Pith. sign in

REVIEW 2 major objections 5 minor 42 references

Learning-based primal-dual optimal control of discrete-time stochastic systems with multiplicative noise

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

Pith's one-line read This paper claims that the multiplicative-noise stochastic LQR can be recast as a non-convex primal-dual optimization problem with strong duality, whose KKT conditions yield a learning algorithm that converges when the drift dynamics are…

desk verdict A solid model-based primal-dual extension of LQR to multiplicative-noise systems, with an unproven partially model-free algorithm that the title oversells. read the letter →

arxiv 2506.02613 v1 pith:CN5CHXFY submitted 2025-06-03 math.OC

classification math.OC MSC 93E2090C4649K4593C55
keywords stochasticlinearquadraticregulatormultiplicativenoiseprimal-dualoptimizationstrongdualityKarush-Kuhn-Tuckerconditionsreinforcementlearningpolicyiterationmodel-freecontrol
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 the infinite-horizon stochastic linear quadratic regulator with multiplicative noise can be solved by a primal-dual optimization method even when the drift dynamics are unknown. The authors reformulate the optimal control problem as a non-convex constrained program in an augmented state, prove strong duality, and derive the KKT conditions. From the KKT system they obtain a model-based primal-dual algorithm whose iterates provably converge to the optimal Q-function and the optimal feedback gain, plus a partially model-free version that estimates the needed covariance-like data from sampled trajectories. The practical payoff would be a Q-learning-style controller for multiplicative-noise systems with rigorous convergence guarantees and no artificial excitation requirement.

What carries the argument

The load-bearing objects are the augmented state $v_k=[x_k;u_k]$ and the generalized Lyapunov operator $\mathcal{D}_{A_F,C_F}X=A_F X A_F' + C_F X C_F'$. Generalized Lyapunov equations in $S$ and $X$ encode, respectively, the state second moment and the Q-function. The stochastic PBH eigenvector test from the paper's cited theory is used to replace the hard stability constraint $F\in\mathcal{F}$ by $S\succeq 0$, which is the step that makes the primal-dual reformulation possible. Strong duality then yields the KKT system, and the equivalence between the dual/primal updates and off-line policy iteration is what transfers the convergence result to the learning algorithms.

What would settle it

Construct a pair $(A,C,Q)$ that is exactly detectable in the finite-window sense of Definition 2 but for which there exists a nonzero symmetric $X$ solving $\mathcal{D}_{A,C}X=\lambda X$ with $CX=0$ and $|\lambda|\ge 1$; finding such an example would invalidate Proposition 2's use of the PBH test, breaking the replacement of the stability constraint by $S\succeq 0$.

Watch

Extended reading notes

Core claim

For the system $x_{k+1}=Ax_k+Bu_k+(Cx_k+Du_k)w_k$ with cost $E\sum_{k=0}^\infty (x_k'Qx_k+u_k'Ru_k)$, the control-design problem is equivalent to a primal-dual pair of matrix optimization problems. Using the augmented vector $v_k=(x_k,u_k)$, the primal problem minimizes $\operatorname{Tr}(\Lambda S)$ subject to the generalized Lyapunov equation $A_F S A_F' + C_F S C_F' + [I;F]Z[I;F]' = S$, with the stabilization constraint replaced by $S\succeq 0$. The paper proves strong duality, and the resulting KKT conditions give the dual Lyapunov equation together with the gain update $F=-(X_{22})^{-1}X_{12}'$. The dual variable $X$ converges to the optimal Q-function $X^*$, and Algorithm 1 is shown to be identical to off-line policy iteration, which transfers the convergence proof. Algorithm 2 solves the same system from data when $A$ and $B$ are unknown but $C$ and $D$ are known, and the included motor-control simulation shows the learned gain and Q-function approaching the model-based optimal values.

Load-bearing premise

The PBH eigenvector test, imported from an established theory for infinite-horizon exact observability and detectability, is used under the paper's relaxed finite-window observability definition without a proof that it still holds.

