Structured Sheaf Learning of Consistent Connection Graphs

Gabriele D'Acunto; Leonardo Di Nino; Paolo Di Lorenzo; Sergio Barbarossa

REVIEW 3 major objections 6 minor 62 references

Structured Sheaf Learning of Consistent Connection Graphs

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proposes SCGL, an algorithm that jointly recovers the topology, node reference frames, and denoised signals of a consistent connection graph from noisy vector-valued observations, with provable convergence to stationary points…

desk verdict A solid and useful extension of structured graph learning to consistent connection graphs, with a real algorithm and good experiments, but the SO(n) parametrization quietly narrows the class of graphs the method can represent. read the letter →

arxiv 2608.08710 v1 pith:7FHM7WD7 submitted 2026-08-09 eess.SP

Leonardo Di Nino , Gabriele D'Acunto , Sergio Barbarossa , Paolo Di Lorenzo This is my paper · ORCID

classification eess.SP

keywords connectiongraphsLaplaciangraphlearningsheafsignalprocessingsynchronizationspectralconstraintsblock-coordinateoptimizationconsistency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes Structured Connection Graph Learning (SCGL), an algorithm that recovers a consistent connection graph—a graph whose edges carry orthogonal rotation matrices aligning vector-valued signals at nodes—from noisy observations. The central claim is that topology, node reference frames, and denoised signals can be estimated jointly by exploiting the spectral characterization of consistency: a consistent connection Laplacian equals a rotated copy of the combinatorial Laplacian tensored with the identity. If correct, this gives signal processing on manifolds and synchronization problems a learning method that controls the spectrum of the inferred operator while guaranteeing a nontrivial space of globally synchronized signals. The paper supports this with a convergence theorem and experiments showing better topology and geometry recovery than competing approaches.

What carries the argument

The load-bearing object is the identity $L = O^\top (\mathcal{L} \otimes I_n) O$, which ties a consistent connection Laplacian to a combinatorial Laplacian $\mathcal{L}$ with each eigenvalue repeated $n$ times. This identity converts the non-Euclidean edge-transport variables into node-wise frames $O_i \in SO(n)$, so the optimization in (P3) runs over $Z$ (denoised signals), $w$ (edge weights), $O$ (frames), and $U,\Lambda$ (eigenvectors and eigenvalues of the combinatorial Laplacian). The algorithm alternates an IIR-filter closed form for $Z$, a projected-gradient step for $w$ using adjoint operators of the Kronecker-structured Laplacian, a splitting-of-orthogonality-constraints closed form for $O$, a projection onto the special orthogonal group for $P$, an eigenproblem for $U$, and an isotonic-regression update for $\Lambda$, with convergence to KKT points guaranteed by block successive upper-bound minimization.

What would settle it

Generate vector-valued signals from a connection Laplacian with a single planted inconsistency—one cycle whose edge rotations compose to a rotation far from the identity—and check whether SCGL's recovered spectrum still equals the combinatorial spectrum with multiplicity $n$ and whether topology F1 stays high; the paper's account predicts a clear degradation as the planted inconsistency grows, and the absence of such a degradation would falsify the claim that consistency is what carries the recovery.

Watch

Extended reading notes

Core claim

The central discovery is that learning a connection graph from smooth signals becomes tractable once consistency is exploited as a parametrization rather than treated as a constraint enforced after the fact. Because a consistent connection Laplacian admits $L = O^\top (\mathcal{L} \otimes I_n) O$, the inverse problem separates into a combinatorial graph part (edge weights and their spectrum) and a geometric part (node frames), and both can be estimated jointly with the denoised signal. SCGL solves the resulting nonconvex problem by block descent, with provable convergence to stationary points; experiments show that it recovers topology and geometry more accurately than the baselines, denoises almost as well as the ground-truth filter, and produces a compression basis that beats the geometrically constructed VDM basis in the sparse-sampling regime.

Load-bearing premise

The method assumes the observed vector-valued signals are actually smooth low-frequency draws from a consistent connection graph—one whose edge rotations compose to the identity around every cycle—so that the equality $L = O^\top(\mathcal{L}\otimes I_n)O$ is exact; for non-consistent geometry such as a curved manifold, the parametrization is misspecified and the spectral guarantees do not apply.

