Pith. sign in

REVIEW 2 major objections 4 minor 46 references

The paper claims that a single total-variation penalized least-squares estimator can jointly recover the parameters of many linear dynamical systems from one short trajectory each, with the mean squared error driven to zero by the graph's c

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 20:39 UTC pith:MXGOG4ME

load-bearing objection Genuinely new TV-penalized joint LDS estimation theory, but Lemma 5 has a factor-of-m scaling error that must be fixed before the stated conditions are certified. the 2 major comments →

arxiv 2511.18737 v3 pith:MXGOG4ME submitted 2025-11-24 math.ST math.OCstat.MLstat.TH

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

Claire Donnat , Olga Klopp , Hemant Tyagi This is my paper
classification math.ST math.OCstat.MLstat.TH MSC 62M1062J0762H12
keywords vector autoregressiontotal variation penaltylinear dynamical systemsfederated learningnetwork graphsnon-asymptotic boundsrestricted eigenvalue conditionsmall-T consistency
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 paper studies a setting in which each of m linear dynamical systems is observed for only a short time, and the systems' unknown matrices are believed to change either smoothly or in a few jumps across the edges of a known graph. It claims a single total-variation penalized least-squares estimator can pool information across the graph, and proves non-asymptotic bounds on the mean squared error that hold with high probability. The headline consequence is that on well-connected graphs, such as the complete graph or a dense random graph, the MSE can go to zero as the number of systems m grows even when the trajectory length T is held constant. A sympathetic reader should care because node-wise estimation from a single short trajectory cannot be consistent on its own; the paper shows graph structure can substitute for long trajectories. The cost is a stiffness assumption: every true system matrix must be uniformly Schur stable with spectral norm bounded away from 1.

Core claim

The central claim is that the TV-penalized joint estimator (1.2)/(2.3), which adds an entry-wise ℓ1 penalty on edge differences to a least-squares fit, satisfies the non-asymptotic bound ∥â−a*∥₂ ≤ (2mλ/T)(1 + 3√|S|/κ_S) + √(8λm/T ∥(eDa*)_{S^c}∥₁) with high probability. Under the uniform stability condition ∥A*_l∥₂ ≤ ρ_max < 1, this yields consistency of the MSE as m→∞ with T fixed for well-connected graphs such as the complete graph and dense Erdős–Rényi graphs, provided the parameter matrices vary smoothly (or have few jumps) across edges; for sparser graphs like the star and 2D grid, T only needs to grow polylogarithmically with m and d. The proof provides, to the authors' knowledge, the f

What carries the argument

The carrying object is the generalized-lasso formulation of the estimator, with block-diagonal design Q = blkdiag(X_lᵀ ⊗ I_d) and penalty λ‖eDa‖₁, where eD is the Kronecker product of the graph incidence matrix with I_{d²}. The argument proceeds through four components: (i) tail bounds for self-normalized vector-valued martingales to control the noise terms (Lemmas 1–2); (ii) a range–nullspace decomposition of the error h = h₁ + h₂ relative to eD, separating graph-edge differences from the global mean (the identity eD†eD = I − eΠ); (iii) a restricted eigenvalue condition for Q obtained by bounding suprema of second-order subgaussian chaos via Talagrand functionals (Lemmas 3–4) and a Hanson–W

Load-bearing premise

That every true system matrix has spectral norm bounded below 1 by a uniform constant, ∥A*_l∥₂ ≤ ρ_max < 1; if any node is marginally stable or unstable, the bounds on β and the Grammians collapse, and the stated rates and small-T consistency are no longer supported by the proof.

What would settle it

Run the estimator on a complete graph with T = 5 and m = 50, 100, 200, where every node's A*_l is a rotation matrix with spectral radius 1 (so condition (2.7) fails), but the matrices still vary smoothly across edges. If the MSE does not tend to zero as m grows — or if it grows — then the uniform Schur-stability assumption is genuinely load-bearing. Equivalently, compute the constants in Corollary 1: as ρ_max approaches 1, the required T ≳ v/Δ² diverges, so the small-T claim should fail in simulation.

