acceptodds
Under review as a conference paper at ICLR 2027

SPD Ricci Flow: Graph Representation Learning through Metric Evolution

13 views

Abstract

We introduce **SPD Ricci Flow**, a graph learning framework in which each node's symmetric positive definite (SPD) representation also serves as its local metric and evolves through Ricci-type dynamics. Unlike previous edge-based graph Ricci flows that evolve scalar edge lengths or weights, our framework directly evolves node-wise metric tensors using a matrix-valued Ricci-type curvature operator constructed from AIRM discrepancies between neighboring SPD metrics. This makes curvature-driven metric evolution itself the mechanism for updating node representations. The curvature also enables interpretable analysis of local geometric variation in node representations. We prove continuous-time energy dissipation and show that, under graph-Laplacian consistency, the linearized continuum limit matches the heat-type principal part of Ricci–DeTurck evolution. Experiments on node classification, molecular property prediction, and financial spillover prediction support metric evolution as a mechanism for graph representation learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural networks, Riemannian geometry, Symmetric positive definite matrices, Ricci flow, Curvature

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_CjhZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SPD Ricci Flow, a graph representation learning framework in which each node’s SPD representation also serves as its local feature-space metric. The method constructs a matrix-valued Ricci-type curvature operator by aggregating discrepancies between neighboring SPD metrics under the AIRM, and uses the resulting metric evolution to update node representations. Experiments on node classification, molecular property prediction, and financial spillover prediction report competitive performance, with ablation studies examining the contributions of graph propagation, adaptive weighting, and the curvature update. Results are competitive across node classification, molecular property prediction, and financial spillover prediction, with ablations testing graph propagation, adaptive weighting, and curvature updates.

Critical strengths
  1. Theorem 3.2 establishes that the Dirichlet-type energy is nonincreasing along the continuous-time SPD flow and characterizes stationary points through zero Ricci-type curvature. This gives a precise account of the idealized metric evolution and is the theoretical contribution.
  2. The experimental evaluation covers node classification, molecular property prediction, and financial spillover prediction. In the molecular experiments, retaining the SPDSheaf architecture while replacing its propagation layers helps isolate the update rule. Table 4 reports the highest mean ROC-AUC on four of seven MoleculeNet benchmarks and a tie on Tox21, providing empirical support beyond a single task family.
Critical weaknesses
  1. Theorem 3.2 proves energy dissipation for a continuous-time flow, but the implemented layer in Eq. (15) adds a finite Euler step, a learnable BiMap and its inverse, and a spectral activation. Positive definiteness of the output alone does not establish energy dissipation. The paper should distinguish properties proved for the idealized flow from those claimed for the trained layer. If monotonicity is intended to explain the latter, sufficient conditions or measured energy trajectories across layers are needed.
  2. The ablations in Tables 2 and 6 compare the proposed curvature update with an alternative update only under uniform weighting. They do not include a comparison that retains adaptive weights while changing only the update rule. In Table 6, the alternative update achieves higher mean scores than the proposed curvature update on Wisconsin (73.60 versus 69.40) and BACE (88.89 versus 87.71). The existing ablations do not sufficiently establish the additional benefit of the proposed curvature update when combined with adaptive weighting in the full model.
Questions for authors
  1. Does the energy-dissipation property established in Theorem 3.2 extend to the propagation layer in Eq. (15)? If so, please specify the required conditions, particularly those concerning the step size and spectral activation.
  2. Can the authors compare the proposed curvature update with the alternative update used in the ablation study while retaining the same adaptive weighting rule and keeping the remaining model and training settings consistent? Please report the same task metrics and variability for both versions.
  3. Section 4.3 requires the learnable BiMap matrix to be invertible, and Eq. (15) explicitly uses its inverse. How is invertibility maintained throughout training? Please specify the matrix parameterization or constraints used in the experiments, and explain how numerical instability is handled if the matrix becomes nearly singular.
3 comments

W1,Q1:

We thank reviewer CjhZ for highlighting the need to link continuous-time energy dissipation to the finite-step propagation layer. We have added necessary and sufficient dissipation conditions and relate them to our experimental implementation (Appendix A.11). Theorem 3.2's dissipation guarantee extends to the layer under step-size conditions. We give two conditions: the activation condition is necessary and sufficient for safe step sizes to exist, while the step-size condition is sufficient for finite-step dissipation.

Consider our model energy Eβ{\cal E}_β for SPD matrices without assuming commutativity. Fix an undirected graph with symmetric nonnegative weights and at least one positive-weight edge. Assume the layer uses an invertible BiMap shared across all nodes, its tied inverse, and a continuous positive-valued spectral activation ϕϕ.

1. Activation condition. Assume that, for all x,y∈Rx,y∈ℝ, ∣log⁡ϕ(ex)−log⁡ϕ(ey)∣≤∣x−y∣.|\logϕ(e^x)-\logϕ(e^y)|≤|x-y|. Then for every input SPD field GG and every shared invertible BiMap WW, there exists η0(G,W)>0η_0(G,W)>0 such that, for all 0<η<η0(G,W)0 < η < η_0(G,W), the finite Euler step preserves positive definiteness and the complete layer satisfies Eβ(G+)≤Eβ(G){\cal E}_β(G^+)≤{\cal E}_β(G).

The activation condition requires that the spectral activation not amplify geometric differences between node representations. More precisely, the scalar condition above is equivalent to dA(σϕ(P),σϕ(Q))≤dA(P,Q)d_A(σ_ϕ(P),σ_ϕ(Q))≤d_A(P,Q) for all SPD matrices P,QP,Q, where dAd_A denotes the AIRM distance. The eigenvalue-clipping activation used in our implementation of Eq. (15), ϕ(λ)=max⁡(λ,δ)ϕ(λ)=\max(λ,δ) with δ>0δ>0, satisfies this condition because log⁡ϕ(ex)=max⁡(x,log⁡δ)\logϕ(e^x)=\max(x,\logδ) does not expand distances on the logarithmic scale. Thus, this activation not only ensures positive-definite outputs but also does not increase the geometric energy used in our model. The condition is not restricted to this activation: another admissible example is scaled clipping, ϕ(λ)=max⁡(cλ,δ)ϕ(λ)=\max(cλ,δ) with c>0c>0.


Next, we control the energy change in the Euler update itself: an initially energy-decreasing direction need not decrease energy for an arbitrarily large step. Under the same graph, BiMap, and activation assumptions, let Zi=W⊤GiWZ_i=W^⊤G_iW be the state after the BiMap and Vi=−RiciSPD(Z)V_i=-{\rm Ric}_i^{\rm SPD}(Z) the update direction computed at ZZ. Holding this direction fixed, define F(s)=Eβ(Z+sV)F(s)={\cal E}_β(Z+sV) as the energy along the Euler update path, where ss is a trial step size. Thus, F(0)F(0) is the energy before the Euler update and F(η)F(η) is the energy after the proposal Zi′=Zi+ηViZ_i'=Z_i+ηV_i. We write D=−F′(0){\cal D}=-F'(0) for the initial rate of energy decrease; Theorem 3.2 guarantees D≥0{\cal D}≥0.

2. Step-size condition. Suppose a step size η>0η>0 and a constant M≥0M≥0, bounding the second derivative of this energy profile, satisfy Zi+ηVi≻0,∀i;F′′(s)≤M≤2Dη,∀s∈[0,η].Z_i+ηV_i≻0,∀i;F''(s)≤M≤\frac{2{\cal D}}η,∀s∈[0,η]. Then the output G+G^+ of the complete layer satisfies Eβ(G+)≤Eβ(G){\cal E}_β(G^+)≤{\cal E}_β(G).


Under these condition, the layer does not increase energy:

(i) Energy nonincrease in the Euler update. Under Condition 2, the Euler proposals Zi′=Zi+ηViZ_i'=Z_i+ηV_i remain positive definite, and the finite-step second-order error does not exceed the first-order energy decrease, giving Eβ(Z′)≤Eβ(Z){\cal E}_β(Z')≤{\cal E}_β(Z).

