acceptodds
Under review as a conference paper at ICLR 2027

CURVSPREAD: CRITICAL CONFINEMENT FOR HYPERBOLIC GRAPH SELF-SUPERVISION

Abstract

Dispersion prevents collapse in self-supervised learning, but hyperbolic space has no finite radius at which spreading naturally stops. We show that whether spreading stops depends on the reference measure of the spreading law. A wrapped law, specified in tangent coordinates and pushed through the exponential map as in tangent-space isotropy objectives, always inherits a finite radial scale from the tangent law. An intrinsic law, specified against hyperbolic volume under linear radial confinement, has a finite equilibrium exactly when the confinement coefficient exceeds the volume-growth rate . The ratio of the two is thus a single critical coordinate across dimensions and curvatures, with the boundary at one. We introduce CurvSpread, which adds the exponential-map Jacobian to a local-spacing entropy surrogate and calibrates confinement to this rate, so that the confinement strength is set by the geometry rather than tuned per dataset. We prove that its finite-sample objective, before clipping, has exactly this threshold for every sample size. Across 25 dimension–curvature settings, the operational midpoints of the trained-encoder response are centered on the predicted boundary (median 1.00), while the sharpness of that response varies with geometry. One calibration, selected on three development graphs, is competitive on ten node-classification benchmarks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hyperbolic Graph Learning, Graph Self-Supervised Learning, Riemannian Geometry, Representation Learning, Hyperbolic Contrastive Learning.

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_LfjcOpenReview ↗
Rating
4
Confidence
1
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies hyperbolic graph contrastive learning, especially focusing on spreading. Different from previous papers, the authors claimed that intrinsic construction is better than wrapped construction, since the former is directly specified in hyperbolic space. Furthermore, they discovered a critical intrinsic regime, a balance between β\beta and hκh_\kappa, and only β>hκ\beta>h_\kappa can guarantee that the optimized embeddings are normalizable. This critical value is more accurate than former wrapped construction.

Critical strengths
  1. This paper studies a key problem in hyperbolic graph contrastive learning, how to avoid embedding spreading in hyperbolic space.
  2. The paper is strongly supported by theory.
Critical weaknesses
  1. In the main experiments, the selected datasets may not present obvious tree-like structure, so the brought improvement looks not significant.
Questions for authors
  1. The main motivation is to avoid unconstrained spreading in hyperbolic space. However, we also use alignment loss in GCL, which may cancels out a part of spreading. Could you further illurstate if this unconstrained spreading is severe in real case?
  2. One step is to find KNN in hyperbolic space, which is a time-consuming operation. Why not directly use the original edges to find neighbors?
  3. To further prove the usefulness of hyperbolic space, a variant could choose Euclidean space with the same other components (e.g., encoder, augmentations, objectives).
3 comments

Thank you for the encouraging assessment and for raising these useful questions. We agree that the distinction between the geometric argument and what happens during training deserves a clearer explanation. We address each point below, including a focused control for the role of hyperbolic geometry.

W1 : Applicability beyond tree-like graphs. The confinement threshold comes from volume growth in hyperbolic space, not from assuming that the input graph is a tree. Our benchmarks include Disease, a tree with 1,044 nodes and 1,043 edges, and Airport, which contains cycles. Table 6 provides a direct test of what intrinsic spreading contributes when the hyperbolic encoder and the rest of the training setup are unchanged:

Accuracy (%), mean ± s.d.CoraDiseaseAirport
Alignment only85.90±0.3592.80±0.5092.60±0.49
CURVSPREAD88.82±0.1295.90±0.3195.76±0.28

The gains are 2.92, 3.10, and 3.16 percentage points, respectively. The improvement on Airport is particularly relevant here: the benefit is not limited to a tree input. Of course, cycles do not rule out latent hierarchical structure. The comparison establishes the value of spreading within a hyperbolic model; Q3 addresses the separate question of whether hyperbolic geometry itself helps.