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

If this is right

  • On complete graphs and dense Erdős–Rényi graphs, a single short trajectory per node (T constant) is sufficient for weak consistency as m grows, provided the true matrices vary smoothly or have few jumps across edges.
  • On star graphs and 2D grids, T need only grow poly-logarithmically with m and d for consistency; the joint estimator's MSE is o(1) while naive node-wise OLS has O(1) MSE in m.
  • The choice of support set S splits into a smooth regime (S=∅) and a few-changes regime (S=supp(eDa*)), with a quantitative switch criterion: ∥eDa*∥₁ ≲ (M/√T)·(s/κ_S²) favors S=∅ when changes are many but tiny, and S=supp(eDa*) when changes are few but large.
  • The bound separates bias from off-support total variation: the term √(8λm/T ‖(eDa*)_{S^c}‖₁) shows explicitly how misspecification of the smoothness (large off-support TV) degrades the rate.
  • The estimator's advantage over Laplacian smoothing is most pronounced for piecewise-constant matrices, where TV preserves sharp edges; experiments confirm this on synthetic graphs and on EPA air-quality monitoring data.

Where Pith is reading between the lines

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

  • An implication left implicit is a design principle for networked time series: when trajectories are short, increasing the number of observed nodes can substitute for longer observation windows, as long as the graph is well connected and the dynamics are stable; this suggests a testable 'more nodes, same T' experimental protocol.
  • The proof structure suggests that the Fiedler eigenvalue λ_{m−1} is the fundamental resource: condition (2.9) essentially requires µ′√log(|E|d/δ) ≲ 1, which for many graphs translates to λ_{m−1} growing with m — indicating that small-T consistency should fail on graphs with bottlenecks even if they are technically connected.
  • A natural extension, not addressed in the paper, is asynchronous or misaligned node trajectories; since the estimator only uses each node's own trajectory and the graph penalty couples parameter estimates, a permutation-invariant variant could be developed, though the martingale arguments would need adjustment if noise is also correlated across nodes.
  • The constants in Example 1 indicate that the smoothness requirement scales with d³/Δ, so for high-dimensional states (large d) the practical benefit of graph pooling diminishes; this could be tested by varying d while keeping m, T, and the graph fixed.

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 / 4 minor

Summary. The paper studies joint estimation of m linear dynamical systems on a graph, using a total-variation penalized least-squares estimator. The main theorem gives high-probability non-asymptotic bounds on the estimation error, expressed through graph-geometry parameters, controllability Grammians, and a dispersion functional. Corollaries for uniformly Schur-stable systems yield rates for smooth and few-jump regimes, and the paper claims weak consistency as m→∞ even with T fixed for well-connected graphs (complete and Erdős–Rényi), while star and grid graphs require only poly-logarithmic growth of T. The theoretical results are complemented by synthetic and EPA-air-quality experiments.

Significance. If the proof issues are fixed, this is a valuable contribution: it provides the first non-asymptotic guarantees for graph-TV-penalized joint LDS estimation with dependent design and dependent noise, and it introduces a range–nullspace decomposition of the RE condition that is tailored to the graph-TV operator. The explicit dependence on µ, µ′, κ_S, β, and Δ_G is informative, and the paper is careful to separate estimator structure from dynamics assumptions. The experiments are extensive and the code is provided. The paper is methodologically serious and the central claims are plausible, but the proof of the RE condition currently contains a load-bearing scaling inconsistency that must be corrected.

