acceptodds
Under review as a conference paper at ICLR 2027

Symmetric Spaces of Groups over Involutive Algebras for Graph Embeddings

Abstract

Symmetric spaces such as the Siegel upper half-space and the Siegel disk domain have proven promising for graph embedding. Recently, they have also been proposed as representation spaces for discriminative neural networks. In this paper, we are interested in generalizations of Siegel spaces for representation learning. In particular, we focus on symmetric spaces associated with symplectic groups over involutive algebras, which have recently been introduced as generalizations of the real symplectic group. Although these symmetric spaces offer the capacity to capture rich geometric structures, core components required for practical implementation of neural networks on these space are missing, making the task of learning such parametric models difficult. To this end, we further develop the theory of symplectic groups over involutive algebras by deriving closed-form formulas for some missing components, i.e., automorphisms and their inverses. We consider realizations of the studied spaces, and demonstrate their effectiveness for graph tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Symmetric spaces, symplectic groups, involutive algebras, graph embeddings, Riemannian neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qv6pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a general abstract framework that unifies several existing symmetric spaces, including hyperbolic spaces, symmetric positive definite (SPD) spaces, and Siegel spaces, for representation learning. The authors develop a theoretical formulation based on symplectic groups over involutive algebras, which enables the construction of core geometric mappings suitable for integration into deep learning models. This theoretical direction appears novel and potentially interesting. However, the current manuscript suffers from weak presentation and remains incomplete in several important aspects, in my view, prevents it from meeting the expected standard of an ICLR submission.

Critical strengths
  • To the best of my knowledge, the mathematical results presented in Section 3, including the mapping φ\varphi, the automorphisms ϕX(Y)\phi_{\mathbf{X}}(\mathbf{Y}), and their inverses, are novel.
Critical weaknesses
  • The authors do not provide any supplementary materials, such as source code or a link to a GitHub repository, making it difficult to assess the reproducibility of the proposed method.

  • The paper has duplicated references (lines 540-545):

    • Galileo Namata, Ben London, Lise Getoor, and Bert Huang. Query-driven Active Surveying for Collective Classification. In 10th International Workshop on Mining and Learning with Graphs, volume 8, pp. 1, 2012a.
    • Galileo Namata, Ben London, Lise Getoor, and Bert Huang. Query-driven Active Surveying for Collective Classification. In Workshop on Mining and Learning with Graphs, 2012b.
  • The formulations of the mapping φ\varphi, the autormorphisms ϕX(Y)\phi_{\mathbf{X}}(\mathbf{Y}) and their reverse are mathematically heavy for both symplectic groups Sp2(A,σ)Sp_2(A, \sigma) and Sp2(AC,σC)Sp_2(A_{\mathbb{C}}, \sigma_{\mathbb{C}}). Such complexity is likely to introduce higher computational overhead than existing geometric learning methods such as HGCN [1] and SPD-GCN [2]. However, the current manuscript does not provide any analysis or empirical comparison of computational cost against these baselines, which makes it difficult to assess the practical efficiency of the proposed framework.

  • Section 5.2, which describes the application of the proposed precompact model framework QUnQU_{n} to Graph Convolutional Networks (GCNs), is currently unclear and insufficiently explained. For instance, in Lines 341–342, the authors state that Z~\mathbf{\tilde{Z}} is transformed, yet the subsequent unnumbered equation in Lines 343–345 does not contain Z~\mathbf{\tilde{Z}} and instead introduces Z′\mathbf{Z'}. The role of this representation appears to be directly related to the paper's mathematical contributions, including the automorphisms ϕX(Y)\phi_{\mathbf{X}}(\mathbf{Y}) and their inverses. However, in the subsequent GCN formulation (Lines 352–366), particularly in the message-passing equations in Lines 362–366, it is unclear how Z~\mathbf{\tilde{Z}}, Z′\mathbf{Z'} or these proposed mappings are incorporated. As a result, the connection between the precompact model QUnQU_n framework and the actual GCN architecture is difficult to follow. More importantly, the paper does not provide a clear interpretation of the proposed geometric mappings in the context of deep learning, unlike existing methods. For examples, HGCN[1] introduces the logarithmic and exponential mappings which convert node features from hyperbolic space to Euclidean space, in order to utilize the fundamental algebra operations. In the current manuscript, however, the functional roles of the mapping φ\varphi, the automorphisms ϕX(Y)\phi_{\mathbf{X}}(\mathbf{Y}), and their inverses within a neural network remain unclear. The authors should explicitly explain what each mapping accomplishes in the forward computation and how these operations contribute to the proposed GCN architecture, followed by a pseudocode or an illustration.