Q1 : Why alignment does not prevent radial escape. Alignment asks two views of the same node to agree. It says nothing about how far their common representation lies from the center: even two identical embeddings at a very large radius incur zero alignment loss. Proposition 3(iii) captures this distinction. At fixed curvature, alignment does not change the condition β>(d−1)√κ for the unclipped finite-sample objective. This is not a claim that a clipped SGD trajectory literally diverges.

We can also see the radial effect in trained models. Table 10 keeps alignment active and varies confinement. On Disease at (d,κ)=(256,2), increasing χ from 0.90 to 1.10 reduces the terminal 95th-percentile radius, normalized by the clipping radius, from 1.000 to 0.406 when measured from the origin, and from 0.637 to 0.156 when measured from the learned center. The latter is useful because clipping about the origin does not directly fix distances from that center. At (64,1), the origin-referenced value also decreases, from 0.516 to 0.356, although more gradually. Thus alignment does not remove the radial-scale effect, and we do not suggest that every geometry exhibits an abrupt training transition.

Q2 : Why embedding-space kNN instead of graph neighbors. These neighborhoods serve different purposes. Graph edges define which nodes exchange messages. The kNN distances in Eq. (7), by contrast, estimate local density in the learned representation distribution for the Kozachenko–Leonenko entropy term. A graph neighbor need not be among a node's nearest neighbors in representation space; substituting adjacency distances would therefore change the estimator, not simply speed it up.

The implementation also avoids a full-graph hyperbolic-geodesic neighbor search. Algorithm 1 log-maps a sample to the tangent space at a common center and computes Euclidean kNN spacings within that sample. With at most M=2,048 embeddings, exact pairwise distances cost O(M²d) time and O(M²) distance storage. We will make this sampling step and its computational overhead clearer in the paper.

Q3 : Matched Euclidean control on Airport. We agree that the Euclidean baselines in Table 1 cannot isolate geometry because several other components differ. We therefore added a targeted comparison on Airport, keeping the data splits, encoder capacity, augmentations, optimizer, training and evaluation budgets, linear probe, and random seeds matched. The Euclidean version uses squared-distance alignment and kNN entropy without the hyperbolic volume-Jacobian term. Its confinement strength is tuned separately with the same validation budget; simply taking κ→0 in the hyperbolic formula would remove confinement and make the comparison unfair.

Airport accuracy (%), mean ± s.d.Result
CURVSPREAD (hyperbolic)95.76±0.28
Matched Euclidean variant94.91±0.53
Difference in reported means+0.85 points

The hyperbolic variant performs better in this controlled Airport comparison. We view that as evidence for the usefulness of the geometry on this benchmark, not as proof of statistical significance or a universal advantage on all graphs. Our theoretical contribution is more specific: it identifies the intrinsic entropy formulation and critical normalizability threshold in hyperbolic space. In the revision, we will also separate the population argument, finite-sample objective, and behavior of clipped training more clearly.

Thanks for the authors' replies. I will keep my positive score.

We sincerely thank the reviewer for their careful consideration of our responses and positive assessment of the revised manuscript.

Reviewer_5hffOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies representation spreading and radial confinement in hyperbolic graph self-supervised learning. It distinguishes between wrapped constructions, which specify a distribution in tangent coordinates, and intrinsic constructions, which specify an energy relative to Riemannian volume. For intrinsic entropy with linear radial confinement, the paper establishes the normalizability threshold β > (d - 1)√κ. The proposed CURVSPREAD regularizer combines a nearest-neighbor spacing entropy surrogate, an exponential-map Jacobian correction, and a geometry-calibrated linear confinement coefficient. Its main theoretical contribution is a finite-sample result showing that the implemented spreading functional has the same critical coefficient as the population free energy. The analysis also addresses a moving Lorentz centroid and the addition of an alignment loss. Experiments examine training responses across dimension–curvature settings, node classification, component ablations, and link prediction. The paper presents a coherent theoretical motivation and carefully distinguishes objective-level guarantees from training behavior. However, the empirical evidence does not yet sufficiently establish the advantage of the proposed spreading principle over well-tuned alternatives under a common implementation and evaluation protocol.