Editorial extensions

If this is right

  • Theorem 3 gives the model-based certificate: the dual and primal iterates obey the same equations as off-line policy iteration, so $\lim_{i\to\infty} X^i = X^*$ and $\lim_{i\to\infty} F^i = F^*$.
  • Because the dual variable solves the same generalized Lyapunov equation as the Q-function, the algorithm supplies a convergence foundation for Q-learning in multiplicative-noise systems.
  • When the drift matrices $A$ and $B$ are unknown, Algorithm 2 needs only the noise-term matrices $C$ and $D$, and the 15-run Monte Carlo example reaches relative errors near $10^{-2}$ for the gain and $10^{-3}$ for the Q-function by the fourth iteration.
  • The deterministic LQR primal-dual results are recovered as a special case, and the same KKT route can be inspected for the fully model-free and indefinite SLQR problems that the paper leaves open.
  • The arm-movement simulation indicates that the learned controller reproduces a plausible sensorimotor strategy, giving an illustrative bridge between the algorithm and biological motor learning.

Reading between the lines

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

  • The finite-window relaxation in Definition 2 is where the argument is most exposed: if the PBH test requires the infinite-horizon observation window, Proposition 2's replacement of the stability constraint by $S\succeq 0$ could fail, and with it the strong-duality and KKT route.
  • Algorithm 2 estimates $\tilde{S}(F^i)$ and $W(F^i)$ from $H$ sample paths over $M$ steps, but the paper gives no sample-complexity or bias bound; a natural test is whether convergence is maintained as $M$ grows and whether truncation bias vanishes.
  • Because the drift matrices $A$ and $B$ never appear explicitly in the data matrices $\tilde{S}(F^i)$ and $W(F^i)$, the same scheme could plausibly be extended to the fully model-free case by estimating the $C$ and $D$ contributions from data; the paper states this case remains unsolved.
  • The motor-control simulation is illustrative rather than a neural claim; treating it as evidence about the central nervous system would require a separate argument that biological circuitry actually solves these KKT equations.
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 the discrete-time stochastic linear quadratic regulator (SLQR) with multiplicative noise from a primal-dual optimization viewpoint. It reformulates the SLQR as a non-convex primal problem over the pair (S,F), introduces a dual problem, proves strong duality, derives KKT conditions, and gives a model-based primal-dual algorithm (Algorithm 1) that is shown to be equivalent to off-line policy iteration and hence convergent. It then proposes a partially model-free algorithm (Algorithm 2) for the case where the drift matrices A and B are unknown, using empirical estimates of second-moment matrices from Monte Carlo sample paths, and illustrates it on a human-arm movement example. The central theoretical claims are the strong duality result, the KKT derivation, and the convergence of the model-based primal-dual iterations; the central practical claim is that Algorithm 2 provides a partially model-free SLQR policy design with a rigorous foundation.

Significance. If fully established, the paper would be a useful extension of the primal-dual LQR framework of [18] to multiplicative-noise stochastic systems, and it would connect the classical off-line policy iteration to the Q-function and to a largely model-free implementation. The model-based results, especially Theorem 3 and its equivalence to Lemma 4, are straightforward and appear sound; the strong duality theorem follows the structure of [18] and is credible modulo the missing support discussed below. The paper also gives a concrete simulation with honest error plots. However, the data-driven Algorithm 2 has no convergence or consistency analysis, and the PBH-based replacement of the stabilizing constraint by a semidefinite condition is not fully justified under the paper's relaxed observability/detectability definition. These gaps concern the two load-bearing elements of the claimed contribution, so the significance of the paper in its present form is constrained.

