Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Joint estimation of smooth graph signals from partial linear measurements

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A smoothness-penalized least squares estimator can jointly recover graph signals even when each vertex is observed through a single coordinate or not at all.

desk verdict A solid, workmanlike extension of Araya et al. to general graphs, built on a genuinely useful eigenvalue lemma; main results look right, but Corollary 3(2) has an exponent typo that must be fixed before the star-graph consistency claim is publishable. read the letter →

arxiv 2505.23240 v2 pith:6GIH5DPU submitted 2025-05-29 math.ST cs.LGstat.MLstat.TH

Hemant Tyagi This is my paper · ORCID
classification math.STcs.LGstat.MLstat.TH MSC 62J0762G0505C50
keywords TikhonovregularizationLaplaciansmoothingsemi-supervisedlearningongraphsmulti-layerrankingmultitaskgraphsignalestimationsparserandommeasurementsErdős-Rényi
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 studies the problem of recovering $T$ latent vectors $x_t \in \mathbb{R}^n$, one attached to each vertex of a connected graph $G$, from noisy linear measurements $y_t = C_t x_t + \eta_t$ where each $C_t$ may be a single row vector or even zero. It proves that the smoothness-penalized least squares estimator, which penalizes the quadratic variation of the estimates across the graph, provides a high-probability bound on the mean squared error whenever the pooled Gram matrix $\lambda_{\min}(\sum_t C_t^\top C_t)$ is positive. The concrete payoff is weak consistency: for complete and star graphs, the normalized mean squared error tends to zero as $T$ grows, even when only one coordinate is measured per sampled vertex and only a vanishingly small fraction of vertices are sampled. The same machinery extends to a multi-layer ranking problem where each layer supplies noisy pairwise differences, showing that many sparse, disconnected comparison graphs can be pooled together for consistent recovery.

What carries the argument

Lemma 1, a new lower bound on $\lambda_{\min}(C^\top C + \mu M^\top M)$. It decomposes any unit vector into a component constant across vertices (the nullspace of $M$) and an orthogonal component, then lower bounds the quadratic form by a max of a term proportional to $\mu$ times the Fiedler eigenvalue $\lambda_{T-1}$ and a term proportional to $\lambda_{\min}(\sum_t C_t^\top C_t)/T$. This spectral gap converts the smoothness penalty into a guaranteed invertibility of the regularized system, and then the bias–variance analysis of Theorem 1 proceeds along standard lines. The proof also shows that graphs with large Fiedler eigenvalue, such as complete and star graphs, are exactly the cases where the bound remains useful under rank-deficient measurements.

What would settle it

Simulate with $n=2$, a complete graph $G$, and every $C_t = e_1^\top$, so that only the first coordinate is ever measured and $\sum_t C_t^\top C_t$ is singular. If the theorem's identifiability condition is genuinely necessary, the MSE for the second coordinate should fail to vanish as $T$ grows. Conversely, running the random basis model with $\theta = c/\sqrt{T}$ for small constant $c$, below the stated $\omega(1/\sqrt{T})$ rate, and checking whether the normalized MSE still vanishes would test the tightness of the sampling threshold.

Watch

Extended reading notes

Core claim

The central claim is a non-asymptotic mean squared error bound for the estimator (1.4), valid even when every $C_t$ is rank deficient. The bound is governed by the smallest eigenvalue of $C^\top C + \mu M^\top M$, where $M$ is the incidence matrix of $G$; the key novel step is Lemma 1, which lower bounds this eigenvalue using the Fiedler eigenvalue of $G$ and the smallest eigenvalue of the pooled Gram matrix $\sum_t C_t^\top C_t$. Under the random measurement model where each row is a uniformly chosen coordinate vector with probability $\theta$, the bound implies that the normalized MSE vanishes as $T \to \infty$ when $\theta = \omega(1/\sqrt{T})$, with $S_T = o(T^2)$ for complete graphs and $S_T = o(T)$ for star graphs. The extension to multi-layer translation synchronization replaces $\lambda_{\min}$ by $\lambda_{n-1}$, so individual measurement graphs may be disconnected; with Erdős–Rényi layers, consistency holds once the cumulative edge probability $p_{\mathrm{sum}}$ grows sufficiently fast, even when each layer is below the connectivity threshold.

