Pith. sign in

REVIEW 3 major objections 5 minor 47 references

A Spectral Filtering Approach to Regret Analysis of Distributed Online Control for Linear Dynamical Systems

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Each agent in a network of identical linear systems can compete with the best centralized hindsight controller, with sublinear regret, using only local costs and neighbor communication.

desk verdict The distributed extension is plausible, but the central regret bound rests on a false inequality in Lemma 8, so the main theorem is unsupported as written. read the letter →

arxiv 2608.02375 v1 pith:LDVHRBGO submitted 2026-08-03 math.OC cs.LGcs.SYeess.SY

Ting-Jui Chang This is my paper · ORCID
classification math.OCcs.LGcs.SYeess.SY MSC 93C0593A1468Q3290C25
keywords distributedonlinecontrolregretminimizationspectralfilteringHankelmatrixlineardynamicalsystemsconvexoptimizationstrongstabilitymulti-agent
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

This paper tries to establish that distributed online control does not force agents to give up centralized performance: on a network of identical linear time-invariant systems with adversarial disturbances and time-varying convex costs, each agent can run the proposed D-OSC algorithm and achieve individual regret of order O~(n√(nT)/((1−β)γ^3)) against the best centralized (κ,γ)-diagonalizably stable linear policy in hindsight. The algorithm works by reparameterizing control as a distributed online convex optimization problem over a fixed parameter set, using spectral controllers formed from the leading eigenvectors of a Hankel matrix, and updating parameters with consensus-based online gradient descent on memory-less surrogate costs. The resulting regret is sublinear, so each agent's average network cost approaches that of the best centralized policy over time. A sympathetic reader would care because it shows that the improved stability-margin scaling of centralized spectral control survives the move to local information and neighbor-to-neighbor communication.

What carries the argument

The spectral controller: a controller of the form u_t = Σ_{i=1}^h σ_i^{1/4} M_i W_{t-1:t-m} φ_i, where (σ_i, φ_i) are the top h eigenpairs of the Hankel matrix H (with H_{ij} = (1−γ)^{i+j−1}/(i+j−1)) and W_{t-1:t-m} stacks the past m disturbances. Its key role is to convert the non-convex class of stable linear policies into a convex, fixed-dimensional parameter set K: the memory-less surrogate cost f_{i,t}(M_{1:h}|A,B,{w}) evaluates the cost that would have accrued had a fixed M been used throughout, so running distributed online gradient descent over K becomes a standard distributed online convex program. The spectral eigenbasis is what keeps the stability-margin dependence polynomial (1/γ

What would settle it

Pick a small LTI system and a (κ,γ)-diagonalizably stable policy K, compute the spectral controller M* defined in the proof of Theorem 11, and measure the cumulative cost gap Σ_t |f_t(x_t^M,u_t^M)−f_t(x_t^K,u_t^K)| over horizon T; if the gap does not scale as εT with the stated h, or the eigenvector inner products |μ_α^T φ_j| do not match Lemma 8, the theorem is false.

Watch

Extended reading notes

Core claim

The paper's central claim, formalized in Theorems 1 and 11, is that Algorithm 1 (D-OSC) yields, for any agent j and time horizon T, the individual regret bound J^j_T(A_1) − min_{K∈S} J_T(K) = O~( n√(nT) / ((1−β)γ^3) ), where S is the set of (κ,γ)-diagonalizably stable linear policies, n is the number of agents, and β < 1 quantifies the network's mixing time. Each agent chooses controls u_{i,t} = Σ_{r=1}^h σ_r^{1/4} M_{i,r}^t W_{t-1:t-m} φ_r, convolving past disturbances with the top h eigenvectors of the Hankel matrix H with entries (1−γ)^{i+j−1}/(i+j−1), and updates the parameters M via a distributed online gradient step over local memory-less surrogate costs f_{i,t}(M_{1:h}|A,B,{w}). The p

Load-bearing premise

The proof relies on unproved spectral approximation lemmas, imported from the centralized spectral-control literature, claiming that any (κ,γ)-stable linear policy can be approximated by a spectral controller with error ε and that the relevant Hankel eigenvector inner products decay at the stated rate; if either claim fails, the regret decomposition does not close.

