acceptodds
Under review as a conference paper at ICLR 2027

Local Global Geometric Insights for Graph Neural Networks via Entropic Curvature

Abstract

Curvature notions on graphs, particularly Ollivier–Ricci and Forman curvatures, have emerged as powerful tools for addressing fundamental issues in Graph Neural Networks (GNNs), such as oversmoothing and oversquashing. While these metrics capture the stability and efficiency of information propagation, they primarily rely on local neighborhood comparisons and may overlook broader global features. In contrast, we introduce Entropic Curvature, which extends the Lott–Sturm–Villani framework to discrete spaces by focusing on the global displacement of probability measures along geodesics in Wasserstein space. By relating entropy, optimal transport, and curvature lower bounds, this framework offers powerful functional-analytic tools, such as transport-entropy and Poincaré-type inequalities, which establish a thermodynamic regularity on the information flow. We prove that positive entropic curvature provides quantitative guarantees against feature collapse and adversarial distribution shifts. Furthermore, we establish a fundamental expansion paradox, proving that sparsity, strong spectral expansion, and positive entropic curvature cannot coexist in large graphs, revealing that the fast communication required to prevent oversquashing inherently forces a move toward negative curvature and instability. Empirically, we introduce a tractable Weak Entropic Curvature proxy and an E-Gate mechanism that modulates message-passing based on this global prior, consistently enhancing training stability and performance across diverse benchmarks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Graph Curvature

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_w98AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this work the authors present a study on entropic curvature, a different type of discrete curvature for graphs, based on optimal transport and the Lott–Sturm-Villani framework. The study is built on strong theoretical foundations and contributes towards numerous perspectives: the discrete reformulation, the anti-oversmoothing guarantee, a generalization bound and the newly proposed GNN architectures. Nevertheless, the paper does not show application on larger real-world datasets.

Critical strengths
  • A real and rigorous curvature notion into GNNs with theoretical foundation
  • Anti-Oversmoothing Guarantee via Poincaré Inequality is one of the first designed under a curvature notion
Critical weaknesses
  • The “local vs. global” pitch is not clear. κ_w to my understanding is just as local as Ollivier-Ricci (2-hop). The global guarantees only apply when the minimum of κ_w over all nodes is positive. Ollivier-Ricci also gives global results under a uniform positive lower bound. The real difference to me is which inequalities you get, not local versus global.
  • The benchmarks are quite weak. The node-classification datasets are Cora, CiteSeer, PubMed, Cornell, Texas and Wisconsin (with 183–251 nodes each). Under 60/20/20 splits, Cornell’s test set is about 37 nodes. The paper’s pitch is also about long-range information flow, yet I saw no long-range benchmark (LRGB) if I am not mistaken.
Questions for authors
  • Elaborate on the local vs global presentation
  • Why not test on larger datasets?
Reviewer_UVihOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces entropic curvature as a theoretical tool to study graph neural networks. It derives several theoretical results that relate this new curvature notion to existing GNN phenomena, most importantly over-smoothing and over-squashing, as well as generalization gaps. Finally, the paper introduces two practical architectures that derive from entropic curvature: an E-Gate mechanism that modulates message-passing aggregation, and Entropic Structural Encodings (ENT) as an alternative to existing positional and structural encodings for GNNs. The authors evaluate their theoretical and architectural innovations on synthetic and real-world datasets, primarily standard node classification tasks.

Critical strengths

While the use of discrete curvature notions to study graph neural networks is hardly novel at this point, the authors provide an extensive treatment of entropic curvature and how it can be connected to well-known GNN phenomena such as over-smoothing and over-squashing.

Critical weaknesses
  • Motivation and structure: as the paper admits, several notions of discrete curvature have been proposed and connected to GNNs over the last few years. In spite of its numerous propositions and extensive theoretical results, the paper somehow manages to never fully address why this new notion of entropic curvature is really needed. This also relates to the overall structure of the paper: the authors introduce a total of five core contributions, including two architectural innovations, yet none of these seem particularly convincing or developed in depth. Fewer contributions that can be covered in full in the main text would a) significantly improve the readability of the manuscript, and b) potentially clarify the authors’ core contributions.
  • Experimental evaluation: this relates to my first point. The two architectural contributions in the paper are squashed into a single section, all details are moved to the appendix, and the experimental results are presented with five relatively small datasets each. It seems that almost all experimental results are statistically insignificant. In addition to the structural re-organization I suggested above, I would also ask the authors to evaluate both architectural contributions on more - and especially larger - datasets. LCP, for example, is evaluated on graph classification tasks from the TU datasets - why not also test ENT on these? The E-Gate mechanism would similarly benefit from more dataset and especially from baseline comparisons, for example against curvature-based rewiring techniques, holding the GNN fixed, or against alternative aggregation mechanisms that also use gating - most importantly GatedGCN.
Questions for authors