Load-bearing premise

The load-bearing premise is that the measurements, pooled over all vertices, collectively span all $n$ signal directions, i.e., $\lambda_{\min}(\sum_t C_t^\top C_t)>0$ (or $\lambda_{n-1}>0$ after centering in the ranking problem); if all measurements are confined to a proper subspace, the smoothness penalty cannot recover the missing coordinates.

Editorial extensions

If this is right

  • Consistent recovery is possible when only one coordinate is sampled per vertex: Corollary 3 shows the normalized MSE goes to zero provided a fraction $\theta = \omega(1/\sqrt{T})$ of vertices are sampled, so roughly $\omega(\sqrt{T})$ vertices suffice.
  • For complete graphs the condition on smoothness is $S_T = o(T^2)$, and for star graphs it is $S_T = o(T)$; the smoother the signal, the sparser the sampling can be.
  • In the multi-layer ranking problem, each layer's measurement graph may be an Erdős–Rényi graph below the connectivity threshold; consistency holds when the sum $p_{\mathrm{sum}}$ of edge probabilities grows fast enough, showing that information pooling across layers replaces connectivity within a single layer.
  • For path graphs, the approach yields meaningful bounds only when each $C_t$ is full rank or each measurement graph $G_t$ is connected, because the Fiedler eigenvalue of a path scales like $1/T^2$.
  • The eigenvalue lower bound in Lemma 1 is stated as potentially useful beyond this paper, since it resolves a conjectured form of lower bound for a related Gaussian-design penalized regression problem.

Reading between the lines

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

  • The paper does not derive a matching lower bound, so the $\theta = \omega(1/\sqrt{T})$ threshold is an upper-bound artifact; a minimax analysis could show either that considerably sparser sampling suffices or that this rate is essentially necessary.
  • The Lemma 1 mechanism suggests a clean interpolation between regimes: graphs whose Fiedler eigenvalue grows with $T$ inherit the complete- and star-graph consistency results, while graphs with vanishingly small Fiedler eigenvalue behave like the path case. This could be tested by running the estimator on a family of expander graphs with varying spectral gap.
  • For the multi-layer ranking problem, a natural extension would be to allow layers to arrive sequentially and to track time-varying item strengths; the spectral-gap view here suggests consistency should persist as long as cumulative pairwise information grows, which would connect the result to online and streaming ranking settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper studies joint estimation of latent graph signals x_t in R^n attached to the vertices of a connected graph G, from noisy partial linear measurements y_t = C_t x_t + η_t. Under a quadratic smoothness assumption on the signals (1.2), the authors analyze the smoothness-penalized least squares estimator (1.3)-(1.4). The main result, Theorem 1, gives a high-probability bound on the squared estimation error, expressed in terms of a lower bound λ(μ) on the smallest eigenvalue of μ M^T M + C^T C (Lemma 1), which is the key technical novelty. Corollaries 1-3 instantiate Theorem 1 for complete and star graphs under a sparse random measurement model (Proposition 2), yielding weak-consistency regimes when the sampling probability θ = ω(1/√T) and the smoothness budget S_T grows sublinearly in the number of edges. Section 3 extends the framework to multi-layer translation synchronization, where each layer t provides noisy pairwise comparisons through a possibly disconnected graph G_t; Theorem 2 and Corollary 4 establish analogous error bounds and weak consistency when the aggregate Gram matrix is well conditioned. The proofs of Theorem 1 and Lemma 1 are detailed, and the paper includes numerical simulations illustrating the behavior of the MSE for complete and star graphs.