(ii) Energy nonincrease during activation. Condition 1 ensures AIRM nonexpansiveness of the spectral map, even for noncommuting SPD matrices, hence Eβ(σϕ(Z′))≤Eβ(Z′){\cal E}_β(σ_ϕ(Z'))≤{\cal E}_β(Z'). Thus, Conditions 2 and 1 control the energy changes from ZZ to Z′Z' and from Z′Z' to σϕ(Z′)σ_ϕ(Z'), respectively. The shared BiMap and its tied inverse preserve AIRM distances and hence energy, so Eβ(G+)=Eβ(σϕ(Z′))≤Eβ(Z′)≤Eβ(Z)=Eβ(G).{\cal E}_β(G^+)={\cal E}_β(σ_ϕ(Z'))≤{\cal E}_β(Z')≤{\cal E}_β(Z)={\cal E}_β(G). (iii) Existence of safe small steps. Condition 1 is necessary and sufficient for every input SPD field GG and every shared invertible BiMap WW to admit sufficiently small safe steps.


Our model's shared invertible BiMap and tied inverse meet the theorem's requirements. The BiMap need not be orthogonal or close to the identity. We take the spherical experiment case (Fig.2) as an easy example to show an direct substitution of its parameters into the bound: gi=e2uiI2,β=0.5,η=0.01,ϕ(λ)=max⁡(λ,10−9).g_i=e^{2u_i}I_2,β=0.5,η=0.01,ϕ(λ)=\max(λ,10^{-9}). This activation satisfies the condition above. For this conformal setting with maximum weighted degree d_\max^w, we can prove that every update preserves positive definiteness and does not increase energy when η≤\frac1{d_\max^w(1+1/\sqrt{2eβd})}. When the actual mesh has maximum degree 66, setting d=2d=2 and β=0.5β=0.5 gives η≤0.11664.η≤0.11664. The experimental step size η=0.01η=0.01 lies within this guaranteed range. Thus, these conditions directly relate to the activations and step sizes used in our experiments. This explicit bound applies to the conformal spherical experiment; the preceding necessary-and-sufficient criterion and finite-step conditions cover general noncommuting SPD learning layers.

W2, Q2:

We thank the reviewer for identifying this missing comparison. We agree that the uniform-weight ablations alone did not establish the additional benefit of the proposed curvature update when combined with adaptive weighting.

We therefore compare the two update rules while retaining the adaptive weighting rule aij=wijexp⁡(−βrij)a_{ij}=w_{ij}\exp(-\beta r_{ij}), where rijr_{ij} is the squared AIRM distance. The encoder, decoder, BiMap parameterization, spectral activation, propagation depth, seeds, and training protocol are kept unchanged. Only the update rule is replaced:

  • Adaptive + SiS_i: Zi+=Zi+ηSiZ_i^+=Z_i+\eta S_i, corresponding to the alternative update used in the original ablation.
  • Adaptive + Ricci: Zi+=Zi+η2(ZiSi+SiZi)Z_i^+=Z_i+\frac{\eta}{2}(Z_iS_i+S_iZ_i), corresponding to the full model.

Both variants compute SiS_i using Eq. (5), with the same adaptive weighting rule evaluated on their respective current SPD states.

Table A. Update-rule comparison under adaptive weighting. Scores are multiplied by 100 and reported as mean ± standard deviation over seeds. The final column reports the difference in means (Ricci minus SiS_i).

Dataset / metricAdaptive + SiS_iAdaptive + RicciDifference
Wisconsin: Accuracy88.40 ± 4.4089.80 ± 3.52+1.40
DBLP: Accuracy85.88 ± 0.4385.90 ± 0.39+0.02
BACE: ROC-AUC87.77 ± 1.7089.70 ± 1.00+1.93
SIDER: ROC-AUC84.85 ± 0.4085.00 ± 0.30+0.15
Financial: Macro-AUPRC33.09 ± 0.2435.24 ± 1.01+2.15
Financial: NDCG@529.77 ± 0.2432.67 ± 1.28+2.90
Financial: Precision@512.46 ± 0.1113.57 ± 0.40+1.11
Financial: Hit@562.15 ± 0.3566.11 ± 1.66+3.96

With adaptive weighting retained, the Ricci update achieves higher mean scores in all eight comparisons. In particular, the ordering highlighted by the reviewer under uniform weighting on Wisconsin and BACE is reversed: the Ricci update now improves their mean scores by 1.40 and 1.93 points, respectively. On the financial task, it improves all four metrics, including Macro-AUPRC by 2.15 points and NDCG@5 by 2.90 points. The differences on DBLP and SIDER are small and should be interpreted cautiously.

These matched comparisons provide evidence that the metric-coupled update contributes beyond adaptive weighting alone, with the clearest mean gains on BACE and the financial task. The contrasting results under uniform weighting also show that its benefit depends on the weighting scheme and dataset, rather than implying uniform superiority of the Ricci update in every setting. We will incorporate these comparisons into the revised Tables 2 and 6 and clarify the corresponding conclusions.

Q3:

We thank the reviewer for requesting this clarification. The invertibility required by Eq. (15) is enforced at every forward pass, rather than assumed from identity initialization. Specifically, each layer learns a matrix Aℓ∈Rd×dA_\ell\in\mathbb{R}^{d\times d}, initialized to zero, and constructs Wℓraw=I+0.08tanh⁡(Aℓ)W_\ell^{\mathrm{raw}}=I+0.08\tanh(A_\ell), with elementwise tanh⁡\tanh. Before applying the BiMaps, we check its smallest singular value. If σmin⁡(Wℓraw)≥0.1\sigma_{\min}(W_\ell^{\mathrm{raw}})\geq0.1, the matrix is used directly. Otherwise, for Wℓraw=Udiag⁡(σi)V⊤W_\ell^{\mathrm{raw}}=U\operatorname{diag}(\sigma_i)V^\top, we replace it with Wℓ=Udiag⁡(max⁡(σi,0.1))V⊤W_\ell=U\operatorname{diag}(\max(\sigma_i,0.1))V^\top. Both the forward BiMap and the inverse in Eq. (15) use this same effective matrix.

This explicitly handles the near-singular case raised by the reviewer: even if the raw matrix approaches singularity during optimization, its unchecked inverse is never used. The effective matrix satisfies σmin⁡(Wℓ)≥0.1\sigma_{\min}(W_\ell)\geq0.1 and ∥Wℓ−1∥2≤10\lVert W_\ell^{-1}\rVert_2\leq10. Furthermore, ∥tanh⁡(Aℓ)∥2≤∥tanh⁡(Aℓ)∥F≤d\lVert\tanh(A_\ell)\rVert_2\leq\lVert\tanh(A_\ell)\rVert_F\leq d bounds the raw matrix norm by 1+0.08d1+0.08d, and singular-value flooring preserves this upper bound. Hence κ2(Wℓ)≤10(1+0.08d)\kappa_2(W_\ell)\leq10(1+0.08d), which equals 22.822.8 for d=16d=16. These constraints control the amplification associated with the inverse BiMap throughout training, not only at initialization. We have clarified the parameterization, the correction procedure, and these conditioning bounds in Appendix A.4 of the revised manuscript.

I appreciate the authors’ response and clarifications. Having considered the rebuttal, I maintain my original assessment and score.

Reviewer_9HdXOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose a graph learning method based on 2 mechanisms

  1. Embedding the input in spd matrices
  2. Updating node values by a specialized closed form operations that keep spd

In my knowledge this approach is new. Presenting the data as SPD matrices on a graph is an interesting idea. Though I believe that the proposed flow has many limitations.

Critical strengths

I have one very important reservation for the paper.

My main problem is that the flow is basically s smoothing over some metric space. As such it is good for problems where smoothing is required. In some problems this is indeed the case. But for many problems smoothing is actually bad. Take for example pattern formation on a graph (say predicting butterfly wings) then a completely different approach is needed. Furthermore. This architecture is sensitive to oversmoothing.

Critical weaknesses

See above

Questions for authors

I need a realistic discussion about smoothing and the limitations. Also I want to see a dataset with pattern formation being worked on

2 comments

We thank reviewer 9HdX for raising this important point.

