Pith. sign in

REVIEW 2 major objections 5 minor 44 references

Standard ridge regularization actively inverts recovered potential orderings in flow-inverse problems, while a gauge-invariant graph Dirichlet energy keeps the ordering intact and stable across a wide range of regularization strength.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Using the graph Dirichlet energy instead of an L2 penalty makes potential recovery from directed-flow divergence stable across four orders of magnitude in regularization strength, while ridge collapses and can reverse the ordering.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Real diagnosis and clean chain proofs, but the synthetic instrument's forward-model consistency is unproven and the 'every λ>0' claim overreaches. the 2 major comments →

arxiv 2607.13609 v1 pith:2MS2BZJ5 submitted 2026-07-15 cs.LG cs.IReess.SPstat.ML

Gauge-Invariant, Parameter-Insensitive Regularization for Potential Recovery from Flow on Directed Graphs

Mohammad Forouhesh This is my paper
classification cs.LG cs.IReess.SPstat.ML MSC 65N2105C5035R30
keywords inverse problemsgraph LaplacianTikhonov regularizationgauge invariancepotential recoverydirected graphsPoisson equationparameter sensitivity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that the default ridge (identity-seminorm) regularizer for recovering a latent potential from directed flow is not merely suboptimal but actively harmful: as its strength grows, it collapses the dynamic range and reverses the recovered ordering, so the only safe setting is no regularization at all. The authors identify the cause as a gauge mismatch—ridge shrinks toward an origin that is meaningless in a Dirichlet problem—and show that replacing it with the graph Dirichlet energy, which penalizes only differences between neighboring nodes, removes the hazard. The result is parameter-insensitivity: the recovered potential remains stable across four orders of magnitude in the regularization parameter, whereas ridge inverts the ordering for every positive setting. If true, this changes how practitioners regularize discrete inverse problems on graphs and connects the classical issue of gauge invariance to modern graph neural network oversmoothing.

Core claim

On a directed Poisson inverse problem with Dirichlet boundaries, the standard ridge penalty (R = I) imposes a preferred mean-zero gauge that fights the boundary conditions, compressing interior potentials toward the abandon sink and eventually reversing the planted ordering (rank correlation drops from +0.81 to about −0.42). The graph Dirichlet energy, defined via the conductance-weighted graph Laplacian as φᵀL_Gφ = Σ w_uv(φ_v − φ_u)², is gauge-invariant because L_G 1 = 0; it penalizes differences rather than amplitude. With this penalty, the recovered potential retains rank correlation +0.807 and NDCG@5 between 0.97 and 1.0 across λ ∈ [10⁻³, 10], while ridge collapses to negative correlatio

What carries the argument

The graph Dirichlet energy (the graph H¹ seminorm) — the quadratic form φᵀ L_G φ = Σ_{u,v} w_uv(φ_v − φ_u)², built from the conductance-weighted undirected Laplacian of the support graph. Because L_G 1 = 0, this penalty is flat along the constant (gauge) mode, so it does not bias the solution toward any origin. It replaces the identity ridge penalty and is the mechanism that delivers gauge invariance and parameter-insensitivity.

Load-bearing premise

The observed flow counts can be treated as conductances in a gradient-flow constitutive law, so the empirical divergence is exactly the Poisson right-hand side and a single scalar potential is the correct latent object; if real flows carry large non-gradient or solenoidal components, the recovered potential and the regularizer contrast are artifacts of the model, not the data.

