unital associative algebra is unique, and the uniqueness of (I−XσˉC(X))−12\left ( \mathbf{I} - \mathbf{X}\bar{\sigma}_{\mathbb{C}}(\mathbf{X}) \right )^{-\frac{1}{2}} (resp. (I−σˉC(X)X)12\left ( \mathbf{I} - \bar{\sigma}_{\mathbb{C}}(\mathbf{X})\mathbf{X} \right )^{\frac{1}{2}}) follows from spectral decompositions (see Alessandrini et al. (2021), Theorem 2.66).

Q2: Is there a purpose (other than mathematical novelty) for deriving the results in Sections 2 and 3 without specifying a specific algebra as done in Section 4?

Our main aim in Sections 2 and 3 is (1) to derive general formulas for a class of mappings and their inverses that map the proposed bounded domain to itself, and (2) to establish a mapping between the proposed upper half-space and the set of σˉC\bar{\sigma}_{\mathbb{C}}-positive elements. All these mapping can be used for the construction of neural networks on the proposed bounded domain. The results in these sections can be applied to other realizations of this domain. Therefore, given the definition of an anti-involution, one obtains the geometric description of the corresponding realization of the bounded domain, along with the associated mappings with minimal additional effort. This provides a convenient framework for implementing a class of neural networks based on different realizations of the bounded domain.

Q3: What is the motivation/intuition behind the specific form of ρ(Z)\rho(\mathbf{Z}) in 3.6? I would imagine there are many ways to draw that connection.

The result in Proposition 3.6 is a generalization of Lemma 6 in Siegel, (1943) where the author derived a mapping of the Siegel upper half-space onto the space of symplectic positive definite matrices. This mapping is built upon the transitive action of the real symplectic group on the Siegel upper half-space. In our work, we generalize this method by identifying the Möbius transformation that maps iI\mathbf{i}\mathbf{I} to Z\mathbf{Z}. Let X0=12(Z+σˉC(Z))\mathbf{X}_0 = \frac{1}{2}(\mathbf{Z}+\bar{\sigma}_{\mathbb{C}}(\mathbf{Z})), Y0=12i(Z−σˉC(Z))\mathbf{Y}_0 = \frac{1}{2\mathbf{i}}(\mathbf{Z}-\bar{\sigma}_{\mathbb{C}}(\mathbf{Z})). Since Z∈HUn\mathbf{Z} \in \mathbb{HU}_n, there exists P∈AC×\mathbf{P} \in A_{\mathbb{C}}^{\times} such that Y0=PσˉC(P)\mathbf{Y}_0 = \mathbf{P}\bar{\sigma}_{\mathbb{C}}(\mathbf{P}). Thus \begin{equation} \begin{pmatrix} \mathbf{P} & \mathbf{X}_0\bar{\sigma}_{\mathbb{C}}(\mathbf{P})^{-1} \\ 0 & \bar{\sigma}_{\mathbb{C}}(\mathbf{P})^{-1} \end{pmatrix}[\mathbf{i}\mathbf{I}] = \mathbf{Z}. \end{equation}

The mapping is then given by \begin{equation} \begin{pmatrix} \mathbf{P} & \mathbf{X}_0\bar{\sigma}_{\mathbb{C}}(\mathbf{P})^{-1} \\ 0 & \bar{\sigma}_{\mathbb{C}}(\mathbf{P})^{-1} \end{pmatrix} \begin{pmatrix} \mathbf{P} & \mathbf{X}_0\bar{\sigma}_{\mathbb{C}}(\mathbf{P})^{-1} \\ 0 & \bar{\sigma}_{\mathbb{C}}(\mathbf{P})^{-1} \end{pmatrix}^T. \end{equation}

Q4: Does the distance function relate to ϕX\phi_{\mathbf{X}}?

The distance function computed in Algorithm 1 is not built upon automorphims ϕX\phi_{\mathbf{X}}. It is built upon the concept of vector-valued distance [C,D]. Algorithm 1 generalizes the method in [D] (see Appendix B.4 of that paper). The resulting distance is the Finsler distance inducing the l1l^1-metric (see Algorithm 1 in [D]). We propose to include the connection of Algorithm 1 with the method in [D] in the revised manuscript.