Part 1: Smoothing and Over-Smoothing Analysis

We agree that smoothing is an inductive bias shared with diffusion-based GNNs (Li et al., 2018). Our goal is not to eliminate smoothing but to introduce curvature-driven evolution of node-wise metrics into graph learning. As shown below, smoothing remains moderate and far from collapse at practical depths.

Combining our geometric update with over-smoothing-resistant architectures, such as sheaf neural networks (Bodnar et al., 2022), is a promising direction for future work. Here, we use a simple backbone reflecting our geometric motivation.

Theoretically, Corollary A.3 characterizes the diffusion equilibrium: for ψ(r)=r/2\psi(r)=r/2 with fixed neighboring metrics, a node stops evolving at their weighted AIRM barycenter. A constant metric field satisfies this condition, but does not constrain learned representations, since our model uses only finitely many propagation steps between a learned encoder and decoder.

We evaluate three complementary diagnostics at the last propagation layer LL:

  • Dirichlet energy (EL/E0E_L/E_0): Eℓ=∑{i,j}∈Ewijψβ(rij(ℓ))E_\ell=\sum_{\{i,j\}\in\mathcal E}w_{ij}\psi_\beta(r_{ij}^{(\ell)}) (Eq. 3), where rij(ℓ)r_{ij}^{(\ell)} is the squared AIRM distance between neighbors, normalized by its pre-propagation value.

  • Dispersion (DL/D0D_L/D_0): Dℓ=1N∑idAIRM2(gi(ℓ),Mℓ)D_\ell=\frac1N\sum_i d_{\mathrm{AIRM}}^2(g_i^{(\ell)},M_\ell), where Mℓ=exp⁡(1N∑ilog⁡gi(ℓ))M_\ell=\exp(\frac1N\sum_i\log g_i^{(\ell)}) is the log-Euclidean mean, normalized by its pre-propagation value.

  • Normalized effective rank: Effective rank of centered log-metric representations divided by min⁡(N,p)\min(N,p), where p=d(d+1)/2p=d(d+1)/2, measuring spectral dimensionality.

These respectively characterize local smoothing, global variation, and spectral dimensionality.

(i) Trained models (Table C). Across molecular and financial tasks, energy and dispersion retain 41–85% and 54–93% of their input values, while normalized rank remains 0.72–0.78. Even MUV (seven layers) retains 41% energy and 54% dispersion, indicating substantial metric-field variation.

(ii) Depth trade-off during training (Table A). On BACE, we train with L∈{2,4,8,16}L\in\{2,4,8,16\}. Energy decreases from 69% to 14% and dispersion from 91% to 36%, while normalized rank remains 0.77–0.79. ROC-AUC ranges from 87.3 to 90.9 for L≤8L\leq8, then drops to 86.6 at L=16L=16 (14% energy). This suggests stronger smoothing may eventually harm prediction, whereas practical depths (L=2L=2–77) preserve substantial metric variation.

(iii) Long-horizon behavior (Table B). Without training, we apply up to 500 propagation layers. Energy decreases monotonically over the evaluated depths, consistent with the continuous-time dissipation property in Thm. 3.2. At 500 layers, energy retains 45% (Cora-Full) and 66% (DBLP), dispersion 82% and 98%, and normalized rank remains nearly unchanged. For L≤8L\leq8, both datasets retain at least 98% energy in this untrained sweep.

Together, these results indicate moderate, finite-time smoothing without observed collapse at practical depths. Stronger smoothing does not necessarily improve prediction, supporting our shallow metric evolution.

Table A. Depth trade-off on BACE (trained models).

LLROC-AUC (%)E/E0E/E_0D/D0D/D_0Normalized rank
289.700.6910.9060.769
487.270.5310.8400.766
890.870.2970.5850.793
1686.590.1360.3600.778

Table B. Depth sweep without training.

LLCora-Full E/E0E/E_0Cora-Full D/D0D/D_0Cora-Full Norm. rankDBLP E/E0E/E_0DBLP D/D0D/D_0DBLP Norm. rank
20.9960.9990.9780.9971.0000.974
40.9920.9980.9780.9941.0000.974
80.9830.9960.9780.9890.9990.974
160.9670.9920.9780.9780.9990.974
320.9350.9840.9780.9580.9980.974
640.8760.9690.9780.9220.9960.97419
5000.4470.8240.9730.6640.9790.97389

Table C. Over-smoothing diagnostics at the last propagation layer of trained models (mean over seeds).

DatasetLLE/E0E/E_0D/D0D/D_0Normalized rank
BACE20.6910.9060.769
BBBP20.8490.9250.748
SIDER20.6520.8570.774
Tox2130.6320.7280.776
ClinTox50.5180.6550.778
HIV50.5570.6190.766
MUV70.4100.5410.783
Financial20.5830.8140.719

References

[1] Li, Q., Han, Z., and Wu, X.-M. (2018). Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning (AAAI).

[2] Bodnar, C., Di Giovanni, F., Chamberlain, B., Liò, P., and Bronstein, M. (2022). Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs (NeurIPS 2022).

Part 2: Pattern Formation

Following the reviewer's suggestion, we evaluate our model on a graph pattern-formation task inspired by Nakao and Mikhailov (2010). Unlike tasks that benefit from smoothing, this task requires predicting the emergence of heterogeneous node states from nearly uniform initial conditions, that is, the neighbors gradually become different. Small differences between neighboring nodes are amplified over time rather than smoothed away, making it a challenging setting for models with a smoothing inductive bias.

We generate 100 Barabási–Albert graphs (256 nodes and 2,505 edges each) and simulate a two-species reaction–diffusion system on each graph:

duidτ=(35+16ui−ui29−vi)ui+0.12∑j∼i(uj−ui)\frac{du_i}{d\tau}=\left(\frac{35+16u_i-u_i^2}{9}-v_i\right)u_i+0.12\sum_{j\sim i}(u_j-u_i)

dvidτ=(ui−1−0.4vi)vi+1.872∑j∼i(vj−vi)\frac{dv_i}{d\tau}=(u_i-1-0.4v_i)v_i+1.872\sum_{j\sim i}(v_j-v_i)

We initialize the system from a nearly uniform state (ui=5u_i=5, vi=10v_i=10), with independent perturbations in [−0.01,0.01][-0.01,0.01]. We use four initial perturbations per graph, yielding 400 trajectories. The diffusion coefficient of vv is 15.6 times that of uu. Through the interaction of local reactions and unequal diffusion rates, initially small perturbations are amplified, leading to increasingly heterogeneous node states. Thus, unlike conventional smoothing tasks, successful prediction requires modeling growing differences between neighboring nodes rather than suppressing them.

Given the states at τ=0,30,…,270\tau=0,30,\ldots,270 (20 input features per node), the task is to predict (ui,vi)(u_i,v_i) at τ=300\tau=300 for every node. Graphs are split into 70/15/15 for training, validation, and testing, ensuring that test graphs are unseen during training. We report test normalized MSE (NMSE) over five seeds.

We compare against two graph models with contrasting inductive biases. GCN repeatedly aggregates neighboring features and can suffer from over-smoothing (Kipf and Welling, 2017; Li et al., 2018). The sheaf neural network (Bodnar et al., 2022) learns how neighboring features should be compared and is designed to accommodate heterophily and mitigate over-smoothing. We also report two parameter-free references: copying the last observed frame and linear extrapolation from the last two frames.

For our model, the 20-dimensional input is mapped to an 8×88\times8 SPD matrix through a bounded matrix-exponential encoding for numerical stability, followed by two propagation layers and a linear decoder. All models are evaluated without additional fine-tuning.

Table D. Pattern formation on reaction–diffusion dynamics. Test NMSE (mean ± standard deviation over five seeds; lower is better).

MethodParamsTest NMSE ↓
Copy last frame00.2985
Linear extrapolation00.4205
GCN9,7940.3350 ± 0.0023
sheaf NN7,2100.1041 ± 0.0061
Ours3,9700.1316 ± 0.0014

As shown in Table D, GCN performs worse than simply copying the last frame, consistent with the limitations of plain smoothing on this task. Our model reduces the error relative to GCN by a factor of approximately 2.5 and achieves an error less than half that of the copy baseline. It also approaches the performance of the sheaf network while using approximately 45% fewer parameters.