  • The manuscript does not provide sufficient theoretical justification or empirical evidence to demonstrate why the proposed highly abstract precompact model, QUnQU_n, is superior to existing geometric frameworks such as hyperbolic, SPD, and Siegel spaces. These established spaces are typically motivated by the structural properties they are able to represent; for example, hyperbolic geometry is well suited to hierarchical and tree-like graphs, whereas SPD spaces can capture more complex relational structures, including combinations of grid-like and tree-like patterns. In contrast, the paper does not clearly identify which intrinsic graph properties are uniquely preserved or modeled by QUnQU_n, nor why existing geometries are insufficient for these structures. The authors should therefore provide a stronger justification for the necessity of QUnQU_n, supported by theoretical analysis or targeted experiments that demonstrate the specific graph structures for which the proposed framework offers advantages over existing geometric approaches.

  • Table 4 omits several accuracy values and is only briefly referenced in Section 6.4 without further explanation.

[1] Ines Chami, Rex Ying, Christopher R, and Jure Leskovec. Hyperbolic Graph Convolutional Neural Networks. CoRR, abs/1910.12933, 2019.

[2] Wei Zhao, Federico Lopez, J. Maxwell Riestenberg, Michael Strube, Diaaeldin Taha, and Steve Trettel. Modeling Graphs Beyond Hyperbolic: Graph Neural Networks in Symmetric Positive Definite Matrices. CoRR, abs/2306.14064, 2023.

Questions for authors
  • Could the authors compare the proposed framework with existing geometric approaches in the GCN setting, such as hyperbolic methods (HGCN[1] and HGAT[2]) and Siegel-space-based models, through both theoretical analysis and empirical evaluation?

  • How is classification performed within the proposed framework? In particular, which linear classifier or regression formulation is used for the final prediction stage?

  • How can the proposed precompact model be extended to other graph architectures, such as Graph Transformers?

[1] Ines Chami, Rex Ying, Christopher R, and Jure Leskovec. Hyperbolic Graph Convolutional Neural Networks. CoRR, abs/1910.12933, 2019.

[2] Y. Zhang, X. Wang, X. Jiang, C. Shi, and Y. Ye, Hyperbolic graph attention network. CoRR, vol. abs/1912.03046, 2019. [Online].Available: http://arxiv.org/abs/1912.03046

3 comments

We thank the Reviewer for carefully examining our paper and for providing valuable feedback that helps us to improve the clarity of the paper and to strengthen it.

Q1: Could the authors compare the proposed framework with existing geometric approaches in the GCN setting, such as hyperbolic methods (HGCN[1] and HGAT[2]) and Siegel-space-based models, through both theoretical analysis and empirical evaluation?

For the graph and graph node classification experiments, we focused on comparing SPD-GCN and our model because the proposed GCN aims to improve SPD-GCN. In constrast, GCNs on Siegel spaces have not been proposed in previous works. On the other hand, the comparison between the studied domain and Siegel spaces was performed for the graph reconstruction task because Siegel models have proven to be strong baselines in López et al. (2021b) for this task. A detailed comparison between SPD-GCN and hyperbolic GCNs is given in the original SPD-GCN's paper in which SPD-GCN generally outperforms the hyperbolic models for the graph and graph node classification tasks. For this reason, our experiments focused on comparing SPD-GCN and our model.

Q2: How is classification performed within the proposed framework? In particular, which linear classifier or regression formulation is used for the final prediction stage?

The mapping from the original node features to QBn\mathbb{QB}_n is applied only once. After the initial embedding Yj0\mathbf{Y}^0_j of node jj is obtained, the feature transform, aggregation, and bias and nonlinearity operations are performed on SPD manifolds. Therefore, the classification is performed on SPD manifolds. In our experiments, we used the linear classifier implemented in the official code of the method in López et al. (2021b).

Q3: How can the proposed precompact model be extended to other graph architectures, such as Graph Transformers?

We thank the Reviewer for raising this interesting question. In order to extend the current method to other graph architectures such as Graph Transformers, further investigation is needed because of several challenges when working on the studied bounded domain. For instance, the question of how to effectively and efficiently compute a geometric mean on this domain remains open. Also, the lack of geometric mappings such as the exponential and logarithmic maps make the construction of fundamental building blocks on this domain difficult compared to other popular representation spaces such as hyperbolic manifolds.

Comment 1: The paper has duplicated references (lines 540-545)

We thank the Reviewer for this valuable observation. We will carefully check all the references and address this issue in the revised paper.

Comment 2: The current manuscript does not provide any analysis or empirical comparison of computational cost against these baselines

The training times of the competing models on the graph reconstruction task are reported in Table L. Our model is computationally more efficient during the training phase than Siegel models. It is important to note that the training time of the best Siegel model is about 3$\times$ that of our model in most cases.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$_6$0.580.0112.3314.034.8010.86
S$^{F_1}_{4}$1.090.0225.8826.017.7821.94
B$^{F_1}_{4}$1.450.0346.5045.5512.2639.74
QB4\mathbb{QB}_40.650.0114.7514.814.9613.06

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

Comment 3: Table 4 omits several accuracy values and is only briefly referenced in Section 6.4 without further explanation.