Editorial extensions

If this is right

Users can directly impose spectral priors on the inferred operator, such as the number of connected components, because consistency transfers the combinatorial spectrum to the connection Laplacian; the RotatedMNIST experiment shows this by recovering exactly two components aligned with digit classes.
The learned Laplacian doubles as an IIR denoising filter whose performance approaches the theoretical bound set by the ground-truth operator even at low SNR.
The method reduces to standard structured graph learning when $n=1$, so it is a strict extension of spectral-constrained graph learning to vector-valued signals.
On curved domains where consistency is violated, the learned eigenbasis still yields the best compression–reconstruction trade-off among data-driven Laplacians, including the geometric VDM basis.
The parametrization reduces the number of geometric parameters from $V^2 n^2$ edge transports to $V n^2$ node frames, making the per-iteration cost $O(V^3 n^3)$ rather than the cost of a full semidefinite program.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension the paper does not run: use the residual of the spectral-fit term $\|\mathcal{L}(w) - U\Lambda U^\top\|_F^2$ as a per-edge or per-node inconsistency diagnostic, turning SCGL into a detector of where the consistency assumption fails.
The same block-coordinate template could be applied to non-orthogonal restriction maps or to time-varying reference frames, widening the method beyond connection graphs; the paper lists only relaxing consistency as future work.
The RotatedMNIST result suggests a quantitative experiment: measure the angular error between learned frames and the true image rotations as a function of sample size, which would expose how much of the clustering gain comes from the geometric prior versus the spectral constraint.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's the short version: this is a genuine contribution to sheaf-based graph learning, and the algorithm is worth a look. The main caveat is that the parametrization L = O^T (L ⊗ I) O with O in SO(n)^V does not cover all consistent connection graphs as defined, because Definition 3 allows O(n) transports. A tree with one reflection edge is consistent but has no SO(n) frame representation. So Theorem 1's third statement as quoted is false under their own definition, and the central claim of recovering 'the' consistent connection graph is overbroad. This doesn't kill the method for rotation-driven applications, but it should be stated.

What's genuinely new: the joint estimation of topology, orthogonal frames, and denoised signals from noisy vector-valued observations, with explicit spectral and topological priors. The derivation from the factor model to (P3) is clean, and the block-coordinate solver—closed-form Z, projected gradient w, SOC-based O update, Kabsch projection onto SO(n)—is credible and efficient. The experiments are well designed: they compare against SDP, SCOP, and their own prior SLGP, and the gains on random geometric graphs are consistent across sampling ratios. Code is available, which I count as real evidence.

Soft spots, in order of softness. First is the SO(n)/O(n) gap above, which I'd want fixed or explicitly circumscribed. Second, the convergence proof in Appendix C is a terse sketch; it leans on external BSUM results and a two-block extension for non-unique updates, and a referee will reasonably ask for details. Third, the experiments supply the number of connected components as prior knowledge, which helps the topology recovery. Fourth, the model-mismatch test on the sphere only evaluates compression, not topology or frame recovery, so the robustness claim is less supported than the rest.

Who is this for? Anyone working in graph learning from smooth signals, sheaf signal processing, or synchronization. It's a solid paper that would benefit from heavy revision, but it deserves referee time rather than a desk reject.

My recommendation: send it out, with the expectation of major comments targeting the SO(n)/O(n) mismatch and the convergence proof.

Referee Report

3 major / 6 minor

Summary. The paper introduces Structured Connection Graph Learning (SCGL), a block-coordinate algorithm that jointly estimates denoised signals, graph topology, and node reference frames for consistent connection graphs from noisy vector-valued observations. The formulation builds on the spectral characterization of consistent connection graphs, parametrizing the connection Laplacian as L = O^T (L ⊗ I_n) O with node frames in SO(n), and couples the spectrum of the learned connection Laplacian to that of a combinatorial Laplacian through a spectral penalty. The authors derive (P3) from a factor-analysis signal model, provide closed-form updates for each block (Z, w, O, P, U, Λ, B), claim convergence to stationary points of (P3), and report experiments on synthetic random graphs, a non-consistent sphere model, and RotatedMNIST showing improved topology, geometry, denoising, and compression over baselines. The paper also presents a complexity analysis and an initialization strategy based on covariance pseudoinverse fitting.