Although the sheaf network achieves the best performance, our model substantially outperforms GCN and the parameter-free baselines. This suggests that a smoothing inductive bias does not necessarily prevent SPD Ricci Flow from learning pattern-formation dynamics. At the same time, the remaining gap to the sheaf network highlights a meaningful limitation and motivates incorporating heterophily-aware propagation in future work.

Summary. Smoothing is an inductive bias of our model, just as the generally diffusion-based GNNs. However, at the depths used in practice, our model retains substantial metric-field variation, and the pattern-formation experiment demonstrates its ability to model tasks in which neighboring states diverge. In the revised manuscript, we will add a limitations discussion addressing this smoothing bias (line 480 section 6), incorporate the over-smoothing diagnostics and pattern-formation experiment in the appendix (Appendix A.9 A.10), and outline extensions using sheaf-based propagation as future work.

References:

[1] Nakao and Mikhailov (2010). Turing Patterns in Network-Organized Activator-Inhibitor Systems. Nature Physics.

[2] Kipf and Welling (2017). Semi-Supervised Classification with Graph Convolutional Networks. ICLR.

[3] Li et al. (2018). Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning. AAAI.

[4] Bodnar et al. (2022). Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs. NeurIPS.

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

The paper proposes to use each node's symmetric positive definite matrix both as its representation and as a local metric on the feature space, so that updating the representation is the same operation as evolving a metric field. Neighbouring metrics are compared with affine-invariant logarithms, these discrepancies are aggregated with distance-decayed weights into a node-wise tensor, and a congruence lift turns that tensor into a symmetric matrix that plays the role of a Ricci-type curvature; the metric then evolves along the negative of that curvature, discretized as an Euler step and wrapped in a learnable bilinear map with an SPD-preserving spectral activation to form a propagation layer. The paper proves that the free flow dissipates a Dirichlet-type energy, and that near a constant background the linearized flow is graph heat diffusion of the log-metric field, matching the principal part of Ricci–DeTurck evolution under a Laplacian-consistency assumption. Experiments cover node classification on five benchmarks, molecular property prediction on seven MoleculeNet datasets using a modified SPDSheaf backbone, and a financial spillover task, together with an analysis of curvature responses around news events and component ablations.

Critical strengths
  1. The architectural idea is clean and, as far as I can tell, new in this exact form: rather than attaching a scalar curvature to edges, the method carries a matrix per node and lets that matrix be the object that both defines local geometry and gets updated. It is also competitive exactly where it needs to be, and against baselines built for SPD representations rather than generic message passing: the full model is best on four of the five node-classification benchmarks and second on the fifth, and on the seven MoleculeNet datasets it is highest on four, within 0.1 of the backbone on two more, and behind on one. The molecular experiments change only the propagation rule inside a fixed backbone, which isolates the thing the paper claims to contribute, and the code is released, so the construction is reimplementable rather than merely described.

  2. The paper is clear about what its theory does and does not establish: it states repeatedly that the flow is consistent at the level of principal parts rather than recovering the full nonlinear Ricci tensor, gives the commuting case as an exact lemma, shows that noncommutativity enters at higher order, and raises in the appendix the one objection that matters most, that the global minimizer of the energy is a constant metric field, together with an argument that the trained network operates in a forced, finite-time regime. The component ablation is reported in the same spirit: it is broken down finely enough to show that a simpler component explains most of the gain. Publishing an ablation that locates the mechanism against the authors' own naming is more informative than reporting the full model against its own parts, and it is part of why I recommend acceptance despite the framing problem in weakness one.

Critical weaknesses
  1. The Ricci-type framing carries less of the contribution than the vocabulary suggests, because both results that support it follow from the definition rather than constraining the model. The curvature is defined as the congruence-symmetrized aggregation of neighbouring log-metric displacements, equals the AIRM gradient of the Dirichlet energy at the identity, and the flow is set to that same gradient, so the dissipation theorem says only that a gradient flow decreases its own energy, the proof being the chain rule plus the coefficient the congruence lift produces. The comparison with Ricci–DeTurck holds at the level of principal symbols, which every graph-Laplacian diffusion satisfies, and the Bochner–Weitzenböck route is closed off by taking the rough connection Laplacian to vanish, the term the identity would need to express curvature as a correction. Why this is consequential: it decides what the paper is credited for. Read as Ricci-type dynamics the content is thin; read as an architectural claim — a metric per node, log-metrics diffused under affine-invariant distance weights — the contribution is real, and the fix is vocabulary, not method.

  2. The ablations attribute most of the gain to the edge weighting rather than the evolution law, and the margins are not shown to lie outside seed variation. On DBLP the four component variants span 85.86 to 85.90, so where the method reports its best result the variant choice is inert; on Wisconsin deleting the flow layer still yields 87.40 against 89.80, while replacing distance-decayed weights with uniform weights collapses to 69.40. The financial ablation shows the same asymmetry: +0.66 on the primary metric and +0.88 on NDCG@5 over uniform weights, against −4.11 for deleting the layer. Classification margins are 0.15 to 1.80 points against standard deviations of 0.35 to 3.52 with no significance test anywhere, and the seven molecular differences are +0.7, +2.4, +0.2, +0.7, 0.0, −0.1 and −1.1, so one dataset carries the result and another regresses within a standard deviation of 4.3. Why this is consequential: the deployed layer is competitive, so this does not undercut acceptance, but the causal claim that metric evolution is the mechanism is not established; a control with learned, non-AIRM edge weights would settle it, and I expect that as follow-up rather than withhold acceptance.

Questions for authors
  1. How much changes if the same equations are presented as an SPD diffusion with distance-decayed weights? I am asking about framing rather than correctness: given that the curvature is defined as the negative energy gradient and that the rough connection Laplacian is assumed to vanish, which conclusions, if any, would read differently under that presentation?

  2. In the component ablations the four DBLP variants span 0.04 points, and on Wisconsin the no-layer variant reaches 87.40 while the uniform-weight variant falls to 69.40. Can you separate the effect of adaptive distance-based weighting from the effect of the curvature update? A variant with learned, SPD-independent edge weights of comparable flexibility would make the attribution convincing.

  3. With standard deviations of up to 3.52 over ten seeds and differences of 0.15 to 1.80 points, are the reported wins statistically distinguishable from the baselines? Reporting paired tests per seed would settle this.

  4. The appendix notes that the global minimizer of the energy is a constant metric field. Can you measure the trade-off during training, for example the Dirichlet energy and a collapse or rank diagnostic per layer, to show that the learned model operates before collapse rather than because of it?

  5. The layer needs an invertible bilinear map and per-edge matrix logarithms. What are the parameter counts and wall-clock cost per epoch relative to the SPD and message-passing baselines at matched width, and how is invertibility enforced during training?

  6. In the node-classification table, GNRF and GEGCN both report 73.86 with a standard deviation of 1.18 on Questions. Is that a transcription of the same number, or two independent runs that coincide to two decimals?

6 comments

W1:

We thank the reviewer for raising this issue and agree that the role and scope of the Ricci-type framing need clarification. However, our Ricci-type framing is not merely a post hoc interpretation of SPD diffusion based on its ability to recover Ricci flow. The contribution of the underlying Ricci-type geometric mechanism should therefore be reassessed in light of the discussion below.

Starting from local metrics, we construct a geometrically motivated update, verify its correspondence with the Ricci–DeTurck heat-type principal part under explicit conditions (Theorem 3.1), and prove energy dissipation for the same nonlinear flow. Together, these results extend our contribution beyond a diffusion model aimed at recovering the intended geometric mechanism by establishing a geometric discretization framework, while grounding the Ricci-type designation in both the modeling objective and the verified dynamical connection.

Even when some modeling steps reduce to simpler formulations in special cases, this does not imply that the overall construction or its underlying geometric rationale is the same.