What would settle it

Run the synthetic instrument with an added solenoidal flow component (e.g., adding a divergence-free circulation to the planted gradient flow) at fixed λ. If the graph-Sobolev estimate's rank correlation against the planted potential drops substantially while ridge's negative plateau remains, the central contrast depends on the gradient assumption. More directly, on any real dataset with a known ground-truth potential, check whether ridge's Spearman correlation crosses zero for small λ: if it stays positive, the inversion claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners solving discrete Poisson inverse problems on directed graphs should avoid ridge regularization; the graph Dirichlet energy is a parameter-insensitive alternative that preserves whatever ordering the data supports.
  • The Poisson residual — the mismatch between forward-predicted and empirical divergence — can reliably localize absorbing boundaries from flow data alone, even in multi-sink graphs where potential ranking fails.
  • The same gauge-invariance principle applies to deep directed graph neural networks: neutralizing the constant mode per layer prevents oversmoothing, holding performance flat as depth increases.
  • Longer chains help the Dirichlet-energy estimate retain dynamic range, whereas ridge collapses range independently of chain length.
  • The recovered gauge-invariant potential is a usable node feature for downstream prediction, adding predictive signal where the ridge-collapsed potential adds none.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The gauge diagnosis likely generalizes beyond this specific forward model: any regularizer whose null space does not contain the gauge mode (the constant vector) will bias the solution toward an arbitrary origin, so the contrast may appear in other inverse problems with translation-invariant forward operators.
  • Edge-preserving penalties that are also gauge-invariant (e.g., total variation or p-Laplacian) might yield the same parameter-insensitivity while additionally preserving sharp discontinuities; the paper mentions this as future work, and it is a natural testable extension.
  • The paper's claim that the topological-sort extraction outperforms the Hodge projection suggests that the acyclic support need not be physically motivated; a cheaper, data-driven ordering may be sufficient, which could simplify deployment on large graphs.
  • If the divergence field is dominated by solenoidal (non-gradient) flow, the entire potential-recovery framing may break down; a stress-test on synthetic flows with controlled curl would map the boundary of the method's validity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper studies regularized recovery of a latent potential from observed flow on a directed graph, formulated as a discrete Poisson inverse problem with Dirichlet boundaries. The authors claim that standard ridge (identity-seminorm Tikhonov) regularization does not merely degrade but actively inverts the recovered potential ordering for every positive regularization strength, whereas a gauge-invariant graph Dirichlet energy penalty preserves the ordering and renders the estimate parameter-insensitive across four orders of magnitude in λ. Theoretical results include gauge invariance of the penalty, an SPD reduced system under a path-to-boundary condition, and exact range preservation on chains. The paper validates the claims on a synthetic instrument with planted ground truth and on three clickstream corpora, and extends the gauge principle to a GNN oversmoothing application. The manuscript includes proofs, reproducible code, and an extraction ablation showing the main regularizer contrast is robust to the orientation method.

Significance. If the central claim holds, this is a significant and counterintuitive finding for graph inverse problems: the default regularizer can be not merely suboptimal but harmful, and a classical gauge-invariant penalty eliminates the pathology and the need to tune λ. The paper is well-organized, the theoretical statements are clearly separated from empirical evidence, and the released code supports reproducibility. The honest ablation showing that the Hodge projection is not load-bearing is a strength. However, the universal claim that ridge inverts for every λ>0 is not established and is contradicted by the paper's own real-data rank-agreement numbers, and the synthetic instrument's generative model is described too loosely to rule out misspecification. These issues currently weaken the central empirical contribution.

major comments (2)
  1. [Section 7, 'Synthetic Validation'] The generative process of the synthetic instrument is underspecified. The text says it 'plants a ground-truth potential, samples flow from it' and defines φ_true as the absorption probability of a funnel, but the estimator uses W_uv = F_uv and the constitutive law Eq. (1). If the flow is produced by simulating a Markov chain, the sampled F does not in general satisfy q_ij = W_ij(φ_j - φ_i) with W = F, so the Poisson solve is misspecified; the Spearman contrast (+0.81→-0.42) could then be an artifact of model mismatch. Please specify the generation mechanism and either verify that the data satisfy Lφ_true = b with W = F, or add an experiment with flow generated explicitly from Eq. (1) plus noise. This is load-bearing because the ridge-inversion claim rests entirely on this instrument.
  2. [Abstract; §5.7; Proposition 4; Table 5] The claim that ridge 'inverts the ordering for every λ>0' is not supported by the paper's own evidence. Proposition 4 only proves range collapse (Δ ≤ C/λ→0), not rank inversion. The inversion is demonstrated on one synthetic configuration (Table 1), and Table 5 contradicts a universal reading: on RetailRocket and Trivago at λ=1, ridge has positive rank agreement with the unregularized solve (1.000 and 0.833), i.e., it preserves the ordering there. The universal phrasing should be restricted to the synthetic instrument (or a class of graphs with stated conditions), or backed by a theorem. As written, the abstract and conclusion overclaim the generality of the main empirical result.
minor comments (5)
  1. [Abstract] The abstract states 'we prove the reduced solve is SPD' without the path-to-boundary condition required by Theorem 1. Please qualify the statement.
  2. [Throughout] The notation for the regularization strength is inconsistent: the abstract and some sections use λ while equations use λ_1. Unify.
  3. [§3] The sentence 'the empirical flow F_uv, used as the conductance W_uv = F_uv' is a strong modeling assumption. A brief justification or a caveat about when this is appropriate would help.
  4. [§8.4] The GNN extension is interesting but tangential. The phrase 'recovering PairNorm' overstates the connection; gauge-centering plus rescaling is essentially a PairNorm variant rather than a recovery from one principle.
  5. [§5.2] The penalty uses the symmetric Laplacian of the undirected support, while the forward operator is the directed Laplacian. The relation between the two null spaces (L1=0 vs. L_G 1=0) and the role of the boundary conditions could be clarified.