major comments (2)
  1. [§3.2, Algorithm 2 and Theorem 3] The paper's central practical claim is a partially model-free algorithm when A and B are unknown, but no convergence or consistency theorem is proved for Algorithm 2. Theorem 3 covers only Algorithm 1, whose dual update (56) and primal update (57) are exact model-based steps; the proof works by identifying Algorithm 1 with the off-line policy iteration of Lemma 4. Algorithm 2 replaces ~S(F_i) and W(F_i) with finite-horizon Monte Carlo averages over H sample paths and solves the perturbed equation (59). There is no analysis of how the estimation error propagates into X^i or F^{i+1}, no proof that F^{i+1} remains in the stabilizing set F, and no statement that F^i converges to F*. The abstract's 'rigorous convergence analysis for RL design' and contribution (2) therefore outrun the proved results. This claim must either be proved for Algorithm 2 or substantially weakened.
  2. [Definition 2, Lemma 2, Proposition 2] Lemma 2 is cited from Theorem 3.7 of [37], where exact observability and exact detectability are defined by requiring y_k ≡ 0 for all k in N, whereas Definition 2 relaxes the condition to k in N_T for a fixed T. The paper gives no proof that the PBH eigenvector test remains valid under this relaxed definition. Proposition 2 then uses Lemma 2 to replace the constraint F in F by S ≽ 0; this replacement is the key step that makes the primal-dual formulation and the KKT framework valid. The same issue affects Proposition 3, which invokes Lemma 2 in a similar way for the pair (A_F, C_F | Λ). Please provide a direct proof of Lemma 2 under Definition 2, or show that finite-horizon and infinite-horizon exact observability/detectability coincide on the second-moment lift, and apply the same fix wherever Lemma 2 is used.
minor comments (5)
  1. [Remark 3] The claim of quadratic convergence of {P(i)} is asserted without proof ('Following the line of Theorem 2.4.1 of [39], it is easy to show'). If this is a contribution, please give a proof or replace the claim by a weaker statement.
  2. [Lemma 2 and Proposition 2] In Lemma 2 the condition 'CX = 0' is inconsistent with the notation in Definition 2 and with the usage in Proposition 2, where the condition is [I; F] Z [I; F]' X = 0; it should presumably read 'QX = 0' (or '\tilde Q X = 0'). Also, '|λ| ≽ 1' for complex λ is undefined and should be '|λ| ≥ 1'.
  3. [Proposition 4 proof] There is a typo in the proof: 'Combing (50), (50) and (51)' should be 'Combining (50) and (51)', and the derivation that X_0* = 0 from trace( ilde S* X_0*) = 0 and ilde S* ≻ 0, X_0* ≽ 0 should be stated explicitly.
  4. [Section 3.1, dual problem statement] The text contains a LaTeX leftover '/suppress L(X, F, S)' before the definition of the Lagrangian, which should be removed.
  5. [Section 4 simulation] Please clarify whether the reported F(7) is the output of Algorithm 2 or of the model-based PI, and state the time-truncation horizon M and the number of sample paths H used in the Monte Carlo experiment, since these are free parameters of Algorithm 2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the primal-dual reformulation and convergence are derived from stated assumptions and background theorems rather than from fitted or assumed conclusions.

full rationale

The paper does not fit parameters to data and then call the same quantities predictions, nor does it define a key object in terms of the conclusion. The reformulation of SLQR as (P1-SLQR) is an exact second-moment lift (Proposition 1), and the KKT conditions in Proposition 4 are obtained from the Lagrangian and the strong-duality theorem proved in the paper. Algorithm 1's convergence is shown in Theorem 3 by identifying its dual/primal updates with the policy evaluation and improvement steps of Lemma 4, which is proved in Section 2.2; this is an equivalence argument, not a tautology. Algorithm 2 is empirical and no convergence theorem is supplied, but that is a missing proof/overclaim, not a circular reduction. The PBH and generalized Lyapunov results are cited from [37] (co-authored by the corresponding author), and the finite-window relaxation in Definition 2 creates a possible gap because Lemma 2 is quoted for the original definition; however, this is a correctness risk, not circularity, since the cited results are stated with assumptions that do not include the SLQR optimality result and are not the renamed output of the present derivation. No equation is shown to reduce to an earlier assumed equation by construction.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper does not introduce new physical entities or fitted physical constants. The only hand-chosen numbers are simulation hyperparameters. The main load-bearing axioms are the standard SLQR assumptions, the unproven validity of the PBH test under the relaxed detectability definition, and the statistical accuracy of the sample-based estimates in the model-free algorithm.