We believe their value for geometric discretization and interpretability should not be overlooked or viewed merely as constraining the model. However, we acknowledge the lack of clarity in our discussion and address the reviewer’s concerns point by point below, with further clarification in the revision.

First, regarding near-identity behavior and the role of the dissipation theorem, we agree that RiciSPD(g)=∇giAIRMEψ(g)\mathrm{Ric}^{\mathrm{SPD}}_i(g) = \nabla^{\mathrm{AIRM}}_{g_i}\mathcal{E}_\psi(g) at gi=Ig_i=I. However, this special-case identity does not mean that our flow is the same as the standard AIRM gradient flow at general SPD states. Details of this difference see answer of Q1. For this reason, Theorem 3.2 is not a restatement of the dissipation result for the standard AIRM gradient flow. Instead, it proves that the same AIRM-based graph energy EψE_\psi remains nonincreasing along our proposed flow, which generally differs from the AIRM gradient flow, and that Si=0S_i=0 at every node characterizes equilibrium, consistent with Ricci-flow curvature smoothing. Thus, using the chain rule in both proofs does not imply identical vector fields or dissipation identities. The value of this analysis is a rigorous dissipation guarantee for our flow, not a new proof technique.

Secondly, we agree that different diffusion models can share the linearized heat-type principal part of Ricci–DeTurck flow. We apologize for not clarifying this distinction. However, we do not aim to construct the only diffusion model recovering this mechanism, and sharing this does not diminish our contribution. A standard graph Laplacian also smooths node metrics but cannot capture our flow's nonlinearity. Our update is nonlinear in the metrics even with fixed edge weights. The full nonlinear dynamics still require separate construction, analysis, and comparison.

Regarding the rough connection Laplacian, we agree that our explanation does not directly assert its vanishing and that its motivation needs clarification. However, its vanishing is geometrically motivated and does not rule out the BW route; in our graph model, it can instead be viewed as a discrete realization of that motivation. This motivation comes from geodesic one-forms extended by parallel transport along radial geodesics. At the base point, their rough Laplacian vanishes, so the one-form BW formula gives the Ricci curvature action from their Laplacian. Viewing SPD features as locally constant metric fields, we model one tensor slot as a geodesic one-form to realize this mechanism and motivate the vanishing-rough-Laplacian condition. Fixing a vector field XX in one slot of the metric gg gives the one-form g(X,⋅)g(X,\cdot); it is geodesic if XX is parallel along radial geodesics. Under this viewpoint, we treat a node neighborhood as locally Euclidean and choose XX locally constant. With this choice implicit, we approximately treat the locally constant metric field as a geodesic one-form, motivating the vanishing rough Laplacian. We will clarify this motivation in the main text and appendix. The BW formula motivates viewing curvature through an operator action, while its geometry motivates modeling a vanishing rough Laplacian; together, these lead to the curvature definition from neighborhood discrepancies in AIRM energy. The significance of the BW-based Ricci-type dynamics remains intact. Beyond a new architectural claim for gradient flow, this provides a complete construction procedure for geometric discretization.

Q1:

We thank the reviewer for this question. Under the same equations and assumptions, describing our method as SPD diffusion with distance-decayed weights would not change its mathematical conclusions or empirical results.

However, this description captures the propagation form without fully conveying our research question. Our work examines how geometry guides node-metric updates, whether these updates realize the intended mechanism, and how the resulting nonlinear flow evolves metrics while reducing neighborhood energy. Together, these analyses establish a geometric discretization framework and provide a geometry-based interpretation of the diffusion that SPD diffusion with distance-decayed weights cannot recover.

The reviewer-cited AIRM gradient in Eq. (4) yields, under the same factor-of-two scaling,

ViAIRM=−2∇giAIRMEψ(g)=2gi1/2Sigi1/2V_i^{\mathrm{AIRM}}=-2\nabla_{g_i}^{\mathrm{AIRM}}\mathcal{E}_\psi(g)=2g_i^{1/2}S_ig_i^{1/2},

whereas our flow in Eq. (7) is

Viours=giSi+SigiV_i^{\mathrm{ours}}=g_iS_i+S_ig_i.

Their difference is:

Viours−ViAIRM=[gi1/2,[gi1/2,Si]]V_i^{\mathrm{ours}}-V_i^{\mathrm{AIRM}}=[g_i^{1/2},[g_i^{1/2},S_i]],

which vanishes when gig_i and SiS_i commute, including at gi=Ig_i=I, but not in general. Thus, equality at the identity does not imply equality of the vector fields at general noncommuting states.

Another difference is from the dissipation result interpretation. For the same current metric gg and neighborhood response SS, consider the previously defined VAIRMV^{\mathrm{AIRM}} and VoursV^{\mathrm{ours}}. Take the eigendecomposition, g=Udiag⁡(λ1,…,λd)U⊤g=U\operatorname{diag}(\lambda_1,\ldots,\lambda_d)U^\top, where U=(u1,…,ud)U=(u_1,\ldots,u_d) is orthogonal and express SS and VV in this eigenbasis as S~=U⊤SU\widetilde{S}=U^\top SU, V~=U⊤VU\widetilde{V}=U^\top VU. Then,

(V~ours)ab=λa+λb2S~ab(\widetilde{V}_{\mathrm{ours}})_{ab}=\frac{\lambda_a+\lambda_b}{2}\widetilde{S}_{ab},

(V~AIRM)ab=λaλb,S~ab(\widetilde{V}_{\mathrm{AIRM}})_{ab}=\sqrt{\lambda_a\lambda_b}\\,\widetilde{S}_{ab}.

Here, aa and bb index rows and columns. The diagonal responses coincide, while difference occurs in the off-diagonal responses use arithmetic and geometric means of the eigenvalues, respectively.

The meaning of this difference becomes clearer in 2D with λ1>λ2>0\lambda_1 > \lambda_2 > 0 and UU representing a rotation by the principal-axis angle θ\theta. For g˙=V\dot{g}=V, we have λ˙a=ua⊤Vua=V~aa\dot{\lambda}_a=u_a^\top Vu_a=\widetilde{V}_{aa} and θ˙=V~12/(λ1−λ2)\dot{\theta}=\widetilde{V}_{12}/(\lambda_1-\lambda_2). Consequently, at the same gg and SS,

λ˙aours=λ˙aAIRM=λaS~aa(a=1,2)\dot{\lambda}_a^{\mathrm{ours}}=\dot{\lambda}_a^{\mathrm{AIRM}}=\lambda_a\widetilde{S}_{aa}\quad(a=1,2),

θ˙ours=λ1+λ22λ1λ2θ˙AIRM\dot{\theta}_{\mathrm{ours}}=\frac{\lambda_1+\lambda_2}{2\sqrt{\lambda_1\lambda_2}}\dot{\theta}_{\mathrm{AIRM}}.

Thus, both updates give the same instantaneous eigenvalue rates λ˙a\dot{\lambda}_a, but ours gives a larger angular-speed magnitude ∣θ˙∣|\dot{\theta}|. This geometric property: a larger ∣θ˙∣|\dot{\theta}| with unchanged λ˙a\dot{\lambda}_a cannot generally be reproduced by adjusting the step size of the standard AIRM update.

The third difference lies in the objective of our theoretical analysis. Beyond a working SPD propagation rule, we seek to understand its dynamical link to Hamilton's Ricci flow. If SPD propagation with distance-decayed weights were the only goal, demonstrating effectiveness and analyzing stability or energy dissipation would suffice, and there would be no need to compare it with classical Ricci flow. It is precisely the Ricci perspective that guides our questions and terminology in geometric discretization, while Theorem 3.1 adds theoretical content to the Ricci-type framing beyond propagation.

For example, without this geometric motivation, the action gi(u,Siv)=uiTSivg_i(u,S_iv)=u^T_iS_iv need not be symmetrized; the sequence “insert a parallel vector field into gg to obtain a one-form → model via the one-form BW formula → leave the locally constant vector field implicit” motivates an additional symmetrization. This also explains our choice of (6) over (4).

Our point is therefore that describing the equations as SPD diffusion is appropriate at the level of their propagation form, but this description alone neither motivates nor fully captures the paper’s research question and theoretical contribution.