Significance. If correct, the results constitute a useful theoretical contribution: they show that smoothness-penalized least squares can consistently recover signals from very partial measurements, including at most one coordinate per vertex on a vanishingly small fraction of vertices, and they extend prior dynamic ranking results from connected per-layer graphs to sparse, disconnected layers. The central eigenvalue lower bound (Lemma 1) is non-trivial and potentially of independent interest, and the explicit scaling conditions in Corollaries 3 and 4 are concrete and falsifiable. The proofs of Theorem 1, Lemma 1, and Proposition 2 are structurally careful, with constants tracked through the variance and bias bounds. The main weakness is a load-bearing typographical error in the star-graph arm of Corollary 3 that, as printed, invalidates the stated weak-consistency claim for star graphs; the complete-graph arm appears consistent, and the error is locally fixable by correcting the exponent and supplying the missing derivation.

major comments (2)
  1. [Corollary 3(2), Remark 1(2)] The regularization parameter choice in Corollary 3(2) for the star graph has the wrong exponent of S_T. The stated formula is μ* = max{σ^{2/3} θ^{1/3} T^{1/3} / S_T^{2/3} − θ/n, c1}, but substituting the Proposition 2 bounds λmin(O_T^T O_T) ≍ θT/n into the star-graph minimizer derived in Appendix D.2 gives μ* = max{σ^{2/3} θ^{1/3} T^{1/3} / S_T^{1/3} − θ/n, c1}. This is not a harmless typo: for n=1, σ=1, S_T=√T, θ=T^{-0.4}, the printed first term tends to 0, forcing μ* = c1, while the correct first term grows as T^{1/30}. With μ = c1, Lemma 1 places λ(μ) in the first regime (λ(μ) ≍ μθ), and the variance term in Theorem 1 is of order T, so the normalized MSE does not vanish. Therefore the bound and weak-consistency conclusion stated in Corollary 3(2) and Remark 1(2) are not valid as printed for the specified parameter choice; the star-graph claim requires the corrected exponent and a written derivation confirming the resulting bounds.
  2. [Corollary 3 (proof omitted)] The proof of Corollary 3 is omitted with the remark that it follows straightforwardly from Corollaries 1-2 and Proposition 2. Given that this corollary carries the paper's headline weak-consistency claims and that the star-graph arm contains an erroneous exponent, the derivation should be supplied in the revision, at least in sketch form, showing explicitly how the Proposition 2 bounds are substituted into the Corollary 1 and Corollary 2 bounds and how the stated μ* choices lead to the displayed error bounds. An explicit derivation would also clarify the role of the constants and would prevent similar exponent errors in the statement.
minor comments (5)
  1. [Appendix D.2] The parenthetical justification for dropping the term 1/(λmin/T + μT)^2 in (D.4) is incorrect as written: the inequality 'λmin/T + μ ≤ λmin/√T + μT^{3/2}' does not imply that the term can be bounded by T/(λmin/T + μ)^2; the needed inequality is (λmin/T + μT)^2 ≥ (λmin/T + μ)^2/T, which does hold for T ≥ 1.
  2. [Corollary 1] The notation in Corollary 1 is confusing: λmin(O_T^T O_T) and λmax(O_T^T O_T) are used both for the eigenvalues of the matrix and for the lower and upper bounds on those eigenvalues. Please introduce separate symbols (e.g., underlined or barred quantities) for the bounds.
  3. [Section 2.3, after (2.3)] The sentence 'The first term in (2.3) is the bias error which increases with μ, while the second term is the variance error which decreases with μ' is slightly loose because the sum in the second term involves λ(μ), which also depends on μ; a brief clarification that the statement holds for fixed graph and measurement parameters would help.
  4. [Remark 1] In Remark 1, the normalized MSE for the complete graph is obtained by dividing the Corollary 3(1) bound by T; the text jumps directly to the normalized display. Stating this normalization explicitly would improve readability.
  5. [Section 5] The simulation section reports MSE averaged over 50 Monte-Carlo runs but does not show error bars or confidence bands; given the small number of runs, adding such bands would make the empirical claims more convincing.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the main bounds are derived from explicit eigenvalue lower bounds with regularization selected from the derived MSE expression; prior self-citations are baselines only.