Significance. If the central claims hold, the paper would provide a useful extension of structured graph learning to connection graphs, with explicit spectral control and a guaranteed nontrivial global-section space. The derivation from the factor model to (P3) is clean, the algorithmic updates are well motivated and mostly closed-form, machine-checkable code is provided, and the experimental validation covers several regimes including a real-world-like task. The empirical gains over SDP and SLGP baselines, and the strong denoising performance, are credible and potentially valuable for sheaf-based signal processing. However, the significance is currently limited by two issues: the parametrization only covers a subclass of consistent connection graphs, and the convergence proof is not sufficiently rigorous to support the theorem as stated.

major comments (3)

[Sec. II-B (Theorem 1), Eq. (7), Sec. IV-A (P-update, Eq. (27))] The parametrization L = O^T (L ⊗ I_n) O in Eq. (7) and the subsequent problems (P2)-(P3) restrict the node frames O to SO(n), but Definition 3 allows edge transports in O(n). Theorem 1, statement 3, asserts that consistency is equivalent to existence of SO(n) frames with O_ij = O_i^T O_j; this is false for O(n) transports. For example, a two-node graph with a single edge and transport O_12 = diag(1,-1), n=2, is consistent (no cycles), and its connection Laplacian has spectrum {0,0,2,2}, matching the combinatorial Laplacian, yet no SO(2) frames can represent it because det(O_1^T O_2) = +1. Consequently, SCGL optimizes over a strict subset of consistent connection graphs and cannot recover reflection-type edge transports; the experiments sample frames from SO(2), so this restriction is not tested. The paper either needs to restrict Definition 3 and all claims to special-orthogonal connection graphs (which would exclude non-orientable geometries), or generalize the parametrization and the P-update to O(n) (e.g., by projecting onto O(n) without the determinant correction) and clearly state what class of consistent CGs the method can recover.
[Appendix C (proof of Theorem 2)] The convergence proof of Theorem 2 is stated at a very high level: it cites the BSUM result [48] and a 'two-block extension' from [32, Thm 7] without verifying the hypotheses for the six-block augmented Lagrangian in (15). In particular, the claim that the non-unique U- and P-updates are allowed because the blocks are 'mutually decoupled' is not sufficient, since the BSUM conditions require a specific structure of the objective and the updates. More importantly, the proof does not establish that the fixed-penalty augmented Lagrangian with dual update (31) converges to a feasible point with O^★ = P^★; the sentence 'The dual update (31) enforces O^★ = P^★' is asserted without a supporting argument. For a fixed ρ, ADMM-type schemes may converge to a point with nonzero primal residual unless additional assumptions hold. Since 'converges to stationary points of the resulting nonconvex problem' is a central advertised claim, the proof needs to be either completed with a rigorous verification of the KKT conditions for (P3), or the theorem must be restated with a weaker and explicitly justified convergence guarantee.
[Sec. V-E vs. Sec. III (Problem (P3))] The RotatedMNIST experiment uses a spectral prior encoding two connected components, and Fig. 5a reports success in recovering exactly two components. However, the formulation in (P3) enforces a single zero eigenvalue: Λ ∈ S_Λ with λ_1 = 0 and U ∈ St(V, V-1), which corresponds to a connected graph. The paper does not describe how the problem or the algorithm is modified for k > 1 connected components, nor does it state that the convergence theorem extends to that case. This leaves a gap between the theory, which is developed for the connected-graph setting, and one of the central real-world demonstrations. Please clarify the multi-component formulation and its convergence status, or otherwise reposition the claim.

minor comments (6)

[Sec. II-B, Definition 5] The indexing formula for d_j in Definition 5 and in Lemma 1 is typeset ambiguously; the intended expression d_j = -j + (j-1)(2V-j)/2 should be written with explicit parentheses to avoid confusion with the alternative reading d_j = -j + (j-1)/2 (2V-j).
[Sec. V-E, third paragraph] The number of equivariant features is stated as T = HJ^2, but the range h = 1, ..., H+1 gives (H+1)J^2 features; either the formula or the range is inconsistent and should be corrected.
[References] Reference [52] (Wax and Kailath) is cited with year 2003; the correct year is 1985.
[Sec. V-A, Fig. 2 caption] The caption for Fig. 2 has empty axis placeholders 'log10( )' for both α and β; the labels should be filled in.
[Sec. III, Eq. (9)] The definition of the rotated signal blocks \tilde{Z}_l in Eq. (9) is imprecise; the text should explicitly state that \tilde{Z}_l is the V × M matrix formed by the l-th coordinate of \tilde{Z} = OZ across all nodes.
[Sec. V-C] The noisy experiments in Fig. 3 do not report the values of α and β; given the sensitivity shown in Sec. V-A, the hyperparameter settings used for the noisy regime should be stated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the consistency parametrization is an explicit modeling assumption backed by an external theorem, and the reported recoveries are validated against external baselines and a mismatch scenario.