Editorial extensions

If this is right

  • If Theorem 1 is correct, every agent's time-averaged network cost approaches that of the best centralized hindsight policy, so local information plus neighbor communication is sufficient for asymptotically optimal distributed online control.
  • The bound O~(n√(nT)/((1−β)γ^3)) shows the stability-margin dependence inherited from centralized spectral control (polynomial in 1/γ) survives the distributed setting.
  • The network term n√n/(1−β) quantifies a trade-off: smaller, better-connected networks enjoy strictly tighter regret guarantees, giving a concrete design target for communication topology.
  • The explicit hyperparameters (ε=1/√T, m=⌈(1/γ)log(·)⌉, h=2 log T log(·), η=γ^2/(m h^2 √T)) yield a ready-to-run tuning rule for practitioners.

Reading between the lines

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

  • The four-term decomposition is modular, and network dependence appears only in the drift and D-OGD terms; this suggests the same spectral parameterization would combine with other online-learning oracles (bandit feedback, strong convexity) to yield analogous regret bounds in those settings.
  • A concrete numerical check of the approximation lemma — computing the spectral controller for a small stable policy and measuring the cost gap as h and m grow — would isolate whether the imported spectral lemmas hold with the required constants, and is a natural falsification test.
  • The identical-agent assumption is convenient but likely not essential; if agents share the same Hankel eigenbasis, the consensus-on-average argument would probably extend to heterogeneous dynamics, making the result applicable to fleets with varied platforms.
  • The bound's explicit dependence on β suggests an experimental measurement of regret across random graphs of varying connectivity could verify the predicted (1−β)^{-1} scaling.
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

3 major / 5 minor

Summary. The paper proposes D-OSC, a distributed online spectral controller for a network of n identical LTI systems driven by adversarial disturbances and time-varying convex local costs. Each agent runs a consensus-based distributed online gradient step over a spectral controller parameterization and applies a controller obtained by convolving past disturbances with the top eigenvectors of a Hankel matrix. The central claim is an individual regret bound of order O~(n sqrt(nT)/((1-beta) gamma^3)) against the best centralized (kappa,gamma)-diagonalizably stable linear policy in hindsight, with explicit choices of the learning rate, window length, and number of spectral filters. The proof decomposes regret into four terms: the gap between the true trajectory and a memory-less surrogate, the distributed online-convex-optimization regret, the approximation error of the optimal linear policy by a spectral controller, and the truncation error in the disturbance window.

Significance. If the main theorem were established, the result would be a meaningful extension of online spectral control to distributed settings: it would show that agents with only local costs and neighbor communication can compete with a centralized linear policy, with sublinear regret and improved dependence on the stability margin compared with disturbance-feedback parameterizations. The distributed OCO part of the argument is a genuine and promising reduction: Theorem 10 and Lemmas 2, 6, 9, 10 provide a real proof sketch of consensus-based online gradient descent with projection, and the paper is explicit about hyperparameter choices and does not fit parameters to make the bound true. However, the central claim depends on spectral approximation lemmas (Lemmas 3, 4, 7, 8) that are stated without proof, and one of them is false as stated. The load-bearing step of Theorem 11 therefore collapses, so the significance is conditional until those lemmas are supplied and corrected.