major comments (2)
  1. [Section 3.5, Eq. (3.37)–(3.38) and Theorem 1 condition (iii)] The displayed bound (3.37) scales as M ≤ dµ/√m Δ_G + cβ²µ√(T/m)log = (µ/√m)(dΔ_G + cβ²√T log). Multiplying by the bracket in (3.31) can only yield a cross-term bound proportional to µ/√m. However, Lemma 5's statement and Eq. (3.38) replace µ/√m by µ√m, and the 'Putting everything together' step uses this larger factor in condition (iii). This is not a harmless typo: Theorem 1 condition (iii) is not proven as stated, and Corollary 3 and Examples 1–4 inherit the wrong m-power. The lemma must be corrected to µ/√m and the proof steps reconciled. Because the corrected factor is smaller, the small-T consistency conclusion is likely preserved, but the stated theorem is not certified as written.
  2. [Section 2.4, (2.11)–(2.12), Corollary 3, Examples 1–4] Once Lemma 5 is corrected, the sample-size side conditions must be re-derived. Condition (C3b) should become T≥CµΦ_S dΔ_G/√m rather than T≥Cµ√mΦ_S dΔ_G. Moreover, the small-T conditions in Corollary 3 do not follow algebraically from the stated (2.11)–(2.12): substituting T≍log²(1/δ)/Δ² into (2.12) gives µΦ_S ≲ log²/(√m dΔ_GΔ²), not the √m log²/(dΔ_GΔ²) printed in Corollary 3(ii). The thresholds in Examples 1 and 2 (e.g., the m^{3/2} condition) are therefore off by a further factor of m. The quantitative small-T claims need a careful re-derivation.
minor comments (4)
  1. [Lemmas 3 and 4] The probability statements are written as P(...) ≤ 2exp(...); from the context and usage, the inequality should be ≥. As printed, the lemmas assert that the desired event has small probability.
  2. [Lemma 5 statement, Eq. (3.8)] The statement uses ∥(Da*)_{S^c}∥₁ while the proof and Theorem 1 use ∥(eDa*)_{S^c}∥₁. Use the tilde consistently.
  3. [Section 3.3] The sentence 'We now set t=G√v for any v≥1, t=G√v for any v≥1' is duplicated.
  4. [Introduction / Section 2.4] The abstract says 'MSE goes to zero ... even when T is constant' without mentioning the uniform Schur-stability condition (2.7). This condition is load-bearing and should be stated more prominently in the abstract or introduction.

Circularity Check

0 steps flagged

No circularity: the MSE bounds are oracle-type bounds in terms of true difficulty parameters, not fitted inputs or self-imported uniqueness claims.

full rationale

Walking the derivation chain: Theorem 1 follows from the optimality inequality (3.1)–(3.6), the restricted-eigenvalue decomposition in (3.7), and Lemmas 3–5. The error bounds depend on true but unknown quantities — β, Δ_G, μ, μ′, κ_S and ∥(eDa*)_{S^c}∥_1 — which are difficulty parameters, not fitted values. No step defines the target through the estimator or fits a parameter and then renames it a prediction. Corollaries 1–3 and Examples 1–4 are instantiations of Theorem 1 using graph-geometric bounds from Table 1, Propositions 1–2 and Lemma 9; they are not obtained by assuming the conclusion. The self-citations ([35], [36], [37], [38]) are contextual comparisons or standard facts; the proof’s load-bearing external tools are independent ([1], [11], [12], [13], [15]). I also note, as a correctness issue rather than a circularity, an apparent factor-of-m inconsistency between Eq. (3.37) and Lemma 5 / Theorem 1 condition (iii): (3.37) gives a cross-term bound of order μ/√m·(dΔ_G + β²√T log), while (3.38) states μ√m·(...). This affects the stated sample-size conditions but does not make the argument circular.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central proof relies on standard probability theory (martingale tail bounds, Hanson-Wright, chaos concentration) and on domain assumptions about the dynamics and graph. No new unobserved entities are postulated. λ is a tuning parameter; the theorem's constants are absolute, not data-fitted.

free parameters (1)
  • regularization parameter λ = selected by validation in experiments; theory requires λ ≥ (2/m)max{F1,F2} and suggests (2.8)
    the estimator's performance depends on λ; the paper provides a valid range but no data-free optimal λ
axioms (5)
  • domain assumption Noise η_{l,t} are i.i.d. centered standard Gaussian (or subgaussian with bounded norms)
    used throughout to apply concentration and martingale bounds; mild but restricts to exogenous stochastic excitation
  • domain assumption Schur stability ∥A*_l∥₂ ≤ ρ_max < 1 for all l (for corollaries)
    required to bound β, F1-F3 and derive consistency; the paper's headline small-T results only under this condition
  • domain assumption Graph G is known, undirected, connected
    the estimator uses G; known graph structure is a premise of the method
  • domain assumption Smoothness/few-jumps of A*_l across edges (small ∥eDa*∥₁ or support S)
    consistency statements require these quantities to be small; they are oracle conditions on the truth
  • domain assumption Initial state x_{l,0}=0 for each l
    simplifies the state representation in the proofs; not load-bearing

pith-pipeline@v1.3.0-alltime-deepseek · 41864 in / 22727 out tokens · 181045 ms · 2026-08-03T20:39:15.278089+00:00 · methodology