Circularity Check

0 steps flagged

No circularity: the regularizer contrast is an empirical result, not a definitional reduction.

full rationale

We walked the derivation chain: empirical flow F defines W and b (Section 3); the Poisson inverse problem L phi = b is a modeling assumption, not a consequence of the definitions of the regularizers. The gauge-invariance of the Dirichlet energy (Proposition 5, 'Since LG 1 = 0...') is a direct algebraic identity, not a self-fulfilling prediction. Theorem 1 (SPD) and Theorem 2 (range preservation on chains) are proved from the stated objectives and do not assume the target result. The headline claim — ridge inverts the recovered ordering while graph-Sobolev is stable — is an empirical observation on a synthetic instrument (Table 1, Figure 1) and is presented as such; Proposition 4 explicitly does not derive the inversion ('Proposition 4 does not claim the rank correlation tends to -1; empirically it settles to a negative plateau'). No parameter is fitted to the planted truth; lambda is swept and the claim is insensitivity. There are no self-citations and no uniqueness theorem imported from the authors' prior work. The strongest caveat — that setting W_uv = F_uv may not match the generative process of the synthetic funnel — is a model-misspecification/correctness risk, not a circular reduction: the recovered potential is not defined as the planted absorption probability. The paper's own scope statement ('The guarantee is preservation, not signal') is a limitation, not a circular step. Accordingly the circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

Central claims depend on: (a) the gradient-flow representation of clickstream flows, (b) acyclicity after extraction, (c) boundary labels defining the gauge, and (d) several hand-set extraction and scoring thresholds. No free parameter is fitted to the planted truth (lambda is swept), but the selected support and the 'determined interior' scoring set affect every reported correlation. No new physical or model entities are introduced; 'gauge' refers to the existing constant null mode of the Laplacian.

free parameters (7)
  • regularization strength lambda_1 = swept 1e-3 to 10; not fitted
    Swept in Table 1 and Figure 1; the parameter-insensitivity claim depends on stability over this range, not on a fitted value.
  • dominance threshold rho = 2
    Appendix A: 'All reported runs use rho=2'; filters reciprocal edges and shapes the acyclic support.
  • top-k pruning k = 10
    Appendix A: retain each node's k heaviest outgoing edges; affects support density.
  • edge retention threshold tau = per experiment (not tabulated)
    Equations (8)-(9) in Appendix A; trades support density against strictness; smaller graphs need lower tau.
  • well-visited state threshold = 100 visits
    Section 5.7: determined interior = states with >=100 visits and a path to boundary (n=95 of 277); scoring on this subset drives headline correlations.
  • Spearman rounding tolerance = numerical tolerance
    Section 5.7: estimates rounded so the saturated near-zero cluster ties deterministically; affects reported Spearman values.
  • penalty trace normalization = tr[(L_G)_II]
    Section 5.6: makes lambda comparable across graphs and enters Theorem 2's exact range formula.
axioms (4)
  • domain assumption Flows on the graph follow a gradient constitutive law q_ij = W_ij(phi_j - phi_i) with conductance W_uv = F_uv; the empirical divergence b is the corresponding Poisson right-hand side.
    Sections 3-4; this is the forward model that makes a scalar potential the right latent. Non-gradient or solenoidal traffic components would be unexplained and would bias recovery.
  • domain assumption After dominance-orientation, the retained support G_delta is acyclic and its transition operator A is nilpotent.
    Propositions 1-2; real session data can contain loops and reciprocal traffic; extraction must remove them for the Poisson setup and nilpotency certificate to apply.
  • ad hoc to paper The gauge is fixed by pinning conversion sinks at 1 and the abandon sink at 0, and the target potential is the harmonic absorption probability.
    Section 4 and Section 7; these boundary labels are application conventions, not forced by the mathematics; the ridge-vs-Sobolev contrast is evaluated relative to this gauge choice.
  • domain assumption The empirical divergence b is an unbiased (noisy) estimator of L phi.
    Section 4 and Appendix D; recovery quality depends on b being a good estimate; low-traffic states are excluded because this fails.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Gauge-Invariant, Parameter-Insensitive Regularization for Potential Recovery from Flow on Directed Graphs." pith.science (2026). https://pith.science/paper/2MS2BZJ5