We apologize for this issue. We used the structure and formatting of another table as a template. However, due to an oversight during the preparation of the paper, the table was not fully updated, and some entries from the original table remained. Regarding the difference between the performances of QB-GCN in Tabs. 2 and 4, we originally intended to compare QB-GCN and QB-GCN-NoAut in a setting where QB-GCN achieves the best accuracy with lowest dimension, which we found to be 10×1010 \times 10. The results of QB-GCN in Tab. 4 were obtained with this setting.

We acknowledge this error and take full responsibility for it. We have carefully reviewed the corrected results and present them in Table M below. As can be observed, applying the transformation ϕP(⋅)\phi_{\mathbf{P}}(\cdot) improves the mean accuracy of QB-GCN-NoAut on 3 of the 4 datasets.

ModelAirportPubmedCoraDisease
QB-GCN-NoAut88.6$\pm$0.978.1$\pm$0.582.9$\pm$0.897.0$\pm$1.6
QB-GCN92.9$\pm$0.679.6$\pm$0.482.8$\pm$0.598.1$\pm$1.9

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

Dear Reviewer,

We have updated our previous responses to provide additional clarification. We would appreciate it if you could kindly refer to the updated version.

Thank you again for your valuable feedback.

Best regards,

Reviewer_M9ybOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends symmetric-space graph representation methods beyond Siegel spaces by studying spaces associated with symplectic groups over involutive algebras. The theoretical contribution derives mappings between upper-half-space and bounded realizations and closed-form automorphisms/inverses. The authors then define a distance function on a quaternionic bounded domain and construct QB-GCN, which maps the proposed geometric representation into an SPD manifold before performing message passing.

Experiments cover graph reconstruction, node classification, and graph classification. The reconstruction results compare against SPD/Siegel spaces, while the classification experiments compare mainly against SPD-GCN.

Critical strengths
  1. The paper contains a nontrivial mathematical development rather than a superficial manifold substitution. Providing explicit automorphisms and inverses for these generalized symmetric spaces is potentially useful infrastructure for future geometric deep-learning work, and the paper makes a serious effort to turn the geometry into an implementable graph model.

  2. The reported results are consistently favorable relative to the chosen SPD/Siegel baselines: QB4 has lower reconstruction distortion in Table 1, QB-GCN substantially improves Airport node classification, and the mean graph-classification accuracy is higher than SPD-GCN on all three reported datasets. These results suggest that the added representation family is worth further investigation.

Critical weaknesses
  1. The empirical comparison is too narrow and capacity-confounded to support the claimed learning advantage. For node/graph classification the main baseline is SPD-GCN, with no strong Euclidean, hyperbolic, or Siegel-based neural baselines under a common protocol. More importantly, the compared representation dimensions do not appear capacity matched: node classification uses QB8 versus Sym+_14, and from the parameterization in Appendix A.2 QB8 has more real scalar degrees of freedom than a 14x14 real SPD point. The reconstruction comparison similarly uses QB4 against lower-DOF SPD/Siegel representations. Since the paper itself attributes gains to “additional degrees of freedom,” a matched-DOF/parameter/compute comparison is necessary to show that the geometry, rather than simply more capacity, is responsible. More controlled experiments that replace the embedding representation within otherwise comparable graph learning architectures would help isolate the contribution of the proposed geometry from other architectural factors.

  2. The ablation/reporting inconsistency materially reduces confidence in the empirical results. Table 4 says it evaluates five node-classification datasets, although Section 6.2 lists four; it includes Citeseer only in the NoAut row, and the QB-GCN row is missing Pubmed, Citeseer, and Disease values. Several NoAut numbers exactly match the SPD-GCN results in Table 2. If removing the automorphism makes the model mathematically identical to SPD-GCN, that equivalence should be stated; otherwise the exact numerical matches plus missing cells look like a reporting error. This is the only component ablation, so its incompleteness is especially problematic. Furthermore, the QB-GCN results in Table 4 differ from those reported in Table 2 for Airport (93.2 vs. 92.9) and Cora (83.8 vs. 82.8), without an explanation of whether different experimental settings were used.

Questions for authors
  1. What are the missing QB-GCN entries in Table 4, and why does Table 4 include Citeseer when it is not listed among the node-classification datasets in Section 6.2 or Table 2? Why do the QB-GCN-NoAut results for Airport, Pubmed, and Cora exactly match the SPD-GCN results in Table 2? Is there a formal equivalence after removing Eq. (5)?
  2. Please report the real scalar degrees of freedom and trainable parameter counts for QB8 and Sym+_14 (and QB4 versus the reconstruction baselines). Were these choices intended to match capacity?
1 comment

Q1: What are the missing QB-GCN entries in Table 4, and why does Table 4 include Citeseer when it is not listed among the node-classification datasets in Section 6.2 or Table 2? Why do the QB-GCN-NoAut results for Airport, Pubmed, and Cora exactly match the SPD-GCN results in Table 2? Is there a formal equivalence after removing Eq. (5)?

