acceptodds
Under review as a conference paper at ICLR 2027

𝕊𝕃(n) Representation Learning: An Intrinsic Mixed-Curvature Space with Higher Curvature Capacities and Deeper Order-Aware Composition

17 views

Abstract

Mixed-curvature representation learning seeks to capture rich geometric structures that cannot be adequately modeled by a single curvature regime. Existing approaches largely rely on product manifolds, which require manually specifying how different curvature spaces are combined and separate their curvature contributions across factors. We introduce the space, a representation geometry defined by the simple constraint and a left invariant Schatten- Finsler structure. Despite this minimal construction, exhibits pointwise negative, zero, and positive flag curvature around a common flagpole, while its mixed-curvature and curvature-coupling capacities are asymptotically maximal relative to the intrinsic geometric upper bound. Beyond geometry, its noncommutative group structure provides inherent order sensitivity, and its non-nilpotent Lie algebra admits nonzero nested Lie brackets at arbitrary depth, enabling deep order-aware composition. Empirically, consistently outperforms a broad range of representation manifold baselines across graph benchmarks at different scales. It reduces average distortion over the strongest baselines by on KEGG and on HumanCyc, and improves Hits@20 by on OGBL-PPA. Experiments on Flickr30k-Order further support its ability to capture higher order dependencies from ordered composition. Together, these results show how a seemingly simple structural constraint can yield unexpectedly rich geometry, capacity, and composition within a unified representation space.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, mixed-curvature geometry, Lie groups, Finsler geometry, noncommutative composition, curvature coupling

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_WBHvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SL(n) as a single latent representation space built on the special linear group with a left-invariant Schatten-p tangent norm. The authors argue that existing mixed-curvature product manifolds suffer from two limitations: the factor decomposition must be predetermined, and product metrics have vanishing mixed sectional curvature so curvature regimes do not interact across factors. They prove SL(n) exhibits pointwise positive, zero, and negative flag curvature around a common flagpole, with asymptotically maximal mixed-curvature and coupling capacities, and its non-nilpotent Lie algebra supports arbitrarily deep nested Lie brackets. Experiments on graph metric reconstruction, large-scale link prediction, and order-sensitive composition report consistent improvements over geometric baselines.

Critical strengths
  1. The paper identifies a real, provable limitation of product metrics (vanishing mixed sectional curvature) and builds a single manifold that circumvents it.

  2. The order-sensitive composition experiment compares SL(4) against Heisenberg and triangular groups with finite composition depth. The performance ordering is consistent with depth ordering, isolating an algebraic property from a geometric one.

Critical weaknesses
  1. The claim about product metrics (vanishing mixed sectional curvature) is correct, but prior work explicitly studies geometry interaction. CurvRec (SIGIR 2023) uses curvature-aware geodesic distances to maximize mutual information between Euclidean and non-Euclidean spaces, stating message passing is hindered without interaction. GIE (AAAI 2022) proposes geometry interaction across Euclidean, hyperbolic, and hyperspherical spaces. MKRec (2024) transfers geometric information between manifold spaces. These operate at the representational level rather than the curvature-operator level, but the paper neither cites them nor distinguishes its "coupling" from their "interaction." A claim of first genuine interaction cannot omit published work on exactly that.

  2. Main experiments use Exp-AdamW for SL(n) while manifold baselines use standard optimization. The Riem-AdamW control shows Exp-AdamW is better for SL(n), not that all geometries were equally optimized. A matched-parameterization comparison is needed. No wall-clock or memory comparison is reported, so the practical significance of the gains cannot be assessed.

  3. No experiment isolates coupling from mixed curvature. The claim that coupling drives gains rests on indirect evidence. A matched-capacity baseline with zeroed cross-sign interaction is needed.

  4. The semidistance is a non-metric surrogate. Local tightness and path-refinement are shown only on a 64-node KEGG subgraph. Adequacy on OGBL-PPA and HumanCyc is unverified. Using a non-metric objective on the largest experiment deserves more than an appendix subsection.

Questions for authors
  1. The paper states product mixed-curvature spaces have “no genuine interaction.” How does this relate to CurvRec (SIGIR 2023), GIE (AAAI 2022), and MKRec (2024), which explicitly model interaction between curved manifolds? If the distinction is representational vs. curvature-theoretic, please state it explicitly in the main text. If these are not considered relevant, explain why.

  2. Can you provide a controlled comparison where all geometries—including product manifolds, SPD, Grassmann, and Siegel—are optimized under a matched parameterization and learning-rate protocol? If rankings change under matched optimization, what does that imply for attribution of gains to geometry?

  3. Report wall-clock time per update, peak GPU memory, and total training time for SL(8) versus the strongest vector baseline on OGBL-PPA.

  4. Provide branch-margin distributions, fraction outside the principal-log domain, and comparison against K=4 or K=8 path approximations for HumanCyc and OGBL-PPA.

  5. Design an experiment that isolates coupling from mixed curvature, e.g., a variant retaining positive and negative eigenspaces but suppressing cross-sign curvature interaction.