Q5: Is Eq. 5 an "almost geodesic" or minimizing geodesic?

We use the concept of almost geodesics proposed in Nguyen & Grozavu (2026). They are not necessarily minimizing geodesics. They are used to update the running statistics in the BN algorithm. Please refer to Definition 3.1 of that paper.

References

[C] M. Kapovich, B. Leeb, and J. Porti. Anosov Subgroups: Dynamical and Geometric Characterizations. CoRR, abs/1703.01647, 2017.

[D] F. López, B. Pozzetti, S. Trettel, M. Strube, and A. Wienhard. Symmetric Spaces for Graph Embeddings: A Finsler-Riemannian Approach. In ICML, pages 7090–7101, 2021.

Thank you for the responses, they mostly answer my questions. The numerical results look promising and the representation is interesting, but the paper is very difficult to read, so I maintain my score. I highly encourage the authors to address the exposition concerns raised in my review when submitting an updated version.

Reviewer_Ymy5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates a generalization of Siegel spaces through symplectic groups over involutive algebras. The authors study associated upper half-space and bounded domain representations and derive explicit expressions for automorphisms and their inverses. These constructions are then used to develop neural network components, including normalization and classification layers, together with an extension of graph neural networks operating on positive definite matrices.

The proposed framework is evaluated on radar classification, action recognition, graph reconstruction, and node classification, with results suggesting that the larger representation spaces can offer advantages over existing Siegel and SPD-based approaches.

Critical strengths

I think the main strength of this paper is its attempt to explore richer geometric representation spaces beyond those commonly used in geometric deep learning. I find this direction interesting because it raises a fundamental question about the relationship between the algebraic structure of a representation space and the classes of functions that can be efficiently learned on it.

Another strength is that the paper does not stop at the mathematical construction. The authors derive explicit transformations that make it possible to implement neural network operations on the considered domains. The connection between the underlying group actions and practical network components is useful, and the experiments across different tasks provide encouraging evidence that these geometric constructions may be beneficial.

Critical weaknesses

My main concern is the novelty and interpretation of the proposed geometric spaces. In particular, the concrete bounded domain considered in the experiments appears to coincide with the classical Cartan bounded symmetric domain of type I, consisting of complex matrices whose operator norm is smaller than one. The associated fractional linear transformations and automorphism structures are well studied. It is therefore unclear which aspects of the mathematical construction are fundamentally new and which are reformulations of existing results. I think this distinction is important because the mathematical contribution is central to the paper.

A second concern is that the experiments do not sufficiently establish whether the improvements come from the geometric structure itself or simply from the additional representational degrees of freedom. The proposed domains are larger than the corresponding Siegel spaces, and the resulting models can have more parameters and higher computational costs. Without controlling for these differences, it is difficult to determine whether the observed improvements reflect a genuine advantage of the proposed geometry.

Finally, although the proposed constructions are mathematically interesting, the relationship between their geometric properties and the learning problems remains somewhat unclear. The experiments demonstrate improved performance in several settings, but provide limited insight into what properties of the data make these particular representation spaces preferable.

Questions for authors
  • How does the proposed bounded domain relate to the classical Cartan domains of type I? In particular, can the authors clarify which aspects of the automorphism constructions and their inverses are new relative to the existing theory of bounded symmetric domains?

--What fundamental properties of the data would make the proposed representation space more appropriate than a Siegel space or an SPD manifold? Is there a way to characterize the class of relational or geometric structures that benefit from the additional degrees of freedom?

-would the performance advantages remain if the competing models were matched in terms of representation dimension, number of trainable parameters, and computational budget? This seems important for separating geometric benefits from increased model capacity.

-to what extent do the improvements arise from the intrinsic geometry of the bounded domain, as opposed to the subsequent mapping into positive definite matrix spaces and the choice of classification architecture?

  • the normalization construction uses interpolation based on the Kobayashi distance, while the Fréchet mean is defined through a different matrix-valued distance construction. Can the authors clarify the relationship between these geometries and whether the resulting normalization procedure has a consistent geometric interpretation?

my assessment : I find the mathematical direction of the paper interesting, and the construction of practical learning operations on these domains is a useful contribution. However, I have reservations about the novelty of the proposed spaces and the extent to which the experiments establish a specific advantage of their geometric structure.