We apologize for the inconsistency in Tab. 4 of the paper. We used the structure and formatting of another table as a template. However, due to an oversight during the preparation of the paper, the table was not fully updated, and some entries from the original table remained. We acknowledge this error and take full responsibility for it. We have carefully reviewed the corrected results and present them in Table F below. We apologize for this oversight and any confusion it may have caused. We appreciate the reviewers' attention to this issue and the opportunity to correct it.

QB-GCN-NoAut is created by removing the transformation ϕP(⋅)\phi_{\mathbf{P}}(\cdot) in Eq. (5) from QB-GCN. As can be observed from Table F, applying the transformation ϕP(⋅)\phi_{\mathbf{P}}(\cdot) improves the mean accuracy of QB-GCN-NoAut on 3 of the 4 datasets.

ModelAirportPubmedCoraDisease
QB-GCN-NoAut88.6$\pm$0.978.1$\pm$0.582.9$\pm$0.897.0$\pm$1.6
QB-GCN92.9$\pm$0.679.6$\pm$0.482.8$\pm$0.598.1$\pm$1.9

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

Q2: Please report the real scalar degrees of freedom and trainable parameter counts for QB8 and Sym+_14 (and QB4 versus the reconstruction baselines). Were these choices intended to match capacity?

The real dimensions of SPD$_{14}$ and QB$_8$ are 105 and 136, respectively. For the node classification experiments, we used the best setting for the dimension of SPD embeddings (when this dimension exceeds 14, we did not observe clear improvement). The numbers of trainable parameters for SPD$_{14}$ and QB$_8$ in these experiments are provided in Table G.

ModelAirportPubmedCoraDisease
SPD$_{14}$339299690282982197584
QB$_8$87163993110175971402

Table G: The number of trainable parameters of SPD$_{14}$ and QB$_8$ on the node classification experiments.

For the graph reconstruction experiments, we used the settings for the dimensions of SPD and Siegel embeddings in López et al., (2021), and set the dimension of embeddings on QB to the dimension of Siegel embeddings. The numbers of trainable parameters for the competing models in these experiments are provided in Table H.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_{6}$510384022680226801344021168
S$^{F_1}_4$7776128034560345602048032256
B$^{F_1}_4$7776128034560345602048032256
QB$_4$8748144038880388802304036288

Table H: Number of trainable parameters for the competing models in the graph reconstruction experiments.

We performed additional experiments to obtain the results of QB$_3$ and SPD$_8$. These results and those from Tab. 1 of the paper are provided in Tables I and J. For each table, the real dimensions of the embeddings are matched.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_6$14.33$\pm$0.045.45$\pm$0.069.07$\pm$0.028.97$\pm$0.022.23$\pm$0.045.31$\pm$0.03
QB$_3$8.85$\pm$0.082.18$\pm$0.045.37$\pm$0.035.40$\pm$0.032.21$\pm$0.063.58$\pm$0.07

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

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_8$14.38$\pm$0.035.44$\pm$0.058.62$\pm$0.038.56$\pm$0.032.39$\pm$0.055.89$\pm$0.04
QB$_4$1.16$\pm$0.140.57$\pm$0.041.96$\pm$0.021.53$\pm$0.050.62$\pm$0.071.10$\pm$0.09

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

Reviewer_A6YyOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper brings together existing symmetric-space constructions into a graph representation model based on Sp2n(C)/USp(2n)\mathrm{Sp}_{2n}(\mathbb C)/\mathrm{USp}(2n). Following the upper half-space and bounded-domain descriptions of Siegel space, it represents this space using quaternionic matrices and provides explicit formulas for the Cayley transform, automorphisms, and distances. The network maps real symmetric node features into the quaternionic domain, applies a learnable automorphism, and converts the output into SPD matrices. A dimension-reducing BiMap connects these representations to an SPD graph convolution framework.

Critical strengths
  1. The loss function is appropriate (A.1.2), which encourages embedding distances to match graph distances and follows existing graph embedding methods.

2.The construction gives Euclidean node features a structured SPD representation and connects it to an existing SPD graph convolution framework through BiMap. Each step has an explicit formula, making the model accessible for implementation and analysis. It provides a concrete way to study the combination of a nonlinear front end and SPD propagation.

Critical weaknesses
  1. Theoretical correctness and contribution:
  • The real symmetric space used in the paper differs from the usual Siegel space. The space and its group action were established in prior work. The contribution of this paper lies in its application on graph learning, explicit implementation, and choice of nonlinear structure. The current derivations do not establish a new theoretical result or principle.
  • The intrinsic Kobayashi interpretation of the distance has mathematical problems, whichn leads the interpretation be flase. Alternating correct interpretation reassess distance's novelty.
  1. Model contribution and experimental evidence:
  • The motivation does not explain the benefit of a richer representation space clearly.
  • The classification experiments do not provide enough detail about parameter budgets and training settings.