free parameters (3)
  • H (Monte Carlo sample paths) = 15
    Chosen for the numerical example in Section 4; not part of the theoretical convergence proof and does not affect the central claim.
  • M (time truncation horizon for S and W) = not specified
    Used in Algorithm 2 to construct finite-horizon estimates of S and W; the value is not given and its effect on accuracy is not analyzed.
  • Initial feedback gain F(0) = 2x6 matrix in Section 4
    Handpicked stabilizing initial gain in the simulation; a different stabilizing F(0) would still be valid, so this is not a fitted constant in the derivation.
assumptions (3)
  • domain assumption Assumption 1: the system [A,B;C,D] is stabilizable, Q >= 0, R > 0, and (A,C|Q) is exactly observable or exactly detectable.
    Standard SLQR assumptions that guarantee the GARE has a unique solution and the optimal feedback gain exists. Used throughout Section 2 and 3.
  • ad hoc to paper The relaxed exact observability/detectability definitions in Definition 2 (finite horizon N_T) preserve the PBH eigenvector test of Lemma 2.
    The paper modifies the definition from [37] in Remark 2 but then cites Lemma 2 from [37] without proving that the PBH test still holds under the relaxed definition. Proposition 2 depends on this to replace the stability constraint with S >= 0.
  • domain assumption The finite-sample estimates of S(F^i) and W(F^i) in Algorithm 2 are accurate enough for the dual update to produce a valid X^i.
    Algorithm 2 relies on replacing exact expectations with Monte Carlo averages over H paths and a finite horizon M. No statistical convergence analysis is provided, so the algorithm's correctness depends on this approximation being good.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-based primal-dual optimal control of discrete-time stochastic systems with multiplicative noise." pith.science (2026). https://pith.science/paper/CN5CHXFY

@misc{pith2026250602613,
  author       = {Pith},
  title        = {Pith review of: Learning-based primal-dual optimal control of discrete-time stochastic systems with multiplicative noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CN5CHXFY}},
  note         = {Machine review of arXiv:2506.02613}
}
read the original abstract

Reinforcement learning (RL) is an effective approach for solving optimal control problems without knowing the exact information of the system model. However, the classical Q-learning method, a model-free RL algorithm, has its limitations, such as lack of strict theoretical analysis and the need for artificial disturbances during implementation. This paper explores the partially model-free stochastic linear quadratic regulator (SLQR) problem for a system with multiplicative noise from the primal-dual perspective to address these challenges. This approach lays a strong theoretical foundation for understanding the intrinsic mechanisms of classical RL algorithms. We reformulate the SLQR into a non-convex primal-dual optimization problem and derive a strong duality result, which enables us to provide model-based and model-free algorithms for SLQR optimal policy design based on the Karush-Kuhn-Tucker (KKT) conditions. An illustrative example demonstrates the proposed model-free algorithm's validity, showcasing the central nervous system's learning mechanism in human arm movement.

Figures

Figures reproduced from arXiv: 2506.02613 by the authors.

Figure 1
Figure 1. The curves of relative error between the learned co [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗
Figure 2
Figure 2. The curves of the relative error between the learne [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. The curves of each element for learned control gain [PITH_FULL_IMAGE:figures/full_fig_p028_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [18]

    Lee, D., & Hu, J. (2019). Primal-dual Q-learning framew ork for LQR design. IEEE Trans- actions on Automatic Control , 64(9), 3756-3763

  2. [37]

    Zhang, W., Xie, L., & Chen, B. S. (2017). Stochastic H2/H∞ Control: A Nash Game Approach. CRC Press

  3. [1]

    Anderson, B. D. O., & Moore, J. B. (1989). Optimal Control-Linear Quadratic Methods . Prentice-Hall, New York

  4. [2]

    P., & Tsitsiklis, J

    Bertsekas, D. P., & Tsitsiklis, J. N. (1996). Neuro-Dynamic Programming. Belmont, MA, USA: Athena Scientific. 29

  5. [3]

    Busoniu, L., de Bruin, T., Tolic, D., Kober, J., & Palunko, I. (2018). Reinforcement learning for control: Performance, stability, and deep approximato rs. Annual Reviews in Control , 46, 8-28