I lean toward weak rejection in the current form. A clearer characterization of the mathematical novelty and more controlled comparisons with existing representation spaces would significantly strengthen the contribution and would change my rating as well.

4 comments

We thank the reviewer for their insightful questions, which provides us an opportunity to clarify and strengthen our work.

Q1: How does the proposed bounded domain relate to the classical Cartan domains of type I? In particular, can the authors clarify which aspects of the automorphism constructions and their inverses are new relative to the existing theory of bounded symmetric domains?

In our paper, we focus on the setting where A=Mat⁡n(R)A=\operatorname{Mat}_n(\mathbb{R}), since it is one of the most relevant settings for applications. In this case, the proposed bounded domain is closely related to the classical Cartan domains of type I. For instance, when the anti-involution σ\sigma is given by σ(Z)=P−1ZTP\sigma(\mathbf{Z}) = \mathbf{P}^{-1}\mathbf{Z}^T\mathbf{P} for any symmetric positive definite matrix P\mathbf{P}, the C\mathbb{C}-antilinear extension σˉC\bar{\sigma}_{\mathbb{C}} is then given by σˉC(Z)=P−1ZHP\bar{\sigma}_{\mathbb{C}}(\mathbf{Z}) = \mathbf{P}^{-1}\mathbf{Z}^H\mathbf{P}. Consequently, the proposed bounded domain is defined as \begin{equation*} \mathfrak{B} = \{ \mathbf{Z} \in A_{\mathbb{C}} : \mathbf{I} - \mathbf{P}^{-1}\mathbf{Z}^H\mathbf{P}\mathbf{Z} \in ( A_{\mathbb{C}}^{\bar{\sigma}_{\mathbb{C}}} )_+ \}. \end{equation*}

Let F(X)=P1/2XP−1/2F(\mathbf{X}) = \mathbf{P}^{1/2} \mathbf{X} \mathbf{P}^{-1/2}, then it is easy to see that the map FF and its inverse are holomorphic and that FF map B\mathfrak{B} to the following domain \begin{equation*} \mathfrak{D} = \{ \mathbf{W} \in \operatorname{Mat}_n(\mathbb{C}) : \mathbf{I} - \mathbf{W}^H\mathbf{W} \in \mathbb{H}_n^+ \}, \end{equation*} which is a classical Cartan domain of type I. Therefore, the domain B\mathfrak{B} is biholomorphic to a classical Cartan domain of type I. For a general anti-involution, further investigation is needed to identify the connection between the proposed bounded domain and the classical Cartan domains of type I.

Although the proposed bounded domain and the classical Cartan domains of type I have a close connection, our work is different from existing works in the following aspects. First, our main results provided in Proposition 3.3, Corollary 3.4, Proposition 3.5, and Proposition 3.6 are extensions of the works in Alessandrini et al. (2021); Huang et al. (2025). Second, to our knowledge, expressions of automorphisms in terms of anti-involutions and their extensions (i.e., those studied in Alessandrini et al. (2021); Huang et al. (2025)) as proposed in our work have not appeared in the classical theory of bounded symmetric domains. Although this theory eventually produces identical expressions of automorphisms in specific realizations as in our work, the core structures required by both approaches are not the same. For instance, in the classical operator-ball theory [A], explicit formulas of automorphisms typically involve adjoint operators constructed from Hilbert-space inner products. Therefore, these constructions of automorphisms depend on Hilbert-space structures. In our approach, the construction of automorphisms depends on anti-involutions and their algebraic properties.

Beside the mathematical constructions, one of our main contributions is to show how to apply the proposed domain to machine learning tasks. The difference between our work and existing works are:

  • López et al. (2021): This work only focuses on graph embedding applications.
  • Nguyen et al. (2025); Nguyen & Grozavu (2026): In addition to graph embedding applications, these works address discriminative tasks where the input data belong to (product of) Siegel spaces.
  • Our work: Compared to the above works, we further introduce methods that learn across the proposed bounded domain and spaces of matrices with positive-definiteness constraints (i.e., SPD and HPD manifolds).