Details:

  1. The authors acknowledge prior work, but they need to identify which theoretical results are inherited and what remains new. The claimed contribution of the transform and its inverse also needs to be examined.

    • The use of upper half-space and bounded models, the transform, moving a point to the basepoint, and decomposition for distance computation follows the symmetric-space graph embedding framework of López et al. [1]. The transform in Proposition 3.4--3.5 are essentially Cayley transform in referenced paper.

    • Symplectic groups over algebras with an anti-involution, the symmetric space used here, and its group action models were established in earlier work [2, 3]. They describe the bounded model and the transitive action of the conjugate group. The formula and group action in Propositions 3.4--3.5 of the submission follow directly from the corresponding projective coordinates and representatives. These two propositions should be viewed as existing results.

    • In QBn\mathbb Q\mathcal B_n background, Proposition 3.6 agrees with the standard matrix-ball formula. For X,Y∈QBnX,Y\in\mathbb Q\mathcal B_n, write X∗=σ1(X)X^*=\sigma_1(X) and denote AA here for the upper-left block. Set D=(I−X∗X)−1/2D=(I-X^*X)^{-1/2}and A=σ0(D)A=\sigma_0(D), one can calculate

      A=(I−XX∗)−1/2,AX=XD,X∗A=DX∗.A=(I-XX^*)^{-1/2},\qquad AX=XD,\qquad X^*A=DX^*.

      The formula therefore becomes

      \phi_X(Y) &=(AY-XD)(-X^*AY+D)^{-1}\\\\ &=(I-XX^*)^{-1/2}(Y-X)\\\\ &\qquad\cdot(I-X^*Y)^{-1}(I-X^*X)^{1/2}. \end{aligned}$$ Ostrovskii, Shulman, and Turowska [4] give the same form of operator-ball transformation and state that changing the sign of the parameter gives its inverse. Taking their parameter to be $-X$ gives the formula above. The paper still needs to prove that this formula preserves the fixed-point subspace of the additional anti-involution and belongs to the specified acting group.
  2. The added value of using the general framework of algebras with an anti-involution is not clear. This algebraic framework already exists. After introducing it, the paper focuses on matrix models in its applications and implementation, and the proofs are largely restricted to the matrix case.

  3. The concrete space in the paper is

    \mathrm{Sp}_{2n}(\mathbb C)/\mathrm{USp}(2n),\qquad \dim_{\mathbb R}=2n^2+n.$$ The usual Kobayashi pseudodistance is defined using complex manifolds and holomorphic discs. Here, the case $n=1$ is a three-dimensional real ball. Even when $n$ is even, the paper provides no compatible complex structure. A quotient of a complex symplectic group by its maximal compact subgroup is not automatically a complex manifold; a complex structure requires a separate definition and justification. The distance in equation (4) needs a precise geometric interpretation. In the matrix realization, suppose that $\\|\cdot\\|_2$ denotes the quaternionic operator norm. Let $s_1,\ldots,s_n$ be the singular values of $\phi_X(Y)$ and set $v_i=\log\frac{1+s_i}{1-s_i}$. Since $\\|\phi_X(Y)\\|_2=\max_i s_i$, (4) becomes $$d(X,Y) =\frac12\max_i\log\frac{1+s_i}{1-s_i} =\frac12\\|v\\|_\infty.$$ This identifies (4) with the $F_\infty$ distance obtained from the Weyl distance vector in the symmetric-space framework of López et al. [1]. The authors should clarify this interpretation and provide the complex structure needed for their claim of an intrinsic Kobayashi distance on $\mathbb Q\mathcal B_n$.
  4. Proposition 3.7 is the standard inverse relation for Proposition 3.6. The defect operators A,DA,D are unchanged when the parameter changes sign, take MX=(A−XD−X∗AD)M_X= \begin{pmatrix} A&-XD\\\\ -X^*A&D \end{pmatrix}, then

    \qquad \phi_X^{-1}=\phi_{-X}.$$ Once domain preservation in Proposition 3.6 is proved, applying the same argument to $-X$ proves that the inverse preserves the domain. This does not constitute a separate theoretical contribution.
  5. The classification model starts from a real symmetric n×nn\times n matrix XX, which has n(n+1)/2n(n+1)/2 independent entries. For a fixed shared parameter PP, the subsequent smooth transformations produce a family of representations of dimension at most n(n+1)/2n(n+1)/2. Thus, the larger dimension of QBn\mathbb Q\mathcal B_n does not by itself establish greater representational capacity for the classification model. Please explain how the proposed transformations improve the representation of the input features.

  6. Page 13 describes MM as an invertible real matrix as in Zhao et al. [5], but they in fact use an orthogonal parameterization. Please specify the constraint used here.