A few clarification requests:

  • In section 4, it says: “A central question in GNN design is whether a graph can simultaneously achieve three desirable properties: sparsity (bounded degree, ensuring scalable message-passing), spectral expansion (fast mixing, preventing oversquashing), and positive entropic curvature (geometric regularity, preventing oversmoothing).” The last point about positive entropic curvature in particular makes this look like a claim that the authors are making. If so, could you provide evidence that graphs satisfying these properties (especially 2 and 3) are indeed desirable? If this is to be read as a summary of the prior GNN literature, could you please provide appropriate references?
  • Also in section 4: “The paradox thus unifies oversmoothing and oversquashing as two manifestations of the same geometric trade-off: they are not independent failure modes, but opposite ends of a single curvature spectrum.” How is this substantially different from the ORC-based results in Nguyen et al.? They also relate both over-smoothing and over-squashing to the same notion of curvature (ORC).
  • At the end of section 4: “our Midpoint-Completion Rewiring (MCR) procedure repeatedly targets the most negatively curved region and adds an edge that creates an additional midpoint for a poorly supported two-hop connection” How does this relate to existing curvature-based rewiring approaches? In particular, why only compare against Ollivier and Forman Ricci curvature approaches on synthetic graphs to demonstrate spectral gap improvements, instead of comparing against existing rewiring techniques such as BORF or FoSR to see how performance in real-world datasets is affected?
Reviewer_hGoPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces the notion of entropic curvature, which adapt prior work on optimal transport-based curvature to discrete spaces, and the weak entropic curvature, which is a efficiently-computable lower bound for the entropic curvature. It then uses these curvature notions to establish several results, including 1) no sufficiently large graphs can be 'perfect' for graph learning, 2) upper bound generalization error, 3) derive bounds for the variation of graph signals; and also to 4) experiment with using them to either modify GNN weights or to use as structural encoding.

Critical strengths
  • The paper touches on many topics of relevance to the geometric deep learning community, including over-smoothing, over-squashing, generalization error, geometric techniques for rewiring and positional encoding.
  • Its use of an efficiently-computable curvature notion in a 2-hop neighborhood instead of just a 1-hop neighborhood via entropic curvature is a nice idea that extends prior work.
Critical weaknesses
  • The paper has a low presentation quality. The sections do not follow a logical flow, e.g., a result from a later section (Theorem 7.1) motivates implementation details in an earlier section (Section 6); a result from a later section and a result from an earlier section is jointly discussed in the current section (see the end of Section 4); even in the appendix, the proofs of the results are not presented in order.
  • Three of the five claimed contributions are not well-substantiated (anti-oversmoothing, expansion paradox, generalization bound) and lack exposition.
  • The supposed proof of Theorem 4.2 is presented more like a literature review than a proof. I could not parse where the proof begin and end among 6 pages of definitions and results citing from prior work.
  • The paper contains at least 1 factually incorrect statement about prior work: [Nguyen et al., 2023] did not derive any result about the spectral gap, contrary to what's claimed on lines 459-460.
  • The paper's message on lines 279-285 is not consistent with prior work, and with its own results.
  • The figures and tables presented are tiny.
Questions for authors
  • The notion of a perfect graph has already been well-established. What is the reasoning behind the attempt to redefine it in Definition 4.1, and then never use it again except for the paragraph right below it?
  • Why is Theorem 4.2 a paradox? Why does it prove that no sufficiently large graph can be perfect? (one of the 3 problematic claimed contributions)
  • Theorem 5.1 and 5.2 assume two different test and train distributions on the same set of vertices and measure their supposed difference in the loss function. How can you interpret this result so that it gives us any new insight on generalization? (one of the 3 problematic claimed contributions)
  • Over-smoothing happens when GNNs have many deep layers. Why does having a bound on the variation of a graph signal inform you about over-smoothing behaviors? (one of the 3 problematic claimed contributions)
  • See the weaknesses above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.