Critical strengths
  1. A meaningful connection between theory and the implemented regularizer. Proposition 3 goes beyond the population normalizability argument by establishing the critical coefficient for the finite-sample functional actually used by the method.
  2. The normalizability argument, Gibbs/KL decomposition, supercritical lower bound, and treatment of logarithmic escape at the critical boundary appear internally consistent. The moving-centroid argument is handled through isometry invariance. I did not identify an obvious error invalidating the central propositions.
Critical weaknesses
  1. The most important controlled comparison is missing. The wrapped Gaussian and Pairwise-U arms are auxiliary diagnostics rather than tuned downstream baselines. The paper also does not compare the actual HyperGCL isotropy regularizer against CURVSPREAD under the same implementation stack. Consequently, the ablations establish that the proposed components help within the chosen configuration, but do not sufficiently establish that intrinsic spreading is preferable to well-tuned wrapped or pairwise alternatives.
Questions for authors
  1. Does CURVSPREAD retain its advantage over tuned wrapped, actual HyperGCL isotropy, and pairwise regularizers under a common implementation and matched tuning budgets?
  2. What are the per-cell bootstrap confidence intervals for Table 9? Are the midpoint conclusions stable after excluding weak responses or changing the sweep endpoints?
  3. How is fixed β instantiated from the search coordinate? Do the ablation conclusions hold after including χ=1.1 and expanding the quadratic-confinement search?
  4. Are validation and test edges removed before representation learning in the link-prediction experiments?
3 comments

Thank you for the thoughtful review. The central question is whether intrinsic spreading still helps when competing objectives are evaluated under the same conditions. We address that comparison first, then the robustness, ablation, and link-prediction concerns.

W1 + Q1 : Matched comparisons with alternative spreading objectives. The wrapped-Gaussian and Pairwise-U experiments in Appendix H.2 were designed as mechanism checks, not tuned classification baselines. Table 7 likewise compared against author-reported HyperGCL results. We therefore evaluated all five spreading objectives in a shared pipeline: the same two-layer GCN, augmentations, alignment, learnable curvature, data splits, optimizer, linear probe, and ten paired seeds. Each method received 16 validation configurations with objective-appropriate ranges. The HyperGCL arms implement the original Normal and Laplace isotropy losses; only the spreading objective changes.

Accuracy (%), mean ± s.d.CoraCiteSeerPubMed
CURVSPREAD88.69±0.2576.71±0.4088.33±0.29
HyperGCL-Normal87.96±0.3175.93±0.4387.69±0.29
HyperGCL-Laplace87.83±0.3575.86±0.4587.61±0.32
Wrapped Gaussian87.58±0.3775.61±0.4887.42±0.33
Pairwise-U87.42±0.4075.49±0.5087.21±0.36

Against the strongest alternative, the gains are 0.73, 0.78, and 0.64 percentage points on Cora, CiteSeer, and PubMed. We bounded the two-sided paired t-test p-values for all twelve contrasts and applied Holm correction. Because this summary does not use the paired-seed covariance, the values below are conservative upper bounds on adjusted p-values, derived from the reported sample standard deviations over ten runs, not exact paired-test p-values.

Holm-adjusted p upper boundCoraCiteSeerPubMed
HyperGCL-Normal0.01560.02310.0206
HyperGCL-Laplace0.01140.02310.0188
Wrapped Gaussian0.00340.01680.0110
Pairwise-U0.00200.01430.0041

All twelve bounds are below 0.05 under the stated t-test assumptions. This directly addresses the reviewer's main concern: in a matched implementation, Riemannian-volume entropy with geometry-calibrated confinement improves downstream accuracy over the alternative spreading objectives.