W2, Q2:

We thank reviewer MJak for highlighting the need to disentangle the effects of adaptive edge weighting and curvature-driven metric evolution. We agree that the original ablations did not fully isolate these two components.

To address this concern, following the reviewer's suggestion, we complete the ablation as a full factorial design that crosses three weightings, namely uniform (β=0\beta=0), learned SPD-independent, and adaptive (ours), with two update rules, namely SiS_i (no metric coupling, Zi+ηSiZ_i+\eta S_i) and Ricci (ours), keeping all other components fixed (Table A). The learned weighting aij=wijexp⁡(−β∥ui−uj∥2)a_{ij}=w_{ij}\exp(-\beta\lVert u_i-u_j\rVert^2), with ui=MLP(xi)u_i=\mathrm{MLP}(x_i), does not depend on the SPD states.

The resulting comparisons allow us to examine the effects of the weighting and update rules separately. The full model achieves the highest mean performance on all evaluated datasets and financial metrics except Hit@5, where the learned-weighting variant is marginally higher (66.15 vs. 66.11).

Table A. Separating the weighting rule from the update rule. SiS_i: Zi+ηSiZ_i+\eta S_i; Ricci: Zi+η2(ZiSi+SiZi)Z_i+\frac{\eta}{2}(Z_iS_i+S_iZ_i). The learned weighting is independent of the SPD states. Scores are multiplied by 100 and reported as mean ± standard deviation.

WeightingWisconsin SiS_iWisconsin RicciDBLP SiS_iDBLP RicciBACE SiS_iBACE RicciSIDER SiS_iSIDER Ricci
Uniform (β=0\beta=0)73.60 ± 4.2769.40 ± 4.1085.89 ± 0.4485.86 ± 0.4488.89 ± 1.1087.71 ± 0.4783.99 ± 0.3984.20 ± 0.34
Learned, SPD-independent87.20 ± 1.6988.20 ± 3.1985.88 ± 0.4385.89 ± 0.4587.62 ± 0.9288.01 ± 1.3084.95 ± 0.4584.57 ± 0.38
Adaptive (ours)88.40 ± 4.4089.80 ± 3.5285.88 ± 0.4385.90 ± 0.3987.77 ± 1.7089.70 ± 1.0084.85 ± 0.4085.00 ± 0.30
No propagation layer87.40 ± 2.5485.88 ± 0.4687.96 ± 1.1384.03 ± 0.19

Financial spillover:

WeightingMacro-AUPRC SiS_iMacro-AUPRC RicciNDCG@5 SiS_iNDCG@5 RicciPrecision@5 SiS_iPrecision@5 RicciHit@5 SiS_iHit@5 Ricci
Uniform (β=0\beta=0)33.10 ± 0.2134.58 ± 0.4329.91 ± 0.3831.79 ± 0.5812.66 ± 0.2113.32 ± 0.2163.12 ± 1.0865.75 ± 1.20
Learned, SPD-independent33.17 ± 0.3334.93 ± 0.4829.99 ± 0.3932.29 ± 0.6312.69 ± 0.0913.49 ± 0.1963.10 ± 0.6166.15 ± 1.09
Adaptive (ours)33.09 ± 0.2435.24 ± 1.0129.77 ± 0.2432.67 ± 1.2812.46 ± 0.1113.57 ± 0.4062.15 ± 0.3566.11 ± 1.66
No propagation layer31.13 ± 0.4527.56 ± 0.3911.71 ± 0.1858.90 ± 1.31

The two components contribute in different regimes:

(i) On the financial task, the curvature update is the dominant factor. Under every weighting, Ricci outperforms SiS_i by 1.5–2.2 points in Macro-AUPRC and 1.9–2.9 points in NDCG@5, whereas changing the weighting has little effect for a fixed update.

(ii) On the Wisconsin graph, the weighting is the dominant factor. Note that the adaptive weighting is derived from the Dirichlet energy through aij=2wijψβ′(rij)a_{ij}=2w_{ij}\psi_\beta'(r_{ij}) (Eqs. 3–5), rather than introduced as an external gate. With uniform weights (β=0\beta=0), the energy is quadratic in the AIRM distance, and highly dissimilar neighbors can exert disproportionately large influences on the update. On this heterophilic graph, both updates fall below the variant without propagation (73.60 for SiS_i and 69.40 for Ricci vs. 87.40). With β>0\beta>0, the weighting e−βrije^{-\beta r_{ij}} attenuates contributions from distant neighbors, and both non-uniform weighting schemes alleviate this degradation (87.20–89.80). With non-uniform weights, Ricci is higher than SiS_i (+1.0 with learned weights and +1.4 with adaptive weights), although these differences lie within the variance of this small dataset. On DBLP, all variants, including the one without propagation (85.88), lie within 0.04 points; as the reviewer noted, this benchmark is insensitive to propagation and cannot support attribution.

(iii) On BACE, the gain appears only when the two components are combined. Adaptive + Ricci reaches 89.70, whereas adaptive + SiS_i (87.77) and learned + Ricci (88.01) remain close to the variant without propagation (87.96). In particular, the learned SPD-independent weighting does not reproduce this gain despite its additional parameters, which suggests that the AIRM-distance weighting contributes beyond generic learned edge weights.

In summary, the ablations suggest that the gains are not solely attributable to adaptive edge weighting: the Ricci update provides benefits under fixed weighting schemes, while the learned SPD-independent control does not reproduce the full model's performance on BACE. These results support the complementary roles of weighting and curvature-driven metric evolution. We will update Tables 2 and 6 accordingly.

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

The paper proposes SPD Ricci Flow, a graph learning framework that evolves node representations through a matrix valued Ricci type operator. The main ingredients are representing each node by an SPD matrix, measuring discrepancies between neighboring SPD matrices using AIRM (Eq 2), and aggregating these discrepancies with adaptive weights (Eq 5). The authors define a matrix valued Ricci type curvature operator (Eq 6), which drives the evolution of node representations (Eq 7). This evolution is then incorporated into a learnable neural network layer (Eq 15).

Mathematically, the mechanism can be understood as nonlinear, distance weighted diffusion of SPD valued node representations. The paper establishes two theoretical results. Theorem 3.1 shows that the linearized continuum limit agrees with the heat type principal part of Ricci DeTurck flow, while Theorem 3.2 establishes continuous time energy dissipation. Experiments on node classification, molecular property prediction, and financial spillover prediction demonstrate generally modest but consistent improvements over existing methods.

Critical strengths

(1) The paper presents a coherent framework that unifies SPD node representations and local metrics, allowing metric evolution to serve directly as the representation update mechanism.

(2) The theoretical analysis establishes continuous time energy dissipation and provides a connection to the linearized dynamics of RicciDeTurck flow.

(3) The experiments cover diverse applications, and the ablation studies provide evidence that the proposed propagation mechanism contributes to predictive performance.

Critical weaknesses

(1) The proposed curvature is constructed from a Dirichlet type aggregation of neighboring metric discrepancies. Its theoretical properties largely characterize the resulting diffusion dynamics. Given existing SPD graph learning and Ricci based feature evolution methods, the contribution appears incremental to me.

(2) Although the ablations demonstrate the importance of graph propagation and matrix valued updates, they do not establish whether the proposed Ricci type operator offers advantages over other nonlinear SPD diffusion operators.

(3)The improvements over strong baselines are generally small, and the benefits of adaptive weighting appear limited in the financial prediction ablations.

Questions for authors

(1) What is the specific advantage of the proposed SPD Ricci type evolution? The method introduces another way to propagate SPD valued node representations. What limitation of existing SPD graph learning or nonlinear diffusion methods does this construction address?

(2) What motivates the particular definition of curvature in Eq (6)? You derive a natural AIRM gradient operator in Eq (4) but choose a different, symmetrized operator in Eq. (6). Does this choice provide a distinctive geometric property or representational capability?

(3) Is the proposed operator empirically preferable to simpler alternatives? Can the authors compare it against alternatives using the same encoder, decoder, and model capacity?