full rationale

The paper's central derivation is self-contained. Theorem 1 follows from Lemma 1, which lower-bounds the smallest eigenvalue of C^T C + mu M^T M using explicit quantities b1, b2, b3, and from standard bias and variance decompositions (Lemmas 2 and 3). The corollaries then choose mu* by minimizing the displayed convex function f(mu) obtained from the derived bound; this is an analytic optimization over a derived expression, not a fit to the data or to the target MSE. Proposition 2 is proved from an external Matrix-Chernoff bound (Tropp), and Proposition 4 uses an external concentration result (Paul and Chen). No uniqueness theorem is invoked, and no modeling assumption is imported from the prior self-cited work: Araya, Karlé and Tyagi (2023) is used only as a baseline for comparison or as a reference for path-graph calculations, and its results are not inputs to the derivation of Theorem 1 or Corollaries 3 and 4. The smoothness class constraint ||M x||_2^2 <= S_T is an assumption, not a consequence of the bound, and the bound is stated uniformly over that class. The possible exponent issue in Corollary 3(2) (star graph) is a correctness or proof-checking concern about the unconstrained minimizer, not a circularity; it does not make the derivation equivalent to its inputs. The manuscript also flags its own limitations, e.g. Remark 1 on path graphs, rather than hiding them. Accordingly, no reduction by construction, no fitted input renamed as a prediction, and no load-bearing self-citation chain is present, so the appropriate circularity score is 0.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

All free parameters and assumptions are listed. The main tuning parameter µ is theoretical; the paper does not provide a data-driven selection method. No new entities are introduced.

free parameters (1)
  • Regularization parameter µ = µ* as defined in Corollaries 1 to 4; depends on S_T, σ, n, T, graph, and sampling parameters
    The theorem holds for any µ>0; the corollaries pick µ* to minimize the derived bias-variance bound. No data-driven selection rule is given, so the rates are conditional on knowing S_T and σ.
assumptions (7)
  • domain assumption Signals are smooth: ∥Mx∥₂² ≤ S_T (equation 1.2)
    Defines the class of signals over which the risk bounds are uniform.
  • domain assumption Noise vectors are centered and σ-subgaussian
    Used in the quadratic-form concentration bound of Hsu et al. in Lemma 3.
  • domain assumption G is connected
    Guarantees λ_{T-1}>0, which is needed for the constant b1 in Lemma 1.
  • domain assumption Aggregate measurement matrix O_T has full column rank (or λ_{n-1}(O_T^T O_T)>0 for Theorem 2)
    Needed for the matrix C^TC + µM^TM to be positive definite in the relevant subspace; without it the problem is not identifiable.
  • domain assumption For multi-layer ranking, each x_t is centered and the G_t are independent Erdős-Rényi graphs
    Centering handles the global shift identifiability; the ER model is used in Proposition 4.
  • standard math External concentration bounds: Matrix Chernoff, Hsu et al., Hoeffding, Weyl, Gershgorin
    Quoted from the literature and used in Propositions 2 and 4 and in Lemma 3.
  • standard math Paul and Chen (2020b) Theorem 1 on concentration of sums of adjacency matrices
    Used in Proposition 4 part 1 to bound the spectral norm of the summed adjacency matrices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Joint estimation of smooth graph signals from partial linear measurements." pith.science (2026). https://pith.science/paper/6GIH5DPU

@misc{pith2026250523240,
  author       = {Pith},
  title        = {Pith review of: Joint estimation of smooth graph signals from partial linear measurements},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6GIH5DPU}},
  note         = {Machine review of arXiv:2505.23240}
}
abstract

