The paper studies learning a control-relevant state from a finite history of observations and actions by predicting cumulative quadratic costs. Two algorithms estimate latent dynamics either directly or through future-cost prediction plus coordinate alignment. The authors claim finite-sample guarantees for stable, partially observed linear–Gaussian systems. Several proof steps and a numerical lemma are incorrect as printed, so the rates below are reported claims rather than independently validated guarantees.
1 Introduction
Observation reconstruction can preserve irrelevant detail; inverse-action prediction targets controllability; cost prediction targets the control objective. This work develops the last approach for stationary, infinite-horizon LQG, extending the time-varying finite-horizon Part I. It connects a classical identification problem to the cost/value-prediction idea behind MuZero.
The key difficulty is statistical dependence within one exploration trajectory. Quadratic regression lifts history vectors into products of Gaussian variables; a new excitation argument and a proposed mixing-based concentration analysis support the claimed rates. The practical insight is that independently learned latent factors can use incompatible coordinates even when they predict identical costs.
This edition covers all 38 pages of v1, dated 8 March 2026. It preserves both complete algorithms, all 184 display containers, the exact assumption/definition/proposition/theorem/lemma statements and all 38 references. There are no empirical experiments or numbered graphical figures in this theoretical paper.
2 Problem setup
The hidden state obeys xₜ₊₁=A*xₜ+B*uₜ+wₜ; observations are yₜ=C*xₜ+vₜ. Process/measurement noises and the initial state are Gaussian. Minimize infinite-horizon time-averaged expected quadratic cost with positive semidefinite state cost Q* and positive definite action cost R*. Dynamics, observation model and noise covariances are unknown; R* is initially treated as known.
Assumption 1 requires open-loop stability, controllability through actions and process noise, observability through measurements and costs, full-rank observation noise, positive action penalties and bounded conditioning. These are substantial restrictions; the result does not cover arbitrary unstable, nonlinear or weakly identifiable systems. Stability/mixing constants still govern history length and sample requirements.
With known parameters, the separation principle combines a Kalman state estimate with the LQR gain obtained from algebraic Riccati equations. The exact optimum uses the recursive filter; the learned finite-history policy will incur truncation error.
is -controllable for some , that is, the controllability matrix
Equation
has rank and .
is -observable for some , that is, the observability matrix
Equation
has rank and .
is -controllable for some .
is -observable for some .
for some ; this can always be achieved by inserting Gaussian noises with full-rank covariance matrices to the observations.
for some .
The operator norms of , , , , , , , and are ,
the singular value lower bounds , , , , , and spectral radii are , where with defined in (2.5).
Equations 2.4–2.8
(2.4)
(2.5)
(2.6)
Original selectable MathML
(2.7)
(2.8)
2.1 Latent model of infinite-horizon time-invariant LQG
Assume the Kalman error covariance is stationary at initialization. Proposition 1 expresses filtered states as a linear system driven by independent innovations; observed cost equals a quadratic function of the estimate plus a policy-independent constant and centered noise. The constant can be ignored when choosing actions, but the noise matters for learning.
Unroll the filter and retain H recent observations and actions. This defines z*ₜ=M*hₜ+δₜ, where δₜ decays with the stable filter dynamics. History dimension is dₕ=H(dᵧ+dᵤ). The policy combines a representation matrix M with latent feedback K. Initial H steps can use zero/Gaussian inputs because the open-loop system is assumed stable. Finite history approximates, rather than equals, the optimal recursive estimator.
Equation 2.9
(2.9)
Proposition 1.
Let be the initial state estimate
and be the state estimates given by the Kalman filter. Then, for ,
Equation
where is independent of and , i.e., the state estimates follow the same linear dynamics with noises .
The cost at step can be reformulated as a function of the state estimates by
Equation
where , and , are both zero-mean subexponential random variables. Moreover, and ; if control for , then we have .
Equations 2.10
(2.10)
3 Method
Collect one trajectory using independent zero-mean Gaussian exploration actions. Algorithm 1 regresses cumulative state costs on quadratic history features, factors the result into a latent representation, estimates dynamics and one-step cost, projects the cost matrix to be positive semidefinite, then solves the latent Riccati equation. Multiple independent trajectories can also supply the data.