6 comments

W1, Q1:

We thank reviewer WnLj for the thoughtful comments on the relationship between our construction and SPD diffusion, and for the opportunity to clarify its novelty. We agree that our curvature arises from a Dirichlet-type aggregation of neighboring metric discrepancies and that our theory characterizes the resulting diffusion dynamics.

This diffusion behavior is, however, not incidental but expected. In harmonic coordinates, or after the DeTurck modification (which alters Ricci flow only by diffeomorphisms), Hamilton's Ricci flow becomes a nonlinear heat-type equation for the metric (Example 1). A faithful discretization of Ricci flow should therefore exhibit diffusion dynamics, and the analysis in Sec. 3.2 verifies exactly this: at the level of linearized principal dynamics, our metric evolution is consistent with Ricci–DeTurck evolution. Where we differ from the reviewer's reading is in the perspective of the construction. We did not design a diffusion and then give it a Ricci interpretation; we derived the update by discretizing the principle of Ricci flow, namely curvature-driven evolution of the metric itself.

This perspective is what distinguishes our work from existing graph Ricci-flow methods. Many existing graph Ricci-flow methods use scalar discrete edge curvatures to evolve scalar edge weights. For example, GNRF starts from ∂twij=−κijwij,\partial_t w_{ij}=-\kappa_{ij}w_{ij}, where the scalar edge curvature κij\kappa_{ij} drives the evolution of edge weight wij,w_{ij}, rather than directly evolving the intrinsic geometry defined by a local metric tensor, as Hamilton's Ricci flow does through ∂tg(t)=−2Ric⁡(g(t))\partial_t g(t) = -2\operatorname{Ric}(g(t)). Accordingly, GNRF analyzes edge-curvature decay and feature Dirichlet-energy bounds, without establishing consistency with metric-tensor evolution.

The same perspective also distinguishes our work from existing SPD graph learning methods, and it determines the main design choices. (i) The node state is the local metric itself, rather than an SPD feature propagated by a separately chosen aggregation rule, such as tangent-space or Fréchet averaging in SPDGNN. (ii) The curvature is a node-wise tensor rather than a scalar edge quantity, so it is direction-resolved. and the curvature is a meaningful geometric quantity in its own right: we use it as a diagnostic of local metric structure in Sec. 4.1 (protein normal modes) and Sec. 5.3 (responses to financial news). Existing SPD aggregation schemes do not provide such a quantity.

Finally, although our update can be interpreted as SPD diffusion with distance-decayed weights, our research objective goes beyond designing such a propagation rule. If our goal were merely to design an SPD propagation rule, demonstrating its effectiveness and analyzing its stability or energy dissipation would be natural objectives, whereas establishing consistency with classical Ricci flow, as in Theorem 3.1, would not be necessary for that purpose. Our Ricci-type framing instead guides both the construction of the curvature-driven metric evolution and the theoretical analysis of its geometric consistency. Importantly, the resulting theoretical guarantees are properties of the proposed equations and remain valid under the diffusion interpretation. To our knowledge, no prior graph method evolves node-wise SPD metrics through a tensor-valued curvature while establishing both consistency with the linearized principal dynamics of Ricci–DeTurck flow (Theorem 3.1) and provable metric-field energy dissipation (Theorem 3.2). The novelty therefore lies in a geometrically grounded metric-evolution mechanism and its theoretically verified properties, rather than in another SPD propagation rule.

W2,Q3:

We thank the reviewer for this suggestion.

1.

We performed the requested comparison and add it to Tables 2 and 6 in the revision. All variants share the same encoder, decoder, BiMap, spectral activation, layer count, weight aija_{ij}, seeds, and training protocol etc. Only the update in Eq. (14) is replaced; SiS_i is the same function in Eq. (5). Writing gi+g_i^+ for the next state, we compare with three alternatives:

  • Log-Euclidean: gi+=exp⁡(log⁡gi+η∑j∼iaij(log⁡gj−log⁡gi))g_i^+=\exp(\log g_i+\eta\sum_{j\sim i}a_{ij}(\log g_j-\log g_i)), i.e., diffusion in matrix-logarithm coordinates.
  • AIRM Euler: gi+=gi+ηgi1/2Sigi1/2g_i^+=g_i+\eta g_i^{1/2}S_i g_i^{1/2} i.e., the AIRM gradient of the Dirichlet energy in Eq. (4). It uses the same whitened tensor SiS_i as ours and differs only in how SiS_i is lifted to the tangent space at gig_i (see our response to Q2).
  • Euclidean: gi+=gi+η∑j∼iaij(gj−gi)g_i^+=g_i+\eta\sum_{j\sim i}a_{ij}(g_j-g_i) i.e., diffusion of the matrix entries.

SPD Ricci Flow has the highest mean on all eight metrics (Tables A–B). LE = Log-Euclidean; AIRM = AIRM Euler; EU = Euclidean. Values with ± report mean ± standard deviation.

Table A. Financial spillover

UpdateMacro-AUPRCNDCG@5Precision@5Hit@5
LE33.57±3.3230.48±4.2812.76±1.4962.58±8.07
AIRM35.19±0.5632.51±0.6413.42±0.2165.61±0.77
EU33.36±0.5229.95±0.8012.59±0.2363.28±2.09
Ours35.24±1.0132.67±1.2813.57±0.4066.11±1.66

Table B. Node classification (accuracy) and molecular prediction (ROC-AUC)

UpdateWisconsinDBLPBACESIDER
LE88.00±3.1085.87±0.4287.72±1.0384.51±0.22
AIRM86.80±3.3785.75±0.3088.13±0.9184.41±0.21
EU88.60±2.8485.87±0.3687.27±1.0884.60±0.15
Ours89.80±3.5285.90±0.3989.70±1.0085.00±0.30

2.

To further examine what the update rule itself contributes, we compare the four updates on a synthetic SPD denoising task with known ground truth, i.e., a denoising problem defined directly on an SPD field, which allows us to see how our update behaves differently from the alternatives (Appendix A.8). The graph is a 32×3232\times32 periodic grid. Each node is assigned a ground-truth SPD matrix defined as $ g^\star(x,y)=\mathrm{diag}(e^{a(x,y)},e^{-a(x,y)}), a(x,y)=\tfrac12\log20+0.3\cos(\pi x/4)\cos(\pi y/4).$ The noisy input is gi(0)=(gi⋆)1/2exp⁡(0.3Ni)(gi⋆)1/2g_i^{(0)}=(g_i^\star)^{1/2}\exp(0.3N_i)(g_i^\star)^{1/2}, where NiN_i is symmetric with independent Gaussian entries, normalized by E∥Ni∥F2=1\mathbb{E}\lVert N_i\rVert_F^2=1.

All models consist of the operator alone, without any learnable parameters (no encoder, decoder, BiMap, or training): the noisy SPD field is given as input, the operator is applied for KK steps to obtain the output, and the output is then compared with the ground-truth SPD matrices. In this way, we test only the operator itself. Step count KK are selected on separate validation noise samples using the same grid: K∈1,…,50K\in\\{1,\ldots,50\\}. Results use 30 independent noise realizations.

As shown in Table C below, our update attains the lowest AIRM error, reducing it by 5.1%, 6.0%, and 10.1% relative to Log-Euclidean, AIRM Euler, and Euclidean diffusion, respectively, and it outperforms all three in each of the 30 noise realizations. The largest difference appears in orientation: the principal-axis error is about 37–41% lower than that of the alternatives.

This is consistent with the coupling analyzed in Q2. Given the same state gg and message SS, the updates 12(gS+Sg)\frac{1}{2}(gS+Sg) (ours) and g1/2Sg1/2g^{1/2}Sg^{1/2} (AIRM Euler) have identical diagonal entries in the eigenbasis of gg. Their off-diagonal entries, which rotate the eigenvectors, are λ1+λ22S12\tfrac{\lambda_1+\lambda_2}{2}S_{12} and λ1λ2,S12\sqrt{\lambda_1\lambda_2}\\,S_{12}, respectively. Our update therefore corrects orientation discrepancies more strongly, by the factor c(κ)=12(κ+1/κ)≥1c(\kappa)=\frac{1}{2}(\sqrt{\kappa}+1/\sqrt{\kappa})\ge 1, while treating eigenvalue discrepancies in the same way.