0 comments
read the original abstract

We consider the problem of joint estimation of the parameters of $m$ linear dynamical systems, given access to single realizations of their respective trajectories, each of length $T$. The linear systems are assumed to reside on the nodes of an undirected and connected graph $G = ([m], \mathcal{E})$, and the system matrices are assumed to either vary smoothly or exhibit small number of ``jumps'' across the edges. We consider a total variation penalized least-squares estimator and derive non-asymptotic bounds on the mean squared error (MSE) which hold with high probability. In particular, the bounds imply for certain choices of well connected $G$ that the MSE goes to zero as $m$ increases, even when $T$ is constant. The theoretical results are supported by extensive experiments on synthetic and real data.

Figures

Figures reproduced from arXiv: 2511.18737 by Claire Donnat, Hemant Tyagi, Olga Klopp.

Figure 1
Figure 1. Figure 1: Examples of coefficients β ∗ l generated, overlaid on the graph. Here, the coef￾ficients (β ∗ l ) m l=1 are chosen to be piecewise constant across the graph, and the maximal jump size is constrained to ∥Dae ∗∥∞ = 0.4. (ii) Smoothly varying. In this case, we compute the 2D spectral embedding for each node in the graph and define the signal as (for ω a parameter) β ∗ l = s(cos(ωt1) × cos(ω t2)), where t1 = l… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of coefficients βl generated, overlaid on the graph. Here, the coeffi￾cients (β ∗ l ) m l=1 are chosen to be smoothly varying across the graph, as per (4.4). We report the parameter and test prediction MSEs, defined as (Parameter MSE) MSE = 1 m Xm l=1 [PITH_FULL_IMAGE:figures/full_fig_p034_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance of the TV regularizer and other competing methods as a function of the maximal jump size ∥Dβ∗∥∞ on a piecewise constant signal (as per (4.3)) [PITH_FULL_IMAGE:figures/full_fig_p035_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance of the TV regularizer as a function of the maximal jump size ∥Dβ∗∥∞ on smooth signal (defined as per (4.4)). is extremely small, as the benefits of the smoothing induced by fitting the estimator over the entire graph are expected to outweigh the bias incurred in this case. The results for the different topologies and signals are reported in Figures 5a and 5b, and confirm our expected findings. … view at source ↗
Figure 5
Figure 5. Figure 5: Parameter and test prediction MSE as a function of training time length Ttrain for different graph topologies. (c) Error vs. number of nodes m. We hold Ttrain fixed (here, Ttrain = 10) and increase the graph size (m). The results are shown in [PITH_FULL_IMAGE:figures/full_fig_p036_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Parameter and test prediction MSE versus m for a piecewise constant signal, for different graph topologies. Lines and dots indicate MSE mean averaged over 15 trials ± 95% CI. 4.3 Real Data Experiments: EPA data We evaluate the proposed estimator on data from the U.S. Environmental Protection Agency (EPA) national air-quality monitoring network. Specifically, we analyze daily arithmetic-mean concentra￾tions… view at source ↗
Figure 7
Figure 7. Figure 7: Air quality stations, along with the induced kNN graph (k = 5). evidenced by the wide spread of coefficients. By contrast, pooled OLS yields a single global co￾efficient shared across stations. Both the graph total-variation (TV) denoiser and the Laplacian regularizer produce coefficients that vary smoothly over the spatial graph. As expected, TV re￾covers piecewise-constant (or piecewise-smooth) structure… view at source ↗
Figure 8
Figure 8. Figure 8: Coefficient stability and test RMSE for predicting NO concentration. Boxplots show the distribution of errors over 15 independent trials (years). (a) Coefficient Stability: PM10 (b) Test MSE: PM10 [PITH_FULL_IMAGE:figures/full_fig_p039_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Coefficient stability and Test RMSE for predicting PM10 concentration. Boxplots show the distribution of errors over 15 independent trials (years). 39 [PITH_FULL_IMAGE:figures/full_fig_p039_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Monitoring stations colored by the value of the estimated LDS coefficient Abl [1, 1] under four methods. Nodes are stations; (optional) gray edges indicate the k-NN graph (k = 5). Acknowledgment The work of O. Klopp was funded by the CY Initiative Grant Investissements d’Avenir Agence Nationale de Recherche-16-Initiatives d’Excellence-0008 and Labex MME-DII Grant ANR11-LBX￾0023-01. H.T was supported by a … 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