4 comments
AuthorsOpenReview ↗

Response to Reviewer WBHv

First and foremost, we sincerely thank the reviewer WBHv for the patient reading of our work and the insightful and constructive comments! We provide our detailed responses below:



Q1(&W1): Relation to Representation Level Geometry Interaction (CurvRec, GIE, MKRec):

Thank you for this question and your broad reading of the literature! We have added citations to these works in the revised version, and we clarify that "coupling" in our paper refers specifically to the curvature operator. The two are complementary rather than competing, and differ in three ways:

  1. Where: These works interact across several spaces, while we couple within the same tangent space of a single manifold.
  2. What: These works exchange feature representations, while we couple positive and negative curvature modes, measured by the coupling capacity.
  3. Source: These works rely on extra modules, and their underlying spaces have no cross sign curvature coupling, while our coupling is a property of the geometry itself.

In addition, these works depend on task specific architectures for recommendation or knowledge graph completion, so they cannot be directly compared in our shallow embedding setting. In contrast, SL(n)\mathbb{SL}(n) is a general representation space and has been shown to be effective on several tasks, including reconstruction, link prediction, node classification, and ordered composition. Our ablation study also shows that the full SL(n)\mathbb{SL}(n) clearly outperforms Flat (removing the nonlinear Lie group structure) and Polar (keeping both curvature signs without coupling, see Q5).


Q2(&W2): Controlled Comparison under Matched Parameterization and Optimization:

Thank you for this question. To satisfy their constraints, all manifolds need to choose which parameters enter the optimizer. SL(n)\mathbb{SL}(n) uses Lie algebra coordinates, which are mapped by the exponential map and then optimized by standard AdamW. Therefore, EXP-AdamW is not a special optimizer, but a general parameterization that allows SL(n)\mathbb{SL}(n) to be trained directly with AdamW. On HumanCyc, OGBL-PPA, and Airport, all geometries use strictly matched hyperparameters and dimensions. Since the matrix manifolds were originally matched by matrix dimension, we additionally provide results on KEGG with matched intrinsic dimension:

ModelIntrinsic Dim.Avg. Distortion ↓q90 ↓High-related ↓Worst-group ↓
S32×H32\mathbb{S}^{32}\times\mathbb{H}^{32}640.0761 ± 0.00000.16670.10040.1773
S32×E32\mathbb{S}^{32}\times\mathbb{E}^{32}640.0846 ± 0.00040.18330.10940.1821
H32×E32\mathbb{H}^{32}\times\mathbb{E}^{32}640.0707 ± 0.00000.15590.09660.1691
Hκ121×E21×Sκ221\mathbb{H}^{21}_{\kappa_1}\times\mathbb{E}^{21}\times\mathbb{S}^{21}_{\kappa_2}630.0596 ± 0.00060.13880.08760.1692
Gr(7,16)\mathrm{Gr}(7,16)630.0888 ± 0.00010.18170.11070.1903
SPD(11)\mathrm{SPD}(11)660.0810 ± 0.00010.17280.10950.1787
Siegel(8)\mathrm{Siegel}(8)720.0590 ± 0.00020.12620.07650.1507
SL(8)\mathbb{SL}(8)630.0329 ± 0.00010.06980.03890.0841

With matched intrinsic dimension, the conclusion does not change: SL(8)\mathbb{SL}(8) remains the best on all metrics.


Q3(&W2): Computational Cost:

Thank you for this question. We measured all geometries on the same single H200 GPU (3 warm up iterations, then 10 measured iterations with CUDA synchronization). Each iteration includes sampling, forward, backward, gradient clipping, and the optimizer step, with an effective positive edge batch size of 4,194,304.

ModelTime per Update (s) ↓Peak Memory (GiB) ↓Positive Edges / s ↑
S32×H31\mathbb{S}^{32}\times\mathbb{H}^{31}7.57 ± 2.501.63554,106
S32×E31\mathbb{S}^{32}\times\mathbb{E}^{31}7.79 ± 2.281.63538,452
H32×E31\mathbb{H}^{32}\times\mathbb{E}^{31}8.00 ± 2.121.63524,020
H21×E21×S21\mathbb{H}^{21}\times\mathbb{E}^{21}\times\mathbb{S}^{21}7.73 ± 2.321.63542,352
Hκ121×E21×Sκ221\mathbb{H}^{21}_{\kappa_1}\times\mathbb{E}^{21}\times\mathbb{S}^{21}_{\kappa_2}7.76 ± 2.651.63540,589
Gr(7,16)\mathrm{Gr}(7,16)7.07 ± 1.891.63593,299
SPD(8)\mathrm{SPD}(8)10.27 ± 1.317.25408,456
Siegel(8)\mathrm{Siegel}(8)29.47 ± 2.238.03142,338
SL(8)\mathbb{SL}(8)9.78 ± 1.836.50428,925

SL(8)\mathbb{SL}(8) is about 1.3 times slower per update than the strongest vector baseline, and is faster and uses less memory than matrix manifolds (SPD(8) and Siegel(8)). The total training time is 63.18 h versus 45.05 h, which also includes evaluation, communication, and I/O. The higher memory comes from intermediate matrix operations, not from parameter storage. Full results are in the revised appendix (p. [19]).