major comments (3)
  1. [Supplementary, Lemma 8; used in Theorem 11, Step 2] Lemma 8's second inequality is false as stated. Since H_m has H_{11}=1-gamma, the largest singular value satisfies sigma_1 >= ||H_m e_1|| >= 1-gamma. Hence sqrt(2/gamma) sigma_1^{1/4} >= (1-gamma)^{1/4} sqrt(2/gamma), which grows like gamma^{-1/2}. The claimed right-hand side is at most 30 sqrt(gamma) log^{1/4}(2/gamma), which grows like gamma^{1/2}. For gamma=10^{-3}, the left side is at least about 44.7, while the right side is about 1.6 for every T>=10. The inequality cannot hold. Theorem 11 uses Lemma 8 to conclude ||M^*_i|| <= kappa^3 sqrt(2/gamma) and then that M^*_{1:h} lies in K; without this, the comparator M^* is not certified to belong to the feasible set K, and the application of Theorem 10 leading to Eq. (21) and Eq. (24) is unjustified. This is a load-bearing mathematical error, not merely a missing derivation.
  2. [Theorem 11, Step 2; Lemmas 3, 4, 7, 8] The proof of the main theorem relies on several spectral facts that are not proved in the manuscript. Lemmas 3, 4, 7, and 8 are stated without proof, and the crucial containment M^*_{1:h} in K is delegated to 'the proof of Lemma 5.2 in Hazan', which is not stated in the paper and does not appear in the reference list. These statements are not stylistic details: Lemmas 3 and 4 control the approximation of the hindsight linear policy by the spectral parameterization, Lemma 7 controls the Hankel spectrum, and Lemma 8 controls the inner products used to bound M^*. Since Lemma 8 is false, the existing text cannot serve as a proof, and the unproved external lemma cannot fill the gap. A journal proof needs complete, correct statements of all load-bearing lemmas, with proofs or precise verifiable references.
  3. [Theorem 1, Theorem 11, and Eq. (24)] The displayed T-scaling of the main regret bound is ambiguous and inconsistent across the paper. Theorem 1 writes \tilde O(n sqrt(n) T /((1-beta) gamma^3)) in the raw text, which can be parsed as n sqrt(n) * T, i.e., linear regret; Theorem 11 and the conclusion use the same ambiguous 'n sqrt(nT)' notation, while the abstract states O(sqrt(T) poly(log T)/gamma^3). The proof's Eq. (24) actually gives a bound of order n sqrt(n) sqrt(T)/gamma^3 plus n sqrt(T) after substituting epsilon = 1/sqrt(T). The exact power of T is part of the central claim, so the authors must state it unambiguously as n sqrt(nT) if that is intended, and reconcile the abstract and theorem statements.
minor comments (5)
  1. [Abstract] The abstract contains a duplicated phrase: 'We formulate this problem this problem as a regret minimization problem.'
  2. [Lemma 9] Lemma 9 defines \bar M^t_{1:h} as 1/m times the sum over n agents, and the proof sometimes uses sqrt(m) in the mixing bound where n is intended. This appears to be a typo (the average should be over n agents), but it should be corrected because it affects the stated network-size dependence.
  3. [Lemma 9 / Theorem 10 statements] The supplement contains placeholder 'Algorithm??' in the statements of Lemma 9 and Theorem 10, and several cross-references such as 'the last inequality is based on Equation.' are missing their equation numbers.
  4. [Theorem 1 hyperparameter h] The formula for h in Theorem 1 does not visually match Lemma 4: the theorem statement contains a fractional log term and a denominator kappa^{5/2}, while Lemma 4 uses 1/epsilon gamma^{5/2} and a different arrangement of constants. These should be aligned so the reader can verify that the stated h satisfies the hypothesis of Lemma 4.
  5. [Algorithm 1, initialization] Algorithm 1 says 'randomly generate the same M^{i,1}_{1:h} in K for all i'. Random generation per agent does not by itself guarantee identical initial iterates unless shared randomness or a deterministic initialization is specified; please make this precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the D-OSC regret proof is an independent distributed-OCO argument over the external spectral parameterization from [4]; self-citations [34,35] are related-work only. Lemma 8's scaling inconsistency and the unproved 'Lemma 5.2 in Hazan' step are correctness/verification gaps, not circularity.

full rationale

The proof chain of Theorem 11 does not reduce to its own inputs. The benchmark K* is an arbitrary (κ,γ)-diagonalizably stable linear policy, external to Algorithm 1; the spectral comparator M* is then constructed from K* using the Hankel eigenbasis, and the claim M* ∈ K is made via Lemma 8 and the proof of Lemma 5.2 in Hazan. Lemma 8 is not fitted to the regret quantity; it is a spectral bound on the Hankel matrix imported from independent prior work [4] by different authors. Even if Lemma 8 is false—and it appears inconsistent as stated, since H_{11}=1−γ forces σ1 ≥ 1−γ so the left side sqrt(2/γ)σ1^{1/4} grows like γ^{-1/2} while the right side is O(√γ polylog)—that is a mathematical error in an auxiliary bound, not an instance of assuming the conclusion. Similarly, Lemmas 3, 4, and 7 are unproved here and load-bearing for the approximation terms, but they are stated as parameter-free spectral/OCO lemmas from [4], not as predictions of this paper and not derived from the target regret. The only self-citations ([34], [35]) concern distributed LQR and appear only in the related-literature section; they are never used to justify the main theorem. There is no fitted parameter renamed as a prediction, no definition of a quantity in terms of the target, and no uniqueness theorem imported from the author's own prior work. The main vulnerability is a verification gap / false auxiliary lemma, which belongs to correctness risk rather than circularity.