    If MM is orthogonal, then the feature transformation and aggregation satisfy

    = \sum_{k\in N(j)}v_{jk} M_l\bigl(\log Y_k^{l-1}\bigr)M_l^T.$$ Thus in logarithmic coordinates, the aggregate operation is essentially linear. The role of the proposed geometry in these operations should be explained.
  7. Section 5.2 maps the quaternionic representation to Z′Z', then to the real matrix Z′′Z'', and applies BiMap before SPD message passing. Positive definiteness is preserved by BiMap, but the paper does not establish its output or the subsequent message-passing layers retain a representation in QBn\mathbb Q\mathcal B_n. Please clarify which geometric structure is preserved at each stage, or if this preserving is necessary.

  8. The results do not adequately separate the benefit of additional adjustable coordinates from the benefit of geometry. By definition, Table 1 compares

    \dim_{\mathbb R}\mathrm{Siegel}_4=20,\qquad \dim_{\mathbb R}\mathrm{SPD}_6=21.$$ In graph reconstruction, each node embedding is optimized directly. The proposed space gives each node more independently adjustable coordinates than the baselines. The reported improvement may therefore come from both the geometry and the additional coordinates. Comparisons at matched dimensions are needed to distinguish these effects.
  9. Under the calculation with BiMap rules, the proposed SPD backend is SPD16\mathrm{SPD}_{16}:

    \to \mathbb Q\mathcal U_8 \to \operatorname{Herm}_{16}^{+}(\mathbb C) \to \operatorname{Sym}_{32}^{+}(\mathbb R) \to \operatorname{SPD}_{16}^{+}(\mathbb R).$$ whereas the baseline uses $\mathrm{SPD}_{14}$.
  10. Table 4 on page 9 is incomplete and differs from Table 2. The Full model has no entries for Pubmed, Citeseer, or Disease. Airport is reported as 92.9±0.692.9\pm0.6 in Table 2 and 93.2±0.793.2\pm0.7 in Table 4; Cora is reported as 82.8±0.582.8\pm0.5 and 83.8±0.483.8\pm0.4, respectively. The NoAut part for Airport, Pubmed, and Cora in Table 4 exactly match those of SPD-GCN in Table 2. Author need to complete experimental result and explain why these phenomenas are reasonable.

  11. The graph construction descriptions and statistics in Table 5 on page 12 are inconsistent. Tree$\diamond$Grid and Grid$\diamond$Tree are listed with 364 nodes and 363 edges, and 63 nodes and 62 edges, respectively. Both satisfy ∣E∣=∣V∣−1|E|=|V|-1, which means the graph is a tree; but a connected rooted product containing a 4×44\times4 grid must contain cycles. One the other side, under the map (t,g)↦(g,t)(t,g)\mapsto(g,t), Tree$\times$Grid and Grid$\times$Tree are isomorphic, so must not have difference.

References

  1. Federico López, Beatrice Pozzetti, Steve Trettel, Michael Strube, and Anna Wienhard. Symmetric Spaces for Graph Embeddings: A Finsler-Riemannian Approach. In Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, vol. 139, pp. 7090--7101, 2021. https://proceedings.mlr.press/v139/lopez21a.html.

  2. Daniele Alessandrini, Arkady Berenstein, Vladimir Retakh, Eugen Rogozinnikov, and Anna Wienhard. Symplectic groups over noncommutative algebras. arXiv:2106.08736v1, 2021. https://arxiv.org/abs/2106.08736v1.

  3. Pengfei Huang, Georgios Kydonakis, Eugen Rogozinnikov, and Anna Wienhard. Symmetric spaces for groups over involutive algebras and applications to Higgs bundles. arXiv:2509.02002v1, 2025. https://arxiv.org/abs/2509.02002v1.

  4. M. I. Ostrovskii, V. S. Shulman, and L. Turowska. Unitarizable representations and fixed points of groups of biholomorphic transformations of operator balls. Journal of Functional Analysis, 257 (2009), pp. 2476--2496. doi:10.1016/j.jfa.2009.06.021. Preprint: arXiv:0811.1759v1, 2008, https://arxiv.org/abs/0811.1759v1. The page and equation cited above refer to this preprint.

  5. Wei Zhao, Federico López, J. Maxwell Riestenberg, Michael Strube, Diaaeldin Taha, and Steve Trettel. Modeling Graphs Beyond Hyperbolic: Graph Neural Networks in Symmetric Positive Definite Matrices. arXiv:2306.14064v1, 2023. https://arxiv.org/abs/2306.14064v1. Public implementation: https://github.com/andyweizhao/SPD4GNNs.

Questions for authors
  1. Which new results go beyond direct substitution or routine expansion of existing theory? What new mathematical obstacle do they resolve? Does the paper prove a result beyond the known abstract framework, or show how the generalization provides a unified proof or an application benefit across different examples? If only the current matrix example is retained, which aspects of the general algebraic framework become trivial?