Q4(&W4): Numerical Validity of the Semidistance on Large-Scale Data:

Thank you very much for this question. We have added verification on all three datasets. For Principal-logarithm domain: All of the millions of relative matrices checked on the three datasets lie in the principal-logarithm domain, with minimum branch margins of 0.83–2.97 rad, and no NaN or Inf values occur. For Path approximation: We perform end-to-end training with K∈{1,2,4,8}K\in\{1,2,4,8\} path segments as the objective, and measure the gap on 100 randomly sampled pairs for each learned embedding. The relative gap between the semidistance and the K=8K=8 reference is only 1.4%–2.1%, and the change in downstream performance is at most 0.5%, which is far smaller than the standard deviation.

These results have been added to the revised version (p. [18–19], Fig. 7).

Theoretical guarantee: As proven in Lemma B.1 of the original paper, DSLD_{\mathbb{SL}} is exactly the length measured along the standard exponential path. It is therefore always an upper bound on the intrinsic distance, and locally it agrees with the intrinsic distance up to second order, with errors only in higher-order terms; the closer the points, the more accurate the approximation. DSLD_{\mathbb{SL}} is nonnegative, symmetric, and separates distinct points, and only the triangle inequality is not guaranteed. Such distances are common in representation learning: for example, the squared Euclidean distance and the squared Lorentzian distance in hyperbolic representation learning [1] do not satisfy the triangle inequality in any space with midpoints, yet they are widely used as optimization objectives. Moreover, as shown above, the relative gap between DSLD_{\mathbb{SL}} and the refined path distance is only about 2%, so the practical effect of the missing triangle inequality is negligible.


Q5(&W3): Isolating Curvature Coupling from Mixed Curvature:

Thank you very much for this question, which is one of the most important questions for our work! We first clarify that the product-manifold baselines in the original paper (e.g., S×H×E\mathbb{S}\times\mathbb{H}\times\mathbb{E}) already have mixed curvature without coupling [3], and SL(n)\mathbb{SL}(n) outperforms them on all tasks. However, their parameterization differs from that of SL(n)\mathbb{SL}(n), so we added a strictly matched decoupled control.

In the ablation, we use the fact that every invertible matrix has a unique polar decomposition Ai=QiPiA_i=Q_iP_i, where QiQ_i is orthogonal and PiP_i is symmetric positive definite [2]. Since det⁡Ai=1\det A_i=1 and det⁡Pi>0\det P_i>0, we have det⁡Qi=det⁡Pi=1\det Q_i=\det P_i=1, i.e., Qi∈SO(n)Q_i\in SO(n) and Pi∈SPD1(n)P_i\in SPD_1(n). We equip SO(n)×SPD1(n)SO(n)\times SPD_1(n) with the product metric. SO(n)SO(n) provides nonnegative curvature and SPD1(n)SPD_1(n) provides nonpositive curvature, so both the positive and negative curvature eigenspaces are kept. However, the curvature tensor of a product metric is block-diagonal, so the sectional curvature across the two factors is always zero and there is no coupling [3]. This is exactly the variant you described. It shares the same parameters, dimension, optimizer, and objective with the full model (p=2p=2), and differs only in whether coupling exists.

DatasetMetricSL(n)\mathbb{SL}(n) w/o Curvature CouplingSL(n)\mathbb{SL}(n) (Full)
KEGGAvg. Distortion ↓0.08050.0366
KEGGHigh-related ↓0.10910.0430
HumanCycAvg. Distortion ↓0.05230.0316
HumanCycHigh-related ↓0.05500.0331
OGBL-PPAHits@20 ↑0.11240.1229

These results show that mixed curvature without coupling does not bring improvement, and the gains come from intrinsic coupling, especially in regions with high curvature mixture. We also performed an additional flat ablation that removes the nonlinear Lie group structure. The full ablation study is on p. [10] of the revised version.



Finally, we thank you again for your patient reading and your insightful and valuable comments! Your comments have made this work more complete, and we hope that this work, improved together with your help, can bring some inspiration and contribution to the community.


[1]Law et al. (2019), Lorentzian Distance Learning for Hyperbolic Representations. ICML 2019.
[2]Higham (2008), Functions of Matrices: Theory and Computation. SIAM.
[3]Sáez de Ocáriz Borde (2026), On the Expressive Power of Mixed-Curvature Representations in Product Manifolds. Workshop on Geometry-grounded Representation Learning and Generative Modeling at ICLR 2026.

Thank you for the detailed response. I have raised my score to 3 (Weak Acceptance).

Dear Reviewer WBHv,

Thank you very much for taking the time to evaluate our rebuttal. We are glad that our responses addressed your concerns, and we are sincerely grateful for your decision to raise the score.

We are also grateful for the time and care you devoted to reviewing our paper and helping us improve it! We will incorporate your suggestions into the final version, and we hope that this work, strengthened by your feedback, will bring some meaningful inspiration and contributions to the community.

Best regards,
The Authors

Reviewer_mTBmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the SL(n)SL(n) space, an intrinsic mixed-curvature geometry designed to capture heterogeneous structural patterns in graphs. Section 2 provides a detailed construction of this space by integrating theoretical foundations from several geometric frameworks, including smooth manifolds, Finsler manifolds, Riemannian manifolds, and Lie groups. The experimental study is extensive and evaluates both the effectiveness of the proposed space and its deep order-aware composition property. Nevertheless, several aspects of the experimental design and the practical feasibility of the proposed geometric framework remain unclear and, in my view, require further clarification.

Critical strengths
  • The paper is well written. The figures are clearly presented and effectively visualized.