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

No data-fitted constants appear. The algorithm's hyperparameters (m, h, eta, epsilon) are set as explicit functions of T and gamma in the theorem statement and are not counted as free parameters because they have no fitted values. The largest unverified burden is the set of spectral-approximation lemmas assumed from the literature.

assumptions (5)
  • domain assumption Assumption 1: ||B|| <= kappa_B and ||w_t|| <= W for all t (bounded disturbance).
    Invoked in every Lipschitz/boundedness bound (Lemmas 2, 6, 9); standard in adversarial online control.
  • domain assumption Assumption 2: each f_{i,t} is convex and smooth with gradient norm <= G D for ||x||,||u|| <= D.
    Required for convexity-based regret bound and Lipschitzness of surrogate costs.
  • domain assumption Assumption 3: the zero policy K=0 is (kappa,gamma)-diagonalizably stable.
    The main theorem states the bound under this assumption; the note that it can be relaxed via a precomputed stable controller is not part of the proof.
  • domain assumption All agents have identical dynamics (A,B) and receive the same disturbance sequence w_t; the communication matrix P is symmetric, doubly stochastic, connected, with positive diagonal and mixing bound (3).
    The distributed proof and consensus error (Lemma 9) rely on these; if agents are heterogeneous or disturbances differ, the averaging analysis does not apply.
  • domain assumption Spectral approximation results Lemmas 3, 4, 7, 8 and 'Lemma 5.2 in Hazan' hold as stated.
    They are stated without proof and imported from prior work; the main theorem's approximation terms depend on them. No independent verification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Spectral Filtering Approach to Regret Analysis of Distributed Online Control for Linear Dynamical Systems." pith.science (2026). https://pith.science/paper/LDVHRBGO

@misc{pith2026260802375,
  author       = {Pith},
  title        = {Pith review of: A Spectral Filtering Approach to Regret Analysis of Distributed Online Control for Linear Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LDVHRBGO}},
  note         = {Machine review of arXiv:2608.02375}
}
abstract

This paper studies the distributed online control problem over a network of linear time-invariant (LTI) systems in the presence of adversarial disturbances and time-varying convex costs. The network cost is characterized by the summation of local cost functions, where each local function is sequentially revealed only to the corresponding agent. The goal of each agent is to generate a control sequence, using only local observations and neighbor communication, that competes with the best {\it centralized} linear policy in hindsight. We extend the recently proposed Online Spectral Control framework from the centralized setting to the distributed setting. In particular, each agent applies a spectral controller obtained by convolving past disturbances with the leading eigenvectors of a Hankel matrix, while the controller parameters are updated through a distributed online gradient descent step over the local surrogate costs. We formulate this problem this problem as a {\it regret} minimization problem based on the spectral parameterization, and under standard assumptions, we establish a sublinear regret bound of $O(\frac{\sqrt{T}\text{poly}(\log T)}{\gamma^3})$, where $T$ is the time horizon and $\gamma$ denotes the stability margin. The resulting bound also captures the dependence on the network size and connectivity.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 3 linked inside Pith

  1. [1]

    Online linear quadratic control,

    A. Cohen, A. Hasidim, T. Koren, N. Lazic, Y . Mansour, and K. Talwar, “Online linear quadratic control,” inInternational Conference on Machine Learning (ICML), 2018, pp. 1029–1038

    2018

  2. [2]

    On- line control with adversarial disturbances,

    N. Agarwal, B. Bullins, E. Hazan, S. M. Kakade, and K. Singh, “On- line control with adversarial disturbances,” inInternational Conference on Machine Learning (ICML), 2019, pp. 154–165

    2019

  3. [3]

    Logarithmic regret for online control,

    N. Agarwal, E. Hazan, and K. Singh, “Logarithmic regret for online control,” inAdvances in Neural Information Processing Systems (NeurIPS), 2019, pp. 10 175–10 184

    2019

  4. [4]

    A new approach to controlling linear dynamical systems,

    A. Brahmbhatt, G. Buzaglo, S. Druchyna, and E. Hazan, “A new approach to controlling linear dynamical systems,”arXiv preprint arXiv:2504.03952, 2025

    arXiv 2025

  5. [5]

    Automated control of aircraft in formation flight,

    L. Buzogany, M. Pachter, and J. D’azzo, “Automated control of aircraft in formation flight,” inGuidance, Navigation and Control Conference, 1993, p. 3852. 5