Given an undirected and connected graph $G$ on $T$ vertices, suppose each vertex $t$ has a latent signal $x_t \in \mathbb{R}^n$ associated to it. Given partial linear measurements of the signals, for a potentially small subset of the vertices, our goal is to estimate $x_t$'s. Assuming that the signals are smooth w.r.t $G$, in the sense that the quadratic variation of the signals over the graph is small, we obtain non-asymptotic bounds on the mean squared error for jointly recovering $x_t$'s, for the smoothness penalized least squares estimator. In particular, this implies for certain choices of $G$ that this estimator is weakly consistent (as $T \rightarrow \infty$) under potentially very stringent sampling, where only one coordinate is measured per vertex for a vanishingly small fraction of the vertices. The results are extended to a ``multi-layer'' ranking problem where $x_t$ corresponds to the latent strengths of a collection of $n$ items, and noisy pairwise difference measurements are obtained at each ``layer'' $t$ via a measurement graph $G_t$. Weak consistency is established for certain choices of $G$ even when the individual $G_t$'s are very sparse and disconnected.

Figures

Figures reproduced from arXiv: 2505.23240 by the authors.

Figure 1
Figure 1. MSE versus T for star graph (top row) and complete graph (bottom row), for fixed θ ∈ {0.2, 0.5} with Ct generated as in Proposition 2. Here, xb is obtained as a solution of (2.1). We set n = 5 and σ = 1. The MSE is averaged over 50 Monte Carlo trials. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. MSE versus T for star graph (top row) and complete graph (bottom row), for the multi-layer translation synchronization model. Here, Gt are independent Erd¨os-Renyi graphs with pt = p for all t, where p ∈ {0.02, 0.004}. Also, n = 50 and σ = 1. The estimate xb is found using (3.3). The MSE is averaged over 50 Monte Carlo trials. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Joint learning of a network of linear dynamical systems via total variation penalization

    math.ST 2025-11 conditional novelty 6.0 of 10

    TV-penalized joint least squares estimates the matrices of m related linear dynamical systems on a graph with MSE bounds that vanish as m grows, even at constant trajectory length T.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Dynamic ranking and translation synchronization

    Ernesto Araya, Eglantine Karlé, and Hemant Tyagi. Dynamic ranking and translation synchronization . Information and Inference: A Journal of the IMA, 12 0 (3): 0 2224--2266, 2023

    work page 2023

  2. [2]

    Belkin, I

    M. Belkin, I. Matveeva, and P. Niyogi. Regularization and semi-supervised learning on large graphs. In Learning Theory, pages 624--638, 2004

    work page 2004

  3. [3]

    Concentration Inequalities: A Nonasymptotic Theory of Independence

    Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, 2013

    work page 2013

  4. [4]

    How fine-tuning allows for effective meta-learning

    Kurtland Chua, Qi Lei, and Jason D Lee. How fine-tuning allows for effective meta-learning. In Advances in Neural Information Processing Systems, volume 34, pages 8871--8884, 2021

    work page 2021

  5. [5]

    Cormen, Charles E

    Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms, Third Edition. The MIT Press, 3rd edition, 2009

    2009

  6. [6]

    Dalalyan, Mohamed Hebiri, and Johannes Lederer

    Arnak S. Dalalyan, Mohamed Hebiri, and Johannes Lederer. On the prediction performance of the lasso. Bernoulli, 23 0 (1): 0 552--581, 2017

    work page 2017

  7. [7]

    On consistency of graph-based semi-supervised learning

    Chengan Du, Yunpeng Zhao, and Feng Wang. On consistency of graph-based semi-supervised learning. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 483--491, 2019

    work page 2019

  8. [8]

    Du, Wei Hu, Sham M

    Simon S. Du, Wei Hu, Sham M. Kakade, Jason D. Lee, and Qi Lei. Few-shot learning via learning the representation, provably. arxiv:2002.09434, 2021

    arXiv 2002

Show all 39 references
  1. [9]

    Adaptive and robust multi-task learning

    Yaqi Duan and Kaizheng Wang. Adaptive and robust multi-task learning . The Annals of Statistics, 51 0 (5): 0 2015 -- 2039, 2023

    2015

  2. [10]

    Minimax optimal regression over sobolev spaces via laplacian regularization on neighborhood graphs