full rationale

The derivation chain is not circular. The paper assumes the observed signals are generated by a consistent connection graph and uses the external spectral characterization of [30] (Theorem 1) to parametrize L = O^T(L⊗I_n)O. This is an explicit modeling assumption, not a conclusion derived from the paper's own outputs. The learning problems (P1)-(P3) follow from a MAP/total-variation formulation and the parametrization, and the convergence result (Theorem 2) is proved via external BSUM results [48] and the two-block extension in [32]; no load-bearing step is justified only by the authors' prior work. The self-references [21] and [35] describe a baseline and a preliminary version, respectively, and are not used to force the central claim. The synthetic experiments do generate data from exactly the consistent-CG model assumed by the parametrization, which is a matched-model validation; however, this is not a case of a fitted parameter being renamed as a prediction, since the ground-truth topology and frames are not used to fit the model, and the paper also includes a non-consistent sphere/VDM mismatch experiment and comparisons to external baselines. A representational limitation exists (Theorem 1 is invoked with SO(n) frames while Definition 3 allows O(n) transports, so reflection-type consistent graphs are outside the parametrization), but this is a scope/correctness issue rather than a circular reduction of the paper's claims to its inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the consistency parametrization from external theorem [30], the standard factor-analysis signal model from [33], and the BSUM convergence framework. Hyperparameters alpha, beta, rho, and the spectral bounds are user-chosen; the number of connected components is a supplied prior in experiments. No new physical or mathematical entities are introduced.

free parameters (5)

alpha (sparsity penalty) = 0.0025 (fixed in Sec. V-B after cross-validation)
Balances edge sparsity against data fit; chosen by cross-validation in the ablation study.
beta (spectral penalty) = 60 (fixed in Sec. V-B after cross-validation)
Controls how strongly L(w) is forced to match U Lambda U^T; chosen by cross-validation.
rho (ADMM penalty) = not reported
Penalty for the splitting constraint O=P in (14); the paper does not specify its value or tuning procedure.
c1, c2 (spectral bounds) = not reported
Box constraints on the nonzero Laplacian eigenvalues in (29); values are not given.
Number of connected components = given as prior in experiments
In Sec. V-B and V-E the number of connected components is supplied to the algorithm, which is a strong structural prior.

assumptions (4)

domain assumption Theorem 1 (spectral characterization of consistent CGs)
The equivalence between consistency, spectrum multiplicity, and factorization through node frames is taken from [30] and is the foundation of the parametrization in (7).
domain assumption Factor-analysis signal model
The model x ~ N(0, L^dag + sigma^2 I) in Eq. (8) is borrowed from smooth graph learning [33]; it assumes the observed signals are low-frequency plus Gaussian noise.
standard math BSUM convergence framework
Theorem 2 relies on the block successive upper-bound minimization results of [48] and the two-block extension in [32].
domain assumption Feasibility of spectral constraints
Theorem 2 assumes the spectral constraints c1 <= lambda_2 <= ... <= lambda_V <= c2 are feasible; for disconnected or degenerate graphs this may fail.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Structured Sheaf Learning of Consistent Connection Graphs." pith.science (2026). https://pith.science/paper/7FHM7WD7

@misc{pith2026260808710,
  author       = {Pith},
  title        = {Pith review of: Structured Sheaf Learning of Consistent Connection Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FHM7WD7}},
  note         = {Machine review of arXiv:2608.08710}
}

read the original abstract

Connection graphs (CGs) extend classical graphs by associating vector-valued signals to nodes and orthogonal transport maps across edges, making them a natural model for synchronization and manifold-based signal processing. Despite their growing use, learning CGs directly from observations remains challenging because the network topology and the underlying geometric structure are coupled through non-Euclidean orthogonality constraints. In this work, we address this inverse problem by learning a consistent connection graph from noisy vector-valued signals. Exploiting the spectral characterization of consistent CGs, we formulate a structured learning problem that jointly estimates a denoised signal, the graph topology, and node-wise local reference frames. The proposed formulation couples the spectrum of the learned connection Laplacian to that of an underlying combinatorial Laplacian, enabling explicit spectral and topological priors while guaranteeing a nontrivial global-section space. We develop Structured Connection Graph Learning (SCGL), a block-coordinate algorithm that combines closed-form updates, manifold projections, and spectral constraints, and converges to stationary points of the resulting nonconvex problem. Numerical experiments show that SCGL improves topology and geometry recovery over competing approaches, while also yielding effective denoising and signal-compression bases.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 51 canonical work pages

[48]

A unified convergence analysis of block successive minimization methods for nonsmooth optimization,