  2. Please clarify the source, proof, and contribution of the distance formula. Why the paper choose F∞F_\infty? Compared with F1F_1 and the Riemannian distance on the same space, what advantage does it offer for the task, theory, or computation?

  3. Please specify the parameterization of MM used in this paper.

  4. Please complete the tables and specify the actual architecture and configuration of NoAut, including which training runs produced each reported result. Please provide the graph generation code, correct the statistics. Please add completation of experiment in Table~4 with details to explain why such phenomena happens as in previous mentioned.

  5. Please add comparisons with matched degrees of freedom, such as QB3\mathbb Q\mathcal B_3 versus SPD6\mathrm{SPD}_6, and QB4\mathbb Q\mathcal B_4 versus SPD8\mathrm{SPD}_8.

I would be happy to raise my score to 2 or 3, depending on the clarifications provided by the authors in the rebuttal. Thank you.

13 comments

Q2: Please clarify the source, proof, and contribution of the distance formula. Why the paper choose F∞F_{\infty}? Compared with F1F_1 and the Riemannian distance on the same space, what advantage does it offer for the task, theory, or computation?

We thank the reviewer for pointing out the connection between our distance construction and the vector-valued distance used in López et al. (2021b). We agree that our distance can be interpreted as a Finsler distance inducing the l∞l^{\infty}-metric.

We would like to clarify the motivation behind the use of the Kobayashi distance. One of the main objectives of our work is to derive a closed form formula of automorphisms on the proposed bounded domain. Inspired by the Kobayashi distance on the Siegel disk domain, we adopt an analogous automorphism-based distance construction. This enables a direct use of the derived automorphisms to measure the dissimilarity between embedded points and to validate our main contribution.

Although QBn\mathbb{QB}_n has a Riemannian symmetric space structure as shown in Alessandrini et al. (2021); Huang et al. (2025), we are not aware of a closed form formula for its Riemannian distance in the quaternionic coordinates used in our work. We performed additional graph reconstruction experiments to compare the performance of the F1F_1 and F∞F_{\infty} distances on QBn\mathbb{QB}_n. Results are provided in Table A.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
QB4F∞\mathbb{QB}^{F_{\infty}}_41.16$\pm$0.140.57$\pm$0.041.96$\pm$0.021.53$\pm$0.050.62$\pm$0.071.10$\pm$0.09
QB4F1\mathbb{QB}^{F_1}_41.02$\pm$0.110.55$\pm$0.061.84$\pm$0.051.47$\pm$0.080.54$\pm$0.090.92$\pm$0.05

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

We propose to clarify our distance construction in Section 5.1 in order to avoid any misunderstanding that QBn\mathbb{QB}_n is naturally endowed with an invariant complex structure and therefore the use of the Kobayashi distance is theoretically justified.

Q3: Please specify the parameterization of MM used in this paper.

We sincerely thank the reviewer for pointing out this issue.

As stated in Appendix A.2.2, our implementation is based on the official code of the method in Zhao et al. (2023). We used the default settings in the configuration file, which parameterizes M\mathbf{M} as a square matrix and uses the orthogonal matrix obtained from the QR decomposition of M\mathbf{M} for the feature transform operation of the proposed GCN. We will address it in the revised manuscript.

We note that the mapping from the original node features to QBn\mathbb{QB}_n is applied only once. After the initial embedding Yj0\mathbf{Y}^0_j of node jj is obtained, the feature transform, aggregation, and bias and nonlinearity operations are performed on SPD manifolds. Therefore, the subsequent message-passing layers do not retain a representation in QBn\mathbb{QB}_n.

Q4: Please complete the tables and specify the actual architecture and configuration of NoAut, including which training runs produced each reported result.

We apologize for the inconsistency in Tab. 4 of the paper. We used the structure and formatting of another table as a template. However, due to an oversight during the preparation of the paper, the table was not fully updated, and some entries from the original table remained. We acknowledge this error and take full responsibility for it. We have carefully reviewed the corrected results and present them in Table B below. We apologize for this oversight and any confusion it may have caused. We appreciate the reviewers' attention to this issue and the opportunity to correct it.

QB-GCN-NoAut is created by removing the transformation ϕP(⋅)\phi_{\mathbf{P}}(\cdot) in Eq. (5) from QB-GCN. As can be observed from Table B, applying the transformation ϕP(⋅)\phi_{\mathbf{P}}(\cdot) improves the mean accuracy of QB-GCN-NoAut on 3 of the 4 datasets.

ModelAirportPubmedCoraDisease
QB-GCN-NoAut88.6$\pm$0.978.1$\pm$0.582.9$\pm$0.897.0$\pm$1.6
QB-GCN92.9$\pm$0.679.6$\pm$0.482.8$\pm$0.598.1$\pm$1.9

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

Q5: Please provide the graph generation code, correct the statistics.