  • The theoretical development of the SL(n)SL(n) space in Section 2 is rigorous and novel. Its key properties are carefully analyzed and formally supported by Theorem 2.1, Corollary 2.3, and Lemma 2.4.

Critical weaknesses
  • Although the paper introduces the novel SL(n)SL(n) space, its practical integration into modern deep learning architectures remains insufficiently explored. In particular, the manuscript does not clearly explain how SL(n)SL(n) representations can be incorporated into standard graph neural networks, such as Graph Convolutional Networks (GCNs) or Graph Transformers (GTs), nor how prediction heads such as linear classifiers can be defined in this space. Furthermore, the empirical study mainly focuses on unsupervised tasks, including metric reconstruction and link prediction. Consequently, it remains unclear whether the proposed framework can be effectively extended to supervised graph-learning problems, such as node and graph classification. Additional methodological discussion and experimental validation on these tasks would strengthen the practical relevance of the proposed space.

  • A further concern is that SL(n)SL(n) relies on matrix-valued embeddings, whereas most existing GCN and Graph Transformer architectures are designed around vector-valued representations. Supporting matrix embeddings may require substantial modifications to fundamental neural-network operations and can introduce considerably higher computational and memory costs than conventional vector embeddings. This raises important questions regarding the scalability and practical feasibility of SL(n)SL(n), particularly when integrated into large architectures such as Graph Transformers. The authors should provide a clearer implementation strategy and a computational analysis comparing matrix-based SL(n)SL(n) representations with conventional vector-based alternatives.

Questions for authors
  • Which tool did the authors use to create the visualizations in Figures 1, 3, 4, and 6?

  • Could the authors provide the intuition or hypothesis for selecting the Schatten order pp? Is pp determined solely through brute-force search?

  • Recent studies have explored pseudo-Riemannian manifolds for learning representations of graphs with mixed topological structures [1, 2, 3, 4]. Such manifolds can intrinsically accommodate heterogeneous geometric relationships among nodes, including hyperbolic, spherical, and Euclidean structures. Moreover, pseudo-Riemannian approaches typically operate on vector-valued embeddings, facilitating their integration with standard graph neural network architectures and achieving promising performance across various graph-learning tasks. Could the authors provide a more detailed comparison between the proposed SL(n)SL(n) space and pseudo-Riemannian approaches, supported by theoretical analysis and/or empirical evaluation?

[1] Law, M.; Stam, J. Ultrahyperbolic representation learning. Advances in neural information processing systems 2020, 33, 1668–1678.

[2] Law, M. Ultrahyperbolic neural networks. Advances in Neural Information Processing Systems 2021, 34, 22058–22069.

[3] Xiong, B.; Zhu, S.; Potyka, N.; Pan, S.; Zhou, C.; Staab, S. Pseudo-riemannian graph convolutional networks. Advances in Neural Information Processing Systems 2022, 35, 3488–3501.

[4] Viet Quan Le and Viet Cuong Ta. Pseudo-riemannian graph transformer. Advances in Neural Information Processing Systems 2025, volume 38, Main Conference, pp. 15951–15983.

4 comments
AuthorsOpenReview ↗

Response to Reviewer mTBm

First and foremost, we sincerely thank the reviewer mTBm for the patient reading of our work and the insightful and constructive comments! We provide our detailed responses below:



Q1: Tools of Figures:
Thank you very much for your appreciation of our work! Figures 4 and 6 were created using Python (Matplotlib, NetworkX, SciPy, etc.), while Figures 1 and 3 combined Python generated graphics with careful layout and arrangement in PowerPoint. Through these visualizations, we aimed to convey our core ideas more clearly and intuitively.


Q2: Intuition and Search Strategy for Schatten-pp:

Thank you for the question! Our SL(n)\mathbb{SL}(n) representation space has two main geometric hyperparameters, nn and pp. The matrix dimension nn controls the representation dimensionality and potential curvature-coupling capacity, while the Schatten order pp controls spectral weighting.

Intuitively, datasets with complex, distributed curvature interactions may benefit from larger nn and smaller pp, accommodating more coupled modes while balancing spectral contributions. In contrast, structures dominated by a few strong spectral directions may favor larger pp. Since the optimal hyperparameters depend on the data structure and task, we typically begin with a validation based search over p∈{2,4,8}p\in\{2,4,8\} and extend the range when necessary.


Q3: About Pseudo-Riemannian Manifolds:

Thank you for pointing out this relevant work! We added a dimension-matched Ultrahyperbolic Representation Learning baseline and searched over the signature parameters p,qp, q.

DatasetRepresentationAvg. Distortion ↓q90 ↓Worst-group ↓MAE ↓
KEGGUltrahyperbolic (47,16)0.23230 ± 0.002560.485200.278890.62137
KEGGSL(8)\mathbb{SL}(8)0.0329 ± 0.00010.06980.08410.0811
HumanCycUltrahyperbolic (142,1)0.22673 ± 0.000250.377390.242700.81491
HumanCycSL(12)\mathbb{SL}(12)0.0313 ± 0.00030.06780.03600.1098

The results show that, in our shallow setting without node features or GNN encoders, SL(n)\mathbb{SL}(n) substantially outperforms the pseudo-Riemannian baseline on both datasets. This is consistent with known limitations of ultrahyperbolic geometry: its dissimilarity is not a metric and vanishes on null cones, so distinct points can have zero distance [1], and not all pairs are joined by geodesics [2]. Moreover, Qp,q\mathcal{Q}^{p,q} has constant curvature, so its heterogeneity comes from the indefinite signature rather than from coupled curvature.