    1993

  6. [6]

    Decentralized controllers for unmanned aerial vehicle formation flight,

    J. Wolfe, D. Chichka, and J. Speyer, “Decentralized controllers for unmanned aerial vehicle formation flight,” inGuidance, Navigation, and Control Conference, 1996, p. 3833

    1996

  7. [7]

    A cooperative hunting behavior by mobile robot troops,

    H. Yamaguchi, “A cooperative hunting behavior by mobile robot troops,” inProceedings. 1998 IEEE International Conference on Robotics and Automation (Cat. No. 98CH36146), vol. 4. IEEE, 1998, pp. 3204–3209

    1998

  8. [8]

    A distributed control scheme for multiple robotic vehicles to make group formations,

    H. Yamaguchi, T. Arai, and G. Beni, “A distributed control scheme for multiple robotic vehicles to make group formations,”Robotics and Autonomous systems, vol. 36, no. 4, pp. 125–147, 2001

    2001

Show all 47 references
  1. [9]

    Collaborative navigation and manipulation of a cable-towed load by multiple quadrupedal robots,

    C. Yang, G. N. Sue, Z. Li, L. Yang, H. Shen, Y . Chi, A. Rai, J. Zeng, and K. Sreenath, “Collaborative navigation and manipulation of a cable-towed load by multiple quadrupedal robots,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 10 041–10 048, 2022

    2022

  2. [10]

    Techsat 21: formation design, control, and simulation,

    R. Burns, C. A. McLaughlin, J. Leitner, and M. Martin, “Techsat 21: formation design, control, and simulation,” in2000 IEEE Aerospace Conference. Proceedings (Cat. No. 00TH8484), vol. 7. IEEE, 2000, pp. 19–25

    2000

  3. [11]

    Spacecraft formation flying control using mean orbit elements,

    H. Schaub, S. R. Vadali, J. L. Junkins, and K. T. Alfriend, “Spacecraft formation flying control using mean orbit elements,”The Journal of the Astronautical Sciences, vol. 48, pp. 69–87, 2000

    2000

  4. [12]

    Online convex programming and generalized in- finitesimal gradient ascent,

    M. Zinkevich, “Online convex programming and generalized in- finitesimal gradient ascent,” inProceedings of the 20th international conference on machine learning (icml-03), 2003, pp. 928–936

    2003

  5. [13]

    Cesa-Bianchi and G

    N. Cesa-Bianchi and G. Lugosi,Prediction, learning, and games. Cambridge university press, 2006

    2006

  6. [14]

    Introduction to online convex optimization,

    E. Hazan, “Introduction to online convex optimization,”Foundations and Trends in Optimization, vol. 2, no. 3-4, pp. 157–325, 2016

    2016

  7. [15]

    Implications of regret on stability of linear dynamical systems,

    A. Karapetyan, A. Tsiamis, E. C. Balta, A. Iannelli, and J. Lygeros, “Implications of regret on stability of linear dynamical systems,”arXiv preprint arXiv:2211.07411, 2022

    2022 arXiv

  8. [16]

    On the relation between dynamic regret and closed-loop stability,

    M. Nonhoff and M. A. M ¨uller, “On the relation between dynamic regret and closed-loop stability,”Systems & Control Letters, vol. 177, p. 105532, 2023

    2023

  9. [17]

    Regret bounds for robust adaptive control of the linear quadratic regulator,

    S. Dean, H. Mania, N. Matni, B. Recht, and S. Tu, “Regret bounds for robust adaptive control of the linear quadratic regulator,” in International Conference on Neural Information Processing Systems (NeurIPS), 2018, pp. 4192–4201

    2018

  10. [18]

    Learning linear-quadratic reg- ulators efficiently with only √ Tregret,

    A. Cohen, T. Koren, and Y . Mansour, “Learning linear-quadratic reg- ulators efficiently with only √ Tregret,” inInternational Conference on Machine Learning (ICML). PMLR, 2019, pp. 1300–1309

    2019

  11. [19]

    Logarithmic regret for learning linear quadratic regulators efficiently,

    A. Cassel, A. Cohen, and T. Koren, “Logarithmic regret for learning linear quadratic regulators efficiently,” inInternational Conference on Machine Learning (ICML). PMLR, 2020, pp. 1328–1337