46 extracted references · 2 linked inside Pith

  1. [1]

    Improved algorithms for linear stochastic bandits

    Yasin Abbasi-yadkori, D´ avid P´ al, and Csaba Szepesv´ ari. Improved algorithms for linear stochastic bandits. InAdvances in Neural Information Processing Systems, volume 24, 2011

    2011

  2. [2]

    Regularized estimation in sparse high-dimensional time series models.The Annals of Statistics, 43(4):1535–1567, 2015

    Sumanta Basu and George Michailidis. Regularized estimation in sparse high-dimensional time series models.The Annals of Statistics, 43(4):1535–1567, 2015

    2015

  3. [3]

    Network granger causality with inherent grouping structure.Journal of Machine Learning Research, 16(13):417–453, 2015

    Sumanta Basu, Ali Shojaie, and George Michailidis. Network granger causality with inherent grouping structure.Journal of Machine Learning Research, 16(13):417–453, 2015

    2015

  4. [4]

    Linearized aerodynamic and control law models of the x-29a airplane and comparison with flight data.National Aeronautics and Space Administration, Office of Management

    John T Bosworth. Linearized aerodynamic and control law models of the x-29a airplane and comparison with flight data.National Aeronautics and Space Administration, Office of Management . . ., 4356, 1992

    1992

  5. [5]

    Oxford University Press, 2013

    St´ ephane Boucheron, G´ abor Lugosi, and Pascal Massart.Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, 2013

    2013

  6. [6]

    Campi and E

    M.C. Campi and E. Weyer. Finite sample properties of system identification methods.IEEE Transactions on Automatic Control, 47(8):1329–1334, 2002

    2002

  7. [7]

    Ospina, Fabio Pasqualetti, and Emiliano Dall’Anese

    Yiting Chen, Ana M. Ospina, Fabio Pasqualetti, and Emiliano Dall’Anese. Multi-task system identification of similar linear time-invariant dynamical systems.2023 62nd IEEE Conference on Decision and Control (CDC), pages 7342–7349, 2023

    2023

  8. [8]

    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

  9. [9]

    Learning sparse dynamical systems from a single sample trajectory

    Salar Fattahi, Nikolai Matni, and Somayeh Sojoudi. Learning sparse dynamical systems from a single sample trajectory. In2019 IEEE 58th Conference on Decision and Control (CDC), pages 2682–2689, 2019

    2019

  10. [10]

    Telesford, Alfred B

    Shi Gu, Fabio Pasqualetti, Matthew Cieslak, Qawi K. Telesford, Alfred B. Yu, Ari E. Kahn, John D. Medaglia, Jean M. Vettel, Michael B. Miller, Scott T. Grafton, and Danielle S. Bassett. Controllability of structural brain networks.Nature Communications, 6, 2014

    2014

  11. [11]

    A tail inequality for quadratic forms of subgaus- sian random vectors.Electronic Communications in Probability, 17:1–6, 2012

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

    2012

  12. [12]

    Optimal rates for total variation denoising

    Jan-Christian H¨ utter and Philippe Rigollet. Optimal rates for total variation denoising. vol- ume 49, pages 1115–1146, 2016

    2016

  13. [13]

    Finite-time identification of stable linear systems opti- mality of the least-squares estimator

    Yassir Jedra and Alexandre Proutiere. Finite-time identification of stable linear systems opti- mality of the least-squares estimator. In2020 59th IEEE Conference on Decision and Control (CDC), pages 996–1001, 2020

    2020

  14. [14]

    Oracle inequalities for high dimensional vector autoregressions.Journal of Econometrics, 186(2):325–344, 2015

    Anders Bredahl Kock and Laurent Callot. Oracle inequalities for high dimensional vector autoregressions.Journal of Econometrics, 186(2):325–344, 2015

    2015

  15. [15]

    Suprema of chaos processes and the restricted isometry property.Communications on Pure and Applied Mathematics, 67(11):1877– 1904, 2014

    Felix Krahmer, Shahar Mendelson, and Holger Rauhut. Suprema of chaos processes and the restricted isometry property.Communications on Pure and Applied Mathematics, 67(11):1877– 1904, 2014. 50