M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence analysis of block successive minimization methods for nonsmooth optimization,”SIAM Journal on Optimization, vol. 23, no. 2, pp. 1126– 1153, 2013

2013
[1]

Goal-oriented and semantic communication in 6G AI-native networks: The 6G-GOALS approach,

E. C. Strinati, P. Di Lorenzoet al., “Goal-oriented and semantic communication in 6G AI-native networks: The 6G-GOALS approach,” in European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), 2024, pp. 1–6

work page 2024
[2]

Discrete signal processing on graphs,

A. Sandryhaila and J. M. Moura, “Discrete signal processing on graphs,” IEEE Trans. on Signal Processing, vol. 61, no. 7, pp. 1644–1656, 2013

work page 2013
[3]

Kernels and regularization on graphs,

A. J. Smola and R. Kondor, “Kernels and regularization on graphs,” in 16th Annual Conference on Learning Theory and 7th Kernel Workshop, Washington, DC, USA, August 24-27, 2003, pp. 144–158

work page 2003
[4]

Graph signal processing: History, development, impact, and outlook,

G. Leus, A. G. Marques, J. M. Moura, A. Ortega, and D. I. Shuman, “Graph signal processing: History, development, impact, and outlook,” IEEE Signal Processing Magazine, vol. 40, no. 4, pp. 49–60, 2023

2023
[5]

A comprehensive survey on graph neural networks,

Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and S. Y . Philip, “A comprehensive survey on graph neural networks,”IEEE transactions on neural networks and learning systems, vol. 32, no. 1, pp. 4–24, 2020

work page 2020
[6]

Network topology inference from spectral templates,

S. Segarra, A. G. Marques, G. Mateos, and A. Ribeiro, “Network topology inference from spectral templates,”IEEE Transactions on Signal and Information Processing over Networks, vol. 3, no. 3, pp. 467–483, 2017

work page 2017
[7]

Graph Topology Inference Based on Sparsifying Transform Learning,

S. Sardellitti, S. Barbarossa, and P. Di Lorenzo, “Graph Topology Inference Based on Sparsifying Transform Learning,”IEEE Transactions on Signal Processing, vol. 67, no. 7, pp. 1712–1727, 2019

work page 2019

Show all 62 references

[8]

Connecting the dots: Identifying network structure via graph signal processing,

G. Mateos, S. Segarra, A. G. Marques, and A. Ribeiro, “Connecting the dots: Identifying network structure via graph signal processing,”IEEE Signal Processing Magazine, vol. 36, no. 3, pp. 16–43, 2019

2019
[9]

Differentiable Graph Module (DGM) for Graph Convolutional Networks,

A. Kazi, L. Cosmo, S.-A. Ahmadi, N. Navab, and M. M. Bronstein, “Differentiable Graph Module (DGM) for Graph Convolutional Networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 2, p. 1606–1617, 2023

2023
[10]

A Survey on Over- smoothing in Graph Neural Networks,

T. K. Rusch, M. M. Bronstein, and S. Mishra, “A Survey on Over- smoothing in Graph Neural Networks,”SAM Research Report, vol. 2023, 2023

2023
[11]

Not too little, not too much: a theoretical analysis of graph (over) smoothing,

N. Keriven, “Not too little, not too much: a theoretical analysis of graph (over) smoothing,” inProceedings of the 36th International Conference on Neural Information Processing Systems, 2022, pp. 2268–2281

2022
[12]

Topological Signal Processing over Simplicial Complexes,

S. Barbarossa and S. Sardellitti, “Topological Signal Processing over Simplicial Complexes,”IEEE Transactions on Signal Processing, vol. 68, pp. 2992–3007, 2020

2020
[13]

Topological signal processing and learning: Recent advances and future challenges,

E. Isufi, G. Leus, B. Beferull-Lozano, S. Barbarossa, and P. Di Lorenzo, “Topological signal processing and learning: Recent advances and future challenges,”Signal Processing, p. 109930, 2025

2025
[14]

Position: Topological Deep Learning is the New Frontier for Relational Learning,

T. Papamarkouet al., “Position: Topological Deep Learning is the New Frontier for Relational Learning,” inProceedings of the 41st International Conference on Machine Learning, vol. 235, 2024, pp. 39 529–39 555

2024
[15]

J. M. Curry,Sheaves, cosheaves and applications. University of Pennsylvania, 2014

2014
[16]