Q2 : Uncertainty and robustness of Table 9. A plateau midpoint is a location descriptor, not necessarily a sharp transition. The 25 reported midpoints have median 1.000, IQR [0.98, 1.03], and range [0.95, 1.07]. The flagged smooth-response setting (d=16, κ=0.1) has midpoint 0.96; removing it changes the descriptive median only to 1.005. In the additional endpoint checks, restricting the χ window to [0.75, 1.45] or [0.80, 1.40] gives medians of 1.002 and 1.007. For uncertainty, the appropriate unit of resampling is the full seed-level response curve: 1,000 bootstrap resamples recompute each midpoint and record censored crossings. The cell-wise intervals and weak-response diagnostics belong in the supplement, rather than a 25-cell table here. The χ=1 boundary itself is proved for the specified objective in Propositions 2–3; it is not a claim that every training curve has a sharp change point.

Q3 : Fixed-β and quadratic-confinement controls. For the constant-coefficient baseline, β_fixed=χ_search(d−1)√κ₀ is set at initialization. CURVSPREAD instead uses β_t=(1+δ)(d−1)√sg(κ_t), giving χ=1.1 for δ=0.1. The earlier fixed-β search omitted 1.1. We added that value and left curvature learnable, so the revised comparison changes the β schedule rather than β and κ together. We also retuned the quadratic penalty on its own coefficient scale instead of reusing a grid intended for linear confinement.

Track A accuracy (%), mean ± s.d.CoraCiteSeerPubMed
CURVSPREAD (Table 6)88.82±0.1276.86±0.3888.47±0.27
Fixed β, learnable curvature88.07±0.3076.09±0.4387.85±0.31
Quadratic, independently tuned87.69±0.3575.75±0.4687.47±0.34

The corrected controls still trail CURVSPREAD. This is an empirical observation, not a consequence of the theory: our analysis establishes a critical threshold for intrinsic linear confinement, not a guaranteed ranking of downstream accuracies.

Q4 : Link-prediction leakage. The split precedes representation learning, not just downstream evaluation. Validation and test positive edges are excluded from the graph used by the encoder; message passing and both augmentations operate on training edges only. Validation links are used for model selection, while test links are reserved for final evaluation. We will state this sequence explicitly in Appendix I.1.

Thanks for the authors' response. It answers what I concern. I will maintain this score.

We sincerely thank the reviewer for their positive feedback and for taking the time to evaluate our responses. We greatly appreciate their consideration and constructive assessment.

Reviewer_4jtAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Based on theoretical insight, the authors propose CURVSPREAD, a geometry-calibrated entropy regularizer, and evaluate its effectiveness through controlled experiments and downstream node-classification benchmarks.

Critical strengths
  1. The paper's contribution is pretty clear and easy to read
  2. I like the theory-guided motivation of this paper, and the author did derive the method from theory
  3. They also give the simulation of theory
Critical weaknesses
  1. I am worried about the performance on large datasets. Because in Table1, the performance on arxiv is not so good.
  2. can you provide the hyperparameters of your method and baselines? and how do you select them?
  3. I do graph but not expert in HYPERBOLIC GNN design, so I am not sure whether the baseline work are recent work .
  4. could you give some intuition or motivation of the theorical part? like why you select this theorical framework
Questions for authors

See weakness

3 comments

We sincerely thank the reviewer for the thoughtful assessment and constructive feedback. We address each concern below, clarifying the empirical findings, experimental protocols, baseline comparisons, and theoretical foundations of CURVSPREAD.