In contrast, SL(n)\mathbb{SL}(n) is complete under a positive-definite Finsler metric, FA(V)=∥A−1V∥Sp>0F_A(V)=\Vert A^{-1}V\Vert_{S_p}>0 for V≠0V\neq 0, so any two points are joined by a minimizing geodesic, and our semidistance DSL(A,B)=∥log⁡(A−1B)∥SpD_{\mathbb{SL}}(A,B)=\Vert\log(A^{-1}B)\Vert_{S_p} is positive for A≠BA\neq B.


W1: Integration into Deep Architectures and Supervised Graph Learning:

Thank you for this important question! As you noted, the SL(n)\mathbb{SL}(n) space has great potential for further applications. Based on this work, deep architectures can be developed in the future. This follows the same path as ultrahyperbolic spaces [1] and mixed-curvature product spaces [4]: the representation space was proposed first, and specific architectures were developed later [2, 3, 5]. Our work also focuses on the space itself and uses shallow embeddings for evaluation, which removes the effect of the architecture.

That said, it is actually very simple to use SL(n)\mathbb{SL}(n) in standard architectures:

  1. General GNNs: Reshape the output vector into a matrix, apply the trace projection, and take the matrix exponential to obtain Ai=exp⁡(Πsl(hi))A_i=\exp(\Pi_{\mathfrak{sl}}(h_i)). No constrained optimization is needed.
  2. GCNs: Aggregate in the Lie algebra, Ai←Aiexp⁡(∑jαijlog⁡(Ai−1Aj))A_i\leftarrow A_i\exp\big(\sum_j\alpha_{ij}\log(A_i^{-1}A_j)\big).
  3. Graph Transformers: Use −DSL(Ai,Aj)-D_{\mathbb{SL}}(A_i,A_j) as the attention score.

For prediction heads, a standard classifier can be applied to the Lie-algebra coordinates, or the distance to class prototypes, −βDSL(Pk,Ai)-\beta D_{\mathbb{SL}}(P_k,A_i), can be used.

We also added a supervised node classification experiment on the widely used Airport dataset (no features or GNNs; a frozen embedding followed by an MLP classifier):

GeometryDim.Avg. Distortion ↓Micro-F1 ↑Macro-F1 ↑
S21×H21×E21\mathbb{S}^{21}\times\mathbb{H}^{21}\times\mathbb{E}^{21}630.058187.2785.10
Gr(2,34)\mathrm{Gr}(2,34)640.063386.4383.49
SPD(11)\mathrm{SPD}(11)660.061891.2389.31
Siegel(8)\mathrm{Siegel}(8)720.049690.1987.94
SL(8)\mathbb{SL}(8)630.042291.4489.84

SL(8)\mathbb{SL}(8) achieves the best results in both reconstruction and classification. The above content has been added to the revised version (p. [8]).


W2: Implementation and Computational Cost of Matrix-Valued Embeddings:

Thank you for raising this practical question! Clarifying the implementation and computational cost helps make SL(n)\mathbb{SL}(n) more practical for real applications. In fact, SL(n)\mathbb{SL}(n) embeddings are essentially vector-based. The network outputs an ordinary vector hi∈Rn2h_i\in\mathbb{R}^{n^2} as usual, which is reshaped into an n×nn\times n matrix HiH_i. The trace projection Πsl(Hi)=Hi−tr(Hi)nI\Pi_{\mathfrak{sl}}(H_i)=H_i-\frac{\mathrm{tr}(H_i)}{n}I then removes the trace and maps HiH_i onto the Lie algebra sl(n)\mathfrak{sl}(n) of traceless matrices. Finally, the matrix exponential gives the group element Ai=exp⁡(Πsl(Hi))∈SL(n)A_i=\exp(\Pi_{\mathfrak{sl}}(H_i))\in\mathbb{SL}(n). Since det⁡exp⁡(X)=etr(X)=1\det\exp(X)=e^{\mathrm{tr}(X)}=1 for any traceless XX, the constraint det⁡Ai=1\det A_i=1 holds automatically, so no constrained optimization is needed, and standard AdamW can be used for training. Therefore, using SL(n)\mathbb{SL}(n) in a GCN or Graph Transformer does not require changing any operations, and the parameter storage is the same as that of a vector of the same dimension (see W1).

The extra cost comes from distance computation (O(n3)O(n^3), compared with O(n2)O(n^2) for a vector of the same dimension). This cost depends only on nn and not on the graph size, and the performance already saturates at n=8n=8 to 1212. On OGBL-PPA, SL(8)\mathbb{SL}(8) takes 9.78 s per update with a peak memory of 6.50 GiB, while the vector baselines take 7.07–8.00 s and 1.63 GiB, so SL(8)\mathbb{SL}(8) is only about 1.2–1.4 times slower. SL(8)\mathbb{SL}(8) is also better in both time and memory than SPD(8) (10.27 s, 7.25 GiB) and Siegel(8) (29.47 s, 8.03 GiB). The higher memory comes from intermediate results of matrix operations, not from parameter storage. The full results are given in the appendix of the revised version (p. [19]).