Sheaf theory: from deep geometry to deep learning,

A. Ayzenberg, T. Gebhart, G. Magai, and G. Solomadin, “Sheaf theory: from deep geometry to deep learning,”preprint arXiv:2502.15476, 2025. 13

2025 arXiv
[17]

Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs,

C. Bodnar, F. D. Giovanni, B. P. Chamberlain, P. Li `o, and M. M. Bronstein, “Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs,” 2023

2023
[18]

Laplacians of Cellular Sheaves: Theory and Applications,

J. Hansen, “Laplacians of Cellular Sheaves: Theory and Applications,” Ph.D. dissertation, University of Pennsylvania, 2020

2020
[19]

Toward a spectral theory of cellular sheaves,

J. Hansen and R. Ghrist, “Toward a spectral theory of cellular sheaves,” Journal of Applied and Computational Topology, vol. 3, no. 4, pp. 315–358, 2019

2019
[20]

Learning Sheaf Laplacians from Smooth Signals,

——, “Learning Sheaf Laplacians from Smooth Signals,” inICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 5446–5450

2019
[21]

Learning Sheaf Laplacian Optimizing Restriction Maps,

L. Di Nino, S. Barbarossa, and P. Di Lorenzo, “Learning Sheaf Laplacian Optimizing Restriction Maps,” in2024 58th Asilomar Conference on Signals, Systems, and Computers, 2024, pp. 59–63

2024
[22]

Vector diffusion maps and the connection Laplacian,

A. Singer and H.-T. Wu, “Vector diffusion maps and the connection Laplacian,”Communications on pure and applied mathematics, vol. 65, no. 8, pp. 1067–1144, 2012

2012
[23]

A Cheeger inequality for the graph connection Laplacian,

A. S. Bandeira, A. Singer, and D. A. Spielman, “A Cheeger inequality for the graph connection Laplacian,”SIAM Journal on Matrix Analysis and Applications, vol. 34, no. 4, pp. 1611–1630, 2013

2013
[24]

The vector heat method,

N. Sharp, Y . Soliman, and K. Crane, “The vector heat method,”ACM Transactions on Graphics (TOG), vol. 38, no. 3, pp. 1–19, 2019

2019
[25]

Tangent bundle convolutional learning: from manifolds to cellular sheaves and back,

C. Battiloro, Z. Wang, H. Riess, P. Di Lorenzo, and A. Ribeiro, “Tangent bundle convolutional learning: from manifolds to cellular sheaves and back,”IEEE Transactions on Signal Processing, 2024

2024
[26]

Sheaf neural networks with connection laplacians,

F. Barbero, C. Bodnar, H. S. de Oc´ariz Borde, M. Bronstein, P. Veliˇckovi´c, and P. Li`o, “Sheaf neural networks with connection laplacians,” inTopol., Algebraic and Geometric Learning Wshp., 2022, pp. 28–36

2022
[27]

Convolutional neural networks on manifolds: From graphs and back,

Z. Wang, L. Ruiz, and A. Ribeiro, “Convolutional neural networks on manifolds: From graphs and back,” in2022 56th Asilomar Conference on Signals, Systems, and Computers, 2022, pp. 356–360

2022
[28]

Sheaf neural networks,

J. Hansen and T. Gebhart, “Sheaf neural networks,”NeurIPS 2020 Workshop TDA and Beyond, 2020

2020
[29]

Sheaf attention networks,

F. Barbero, C. Bodnar, H. S. de Oc´ariz Borde, and P. Lio, “Sheaf attention networks,” inNeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations, 2022

2022
[30]

Ranking and sparsifying a connection graph,

F. Chung, W. Zhao, and M. Kempton, “Ranking and sparsifying a connection graph,”Internet Mathematics, vol. 10, no. 1-2, pp. 87–115, 2014

2014
[31]

Bundle Neural Networks for message diffusion on graphs,

J. Bamberger, F. Barbero, X. Dong, and M. M. Bronstein, “Bundle Neural Networks for message diffusion on graphs,” inICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling

2024
[32]

A unified framework for structured graph learning via spectral constraints,

S. Kumar, J. Ying, J. V . d. M. Cardoso, and D. P. Palomar, “A unified framework for structured graph learning via spectral constraints,”Journal of Machine Learning Research, vol. 21, no. 22, pp. 1–60, 2020

2020
[33]

Learning Lapla- cian matrix in smooth graph signal representations,

X. Dong, D. Thanou, P. Frossard, and P. Vandergheynst, “Learning Lapla- cian matrix in smooth graph signal representations,”IEEE Transactions on Signal Processing, vol. 64, no. 23, pp. 6160–6173, 2016