    1904

  16. [16]

    Asymptotic properties of general autoregressive models and strong consistency of least-squares estimates of their parameters.Journal of Multivariate Analysis, 13(1):1–23, 1983

    T.L Lai and C.Z Wei. Asymptotic properties of general autoregressive models and strong consistency of least-squares estimates of their parameters.Journal of Multivariate Analysis, 13(1):1–23, 1983

    1983

  17. [17]

    Extended least squares and their applications to adaptive control and prediction in linear systems.IEEE Transactions on Automatic Control, 31(10):898–906, 1986

    Tze Lai and Ching-Zong Wei. Extended least squares and their applications to adaptive control and prediction in linear systems.IEEE Transactions on Automatic Control, 31(10):898–906, 1986

    1986

  18. [18]

    Least Squares Estimates in Stochastic Regression Models with Applications to Identification and Control of Dynamic Systems.The Annals of Statistics, 10(1):154 – 166, 1982

    Tze Leung Lai and Ching Zong Wei. Least Squares Estimates in Stochastic Regression Models with Applications to Identification and Control of Dynamic Systems.The Annals of Statistics, 10(1):154 – 166, 1982

    1982

  19. [19]

    Variable fusion: A new adaptive signal regression method.Dept

    Stephanie R Land and Jerome H Friedman. Variable fusion: A new adaptive signal regression method.Dept. Statistics, Carnegie Mellon Univ. Pittsburgh, Pittsburgh, PA, USA, Rep, 656, 1997

    1997

  20. [20]

    Graph-based regularization for regression problems with alignment and highly corre- lated designs.SIAM Journal on Mathematics of Data Science, 2(2):480–504, 2020

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

    2020

  21. [21]

    Wainwright

    Po-Ling Loh and Martin J. Wainwright. High-dimensional regression with noisy and missing data: Provable guarantees with nonconvexity.The Annals of Statistics, 40(3):1637 – 1664, 2012

    2012

  22. [22]

    Estimating structured vector autoregressive models

    Igor Melnyk and Arindam Banerjee. Estimating structured vector autoregressive models. In Proceedings of The 33rd International Conference on Machine Learning, pages 830–839, 2016

    2016

  23. [23]

    Joint learning of linear time-invariant dynamical systems.Automatica, 164:111635, 2024

    Aditya Modi, Mohamad Kazem Shirani Faradonbeh, Ambuj Tewari, and George Michailidis. Joint learning of linear time-invariant dynamical systems.Automatica, 164:111635, 2024

    2024

  24. [24]

    Prediction bounds for higher order total variation regularized least squares.The Annals of Statistics, 49(5):2755–2773, 2021

    Francesco Ortelli and Sara van de Geer. Prediction bounds for higher order total variation regularized least squares.The Annals of Statistics, 49(5):2755–2773, 2021

    2021

  25. [25]

    Wainwright, and Bin Yu

    Garvesh Raskutti, Martin J. Wainwright, and Bin Yu. Restricted eigenvalue properties for correlated gaussian designs.J. Mach. Learn. Res., 11:2241–2259, 2010

    2010

  26. [26]

    Total variation classes beyond 1d: Minimax rates, and the limitations of linear smoothers.Advances in Neural Information Processing Systems, 29:3521–3529, 2016

    Veeranjaneyulu Sadhanala, Yu-Xiang Wang, and Ryan J Tibshirani. Total variation classes beyond 1d: Minimax rates, and the limitations of linear smoothers.Advances in Neural Information Processing Systems, 29:3521–3529, 2016

    2016

  27. [27]

    Near optimal finite time identification of arbitrary linear dynamical systems

    Tuhin Sarkar and Alexander Rakhlin. Near optimal finite time identification of arbitrary linear dynamical systems. InProceedings of the 36th International Conference on Machine Learning, ICML, volume 97, pages 5610–5618, 2019

    2019

  28. [28]

    Finite time identification in unstable linear systems.Automatica, 96:342–353, 2018

    Mohamad Kazem Shirani Faradonbeh, Ambuj Tewari, and George Michailidis. Finite time identification in unstable linear systems.Automatica, 96:342–353, 2018

    2018

  29. [29]