Overall, we view this experiment as a complement that helps clarify how the operators differ; for the comparison on real data, see Tables A and B above.

Table C. SPD denoising; lower is better.

For recovered g^i\hat g_i and ground truth gi⋆g_i^\star, errors are averaged over nodes:

  • AIRM: ∥log⁡((gi⋆)−1/2g^i(gi⋆)−1/2)∥F\lVert\log((g_i^\star)^{-1/2}\hat g_i(g_i^\star)^{-1/2})\rVert_F.
  • Axis (rad): arccos⁡∣⟨u1(g^i),u1(gi⋆)⟩∣\arccos\lvert\langle u_1(\hat g_i),u_1(g_i^\star)\rangle\rvert, where u1u_1 is the leading eigenvector.
  • Anisotropy: ∣log⁡κ(g^i)−log⁡κ(gi⋆)∣\lvert\log\kappa(\hat g_i)-\log\kappa(g_i^\star)\rvert, κ=λmax⁡/λmin⁡\kappa=\lambda_{\max}/\lambda_{\min}.
  • Log-det: ∣log⁡det⁡g^i−log⁡det⁡gi⋆∣\lvert\log\det\hat g_i-\log\det g_i^\star\rvert.
MethodAIRMAxis (rad)AnisotropyLog-det
Noisy input0.2770±0.0034———
LE0.1136±0.00200.007210.11200.0619
AIRM0.1147±0.00210.007220.11240.0638
EU0.1200±0.00230.007660.10710.0797
Ours0.1078±0.00230.004530.11310.0655

W3:

We thank the reviewer for the thoughtful comments on the empirical gains and the role of adaptive weighting. We agree that the margins over the strongest baselines are modest on several benchmarks, and we do not claim state-of-the-art performance as a primary contribution (Line 86). As discussed in our response to W1, our main contribution is theoretical: a curvature-driven evolution of the metric itself, whose linearized principal dynamics are consistent with Hamilton's Ricci flow (Sec. 3.2). As a general propagation layer without task-specific components, SPD Ricci Flow still improves consistently across three task families: it obtains the highest mean on 4 of 5 node-classification benchmarks, is best or tied-best on 5 of 7 molecular benchmarks, and is best on all four financial metrics (Tables 1–2, Fig. 3).

We also agree that the effect of adaptive weighting on the financial task appears limited. But this does not affect our main contribution, because the theory does not depend on this particular exponential adaptive weighting (Eq.5). The weights aij=wije−βrija_{ij}=w_{ij}e^{-\beta r_{ij}} are induced by the Dirichlet energy through aij=2wijψβ′(rij)a_{ij}=2w_{ij}\psi_\beta'(r_{ij}) (Eqs. 3–5), and β=0\beta=0 recovers uniform weights with the standard energy ψ0(r)=r/2\psi_0(r)=r/2. Thm. 3.2 holds for any ψ\psi with ψ′≥0\psi'\ge 0, Prop. 3.1 for any β≥0\beta\ge 0, and Lemma 2 for any nonnegative scalar weights. The adaptive weighting is therefore one instantiation permitted by our framework rather than the source of its contribution.

Since the framework admits any such weighting, a natural question is whether our geometric choice is meaningful, or whether a more flexible learned weighting would perform better. To examine this, we replace the AIRM distance with a learned, SPD-independent embedding distance:

a_{ij}=w_{ij}\exp\\!\big(-\beta\lVert u_i-u_j\rVert^2\big).$$ This gate adds 2.5% trainable parameters. As the table below shows, it yields slightly lower means than our parameter-free weighting on three of the four metrics. We therefore retain the adaptive form because it adds no parameters, is geometrically interpretable as a decreasing function of the AIRM distance, and adapts to the evolving state while preserving the energy structure behind Thm. 3.2. In the revision, we will clarify that the weighting is an optional component rather than a requirement of the theory (line 175), and place the emphasis on the definition of the SPD Ricci-type curvature and the resulting metric update, which constitute the core of our contribution. **Financial spillover: effect of the weighting rule (curvature update fixed).** | Weighting | Params. | Macro-AUPRC | NDCG@5 | Precision@5 | Hit@5 | |:---|---:|---:|---:|---:|---:| | Uniform ($\beta=0$) | 54,411 | $34.58_{0.43}$ | $31.79_{0.58}$ | $13.32_{0.21}$ | $65.75_{1.20}$ | | Learned, SPD-independent | 55,771 | $34.93_{0.48}$ | $32.29_{0.63}$ | $13.49_{0.19}$ | $\mathbf{66.15}_{1.09}$ | | **Adaptive (ours)** | 54,411 | $\mathbf{35.24}_{1.01}$ | $\mathbf{32.67}_{1.28}$ | $\mathbf{13.57}_{0.40}$ | $66.11_{1.66}$ |

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary

Dear Reviewers, ACs, and PCs,

We sincerely thank all four reviewers for their careful reading and constructive feedback. We thank Reviewer WnLj for recognizing the coherent framework connecting SPD representations with local metrics, and Reviewer MJak for appreciating the architectural idea of evolving the metric itself and our explicit discussion of the scope of the theoretical results. We thank Reviewer CjhZ for recognizing the energy-dissipation analysis and the empirical evaluation across multiple task families. We also thank Reviewer 9HdX for raising important questions about smoothing and pattern formation, which prompted us to examine the model’s capabilities and limitations more directly.

The reviews prompted us to clarify the geometric contribution, isolate the effects of metric evolution and adaptive weighting, connect the theory to the implemented layer, and examine smoothing-related limitations. Our main responses and additions are summarized below.

  • Geometric contribution [WnLj; MJak]. We clarify the motivation for the symmetric curvature construction and show how our flow differs from standard AIRM gradient flow at noncommuting states, including its distinct orientation response. We also clarify the scope of the Ricci–DeTurck principal-part correspondence and the dissipation guarantee for our nonlinear flow.

  • Controlled comparisons [WnLj; MJak; CjhZ]. With other components fixed, our update achieves the highest mean scores across the eight reported metrics when compared with Log-Euclidean, AIRM Euler, and Euclidean diffusion. Expanded ablations separate the effects of three weighting schemes and two update rules, showing complementary, task-dependent contributions rather than attributing all gains to either component.

  • Finite-step guarantees and implementation [CjhZ; MJak]. We provide explicit activation and step-size conditions for energy nonincrease in the complete layer, together with a concrete bound for the spherical experiment. We also document how BiMap invertibility is enforced throughout training and report parameter counts and runtime.

  • Smoothing and pattern formation [9HdX; MJak]. Energy, dispersion, and effective-rank diagnostics show substantial metric variation at practical depths, while deeper propagation can reduce performance. On the added pattern-formation task, our model outperforms GCN, indicating its potential for modeling heterogeneous node states. We discuss both this capability and the remaining limitation.

The changes are highlighted in the revised PDF. They include clarified performance claims (line 86), the optional role of adaptive weighting (line 175), and the curvature construction (line 182); expanded comparisons and ablations (Tables 2 and 6); SPD denoising (Appendix A.8), oversmoothing diagnostics (Appendix A.9), and pattern formation (Appendix A.10); and a strengthened limitations discussion (Section 6, line 480). We also provide proofs of the finite-step dissipation guarantees (Appendix A.11), details of how BiMap invertibility is enforced (Appendix A.4), and an analysis of computational costs (Appendix A.12). Supporting evidence is provided in our individual responses.

We sincerely thank the reviewers again and we are grateful for the reviewers’ engagement with our work. The questions prompted new analyses and experiments that sharpen our account of what the proposed metric evolution contributes and where its limitations lie! We hope these additions address the concerns raised and provide a stronger basis for an updated assessment of the paper. Thank you for considering the revisions; we would be happy to clarify any points that remain unresolved.

Best regards,
The Authors

Dear Reviewers,

Please respond to the authors' rebuttal. Thank you.

Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.