    2020

  12. [20]

    Naive exploration is optimal for online lqr,

    M. Simchowitz and D. Foster, “Naive exploration is optimal for online lqr,” inInternational Conference on Machine Learning (ICML). PMLR, 2020, pp. 8937–8948

    2020

  13. [21]

    Rein- forcement learning with fast stabilization in linear dynamical systems,

    S. Lale, K. Azizzadenesheli, B. Hassibi, and A. Anandkumar, “Rein- forcement learning with fast stabilization in linear dynamical systems,” inInternational Conference on Artificial Intelligence and Statistics. PMLR, 2022, pp. 5354–5390

    2022

  14. [22]

    The power of predictions in online control,

    C. Yu, G. Shi, S.-J. Chung, Y . Yue, and A. Wierman, “The power of predictions in online control,”Advances in Neural Information Processing Systems (NeurIPS), vol. 33, 2020

    2020

  15. [23]

    On the regret analysis of online lqr control with predictions,

    R. Zhang, Y . Li, and N. Li, “On the regret analysis of online lqr control with predictions,” inAmerican Control Conference (ACC), 2021, pp. 697–703

    2021

  16. [24]

    Improper learning for non-stochastic control,

    M. Simchowitz, K. Singh, and E. Hazan, “Improper learning for non-stochastic control,” inConference on Learning Theory (COLT). PMLR, 2020, pp. 3320–3436

    2020

  17. [25]

    Optimal rates for bandit nonstochastic control,

    Y . J. Sun, S. Newman, and E. Hazan, “Optimal rates for bandit nonstochastic control,”Advances in Neural Information Processing Systems, vol. 36, pp. 21 908–21 919, 2023

    2023

  18. [26]

    Second order methods for bandit optimization and control,

    A. Suggala, Y . J. Sun, P. Netrapalli, and E. Hazan, “Second order methods for bandit optimization and control,” inThe Thirty Seventh Annual Conference on Learning Theory. PMLR, 2024, pp. 4691– 4763

    2024

  19. [27]

    Distributed lqr design for identical dynamically decoupled systems,

    F. Borrelli and T. Keviczky, “Distributed lqr design for identical dynamically decoupled systems,”IEEE Transactions on Automatic Control, vol. 53, no. 8, pp. 1901–1912, 2008

    1901

  20. [28]

    Synchronization of autonomous agents by an optimal networked controller,

    A. Mosebach and J. Lunze, “Synchronization of autonomous agents by an optimal networked controller,” inEuropean Control Conference (ECC), 2014, pp. 208–213

    2014

  21. [29]

    Optimal linear-consensus algorithms: An lqr perspective,

    Y . Cao and W. Ren, “Optimal linear-consensus algorithms: An lqr perspective,”IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 40, no. 3, pp. 819–830, 2010

    2010

  22. [30]

    A suboptimality approach to distributed linear quadratic optimal control,

    J. Jiao, H. L. Trentelman, and M. K. Camlibel, “A suboptimality approach to distributed linear quadratic optimal control,”IEEE Trans- actions on Automatic Control, vol. 65, no. 3, pp. 1218–1225, 2020

    2020

  23. [31]

    Distributed q-learning for dynam- ically decoupled systems,

    S. Alemzadeh and M. Mesbahi, “Distributed q-learning for dynam- ically decoupled systems,” inAmerican Control Conference (ACC), 2019, pp. 772–777

    2019

  24. [32]

    Efficient learning of distributed linear-quadratic control policies,

    S. Fattahi, N. Matni, and S. Sojoudi, “Efficient learning of distributed linear-quadratic control policies,”SIAM Journal on Control and Op- timization, vol. 58, no. 5, pp. 2927–2951, 2020

    2020

  25. [33]

    Learning the globally optimal distributed lq regulator,

    L. Furieri, Y . Zheng, and M. Kamgarpour, “Learning the globally optimal distributed lq regulator,” inLearning for Dynamics and Control (L4DC), 2020, pp. 287–297

    2020

  26. [34]

    Distributed online linear quadratic control for linear time-invariant systems,

    T.-J. Chang and S. Shahrampour, “Distributed online linear quadratic control for linear time-invariant systems,” inAmerican Control Con- ference (ACC), 2021, pp. 923–928

    2021

  27. [35]