The printed data range ends its costs at T+H−1, but its final dₓ-step target needs costs through T+H+dₓ−2; CoSysId’s shifted target needs one further step. Implementation therefore requires a longer trajectory or a shortened valid regression range. The original algorithms are retained without silently changing their indices.
Algorithm 1 Cost-driven state representation learningOriginal selectable source
1:Input: length , history length , noise magnitude
2:
Collect trajectories of length using , for , to obtain
Equation 3.1
(3.1)
3:
Estimate the state representation function and cost constants by solving
Equation 3.2
(3.2)
where and
4:
Find
5:
Compute for all , so that the data are converted to :
Equation
6:
Run SysId (3.4) or CoSysId (Algorithm 2) to obtain system dynamics matrices
7:
Estimate the cost function by solving
Equation 3.3
(3.3)
8:
Truncate negative eigenvalues of to to obtain
9:
Find feedback gain from by solving DARE (2.8) and (2.6)
10:Return: policy
3.1 Cost-driven representation function learning
The dₓ-step cost observability Gramian is positive definite under cost observability. In normalized coordinates it becomes identity, so regressing cumulative costs estimates N*=M*ᵀM*. Subtract known action penalties to reduce target variance; the paper proposes learning R* later if it is unknown.
Factor the estimated symmetric matrix using its largest dₓ nonnegative eigenvalues. The resulting representation is identifiable only up to an orthogonal transformation. Proposition 2 uses action controllability to lower-bound latent covariance and argues that sufficiently long history makes M* full row rank, with smallest singular value of order ν/√H. An eigenvalue threshold can then reveal latent rank once estimation error is sufficiently small.
The text says dₓ need not be known, but the printed cumulative-cost horizon and factorization already use dₓ. An upper bound or adaptive horizon procedure would be needed to operationalize that claim; it is not specified. The normalization also changes coordinates, so raw-state recovery is not the target.
Equations
Proposition 2.
Under i.i.d. actions for , for , where is defined in Assumption 1.2.
Recall that for a square matrix , we define .
There exists a dimension-free constant , such that as long as , has rank and .
Equations
3.2 Explicit learning of system dynamics
CoReL-E fits A and B by least squares from estimated latent states and actions to the next estimated state. This follows the explicit transition-prediction route. Both input and output estimates contain learned-representation and truncation errors; Lemma 5 is intended to propagate those errors through the regression. No observation-reconstruction matrix C* needs to be learned.
Equation 3.4
(3.4)
3.3 Implicit learning of system dynamics (MuZero-style)
CoReL-I instead predicts current and one-step-ahead cumulative costs. Lifting MᵀM and [AM,B]ᵀ[AM,B] turns each cost fit into quadratic-feature linear regression. Separate intercepts decouple the two regressions. After factorization, estimate an alignment matrix between the current-state and predicted-next-state coordinates, then use aligned dynamics in the controller.
This alignment is necessary because cost norms are unchanged by orthogonal rotations, and the two factorizations need not choose the same rotation. Longer future-cost prediction is a suggested alternative, but the analyzed algorithm explicitly performs alignment rather than proving that arbitrary multi-step MuZero training resolves it.
The discussion rejecting a lower-dimensional regression contains a rank error: controllability of (A*,B*) implies [A*,B*] has full row rank. If M* also has full row rank, it has a right inverse and [A*M*,B*] has the same rank as [A*,B*]. The claimed rank distinction cannot justify choosing one regressor over the other under these assumptions.
Equations 3.5–3.6
(3.5)
(3.6)
Algorithm 2CoSysId: Cost-driven system identificationOriginal selectable source
1:Input: data from Algorithm 1, representation function parameter
2:
Estimate the system dynamics by
Equation 3.7
(3.7)
3:
Find
4:
Split to at column and set .