@misc{pith2026260713609,
  author       = {Pith},
  title        = {Pith review of: Gauge-Invariant, Parameter-Insensitive Regularization for Potential Recovery from Flow on Directed Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MS2BZJ5}},
  note         = {Machine review of arXiv:2607.13609}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Recovering a latent potential from observed flow on a directed graph (a discrete Poisson problem with Dirichlet boundaries) is ill-posed, and the standard fix backfires: ridge regularization shrinks toward a gauge-meaningless origin, collapsing and reversing the recovered ordering ($+0.81\to-0.42$ rank correlation against a planted ground truth). The gauge-invariant graph Dirichlet energy removes the hazard and delivers parameter-insensitivity: the estimate is stable across four orders of magnitude in $\lambda$, whereas ridge inverts the ordering for every $\lambda>0$. We prove the reduced solve is SPD and preserves dynamic range exactly where ridge collapses it, and localize absorbing boundaries from flow alone via a Poisson residual. The $H^1$ seminorm is classical; what is new is the gauge diagnosis, the parameter-insensitivity it buys, and an ablation showing the result is robust to the extraction method. On three public clickstream corpora the gauge-invariant estimate retains $28$--$41\%$ of the interior dynamic range while ridge collapses to as little as $0.2\%$. The same gauge invariance carries into graph neural networks -- neutralizing the constant mode per layer prevents the oversmoothing that collapses a deep directed GCN -- linking this classical inverse problem to a central question in graph learning.

Figures

Figures reproduced from arXiv: 2607.13609 by Mohammad Forouhesh.

Figure 1
Figure 1. Figure 1: Recovery of the planted potential versus λ1. Graph-Sobolev (blue) preserves both rank and linear correlation across four orders of magnitude; Tikhonov (red) inverts the rank correlation to ≈ − 0.42 and collapses the linear correlation to zero. Determined interior, n = 95. 5.6 Trace normalization and cost To make λ1 comparable across graphs, we normalize the penalty block by its trace, λ1(LG)II 7→ λ1(LG)II … view at source ↗
Figure 2
Figure 2. Figure 2: Regularizer contrast on real data (λ1 = 1) versus the unregularized solve. Left: interior dynamic range retained. Right: rank agreement with the base solve. Ridge’s collapse deepens as the graph grows, reaching near-total loss on OTTO; graph-Sobolev preserves both throughout. Exact values in [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Recovered ϕˆ versus planted ϕtrue at λ1 = 1 (synthetic instrument). Graph-Sobolev (left) retains a positive relationship with the planted field; Tikhonov (right) collapses every interior state into a flat band near the abandon origin, the range collapse of Proposition 4. 0 10 2 10 1 10 0 10 1 regularization strength 1 0.25 0.50 0.75 1.00 1.25 1.50 1.75 interior d y n a mic ra n g e Chain saturation (L = 30… view at source ↗
Figure 4
Figure 4. Figure 4: Interior dynamic range ∆ϕˆ on a length-30 chain versus λ1. Graph-Sobolev stays near the unit harmonic target; Tikhonov overshoots at small λ1 then collapses toward zero [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Recovered graph-Sobolev potential (λ1 = 1) on RetailRocket and Trivago. Grey bars are pinned sinks; blue bars are recovered interior states. The ordering reproduces the known funnel; the negative interaction item image state on Trivago is discussed in Section 8.1. clicks states carts states 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0 re c o v ere d orders carts abandon OTTO: recovered potential by event type (105-… view at source ↗
Figure 6
Figure 6. Figure 6: OTTO recovered potential by event type (105-state multi-sink graph). Each point is an interior state; black bars are event-type means; dashed lines are the three Dirichlet sink levels. The solve separates carts states (mean +0.26, straddling the cart sink) from clicks states (mean −0.01, near abandon). 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references

  1. [1]

    1977 , publisher =

    Solutions of Ill-Posed Problems , author =. 1977 , publisher =

    1977

  2. [2]

    1998 , publisher =

    Rank-Deficient and Discrete Ill-Posed Problems: Numerical Aspects of Linear Inversion , author =. 1998 , publisher =