W1: Performance and scalability on large graphs. We agree that CURVSPREAD does not achieve the highest mean accuracy on ogbn-arxiv. Table 1 reports 71.63±0.12% on 169,343 nodes, compared with HGCL (71.35%), HyperGCL-Normal (71.14%), HyperGCL-Laplace (71.34%), and the highest listed mean, Dual-Kernel GCCL (72.20%). The last gap is 0.57 percentage points, and we make no claim of statistical superiority from these heterogeneous reported/reproduced cells. CURVSPREAD was evaluated on ogbn-arxiv with the same development-selected configuration used elsewhere, and the released evaluation code loads frozen hyperparameters instead of running an arxiv-specific search. Moreover, the exact kNN-spacing regularizer uses a subset of at most M=2,048 nodes per step; its additional distance calculation takes O(M²d) time and O(M²) memory rather than forming an N×N matrix (Appendix G.1). This supports feasibility on this benchmark, not a claim of superior end-to-end runtime or unlimited scaling, since encoder and alignment costs remain.

W2: Selection and reproducibility of hyperparameters. Table 4 and Appendix H.3 specify a validation-only, two-stage development search on Cora, Amazon Photo, and Coauthor CS (run IDs 1–3). Stage A searches δ∈{0.05,0.10,0.20,0.50} × η∈{0.1,0.3,1,3}, at k=5, M=2048. Stage B freezes (δ,η) and searches k∈{1,3,5,10} × M∈{1024,2048,4096}. The reported selection is (δ,η,k,M)=(0.10,1.0,5,2048) and is reused for CURVSPREAD benchmarks. The implementation suppresses test evaluation during development search and loads the frozen configuration for arxiv. Regarding comparison fairness, Appendix H.3 distinguishes baseline cells designated as official-code reproductions from literature-reported values; Table 1 combines these sources and must not be interpreted as a universally paired or uniformly tuned ranking. We will present this provenance alongside the results and avoid treating literature values as matched controls.

W3: Baseline relevance and recency. The closest geometric reference is HyperGCL (KDD 2025), which addresses hyperbolic representation collapse through tangent-space Normal/Laplace isotropy; HGCL provides an earlier hyperbolic baseline. Table 1 also includes recent general graph-SSL approaches (GraphECL and S3GCL, 2024; GCL-JAM and GALE, 2025; EDA-GCL and Dual-Kernel GCCL, 2026). These methods provide broader downstream-accuracy context, not geometry-matched ablations. The mechanism distinguishing CURVSPREAD from the closest prior is its entropy with respect to Riemannian volume, together with a confinement coefficient calibrated to the manifold's asymptotic volume growth. We will make the separation between geometric nearest prior and wider benchmark context explicit.

W4: Why the geometric theory is necessary. In d-dimensional hyperbolic space of curvature −κ, radial volume grows asymptotically at rate hκ=(d−1)√κ. Under linear intrinsic confinement βr, the radial density has tail proportional to exp[−(β−hκ)r], which is integrable if and only if β>hκ (Proposition 2). Consequently χ=β/hκ gives a dimension- and curvature-normalized critical coordinate, with boundary χ=1. Quadratic confinement is normalizable for any positive coefficient and does not expose this positive linear threshold. Because tangent-space kNN spacings estimate entropy relative to Euclidean Lebesgue measure, the exponential-map log-Jacobian is essential to obtain entropy relative to hyperbolic volume. CURVSPREAD therefore uses β=(1+δ)(d−1)√sg(κ), yielding χ=1+δ on each forward pass while preventing a spurious derivative through the curvature-dependent coefficient. Crucially, Proposition 3 extends the critical coefficient to the unclipped finite-sample objective, including the moving centroid and the alignment term at fixed curvature. It does not assert SGD divergence/convergence, a specific learned radius, or a sharp transition in clipped training. This precise scope is the theoretical contribution, rather than an empirical claim of universal dominance.

Thanks for the authors' replies. I will keep my score.

We sincerely thank the reviewer for their positive evaluation and constructive feedback.

Comments on the paper

1 comment

Hi Reviewers,

We are now near the end of the rebuttal period. If you haven't done so yet, please review the authors' responses and post a reply addressing whether their clarifications or new results resolve your concerns. Even if your score or recommendation remains unchanged, a brief note confirming what did or didn't shift your assessment—and why—is essential for the discussion and final decision. Thank you for your time and engagement.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.