5:
Find alignment matrix by solving
Equation 3.8
(3.8)
6:Return: system dynamics estimate
Equation
4 Theoretical guarantees and proofs
Theorem 1 claims an O(poly(H,dₓ,dᵧ,dᵤ,log(T/p))/T) average-cost gap with probability at least 1−p, after a polynomial sample threshold and H growing roughly as log(T/p)/log(1/ρ). The intended estimation errors scale as T⁻¹ᐟ² up to conditioning, dimension and logarithmic factors; the cost gap is quadratic in small controller error.
Compared with observation-based system identification, the authors acknowledge worse dimension dependence and a longer burn-in because scalar quadratic costs require a much larger feature space. The result is not a dimension-free practical sample budget. The proof issues documented below prevent treating the printed statement as fully established by this version.
Theorem 1.
Given an unknown LQG control problem defined by (2.1) and (2.2), under Assumption 1,
for a given , if we run CoReL-E (Algorithm 1 with (3.4)) or CoReL-I (Algorithm 1 with Algorithm 2) for , ,
and , then with probability at least , the learned representation function is -optimal,
and the overall output policy satisfies
Equation
4.1 Proposition on multi-step cumulative costs
Proposition 3 aims to decompose dₓ-step cumulative state cost into squared latent norm, a constant, exponentially small history-truncation error and centered subexponential noise. It then seeks a √T-type bound for the sum of quadratic features times noise despite temporal dependence.
The proposed argument splits distant time indices into groups, conditionally centers each feature/noise product, and bounds the remaining conditional mean through mixing. However, the first cost expansion drops cross terms pointwise because the summands are independent. Independence removes their expectation, not their realized values; these terms must be included and controlled in the noise decomposition.
Conditional centering also creates martingale differences, not automatically independent variables as the proof states. Its cited concentration step needs a matching dependence argument. A later error recursion prints +vₜ₊₁ where direct algebra gives −L*vₜ₊₁. All source displays are retained; these observations identify gaps in the printed derivation rather than supplying a repaired proof.
Proposition 3.
Given an LQG control problem satisfying Assumption 1, let parameterize the optimal state representation function under the normalized parameterization.
Recall that and .
If we apply , then for any ,
Equation
where is a small error term, is a positive constant, and is a zero-mean subexponential random variable with . Moreover, let . There exists a dimension-free constant such that for a given , as long as , satisfy
Equation
Equations 4.1–4.5
(4.1)
(4.2)
(4.3)
(4.4)
(4.5)
4.2 Persistency of excitation
Quadratic regression is linear in the lifted feature [svec(hhᵀ);1], whose dimension grows quadratically in history length. The block martingale small-ball method asks whether every feature direction has a nontrivial chance of being excited over a time block. Lemma 1 claims a Gram-matrix lower bound of order T/dₕ⁹ after T of order H dₕ⁸ log(dₕ/p), subject to history/mixing conditions.
The proof conditions future histories on the past, controls their nonzero conditional mean, applies a lower bound for Gaussian quadratic forms, then uses Paley–Zygmund and a covering argument. Its conversion of high-probability bounds into the almost-sure small-ball condition and its dimension-free Gaussian-norm moment estimates require additional justification.
Lemma 2’s numeric constant is false already in one dimension. For standard normal Z and unit vector (1,−1)/√2, E|Z²−1|/√2=4φ(1)/√2≈0.6844, below the claimed 0.8; φ is the standard-normal density. Its proof incorrectly lower-bounds an absolute first moment by a variance. This counterexample concerns the stated bound, not a proof that every possible lower bound or final rate is impossible.
Lemma 3 generalizes to arbitrary covariance. Its proof ends with a factor min(λmin(Σ),1), but its statement omits the cap. Taking the quadratic matrix to be zero and the intercept to be one shows why a uniform lower bound growing without limit with λmin(Σ) cannot hold. The original statements and displays remain visible below.
Let be a stochastic process in adapted to the filtration . We say satisfies the -BMSB condition for , and , if for any , for any fixed unit vector ,
almost surely.
Lemma 1.
Let be the -step history at time step in system (2.1) with for .
Define filtration with
.
Define and , adapted to .
Recall that for square matrix , .
There exist dimension-free constants , such that for a given , as long as and , we have
Equation
Equations
Lemma 2.
Let be independent standard Gaussian random variables. Let be a -dimensional unit vector. Then,
Equation
Equations
Lemma 3.