    Alden Green, Sivaraman Balakrishnan, and Ryan Tibshirani. Minimax optimal regression over sobolev spaces via laplacian regularization on neighborhood graphs. In Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, pages 2602--2610, 2021

    2021

  3. [11]

    van de Geer

    Mohamed Hebiri and Sara A. van de Geer. The smooth-lasso and other _1 + _2 -penalized methods. Electronic Journal of Statistics, 5: 0 1184--1226, 2011

    2011

  4. [12]

    Horn and Charles R

    Roger A. Horn and Charles R. Johnson. Matrix Analysis. Cambridge University Press, 1985

    1985

  5. [13]

    A tail inequality for quadratic forms of subgaussian random vectors

    Daniel Hsu, Sham Kakade, and Tong Zhang. A tail inequality for quadratic forms of subgaussian random vectors. Electronic Communications in Probability, 17: 0 1--6, 2012

    2012

  6. [14]

    Translation synchronization via truncated least squares

    Xiangru Huang, Zhenxiao Liang, Chandrajit Bajaj, and Qixing Huang. Translation synchronization via truncated least squares. Advances in neural information processing systems, 30, 2017

    2017

  7. [15]

    Optimal rates for total variation denoising

    Jan-Christian Hütter and Philippe Rigollet. Optimal rates for total variation denoising. In 29th Annual Conference on Learning Theory, pages 1115--1146, 2016

    2016

  8. [16]

    Kirichenko and H

    A. Kirichenko and H. van Zanten. Estimating a smooth function on a large graph by bayesian laplacian regularisation. Electron. J. Statist., 11 0 (1): 0 891--915, 2017

    2017

  9. [17]

    Kirichenko and H

    A. Kirichenko and H. van Zanten. Minimax lower bounds for function estimation on graphs. Electron. J. Statist., 12 0 (1): 0 651--666, 2018. doi:10.1214/18-EJS1407

    2018 doi

  10. [18]

    Co-regularized multi-view spectral clustering

    Abhishek Kumar, Piyush Rai, and Hal Daume. Co-regularized multi-view spectral clustering. In Advances in Neural Information Processing Systems 24, pages 1413--1421. 2011

    2011

  11. [19]

    Graph-based regularization for regression problems with alignment and highly correlated designs

    Yuan Li, Benjamin Mark, Garvesh Raskutti, Rebecca Willett, Hyebin Song, and David Neiman. Graph-based regularization for regression problems with alignment and highly correlated designs. SIAM Journal on Mathematics of Data Science, 2 0 (2): 0 480--504, 2020

    2020

  12. [20]

    Locally adaptive regression splines

    Enno Mammen and Sara Van De Geer. Locally adaptive regression splines. The Annals of Statistics, 25 0 (1): 0 387--413, 1997

    1997

  13. [21]

    Probability and computing: Randomization and probabilistic techniques in algorithms and data analysis

    Michael Mitzenmacher and Eli Upfal. Probability and computing: Randomization and probabilistic techniques in algorithms and data analysis. Cambridge university press, 2017

    2017

  14. [22]

    Semi-supervised learning with the graph laplacian: the limit of infinite unlabelled data

    Boaz Nadler, Nathan Srebro, and Xueyuan Zhou. Semi-supervised learning with the graph laplacian: the limit of infinite unlabelled data. In Proceedings of the 23rd International Conference on Neural Information Processing Systems, page 1330–1338, 2009

    2009

  15. [23]

    Roula Nassif, Stefan Vlaski, Cedric Richard, Jie Chen, and Ali H. Sayed. Multitask learning over graphs: An approach for distributed, streaming machine learning. IEEE Signal Processing Magazine, 37 0 (3): 0 14--25, 2020 a

    2020

  16. [24]

    Roula Nassif, Stefan Vlaski, Cédric Richard, and Ali H. Sayed. Learning over multitask graphs – part i: Stability analysis. Signal Processing, IEEE Open Journal on, 1: 0 28--45, 2020 b