    Regret analysis of distributed online lqr control for unknown lti systems,

    ——, “Regret analysis of distributed online lqr control for unknown lti systems,”IEEE Transactions on Automatic Control, 2023

    2023

  28. [36]

    Learning linear dynamical systems via spectral filtering,

    E. Hazan, K. Singh, and C. Zhang, “Learning linear dynamical systems via spectral filtering,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017, pp. 6702–6712

    2017

  29. [37]

    Spectral filtering for general linear dynamical systems,

    E. Hazan, H. Lee, K. Singh, C. Zhang, and Y . Zhang, “Spectral filtering for general linear dynamical systems,”Advances in Neural Information Processing Systems, vol. 31, 2018

    2018

  30. [38]

    Dimension-free regret for learning asym- metric linear dynamical systems,

    A. Marsden and E. Hazan, “Dimension-free regret for learning asym- metric linear dynamical systems,”arXiv e-prints, pp. arXiv–2502, 2025

    2025

  31. [39]

    Spectral state space models,

    N. Agarwal, D. Suo, X. Chen, and E. Hazan, “Spectral state space models,”arXiv preprint arXiv:2312.06837, 2023

    2023 arXiv

  32. [40]

    Towards provable control for unknown linear dynamical systems,

    S. Arora, E. Hazan, H. Lee, K. Singh, C. Zhang, and Y . Zhang, “Towards provable control for unknown linear dynamical systems,” 2018

    2018

  33. [41]

    Near-optimal distributed linear-quadratic regulator for networked systems,

    S. Shin, Y . Lin, G. Qu, A. Wierman, and M. Anitescu, “Near-optimal distributed linear-quadratic regulator for networked systems,”SIAM Journal on Control and Optimization, vol. 61, no. 3, pp. 1113–1135, 2023

    2023

  34. [42]

    Economic optimization of distributed embedded battery units for large-scale heating, ventilation, and air conditioning applications,

    N. R. Patel, J. B. Rawlings, M. J. Ellis, M. J. Wenzel, and R. D. Turney, “Economic optimization of distributed embedded battery units for large-scale heating, ventilation, and air conditioning applications,” AIChE Journal, vol. 65, no. 7, p. e16576, 2019

    2019

  35. [43]

    Der entwurf linearer regelungssysteme im zustand- sraum,

    J. Ackermann, “Der entwurf linearer regelungssysteme im zustand- sraum,”at-Automatisierungstechnik, no. 7, pp. 297–300, 1972. 6 V. SUPPLEMENTARY A. Complementary materials Lemma 2:Suppose Assumptions 1, 2 and 3 hold. Then by running Algorithm withη= γ2 mh2√ T , we have∀i, j, t...

    1972

  36. [44]

    Based on Lemma 2, we have |fi,t(xj,A1 t ,u j,A1 t )−f i,t(Mj,t 1:h|A,B,{w})| ≤144G2κ2 Bκ10W 4√n γ3√ T(1−β) log1/2 2 γ . (20)

  37. [45]

    Then by settingM ∗ i =σ 1/4 i K∗H Pd i=1 ϕ⊤ i µαj eje⊤ j

    AsK ∗ ∈ S, we haveA+BK ∗ =HLH −1, whereLis diagonal and can be written asL= Pd l=1 αlele⊤ l , αl ≤ 1 3 , l∈[d]. Then by settingM ∗ i =σ 1/4 i K∗H Pd i=1 ϕ⊤ i µαj eje⊤ j . Then by Lemma 8, we have∥M ∗ i ∥ ≤κ3 q 2 γ , 11 which implies∥M ∗ 1:h∥ ≤κ3 q 2h γ . In addition, based on ...

  38. [46]

    Based on our choice ofhand Lemma 4, we have∀i TX t=1 h fi,t(M∗ 1:h|A,B,{w})−f i,t(xK∗,m t ,u K∗,m t ) i ≤ ϵT 2 . (22)

  39. [47]

    With the selection ofmand Lemma 3, we have∀i TX t=1 h fi,t(xK∗,m t ,u K∗,m t )−f i,t(xK∗ t ,u K∗ t ) i ≤ ϵT 2 (23) Substituting Equations (20), (21), (22) and (23) into Equation (19), we get J j T (A1)−min K JT (K) ≤ n 2η M1 1:h −M ∗ 1:h 2 +T 8ηn√n (1−β) + 2η 6GκBκ5W 2√mh γ2 l...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.