Let be a -dimensional zero-mean Gaussian random vector with covariance . For any symmetric matrix and constant that satisfy , there exists an absolute constant , such that .
Equations
4.3 Quadratic regression bound
Lemma 4 treats quadratic regression with perturbed Gaussian covariates, allowing the perturbations to correlate with the original variables. It assumes an excitation lower bound and explicitly retains the accumulated feature–noise product instead of assuming independent noise. The intended error separates regressor perturbation, perturbation–noise interaction and the ordinary regression noise term.
The proof writes normal equations in the lifted space, uses singular-value perturbation to keep the sample Gram matrix invertible, and exploits the rank-two difference between two outer products. Its role is modular: a valid concentration bound for the retained noise sum is still required from the preceding analysis.
Lemma 4.
Let be a sequence of -dimensional Gaussian random vectors adapted to the filtration with for all .
Define random variable , where is a positive semidefinite matrix and is a constant.
Assume and are .
Define , where the perturbation vector can be correlated with and its -norm is sub-Gaussian with , .
Define and .
Assume that satisfies for .
Consider
Equation 4.6
(4.6)
There exsits an absolute constant , such that for a given , under the condition that ,
with probability at least ,
Equation
Original selectable MathML
Equations 4.7–4.9
(4.7)
(4.8)
(4.9)
4.4 Perturbed linear regression bound
Lemma 5 extends linear identification to errors in both inputs and outputs. With small correlated perturbations, a small-ball condition and conditionally centered Gaussian innovation, it bounds estimation error by perturbation terms plus a T⁻¹ᐟ² noise term. The proof first bounds the estimated operator coarsely, then refines the normal-equation estimate.
The assumptions on perturbation size, sample count and the conditioning constants are essential. In applying this lemma, the conditioning filtration must match the claim that the innovation has zero conditional mean.
Lemma 5.
Let be a sequence of -dimensional Gaussian random vectors adapted to the filtration
with for all .
Define , where and is Gaussian with zero mean and .
Define and , where the perturbation vectors and can be correlated with and , and their -norms are sub-Gaussian with , and , .
Assume that satisfies the -BMSB condition, and that are and are .
Consider
Equation 4.10
(4.10)
Then, there exist absolute constants , such that for a given , under the condition that , and ,
with probability at least ,
Equation
Equations 4.11–4.12
(4.11)
(4.12)
Original selectable MathML
4.5 Stable linear system under small perturbations
Lemma 6 controls the stationary covariance of a stable linear system under small zero-mean perturbations that may correlate with the state but remain independent of fresh process noise. A Lyapunov recursion and induction bound covariance uniformly, then show an O(ε) stationary deviation after initial transients decay.
The permitted perturbation depends on the stability margin, spectral-transient factor and system norm. This is a local robustness result for the stated linear–Gaussian setting, not unconditional stability of arbitrary learned controllers.
Lemma 6.
Consider a linear system , , where is a zero-mean Gaussian random vector and are i.i.d. sampled from with independent of for each .
For , perturbation is a zero-mean Gaussian random vector that is independent of and may be correlated with ; let denote the covariance matrices of .
Assume that the operator norms of , , and are and ; assume that for , , where and satisfies .
Let denote the covariance matrix of the stationary distribution of without perturbations.
Recall that for square matrix , we define .
Then,
for large enough such that , we have .
Equations
4.6 Proof of Theorem 1
The intended proof pipeline is: use excitation and cumulative-cost concentration to estimate N*, factor it to recover M* up to rotation, fit the latent cost and dynamics, align implicit-model coordinates, then apply certainty-equivalent LQR perturbation analysis. Positive-semidefinite projection preserves a usable cost matrix. Pseudoinverse perturbation invokes Wedin; factor recovery invokes Procrustes-flow analysis.
For explicit dynamics, the proof correctly switches from a filtration containing the hidden state to one containing observations/actions, so Kalman innovations are conditionally centered. For implicit dynamics, however, A*zₜ+B*uₜ predicts the conditional next state and omits the fresh innovation. The alignment argument treats its difference from zₜ₊₁ as vanishing estimation error and sets regression noise to zero; that nonvanishing innovation needs separate treatment.