    Jordan, and Benjamin Recht

    Max Simchowitz, Horia Mania, Stephen Tu, Michael I. Jordan, and Benjamin Recht. Learning without mixing: Towards a sharp analysis of linear system identification. InProceedings of the 31st Conference On Learning Theory, volume 75, pages 439–473, 2018. 51

    2018

  30. [30]

    Kim, Harang Ju, Dale Zhou, Cassiano Becker, Fabio Pasqualetti, George J

    Pragya Srivastava, Erfan Nozari, Jason Z. Kim, Harang Ju, Dale Zhou, Cassiano Becker, Fabio Pasqualetti, George J. Pappas, and Danielle S. Bassett. Models of communication and control for brain networks: distinctions, convergence, and future outlook.Network Neuroscience, 4(4):1122–1159, 2020

    2020

  31. [31]

    Wellesley-Cambridge Press, Philadel- phia, PA, 2007

    Gilbert Strang.Computational Science and Engineering. Wellesley-Cambridge Press, Philadel- phia, PA, 2007

    2007

  32. [32]

    Springer, 2005

    Michel Talagrand.The Generic Chaining: Upper and Lower Bounds of Stochastic Processes. Springer, 2005

    2005

  33. [33]

    Springer, 2014

    Michel Talagrand.Upper and lower bounds for stochastic processes, volume 60. Springer, 2014

    2014

  34. [34]

    Stanford University, 2011

    Ryan J Tibshirani.The solution path of the generalized lasso. Stanford University, 2011

    2011

  35. [35]

    Huy Tran, Sansen Wei, and Claire Donnat. The generalized elastic net for least squares regression with network-aligned signal and correlated design.IEEE Transactions on Signal and Information Processing over Networks, 2025

    2025

  36. [36]

    Joint estimation of smooth graph signals from partial linear measurements

    Hemant Tyagi. Joint estimation of smooth graph signals from partial linear measurements. arXiv:2505.23240, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Joint learning of linear dynamical systems under smoothness constraints

    Hemant Tyagi. Joint learning of linear dynamical systems under smoothness constraints. Information and Inference: A Journal of the IMA, 14(3):iaaf026, 2025

    2025

  38. [38]

    Learning linear dynamical systems under convex constraints

    Hemant Tyagi and Denis Efimov. Learning linear dynamical systems under convex constraints. arxiv:2303.15121, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Cambridge University Press, 2025

    Roman Vershynin.High-Dimensional Probability: An Introduction with Applications in Data Science (2nd ed.). Cambridge University Press, 2025

    2025

  40. [40]

    Vidyasagar and R.L

    M. Vidyasagar and R.L. Karandikar. A learning theory approach to system identification and stochastic adaptive control.Journal of Process Control, 18(3):421–430, 2008

    2008

  41. [41]

    Di Wang and Ruey S. Tsay. Rate-optimal robust estimation of high-dimensional vector au- toregressive models.The Annals of Statistics, 51(2):846 – 877, 2023

    2023

  42. [42]

    Fedsysid: A federated approach to sample-efficient system identification

    Han Wang, Leonardo Felipe Toso, and James Anderson. Fedsysid: A federated approach to sample-efficient system identification. InProceedings of The 5th Annual Learning for Dynamics and Control Conference, volume 211, pages 1308–1320, 2023

    2023

  43. [43]

    Trend filtering on graphs.Journal of Machine Learning Research, 17(105):1–41, 2016

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

    2016

  44. [44]

    Learning the dynamics of autonomous linear systems from multiple trajectories

    Lei Xin, George Chiu, and Shreyas Sundaram. Learning the dynamics of autonomous linear systems from multiple trajectories. In2022 American Control Conference (ACC), pages 3955– 3960, 2022

    2022

  45. [45]

    Chiu, and Shreyas Sundaram

    Lei Xin, Lintao Ye, George T.-C. Chiu, and Shreyas Sundaram. Identifying the dynamics of a system by leveraging data from similar systems.2022 American Control Conference (ACC), pages 818–824, 2022

    2022

  46. [46]

    Non-asymptotic identification of linear dynamical systems using multiple trajectories.IEEE Control Systems Letters, 5(5):1693–1698, 2020

    Yang Zheng and Na Li. Non-asymptotic identification of linear dynamical systems using multiple trajectories.IEEE Control Systems Letters, 5(5):1693–1698, 2020. 52

    2020