    1998

  3. [3]

    Physica D: Nonlinear Phenomena , volume =

    Nonlinear total variation based noise removal algorithms , author =. Physica D: Nonlinear Phenomena , volume =

  4. [4]

    1984 , publisher =

    Random Walks and Electric Networks , author =. 1984 , publisher =

    1984

  5. [5]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Random walks for image segmentation , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

  6. [6]

    Statistical ranking and combinatorial

    Jiang, Xiaoye and Lim, Lek-Heng and Yao, Yuan and Ye, Yinyu , journal =. Statistical ranking and combinatorial

  7. [7]

    Lim, Lek-Heng , journal =

  8. [8]

    and Benson, Austin R

    Schaub, Michael T. and Benson, Austin R. and Horn, Paul and Lippner, Gabor and Jadbabaie, Ali , journal =. Random walks on simplicial complexes and the normalized

  9. [9]

    IEEE Signal Processing Magazine , volume =

    The emerging field of signal processing on graphs , author =. IEEE Signal Processing Magazine , volume =

  10. [10]

    Belkin, Mikhail and Niyogi, Partha , journal =

  11. [11]

    Semi-supervised learning using

    Zhu, Xiaojin and Ghahramani, Zoubin and Lafferty, John , booktitle =. Semi-supervised learning using

  12. [12]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Learning with local and global consistency , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  13. [13]

    Journal of Machine Learning Research , volume =

    Trend filtering on graphs , author =. Journal of Machine Learning Research , volume =

  14. [14]

    Bodnar, Cristian and Di Giovanni, Francesco and Chamberlain, Benjamin P. and Li. Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  15. [15]

    Zhang, Xitong and He, Yixuan and Brugnone, Nathan and Perlmutter, Michael and Hirn, Matthew , journal =

  16. [16]

    He, Yixuan and Perlmutter, Michael and Reinert, Gesine and Cucuringu, Mihai , booktitle =

  17. [17]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , pages =

    Graph-based semi-supervised and active learning for edge flows , author =. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , pages =

  18. [18]

    IEEE Global Conference on Signal and Information Processing (GlobalSIP) , pages =

    Flow smoothing and denoising: Graph signal processing in the edge-space , author =. IEEE Global Conference on Signal and Information Processing (GlobalSIP) , pages =

  19. [19]

    Learning on Graphs Conference (LoG) , series =

    Edge directionality improves learning on heterophilic graphs , author =. Learning on Graphs Conference (LoG) , series =

  20. [20]

    IEEE Signal Processing Magazine , volume =

    Signal processing on directed graphs: The role of edge directionality when processing and learning from network data , author =. IEEE Signal Processing Magazine , volume =

  21. [21]

    Zhao, Lingxiao and Akoglu, Leman , booktitle =

  22. [22]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Dirichlet Energy Constrained Learning for Deep Graph Neural Networks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  23. [23]

    L aplacian eigenmaps for dimensionality reduction and data representation

    Mikhail Belkin and Partha Niyogi. L aplacian eigenmaps for dimensionality reduction and data representation. Neural Computation, 15 0 (6): 0 1373--1396, 2003

    2003

  24. [24]

    Chamberlain, Pietro Li \`o , and Michael M

    Cristian Bodnar, Francesco Di Giovanni, Benjamin P. Chamberlain, Pietro Li \`o , and Michael M. Bronstein. Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in GNN s. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 18527--18541, 2022

    2022

  25. [25]

    Doyle and J

    Peter G. Doyle and J. Laurie Snell. Random Walks and Electric Networks. Carus Mathematical Monographs. Mathematical Association of America, 1984

    1984

  26. [26]

    Random walks for image segmentation

    Leo Grady. Random walks for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28 0 (11): 0 1768--1783, 2006

    2006

  27. [27]

    Rank-Deficient and Discrete Ill-Posed Problems: Numerical Aspects of Linear Inversion

    Per Christian Hansen. Rank-Deficient and Discrete Ill-Posed Problems: Numerical Aspects of Linear Inversion. SIAM, Philadelphia, 1998

    1998

  28. [28]

    MSGNN : A spectral graph neural network based on a novel magnetic signed L aplacian

    Yixuan He, Michael Perlmutter, Gesine Reinert, and Mihai Cucuringu. MSGNN : A spectral graph neural network based on a novel magnetic signed L aplacian. In Learning on Graphs Conference (LoG), volume 198 of PMLR, 2022