A further controller step replaces history h with M*†(z*−δ). For rectangular full-row-rank M*, M*†M* is a projector, not the identity. The omitted component (M̂−SM*)(I−M*†M*)h need not vanish. Consequently the displayed effective-feedback rewrite is not justified by full row rank alone.
The final claimed bound is O(H²² dₓ²³(dᵧ+dᵤ)²¹dᵤ T⁻¹ log⁵(T/p)), with exponential truncation made smaller by choosing H. Those large powers and the unresolved proof steps should accompany the headline O(1/T) claim. No corrected exponents or theorem are asserted here.
Equations 4.13–4.18
(4.13)
(4.14)
(4.15)
(4.16)
(4.17)
(4.18)
5 Additional discussion on MuZero
The analogy to MuZero is selective: both use history-based representations and predict task-relevant future quantities through latent dynamics. Here histories contain actions as well as observations; exploration is fixed Gaussian noise; the target is cumulative cost under that exploration; and policy computation uses Riccati equations rather than Monte Carlo tree search. Optimal-action/value supervision from online search is not analyzed.
Value prediction networks and earlier frame-stacking approaches provide related precedents. This paper is a linear-control case study, not a finite-sample guarantee for MuZero on games, images or general nonlinear systems.
6 Concluding remarks
The proposed framework separates cost-driven state learning, explicit/implicit dynamics estimation and control synthesis. Coordinate compatibility is a useful design issue even when scalar costs are fit accurately. The statistical theorem needs correction or further justification at the identified steps.
The authors leave recursive representations, nonlinear observations/dynamics, transfer between cost functions and changing or multiple camera views to future work. The paper supplies no empirical evidence that these extensions work or that its very large worst-case polynomial bounds predict practical performance.
Acknowledgment
The authors acknowledge NSFTRIPODS, Simons-Berkeley, the U.S. Army Research Office and an NSFCAREER award.
38 references
Bertsekas (2012)
Dimitri Bertsekas.
Dynamic Programming and Optimal Control: Volume I, volume 1.
Athena Scientific, 2012.
Fazel et al. (2018)
Maryam Fazel, Rong Ge, Sham Kakade, and Mehran Mesbahi.
Global convergence of policy gradient methods for the linear quadratic regulator.
In International Conference on Machine Learning, pages 1467–1476. PMLR, 2018.
Fu et al. (2021)
Xiang Fu, Ge Yang, Pulkit Agrawal, and Tommi Jaakkola.
Learning task informed abstractions.
In International Conference on Machine Learning, pages 3480–3491. PMLR, 2021.
Hafner et al. (2019a)
Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi.
Dream to control: Learning behaviors by latent imagination.
arXiv preprint arXiv:1912.01603, 2019a.
Hafner et al. (2019b)
Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson.
Learning latent dynamics for planning from pixels.
In International conference on machine learning, pages 2555–2565. PMLR, 2019b.
Hafner et al. (2020)
Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba.
Mastering Atari with discrete world models.
arXiv preprint arXiv:2010.02193, 2020.
Hafner et al. (2023)
Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap.
Mastering diverse domains through world models.
arXiv preprint arXiv:2301.04104, 2023.
Jadbabaie et al. (2021)
Ali Jadbabaie, Horia Mania, Devavrat Shah, and Suvrit Sra.
Time varying regression with hidden linear dynamics.
arXiv preprint arXiv:2112.14862, 2021.
Komaroff (1994)
N Komaroff.
Iterative matrix bounds and computational solutions to the discrete algebraic Riccati equation.
IEEE Transactions on Automatic Control, 39(8):1676–1678, 1994.
Lale et al. (2020)
Sahin Lale, Kamyar Azizzadenesheli, Babak Hassibi, and Anima Anandkumar.
Logarithmic regret bound in partially observable linear dynamical systems.
Advances in Neural Information Processing Systems, 33:20876–20888, 2020.
Lale et al. (2021)
Sahin Lale, Kamyar Azizzadenesheli, Babak Hassibi, and Anima Anandkumar.
Adaptive control and regret minimization in linear quadratic Gaussian (LQG) setting.