2016
[34]

F. R. Chung,Spectral graph theory. American Mathematical Soc., 1997, vol. 92

1997
[35]

Learning the Structure of Connection Graphs,

L. Di Nino, G. D’Acunto, S. Barbarossa, and P. Di Lorenzo, “Learning the Structure of Connection Graphs,” inProc. of ICASSP, 2026, pp. 76–80

2026
[36]

Distributed optimization with sheaf ho- mological constraints,

J. Hansen and R. Ghrist, “Distributed optimization with sheaf ho- mological constraints,” in2019 57th annual allerton conference on communication, control, and computing (allerton), 2019, pp. 565–571

2019
[37]

Opinion dynamics on discourse sheaves,

——, “Opinion dynamics on discourse sheaves,”SIAM Journal on Applied Mathematics, vol. 81, no. 5, pp. 2033–2060, 2021

2021
[38]

Tackling Feature and Sample Heterogeneity in Decentralized Multi-Task Learning: A Sheaf- Theoretic Approach,

C. B. Issaid, P. Vepakomma, and M. Bennis, “Tackling Feature and Sample Heterogeneity in Decentralized Multi-Task Learning: A Sheaf- Theoretic Approach,”Transactions on Machine Learning Research
[39]

SheafAlign: A Sheaf-theoretic Framework for Decentralized Multimodal Alignment,

A. Ghalkha, Z. Tian, C. B. Issaid, and M. Bennis, “SheafAlign: A Sheaf-theoretic Framework for Decentralized Multimodal Alignment,” IEEE Communications Letters, 2026

2026
[40]

A Sheaf-Theoretic Framework for Distributed Multi- Site Channel Charting,

E. Grimaldi, L. Di Nino, M. E. Pandolfo, G. D’Acunto, S. Barbarossa, and P. D. Lorenzo, “A Sheaf-Theoretic Framework for Distributed Multi- Site Channel Charting,” 2026

2026
[41]

N. E. Steenrod,The topology of fibre bundles. Princeton university press, 1999, vol. 14

1999
[42]

D. J. Bartholomew, M. Knott, and I. Moustaki,Latent variable models and factor analysis: A unified approach. John Wiley & Sons, 2011

2011
[43]

How to learn a graph from smooth signals,

V . Kalofolias, “How to learn a graph from smooth signals,” inArtificial intelligence and statistics, 2016, pp. 920–929

2016
[44]

A splitting method for orthogonality constrained problems,

R. Lai and S. Osher, “A splitting method for orthogonality constrained problems,”Journal of Scientific Computing, vol. 58, no. 2, pp. 431–449, 2014

2014
[45]

Distributed optimization and statistical learning via the alternating direction method of multipliers,

S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Ecksteinet al., “Distributed optimization and statistical learning via the alternating direction method of multipliers,”Foundations and Trends® in Machine learning, vol. 3, no. 1, pp. 1–122, 2011

2011
[46]

Graph filters for signal processing and machine learning on graphs,

E. Isufi, F. Gama, D. I. Shuman, and S. Segarra, “Graph filters for signal processing and machine learning on graphs,”IEEE Transactions on Signal Processing, vol. 72, pp. 4745–4781, 2024

2024
[47]

Majorization-Minimization Algo- rithms in Signal Processing, Communications, and Machine Learning,

Y . Sun, P. Babu, and D. P. Palomar, “Majorization-Minimization Algo- rithms in Signal Processing, Communications, and Machine Learning,” IEEE Trans. on Signal Processing, vol. 65, no. 3, pp. 794–816, 2017

2017
[49]

Boyd and L

S. Boyd and L. Vandenberghe,Convex Optimization. Cambridge University Press, 2004

2004
[50]

A solution for the best rotation to relate two sets of vectors,

W. Kabsch, “A solution for the best rotation to relate two sets of vectors,” Foundations of Crystallography, vol. 32, no. 5, pp. 922–923, 1976

1976
[51]

Optimization algorithms on matrix manifolds

P.-A. Absil, R. Mahony, and R. Sepulchre, “Optimization algorithms on matrix manifolds.” Princeton University Press, 2009

2009
[52]

Detection of signals by information theoretic criteria,

M. Wax and T. Kailath, “Detection of signals by information theoretic criteria,”IEEE Transactions on acoustics, speech, and signal processing, vol. 33, no. 2, pp. 387–392, 2003

2003
[53]

Boumal,An introduction to optimization on smooth manifolds

N. Boumal,An introduction to optimization on smooth manifolds. Cambridge University Press, 2023

2023
[54]