Regarding the graph generation code, as stated in Appendix A.1.2 (line 619), our experiments were based on the official code of the method in López et al. (2021b). We used the graph generation code included in their code to create our datasets. We would like to confirm that the statistics presented in Tab. 5 (see Appendix A.1.2) of the original submission have errors. We sincerely thank the reviewer for carefully examining our experimental results and pointing out this inconsistency in the dataset statistics. 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 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, will be 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 C 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 C: 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.

Q6: Please add comparisons with matched degrees of freedom, such as QB3\mathbb{QB}_3 versus SPD$_6$, and QB4\mathbb{QB}_4 versus SPD$_8$.

We performed additional experiments to obtain the results of QB3\mathbb{QB}_3 and SPD$_8$. These results along with those from Tab. 1 of the paper are provided in Tables D and E.

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_6$14.33$\pm$0.045.45$\pm$0.069.07$\pm$0.028.97$\pm$0.022.23$\pm$0.045.31$\pm$0.03
QB3\mathbb{QB}_38.85$\pm$0.082.18$\pm$0.045.37$\pm$0.035.40$\pm$0.032.21$\pm$0.063.58$\pm$0.07

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

SpaceGridTreeTree ×\times GridGrid ×\times TreeTree ⋄\diamond GridGrid ⋄\diamond Tree
SPD$^R_8$14.38$\pm$0.035.44$\pm$0.058.62$\pm$0.038.56$\pm$0.032.39$\pm$0.055.89$\pm$0.04
QB4\mathbb{QB}_41.16$\pm$0.140.57$\pm$0.041.96$\pm$0.021.53$\pm$0.050.62$\pm$0.071.10$\pm$0.09

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

The results clearly demonstrate the superiority of QB3\mathbb{QB}_3 over SPD$_6$ and of QB4\mathbb{QB}_4 over SPD$_8$.

Comment 1: The formula and group action in Propositions 3.4--3.5 of the submission follow directly from the corresponding projective coordinates and representatives. These two propositions should be viewed as existing results.

We acknowledge that the mappings between the upper half-space model and the precompact model given in Proposition 3.4 can be deduced via the mapping from the projective model to the upper half-space model and the mapping from the projective model to the precompact model, both of which are given in Alessandrini et al. (2021); Huang et al. (2025). However, we did not find explicit mappings given in Proposition 3.4 in these papers.

Note that the mappings in Proposition 3.4 are given as \begin{equation*} \varphi(\mathbf{Z}) = (\mathbf{Z} - \mathbf{i}\mathbf{I})(\mathbf{I} - \mathbf{i}\mathbf{Z})^{-1}. \end{equation*} \begin{equation*} \varphi^{(-1)}(\mathbf{Z}_0) = (\mathbf{Z}_0 + \mathbf{i}\mathbf{I})(\mathbf{i}\mathbf{Z}_0 + \mathbf{I})^{-1}, \end{equation*} while the standard Cayley transform and its inverse are given as \begin{equation*} \varphi(\mathbf{Z}) = (\mathbf{Z} - \mathbf{i}\mathbf{I})(\mathbf{Z} + \mathbf{i}\mathbf{I})^{-1}. \end{equation*} \begin{equation*} \varphi^{(-1)}(\mathbf{Z}_0) = \mathbf{i}(\mathbf{I} + \mathbf{Z}_0)(\mathbf{I} - \mathbf{Z}_0)^{-1}. \end{equation*}

While both formulations are valid, our formulation has the advantage of remaining consistent with the precompact model and its coordinates as defined in Alessandrini et al. (2021); Huang et al. (2025). For the sake of clarity, we therefore include these mappings in the form of a proposition.

Regarding Proposition 3.5, we acknowledge that the result has been stated, without proof, in Remark 6.17 of Alessandrini et al. (2021). We state it and provide a proof in the matrix case. We will clarify these points in the revised paper.

Comment 2: ...The paper still needs to prove that this formula preserves the fixed-point subspace of the additional anti-involution and belongs to the specified acting group.

In our proof of Proposition 3.5, we show (lines 902-903) that the Möbius transformation associated with M0\mathbf{M}_0 (defined in line 855) maps any point in B\mathfrak{B} to a point in B\mathfrak{B} and that this action is transitive. The equation that characterizes the group acting on B\mathfrak{B} via Möbius transformations is σ0(M)TΩM=Ω\sigma_0(\mathbf{M})^T \Omega \mathbf{M} = \Omega. The matrix M\mathbf{M} defined in line 911 satisfies this equation. Applying the Möbius transformation associated with this matrix to a point Y∈B\mathbf{Y} \in \mathfrak{B} results in the mapping in Proposition 3.6. Therefore, this mapping preserves the fixed-point subspace of the additional anti-involution and belongs to the specified acting group.

Comment 3: Once domain preservation in Proposition 3.6 is proved, applying the same argument to -X proves that the inverse preserves the domain. This does not constitute a separate theoretical contribution.

We thank the reviewer for suggesting this simpler proof of Proposition 3.7. We will adopt this suggestion in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.