    2022

  29. [29]

    Schaub, Santiago Segarra, and Austin R

    Junteng Jia, Michael T. Schaub, Santiago Segarra, and Austin R. Benson. Graph-based semi-supervised and active learning for edge flows. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 761--771, 2019

    2019

  30. [30]

    Statistical ranking and combinatorial H odge theory

    Xiaoye Jiang, Lek-Heng Lim, Yuan Yao, and Yinyu Ye. Statistical ranking and combinatorial H odge theory. Mathematical Programming, 127 0 (1): 0 203--244, 2011

    2011

  31. [31]

    H odge L aplacians on graphs

    Lek-Heng Lim. H odge L aplacians on graphs. SIAM Review, 62 0 (3): 0 685--715, 2020

    2020

  32. [32]

    Marques, Santiago Segarra, and Gonzalo Mateos

    Antonio G. Marques, Santiago Segarra, and Gonzalo Mateos. Signal processing on directed graphs: The role of edge directionality when processing and learning from network data. IEEE Signal Processing Magazine, 37 0 (6): 0 99--116, 2020

    2020

  33. [33]

    Bronstein

    Emanuele Rossi, Bertrand Charpentier, Francesco Di Giovanni, Fabrizio Frasca, Stephan G \"u nnemann, and Michael M. Bronstein. Edge directionality improves learning on heterophilic graphs. In Learning on Graphs Conference (LoG), volume 231 of PMLR, 2023

    2023

  34. [34]

    Rudin, Stanley Osher, and Emad Fatemi

    Leonid I. Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: Nonlinear Phenomena, 60 0 (1--4): 0 259--268, 1992

    1992

  35. [35]

    Schaub and Santiago Segarra

    Michael T. Schaub and Santiago Segarra. Flow smoothing and denoising: Graph signal processing in the edge-space. In IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 735--739, 2018

    2018

  36. [36]

    Schaub, Austin R

    Michael T. Schaub, Austin R. Benson, Paul Horn, Gabor Lippner, and Ali Jadbabaie. Random walks on simplicial complexes and the normalized H odge 1- L aplacian. SIAM Review, 62 0 (2): 0 353--391, 2020

    2020

  37. [37]

    Shuman, Sunil K

    David I. Shuman, Sunil K. Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of signal processing on graphs. IEEE Signal Processing Magazine, 30 0 (3): 0 83--98, 2013

    2013

  38. [38]

    Tikhonov and Vasiliy Y

    Andrey N. Tikhonov and Vasiliy Y. Arsenin. Solutions of Ill-Posed Problems. Winston & Sons, Washington, D.C., 1977

    1977

  39. [39]

    Smola, and Ryan J

    Yu-Xiang Wang, James Sharpnack, Alexander J. Smola, and Ryan J. Tibshirani. Trend filtering on graphs. Journal of Machine Learning Research, 17 0 (105): 0 1--41, 2016

    2016

  40. [40]

    M ag N et: A neural network for directed graphs

    Xitong Zhang, Yixuan He, Nathan Brugnone, Michael Perlmutter, and Matthew Hirn. M ag N et: A neural network for directed graphs. Advances in Neural Information Processing Systems (NeurIPS), 34: 0 27003--27015, 2021

    2021

  41. [41]

    PairNorm : Tackling oversmoothing in GNN s

    Lingxiao Zhao and Leman Akoglu. PairNorm : Tackling oversmoothing in GNN s. In International Conference on Learning Representations (ICLR), 2020

    2020

  42. [42]

    Lal, Jason Weston, and Bernhard Sch \"o lkopf

    Dengyong Zhou, Olivier Bousquet, Thomas N. Lal, Jason Weston, and Bernhard Sch \"o lkopf. Learning with local and global consistency. In Advances in Neural Information Processing Systems (NeurIPS), volume 16, pages 321--328, 2004

    2004

  43. [43]

    Dirichlet energy constrained learning for deep graph neural networks

    Kaixiong Zhou, Xiao Huang, Daochen Zha, Rui Chen, Li Li, Soo-Hyun Choi, and Xia Hu. Dirichlet energy constrained learning for deep graph neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  44. [44]

    Semi-supervised learning using G aussian fields and harmonic functions

    Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty. Semi-supervised learning using G aussian fields and harmonic functions. In Proceedings of the 20th International Conference on Machine Learning (ICML), pages 912--919, 2003

    2003

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.