The Fibonacci lattice,

R. P. Stanley, “The Fibonacci lattice,”The Fibonacci Quarterly, vol. 13, no. 3, pp. 215–232, 1975

1975
[55]

Diffusion by a random velocity field,

R. H. Kraichnan, “Diffusion by a random velocity field,”The physics of fluids, vol. 13, no. 1, pp. 22–31, 1970

1970
[56]

GSTools v1. 3: a toolbox for geostatistical modelling in Python,

S. M ¨uller, L. Sch ¨uler, A. Zech, and F. Heße, “GSTools v1. 3: a toolbox for geostatistical modelling in Python,”Geoscientific Model Development, vol. 15, no. 7, pp. 3161–3182, 2022

2022
[57]

Vetterli, J

M. Vetterli, J. Kova ˇcevi´c, and V . K. Goyal,Foundations of signal processing. Cambridge University Press, 2014

2014
[58]

An empirical evaluation of deep architectures on problems with many factors of variation,

H. Larochelle, D. Erhan, A. Courville, J. Bergstra, and Y . Bengio, “An empirical evaluation of deep architectures on problems with many factors of variation,” inProceedings of the 24th international conference on Machine learning, 2007, pp. 473–480

2007
[59]

Harmonic networks: Deep translation and rotation equivariance,

D. E. Worrall, S. J. Garbin, D. Turmukhambetov, and G. J. Brostow, “Harmonic networks: Deep translation and rotation equivariance,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5028–5037

2017
[60]

Learning steerable filters for rotation equivariant cnns,

M. Weiler, F. A. Hamprecht, and M. Storath, “Learning steerable filters for rotation equivariant cnns,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 849–858

2018
[61]

Graph Structure Learning with Inter- pretable Bayesian Neural Networks,

M. Wasserman and G. Mateos, “Graph Structure Learning with Inter- pretable Bayesian Neural Networks,”Transactions on machine learning research, 2024

2024
[62]

The design and use of steerable filters,

W. Freeman and E. Adelson, “The design and use of steerable filters,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 13, no. 9, pp. 891–906, 1991

1991

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [48]

A unified convergence analysis of block successive minimization methods for nonsmooth optimization,

M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence analysis of block successive minimization methods for nonsmooth optimization,”SIAM Journal on Optimization, vol. 23, no. 2, pp. 1126– 1153, 2013

2013

[2] [1]

Goal-oriented and semantic communication in 6G AI-native networks: The 6G-GOALS approach,

E. C. Strinati, P. Di Lorenzoet al., “Goal-oriented and semantic communication in 6G AI-native networks: The 6G-GOALS approach,” in European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), 2024, pp. 1–6

work page 2024

[3] [2]

Discrete signal processing on graphs,

A. Sandryhaila and J. M. Moura, “Discrete signal processing on graphs,” IEEE Trans. on Signal Processing, vol. 61, no. 7, pp. 1644–1656, 2013

work page 2013

[4] [3]

Kernels and regularization on graphs,

A. J. Smola and R. Kondor, “Kernels and regularization on graphs,” in 16th Annual Conference on Learning Theory and 7th Kernel Workshop, Washington, DC, USA, August 24-27, 2003, pp. 144–158

work page 2003

[5] [4]

Graph signal processing: History, development, impact, and outlook,

G. Leus, A. G. Marques, J. M. Moura, A. Ortega, and D. I. Shuman, “Graph signal processing: History, development, impact, and outlook,” IEEE Signal Processing Magazine, vol. 40, no. 4, pp. 49–60, 2023

2023

[6] [5]

A comprehensive survey on graph neural networks,

Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and S. Y . Philip, “A comprehensive survey on graph neural networks,”IEEE transactions on neural networks and learning systems, vol. 32, no. 1, pp. 4–24, 2020

work page 2020

[7] [6]

Network topology inference from spectral templates,

S. Segarra, A. G. Marques, G. Mateos, and A. Ribeiro, “Network topology inference from spectral templates,”IEEE Transactions on Signal and Information Processing over Networks, vol. 3, no. 3, pp. 467–483, 2017

work page 2017

[8] [7]

Graph Topology Inference Based on Sparsifying Transform Learning,

S. Sardellitti, S. Barbarossa, and P. Di Lorenzo, “Graph Topology Inference Based on Sparsifying Transform Learning,”IEEE Transactions on Signal Processing, vol. 67, no. 7, pp. 1712–1727, 2019

work page 2019