In 2021 American Control Conference (ACC), pages 2517–2522. IEEE, 2021.
Lamb et al. (2022)
Alex Lamb, Riashat Islam, Yonathan Efroni, Aniket Didolkar, Dipendra Misra, Dylan Foster, Lekan Molu, Rajan Chari, Akshay Krishnamurthy, and John Langford.
Guaranteed discovery of controllable latent states with multi-step inverse models.
arXiv preprint arXiv:2207.08229, 2022.
Mania et al. (2019)
Horia Mania, Stephen Tu, and Benjamin Recht.
Certainty equivalence is efficient for linear quadratic control.
Advances in Neural Information Processing Systems, 32, 2019.
Mhammedi et al. (2020)
Zakaria Mhammedi, Dylan J Foster, Max Simchowitz, Dipendra Misra, Wen Sun, Akshay Krishnamurthy, Alexander Rakhlin, and John Langford.
Learning the linear quadratic regulator from nonlinear observations.
Advances in Neural Information Processing Systems, 33:14532–14543, 2020.
Mnih et al. (2013)
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller.
Playing Atari with deep reinforcement learning.
arXiv preprint arXiv:1312.5602, 2013.
Mnih et al. (2015)
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al.
Human-level control through deep reinforcement learning.
nature, 518(7540):529–533, 2015.
Oymak and Ozay (2019)
Samet Oymak and Necmiye Ozay.
Non-asymptotic identification of LTI systems from a single trajectory.
In 2019 American control conference (ACC), pages 5655–5661. IEEE, 2019.
Pathak et al. (2017)
Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell.
Curiosity-driven exploration by self-supervised prediction.
In International conference on machine learning, pages 2778–2787. PMLR, 2017.
Schrittwieser et al. (2020)
Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, et al.
Mastering Atari, Go, chess and shogi by planning with a learned model.
Nature, 588(7839):604–609, 2020.
Silver et al. (2016)
David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al.
Mastering the game of Go with deep neural networks and tree search.
nature, 529(7587):484–489, 2016.
Silver et al. (2017)
David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al.
Mastering the game of Go without human knowledge.
nature, 550(7676):354–359, 2017.
Silver et al. (2018)
David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al.
A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play.
Science, 362(6419):1140–1144, 2018.
Simchowitz et al. (2018)
Max Simchowitz, Horia Mania, Stephen Tu, Michael I Jordan, and Benjamin Recht.
Learning without mixing: Towards a sharp analysis of linear system identification.
In Conference On Learning Theory, pages 439–473. PMLR, 2018.
Simchowitz et al. (2019)
Max Simchowitz, Ross Boczar, and Benjamin Recht.
Learning linear dynamical systems with semi-parametric least squares.
In Conference on Learning Theory, pages 2714–2802. PMLR, 2019.
Subramanian et al. (2020)
Jayakumar Subramanian, Amit Sinha, Raihan Seraj, and Aditya Mahajan.
Approximate information state for approximate planning and reinforcement learning in partially observed systems.
arXiv preprint arXiv:2010.08843, 2020.
Tian et al. (2022)
Yi Tian, Kaiqing Zhang, Russ Tedrake, and Suvrit Sra.
Cost-driven representation learning for linear quadratic Gaussian control: Part I.
arXiv preprint arXiv:2212.14511v4, 2022.
Tsiamis et al. (2022)
Anastasios Tsiamis, Ingvar Ziemann, Nikolai Matni, and George J Pappas.
Statistical learning theory for control: A finite sample perspective.
arXiv preprint arXiv:2209.05423, 2022.
Tu et al. (2016)
Stephen Tu, Ross Boczar, Max Simchowitz, Mahdi Soltanolkotabi, and Ben Recht.
Low-rank solutions of linear matrix equations via Procrustes flow.
In International Conference on Machine Learning, pages 964–973. PMLR, 2016.
Ye et al. (2021)
Weirui Ye, Shaohuai Liu, Thanard Kurutach, Pieter Abbeel, and Yang Gao.
Mastering Atari games with limited data.
Advances in Neural Information Processing Systems, 34:25476–25488, 2021.