Finally, we thank you again for your patient reading and your insightful and valuable comments! Your comments have made this work more complete, and we hope that this work, improved together with your help, can bring some inspiration and contribution to the community.


[1] Law & Stam (2020), Ultrahyperbolic Representation Learning. NeurIPS 2020.
[2] Law (2021), Ultrahyperbolic Neural Networks. NeurIPS 2021.
[3] Xiong et al. (2022), Pseudo-Riemannian Graph Convolutional Networks. NeurIPS 2022.
[4] Gu et al. (2019), Learning Mixed-Curvature Representations in Product Spaces. ICLR 2019.
[5] Bachmann et al. (2020), Constant Curvature Graph Convolutional Networks. ICML 2020.

Thank you for the detailed response. I am satisfied with the clarifications and am increasing my score to a weak accept. Please ensure these revisions are incorporated into the final manuscript.

Dear Reviewer mTBm,

Thank you very much for taking the time to evaluate our rebuttal. We are glad that our responses addressed your concerns, and we are sincerely grateful for your decision to raise the score.

We are also grateful for the time and care you devoted to reviewing our paper and helping us improve it! We will incorporate your suggestions into the final version, and we hope that this work, strengthened by your feedback, will bring some meaningful inspiration and contributions to the community.

Best regards,
The Authors

Reviewer_3JpYOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SL(n) as a unified representation space equipped with a left-invariant Schatten-p Finsler structure. The authors establish theoretical results concerning intrinsic mixed flag curvature, curvature-coupling capacity, and unbounded order depth arising from the noncommutative Lie group structure. Experiments on graph reconstruction, link prediction, and order-sensitive multimodal representation demonstrate improvements over several geometric baselines.

Critical strengths

S1. Theorem 2.1 and Corollary 2.3 provide a mathematical characterization of intrinsic mixed flag curvature and asymptotically maximal curvature-coupling capacity. These results go beyond simply combining constant-curvature manifolds and constitute the strongest theoretical contribution.

S2. The paper evaluates multiple geometric baselines across graph reconstruction, large-scale link prediction, and ordered composition. The improvements on KEGG, HumanCyc, and OGBL-PPA are substantial, and the additional sensitivity and optimization studies strengthen the empirical evaluation.

Critical weaknesses

W1. The connection between theoretical curvature properties and empirical improvements remains insufficiently established.

The central claim is that intrinsic curvature coupling provides a representation advantage over product manifolds. However, the experiments do not isolate curvature coupling from other factors, including the Schatten-p norm, matrix-valued parameterization, and nonlinear distance computation.

In particular, the practical objective uses the matrix-logarithm semidistance rather than the intrinsic Finsler distance. Although the appendix establishes a local approximation and provides numerical path-refinement results, these do not demonstrate that curvature coupling is the source of the performance gains. The improvements in high curvature-mixture regions are suggestive but remain correlational.

A dimension-matched flat Lie-algebra baseline using the same Schatten-p norm would be particularly important for evaluating this claim.

W2. The claimed advantage of unbounded order depth is not convincingly distinguished from ordinary noncommutative composition.

Lemma 2.4 establishes infinite order depth through the non-nilpotent structure of sl(n). However, this is an algebraic property shared by non-nilpotent Lie algebras and does not directly establish superior effective compositional capacity in learned representations.

The Flickr30k-Order experiments demonstrate the usefulness of ordered group multiplication, but they do not conclusively isolate the benefits of unbounded order depth from other architectural and parameterization differences. The absence of stronger noncommutative and conventional order-aware neural baselines further limits the significance of this conclusion.

Questions for authors

Q1. Can the authors introduce a dimension- and parameter-matched flat Lie-algebra baseline using the same Schatten-p norm, replacing the group-based semidistance with a tangent-space distance? This would help determine whether the improvements arise from intrinsic curvature coupling or from the matrix-valued norm and parameterization.

Q2. Can the authors provide direct empirical evidence connecting their proposed curvature-coupling capacity to representation quality? In particular, can they demonstrate that the learned geometric structure, rather than simply higher representational flexibility, explains the improvements on high curvature-mixture regions?

Q3. What evidence establishes that unbounded order depth provides advantages beyond finite-depth noncommutative composition? Could the authors include stronger parameter-matched controls and order-aware neural baselines to isolate this contribution?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3JpY

First and foremost, we sincerely thank the reviewer 3JpY for the patient reading of our work and the insightful and constructive comments! We provide our detailed responses below:



Q1(&W1): Dimension- and Parameter-Matched Flat Lie-Algebra Baseline:

Thank you very much for this question, which is key to identifying the source of the gains! Following your suggestion, we constructed a flat Lie-algebra baseline. It shares the same Lie-algebra parameters Xi∈sl(n)X_i\in\mathfrak{sl}(n), the same Schatten-pp norm with the same pp, dimension, optimizer, and objective as the full model, and only replaces the group-based semidistance with the distance in the Lie algebra (tangent space), Dflat=∥Xj−Xi∥SpD_{\mathrm{flat}}=\Vert X_j-X_i\Vert_{S_p}. By the BCH formula, the semidistance of the full model can be written as DSL=∥Xj−Xi−12[Xi,Xj]+⋯∥Sp.D_{\mathbb{SL}}=\Big\Vert X_j-X_i-\tfrac{1}{2}[X_i,X_j]+\cdots\Big\Vert_{S_p}.