References

[A] Kaup, Wilhelm, and Upmeier, Harald. "Jordan Algebras and Symmetric Siegel Domains in Banach Spaces.." Mathematische Zeitschrift 157 (1977): 179-200.

Q2: What fundamental properties of the data would make the proposed representation space more appropriate than a Siegel space or an SPD manifold?...

When the data not only have a Riemannian structure but also additional structures (e.g., a complex structure or a symplectic structure), the proposed representation space might be more appropriate than SPD manifolds because these structures (in addition to the Riemannian structure) can be exploited for the construction of neural networks. For instance, our BN layers are built upon the automorphisms of the proposed bounded domain, which arise from its complex structure.

Even when the data can be naturally represented on an SPD manifold, mapping them to the proposed representation space and learning transformations in this space may be beneficial for certain tasks. This is demonstrated by our human action recognition experiments where our model outperforms the SPD models across all the datasets.

The choice of a representation space might be nontrivial when the data can naturally be represented in different families of spaces. In the specific case of radar clutter classification experiments, while both HBn\mathbb{HB}_n and SDn\mathbb{SD}_n are appropriate representation spaces, the product space of the input data originally involves copies of HBn\mathbb{HB}_n but not copies of SDn\mathbb{SD}_n [B] (note that we convert the related parts of the input data to SDn\mathbb{SD}_n). Therefore, from a theoretical perspective, HBn\mathbb{HB}_n is a more appropriate representation space than SDn\mathbb{SD}_n for these experiments.

References

[B] Ben Jeuris and Raf Vandebril. The Kahler mean of Block-Toeplitz matrices with Toeplitz structured blocks. SIAM Journal on Matrix Analysis and Applications 37(3):1151-1175, 2016.

Q3:would the performance advantages remain if the competing models were matched in terms of representation dimension, number of trainable parameters, and computational budget? This seems important for separating geometric benefits from increased model capacity.

We agree that this is important to evaluate the real geometric benefits of the proposed domain.

For the radar clutter classification and human action recognition experiments, our model consists of a BN layer and a classification layer, which is a minimal architecture and cannot be further simplified. This is also the case of the Siegel models (i.e., SiegelNetFC and SiegelNetBN). For the SPD models (i.e., SPDNetBN and LieBN), we used the best architectures in the original papers. Thus, matching the number of trainable parameters is not necessarily a meaningful way to ensure a fair comparison because it would require artificially constraining our model or the baselines without obtaining a better interpretation of the final results. Note that for these experiments, we cannot change the dimensions of input matrices for the models (so that the models are matched in terms of representation dimension as in the graph reconstruction and node classification experiments) because these dimensions are determined by the dimensions of input time series and of video sequences.

For the node classification experiments, our model has fewer learnable parameters than the SPD model (see Appendix A.4.3). We performed additional experiments where our model and the SPD model are matched in terms of representation dimension. Specifically, we learn embeddings on HB8\mathbb{HB}_8 and Sym⁡16+\operatorname{Sym}^+_{16}. The real dimensions of HB8\mathbb{HB}_8 and Sym⁡16+\operatorname{Sym}^+_{16} are 128 and 136, respectively. The results are provided in Table B. We can observe that increasing the dimension of the SPD embeddings has little impact on mean accuracy of the SPD model.

ModelAirportPubmedCoraDisease
SPD-GCN82.8$\pm$0.978.5$\pm$0.482.1$\pm$0.696.6$\pm$2.5
HB-GCN91.1$\pm$0.779.2$\pm$0.482.6$\pm$0.597.2$\pm$1.4

Table B: Node classification results (mean accuracy ±\pm standard deviation) across ten runs.