    2020

  17. [25]

    Spectral and matrix factorization methods for consistent community detection in multi-layer networks

    Subhadeep Paul and Yuguo Chen. Spectral and matrix factorization methods for consistent community detection in multi-layer networks. Ann. Statist., 48 0 (1): 0 230--250, 2020 a

    2020

  18. [26]

    Spectral and matrix factorization methods for consistent community detection in multi-layer networks

    Subhadeep Paul and Yuguo Chen. Spectral and matrix factorization methods for consistent community detection in multi-layer networks . The Annals of Statistics, 48 0 (1): 0 230 -- 250, 2020 b

    2020

  19. [27]

    Tibshirani

    Veeranjaneyulu Sadhanala, Yu-Xiang Wang, and Ryan J. Tibshirani. Total variation classes beyond 1d: Minimax rates, and the limitations of linear smoothers. NIPS'16, page 3521–3529, 2016

    2016

  20. [28]

    D. I. Shuman , S. K. Narang , P. Frossard , A. Ortega , and P. Vandergheynst . The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE Signal Processing Magazine, 30 0 (3): 0 83--98, 2013

    2013

  21. [29]

    Analysis of \ p\ -laplacian regularization in semisupervised learning

    Dejan Slep c ev and Matthew Thorpe. Analysis of \ p\ -laplacian regularization in semisupervised learning. SIAM Journal on Mathematical Analysis, 51 0 (3): 0 2085--2120, 2019

    2019

  22. [30]

    Spectral and algebraic graph theory

    D.A Spielman. Spectral and algebraic graph theory. incomplete draft (2025), available at http://cs-www.cs.yale.edu/homes/spielman/sagt/sagt.pdf , 2025

    2025

  23. [31]

    Learning from similar linear representations: Adaptivity, minimaxity, and robustness

    Ye Tian, Yuqi Gu, and Yang Feng. Learning from similar linear representations: Adaptivity, minimaxity, and robustness. arxiv:2303.17765, 2025

    2025 arXiv

  24. [32]

    The generalized elastic net for least squares regression with network-aligned signal and correlated design

    Huy Tran, Sansen Wei, and Claire Donnat. The generalized elastic net for least squares regression with network-aligned signal and correlated design. arXiv:2211.00292, 2022

    2022

  25. [33]

    Provable meta-learning of linear representations

    Nilesh Tripuraneni, Chi Jin, and Michael Jordan. Provable meta-learning of linear representations. In Proceedings of the 38th International Conference on Machine Learning, volume 139, pages 10434--10443, 2021

    2021

  26. [34]

    Joel A. Tropp. An introduction to matrix concentration inequalities. Foundations and Trends in Machine Learning, 8 0 (1-2): 0 1--230, 2015

    2015

  27. [35]

    High-Dimensional Probability: An Introduction with Applications in Data Science

    Roman Vershynin. High-Dimensional Probability: An Introduction with Applications in Data Science. Number 47 in Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press , 2018

    2018

  28. [36]

    Trend filtering on graphs

    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

  29. [37]

    Das asymptotische verteilungsgesetz der eigenwerte linearer partieller differentialgleichungen (mit einer anwendung auf die theorie der hohlraumstrahlung)

    Hermann Weyl. Das asymptotische verteilungsgesetz der eigenwerte linearer partieller differentialgleichungen (mit einer anwendung auf die theorie der hohlraumstrahlung). Mathematische Annalen, 71 0 (4): 0 441--479, 1912

    1912

  30. [38]

    Learning from labeled and unlabeled data on a directed graph

    Dengyong Zhou, Jiayuan Huang, and Bernhard Sch\" o lkopf. Learning from labeled and unlabeled data on a directed graph. In Proceedings of the 22nd International Conference on Machine Learning, page 1036–1043, 2005

    2005

  31. [39]

    Semi-supervised learning using gaussian fields and harmonic functions

    Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the Twentieth International Conference on International Conference on Machine Learning, page 912–919, 2003

    2003

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.