The flat baseline is exactly its zeroth-order truncation. It keeps the matrix-valued norm and parameterization, and only removes the nonlinear group structure introduced by the Lie brackets. Therefore, any gap between the two can only come from the group geometry itself.

DatasetMetricSL(n)\mathbb{SL}(n) Flat BaselineSL(n)\mathbb{SL}(n) (Full)
KEGGAvg. Distortion ↓0.08050.0366
HumanCycAvg. Distortion ↓0.05140.0316
OGBL-PPAHits@20 ↑0.09620.1229

With the same matrix-valued norm and parameterization, the full SL(n)\mathbb{SL}(n) clearly outperforms the flat baseline on all datasets, showing that the gains do not come from the norm or the parameterization.


Q2(&W1): Direct Evidence Linking Curvature Coupling to Representation Quality

Thank you very much for this question, which is one of the most important questions for our work! We first clarify that the product-manifold baselines in the original paper (e.g., S×H×E\mathbb{S}\times\mathbb{H}\times\mathbb{E}) already have mixed curvature without coupling [2], and SL(n)\mathbb{SL}(n) outperforms them on all tasks. However, their parameterization differs from that of SL(n)\mathbb{SL}(n), so to rule out differences in representational flexibility, we added a strictly matched decoupled control.

Every invertible matrix has a unique polar decomposition Ai=QiPiA_i=Q_iP_i, where QiQ_i is orthogonal and PiP_i is symmetric positive definite [1]. Since det⁡Ai=1\det A_i=1 and det⁡Pi>0\det P_i>0, we have det⁡Qi=det⁡Pi=1\det Q_i=\det P_i=1, i.e., Qi∈SO(n)Q_i\in SO(n) and Pi∈SPD1(n)P_i\in SPD_1(n). This gives a one-to-one correspondence between SL(n)\mathbb{SL}(n) and SO(n)×SPD1(n)SO(n)\times SPD_1(n), so the two models represent exactly the same set of points with the same parameters, and thus have the same representational flexibility. We equip SO(n)×SPD1(n)SO(n)\times SPD_1(n) with the product metric. SO(n)SO(n) provides nonnegative curvature and SPD1(n)SPD_1(n) provides nonpositive curvature, so both the positive and negative curvature eigenspaces are kept. However, the curvature tensor of a product metric is block-diagonal, so the sectional curvature across the two factors is always zero and there is no coupling [2]. The two models share the same parameters, dimension, optimizer, and pp, and differ only in whether coupling exists.

DatasetMetricSL(n)\mathbb{SL}(n) w/o Curvature CouplingSL(n)\mathbb{SL}(n) (Full)Improvement
KEGGAvg. Distortion ↓0.08050.036654.5%
KEGGHigh-related ↓0.10910.043060.6%
KEGGLow–Low ↓0.05310.036331.6%
HumanCycAvg. Distortion ↓0.05230.031639.6%
HumanCycHigh-related ↓0.05500.033139.8%
HumanCycLow–Low ↓0.04800.031035.4%
OGBL-PPAHits@20 ↑0.11240.12299.3%
OGBL-PPAMRR ↑0.01650.019317.0%
OGBL-PPARank90 ↓12767988422.6%

Since the two models have the same representational flexibility, the gains cannot be explained by flexibility, but come from the learned coupled geometry. Moreover, the reduction is largest in high curvature-mixture regions (60.6% vs. 31.6% for low–low pairs on KEGG), which directly connects curvature coupling to representation quality where mixed curvature is most prevalent. The full ablation study is on p. [10] of the revised version.


Q3(&W2): Advantage of Unbounded Order Depth over Finite-Depth Noncommutative Composition:

Thank you for pointing out that non-nilpotency alone does not directly establish stronger compositional capacity in learned models. We provide additional evidence from both theory and experiments.

From a theoretical perspective, prior studies show that there is a strict capability gap between finite-depth (i.e., nilpotent) composition and non-nilpotent composition. Take the classic TCS problem, the A5A_5 word problem, which is a standard hard problem for state tracking [3]. From the classic results of Mal'cev and Neumann–Taylor on nilpotent groups [4, 5], it follows that ordered multiplication in a nilpotent group, under any parameterization, cannot exactly solve this problem for inputs of arbitrary length. In contrast, SL(n)\mathbb{SL}(n) contains A5A_5 for n≥3n\ge3 and can solve it exactly. Similarly, Grazzi et al. [6] prove that linear recurrences whose eigenvalues are all 1 cannot even solve parity for arbitrary lengths, and the Heisenberg and unitriangular groups belong to this class. The above theory characterizes the limit of expressive power, and the experiments below further verify that this advantage also holds in actual training.