We now compare our model, the Siegel models, and the SPD model on the graph reconstruction task where they are matched in terms of representation dimension. Specifically, we learn embeddings on HB6\mathbb{HB}_6, SH8\mathbb{SH}_8, SD8\mathbb{SD}_8, and Sym⁡12+\operatorname{Sym}^+_{12}. The real dimensions of HB6\mathbb{HB}_6, SH8\mathbb{SH}_8, SD8\mathbb{SD}_8, and Sym⁡12+\operatorname{Sym}^+_{12} are 72, 72, 72, and 78, respectively. The results are provided in Table C.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_{12}$14.48$\pm$0.035.81$\pm$0.048.43$\pm$0.028.56$\pm$0.021.84$\pm$0.031.59$\pm$0.03
S$^{F_1}_{8}$0.01$\pm$0.01.02$\pm$0.080.44$\pm$0.050.43$\pm$0.040.52$\pm$0.060.48$\pm$0.05
B$^{F_1}_{8}$0.01$\pm$0.00.82$\pm$0.040.41$\pm$0.030.42$\pm$0.030.40$\pm$0.030.43$\pm$0.03
HB6F1\mathbb{HB}^{F_1}_60.01$\pm$0.00.80$\pm$0.040.50$\pm$0.030.48$\pm$0.020.37$\pm$0.020.40$\pm$0.04

Table C: Graph reconstruction results (mean distortion ±\pm standard deviation) across five runs.

In terms of mean distortion, our model and the Siegel models significantly outperform the SPD model, and our model achieves mean accuracy comparable to or better than the best Siegel model (i.e., B$^{F_1}_{8}$) on 4 of the 6 datasets.

Finally, the training times of the competing models on the graph reconstruction task are reported in Table D. For comparable representation dimensions, our model is computationally more efficient during the training phase than the competing models. It is important to note that the training time of the best Siegel model is about 4$\times$ that of our model.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_{12}$1.080.0325.5325.478.2820.96
S$^{F_1}_{8}$1.950.0548.7349.3615.8043.15
B$^{F_1}_{8}$2.920.0783.3283.4424.5669.94
HB6F1\mathbb{HB}^{F_1}_60.940.0220.4720.366.9417.68

Table D: Training times (seconds) per epoch on the graph reconstruction task.

Q4: to what extent do the improvements arise from the intrinsic geometry of the bounded domain, as opposed to the subsequent mapping into positive definite matrix spaces and the choice of classification architecture?

For the radar clutter classification and action recognition experiments, our model consists of a BN layer and a classification layer. No additional layers are used except those proposed in our work. The mapping from the bounded domain to an HPD manifold along with the HPD MLR are part of the proposed framework through which the representation becomes suitable for classification. Therefore, independently evaluating the bounded domain without the subsequent mapping and the HPD MLR would require developing additional building blocks (for classification on the domain) which do not necessarily respect its intrinsic geometry. The improvements obtained in these experiments could be considered as arising from the joint contribution of the proposed components. In our ablation study, we provide experimental results demonstrating the impact of the automorphisms. Since the automorphisms arise from the complex structure of the bounded domain, and this complex structure is an intrinsic aspect of the geometry of the domain, the improvements resulting from the intrinsic geometry of the domain are partly assessed by this ablation study.

For the node classification task, the mapping from the bounded domain to an HPD manifold and then to an SPD manifold is not an additional processing step, but an integral part of the proposed framework that enables the subsequent operations, i.e., feature transform, aggregation, and bias and nonlinearity at each layer of the GCN. We conducted an additional ablation study to investigate the impact of the automorphisms on the performance of the proposed GCN by removing the mapping ϕP\phi_{\mathbf{P}} from the right-hand side of Eq. (9). On Airport dataset, the resulting GCN gives a mean accuracy of 88.6% with a standard deviation of 0.8. From Tab. 4, HB-GCN gives a mean accuracy of 91.1% with a standard deviation of 0.7. Thus, applying the automorphisms results in a 2.5% improvement in mean accuracy. When combining the automorphisms and the mapping from the bounded domain to an SPD manifold, we achieve a 8.5% improvement in mean accuracy (see Tab. 4).

Q5: the normalization construction uses interpolation based on the Kobayashi distance, while the Fréchet mean is defined through a different matrix-valued distance construction. Can the authors clarify the relationship between these geometries and whether the resulting normalization procedure has a consistent geometric interpretation?

The distance used for computing the Fréchet mean is the invariant Riemannian distance. In certain bounded domains (e.g., the complex unit ball), this distance and the Kobayashi pseudodistance coincide (up to scaling). In HBn\mathbb{HB}_n, these distances are not identical. Nevertheless, they share the same property: they are invariant under all biholomorphic automorphisms of the domain. The relationship between these distances can be expressed (Nielsen, 2020) by the following equations: \begin{equation} d_{\mathbb{HB}_n}(\mathbf{X},\mathbf{Y}) = \sqrt{\sum_{i=1}^n \left ( \frac{1}{2} \log \frac{1 + \sqrt{\lambda_i}}{1 - \sqrt{\lambda_i}} \right )^2 }, \hspace{5mm} d^K_{\mathbb{HB}_n}(\mathbf{X},\mathbf{Y}) = \frac{1}{2} \log \frac{1 + \sqrt{\lambda_{max}}}{1 - \sqrt{\lambda_{max}}}, \end{equation} where λi,i=1,…,n\lambda_i,i=1,\ldots,n are the eigenvalues of ϕX(Y)ϕX(Y)H\phi_{\mathbf{X}}({\mathbf{Y}}) \phi_{\mathbf{X}}({\mathbf{Y}})^H, and λmax\lambda_{max} is their maximum value.

The geometric interpretation of the normalization procedure can be given as follows. The Riemannian distance is used to compute the intrinsic statistics within each batch, while the Kobayashi pseudodistance provides a geometry-aware mechanism for incorporating new batch statistics into the running statistics.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response

We thank the reviewers for their insightful comments and suggestions for improving our work. First, we would like to clarify that the statistics presented in Tab. 10 (see Appendix A.3) of the original submission has errors. The errors were caused by our incorrect interpretation of some arguments passed to the Python script for generating synthetic data, and by some mistakes we made when preparing the table. The details are given below:

  • The graph reconstruction datasets were created using the official code of the method in López et al. (2021). The command for generating a grid has two arguments griddimsgrid_dims and nodesnodes that indicate the number of dimensions and the number of nodes in the grid. According to the repository's README file, this command creates a grid of griddimsgrid_dims dimensions with ⌊nodes(1/griddims)⌋\lfloor nodes^{(1/grid_dims)} \rfloor nodes along each dimension. Therefore, the grid layout of the grid graph should be 3×3×3×3×33 \times 3 \times 3 \times 3 \times 3 (5 dimensions with ⌊243(1/5)⌋\lfloor 243^{(1/5)} \rfloor nodes along each dimension) instead of 5×2435 \times 243. Similarly, the grid layout of Tree ×\times Grid and Grid ×\times Tree, which are created by setting griddims=3grid_dims=3 and nodes=27nodes=27, is 3×3×33 \times 3 \times 3.
  • The Tree Height value of Tree ⋄\diamond Grid was incorrect. The correct value was 3.
  • For Tree ⋄\diamond Grid and Grid ⋄\diamond Tree, the numbers of nodes, edges, and triples were incorrect. We obtained the correct statistics by inspecting the generated graphs.

We would like to provide the correct statistics in Table A below.

GraphNodesEdgesTriplesGrid LayoutTree BranchingTree Height
Grid243810294033 ×\times 3 ×\times 3 ×\times 3 ×\times 3
Tree403978033
Tree ×\times Grid108032135826603 ×\times 3 ×\times 333
Grid ×\times Tree108032135826603 ×\times 3 ×\times 333
Tree ⋄\diamond Grid6409992044804 ×\times 433
Grid ⋄\diamond Tree100815745075284 ×\times 425

Table A: Statistics of the graph reconstruction datasets.

We apologize for this issue. We confirm that this correction does not affect the experimental results or the conclusions drawn in the paper.

Dear reviewers, the authors responded to your reviews. Could you please go over the responses, see if you have any further questions, and if appropriate, reassess your evaluation of the paper? Note that the deadline for all reviewer-author discussions is November 18th. Thanks, AC

Dear Reviewer,

We have updated the table (Tab. 10) regarding the statistics of the graph reconstruction datasets in the manuscript and the general response accordingly. We would like to clarify that the grid layout of the grid graph is 3×3×3×3×33 \times 3 \times 3 \times 3 \times 3.

Thank you again for your valuable feedback.

Best regards,

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.