From a experiments perspective, On the same frozen CLIP features, we compare composition models with the same dimension of 15, and add a 15-dimensional Gated Recurrent Unit (GRU) :

ModelDim.Order DepthOrderAcc ↑HardAcc ↑MRR ↑
SL(4)\mathbb{SL}(4) w/o Ordered Composition15017.973.540.272
Heisenberg H7H^715194.2882.320.906
UT(6)\mathrm{UT}(6)15496.3089.110.942
SL(4)\mathbb{SL}(4)15∞97.46‾\underline{97.46}92.87‾\underline{92.87}0.962‾\underline{0.962}
GRU (neural network)15—98.6395.500.976

With the same dimension, performance increases monotonically with order depth, showing that the gains come from order depth rather than parameterization or capacity. Even UT(15)\mathrm{UT}(15) (97.25), with 7 times more parameters, still falls behind SL(4)\mathbb{SL}(4). GRU achieves higher accuracy through nonlinear gating and recurrent parameters, while SL(4)\mathbb{SL}(4) introduces no recurrent parameters and reaches a close level using group multiplication alone.



Finally, we thank you again for your patient reading and your insightful and valuable comments! Your comments have made this work more complete, and we hope that this work, improved together with your help, can bring some inspiration and contribution to the community.


[1] Higham (2008), Functions of Matrices: Theory and Computation. SIAM.

[2] Sáez de Ocáriz Borde (2026), On the Expressive Power of Mixed-Curvature Representations in Product Manifolds. Workshop on Geometry-grounded Representation Learning and Generative Modeling at ICLR 2026.

[3] Barrington (1989), Bounded-Width Polynomial-Size Branching Programs Recognize Exactly Those Languages in NC¹. Journal of Computer and System Sciences.

[4] Mal'cev (1953), Nilpotent Semigroups. Uchenye Zapiski Ivanovskogo Gosudarstvennogo Pedagogicheskogo Instituta.

[5] Neumann & Taylor (1963), Subsemigroups of Nilpotent Groups. Proceedings of the Royal Society A.

[6] Grazzi et al. (2025), Unlocking State-Tracking in Linear RNNs Through Negative Eigenvalues. ICLR 2025.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns. The flat Lie-algebra baseline provides useful evidence that the nonlinear group structure contributes beyond the Schatten norm and matrix parameterization. The polar-decomposition control is also informative. I therefore raise my rating to 4 (Clear Acceptance).

Dear Reviewer 3JpY,

Thank you very much for taking the time to carefully evaluate our rebuttal and additional experiments. We are delighted that our responses addressed your concerns, and we sincerely appreciate your recognition and decision to raise the score.

We are especially grateful for your constructive feedback and the important ablation studies you suggested, which helped us more clearly demonstrate the benefits of our method to readers. Your thoughtful comments and support mean a great deal to us.

We will incorporate your valuable suggestions and additional experimental findings into the final version. We sincerely hope that our work, strengthened by your feedback, will make some inspiration and meaningful contribution to the community.

Best regards,
The Authors

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

First and foremost, we sincerely thank all the reviewers and the AC for the patient reading and the insightful and constructive comments! The revised version adds ablations of group nonlinearity and curvature coupling (3JpY, WBHv), supervised classification task (mTBm), computational cost analysis (WBHv, mTBm), large scale numerical validity checks (WBHv), and a comparison with pseudo-Riemannian methods (mTBm) and geometry interaction methods (WBHv).

We believe the significance of this work lies in the following points:

  1. Breaking a fundamental limitation. In product manifolds, positive and negative curvature lie in separate factors without coupling [1], which makes it hard to represent data where hierarchical, cyclic and flat patterns share common features. SL(n)\mathbb{SL}(n) realizes intrinsically coupled mixed curvature within a single space. Reviewer WBHv noted that the paper identifies a real, provable limitation of product manifolds and overcomes it.

  2. Simple construction, rich structure. A single constraint det⁡A=1\det A=1 yields both asymptotically maximal curvature coupling and unbounded order depth. The reviewers found the theory rigorous and novel (3JpY, mTBm), and noted that the order-depth experiment isolates the role of the algebraic structure (WBHv).

  3. Strong ablation evidence. Strictly matched ablations (3JpY, WBHv) isolate and verify the contribution of curvature coupling itself, which Reviewer 3JpY found useful and informative.

  4. Substantial and consistent gains. On real-world graphs with complex mixed-curvature structures, ranging from 10310^3 to 10710^7 edges, SL(n)\mathbb{SL}(n) improves over the strongest baselines by more than 40%, which the reviewers found substantial and consistent (3JpY, WBHv).

  5. No combinatorial search. Product manifolds require the curvature combination and dimensions to be fixed in advance, with more than 101010^{10} candidates already at dimension 64, while SL(n)\mathbb{SL}(n) only needs a few hyperparameters to be tuned.

  6. Broad potential. SL(n)\mathbb{SL}(n) can be directly plugged into existing models as a shallow embedding, and also provides a geometric foundation for deep neural architectures. Our experiments cover unsupervised, self-supervised, and supervised tasks (mTBm).

We thank all reviewers and the AC again for their valuable suggestions! We believe that the general SL(n)\mathbb{SL}(n) representation space opens a promising new direction for mixed-curvature representation learning, with potential applications to graphs with complex coupled structures, knowledge graphs, sequential and multimodal data, and scientific domains such as biological networks. We hope that this work, strengthened by the reviewers' feedback, can bring some useful insights and contributions to the community.

Best regards,
The Authors


[1] Sáez de Ocáriz Borde (2026). On the Expressive Power of Mixed-Curvature Representations in Product Manifolds. Workshop on Geometry-grounded Representation Learning and Generative Modeling at ICLR 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.