  6. [4]

    Boyd, B., & Vandenberghe, V. (2004). Convex Optimization . Cambridge University Press

  7. [5]

    Cui, L., Pang, B., Krstic, M., & Jiang, Z.-P. (2025). Lear ning-based adaptive optimal control of linear time-delay systems: A value iteration app roach. Automatica, 171, 111944

  8. [6]

    El Bouhtouri, A., Hinrichsen, D., & Pritchard, A. J. (199 9). H∞-type control for discrete- time stochastic systems. Int. J. Robust Nonlinear Control , 9, 923-948

Show all 42 references
  1. [7]

    Chen, S., Li, X., & Zhou, X. (1998). Stochastic linear qua dratic regulators with indefinite control weight costs. SIAM J. Contr. Optim., 36, 1685-1702

  2. [8]

    Dombrovskii, V. V. & Lyashenko, E. A. (2003). A linear qua dratic control for discrete systems with random parameters and multiplicative noise and its application to investment portfolio optimization. Automat. Remote Control , 64, 1558-1570

  3. [9]

    Farjadnasab, M., & Babazadeh, M. (2022). Model-free LQR design by Q-function learning. Automatica, 137, 110060

  4. [10]

    M., & Mesbahi, M

    Fazel, M., Ge, R., Kakade, S. M., & Mesbahi, M. (2018). Gl obal convergence of policy gradient methods for the linear quadratic regulator. In International conference on machine learning, 1467-1476. Stockholm, Sweden

  5. [11]

    Huang, Y., Zhang, W., & Zhang, H. (2008). Infinite horizon linear quadratic optimal control for discrete time stochastic systems. Asian Journal of Control , 10(5), 608-615

  6. [12]

    Jiang, X., Wang, Y., Zhao, D., & Shi, L. (2024). Online Pareto optimal control of mean-field stochastic multi-player systems using policy iteration. Science China Information Sciences , 67(4), 140202:1-140202:17

  7. [13]

    Kleinman, D. L. (1968), On an iterative technique for Riccati equation computations, IEEE Transactions on Automatic Control , 13, 114-115

  8. [14]

    Kalman, R. E. (1960). Contributions to the theory of opt imal control. Bol. Soc. Mat. Mex. , 5(2), 102-119. 30

  9. [15]

    Karl, K. & Tu, S. (2019). Finite-time analysis of approx imate policy iteration for the linear quadratic regulator. In International conference on machine learning , 8514-8524. Vancou- ver, Canada

  10. [16]

    L., & Jiang, Z

    Kiumarsi, B., Lewis, F. L., & Jiang, Z. P. (2017). H∞ control of linear discrete-time systems: Off-policy reinforcement learning. Automatica, 78, 144-152

  11. [17]

    Lewis, F. L. (1986). Optimal Control. John Wiley & Sons

  12. [19]

    Lai, J., Xiong, J., & Shu, Z. (2023). Model-free optimal control of discrete-time systems with additive and multiplicative noises. Automatica, 147, 110685

  13. [20]

    X., Wang, Y., & Kang, Y

    Li, M., Qin, J., Zheng, W. X., Wang, Y., & Kang, Y. (2022). Model-free design of stochastic LQR controller from a primal-dual optimization perspectiv e. Automatica, 140, 110253

  14. [21]

    Liu, D., & Todorov, E. (2007). Evidence for the flexible s ensorimotor strategies predicted by optimal feedback control. The Journal of Neuroscience , 27(35), 93354-9368

  15. [22]

    Oura, R., Ushio, T., & Sakakibara, A. (2024). Bounded synthesis and reinforcement learning of supervisors for stochastic discrete event systems with L TL specifications. IEEE Trans- actions on Automatic Control , 69(10), 6668–6683

  16. [23]

    Pang, B., & Jiang, Z. P. (2022). Reinforcement learning for adaptive optimal stationary control of linear stochastic systems. IEEE Transactions on Automatic Control , 68(4), 2383- 2390

  17. [24]

    Pang, B., Cui, L., & Jiang, Z. P. (2022). Human motor lear ning is robust to control- dependent noise. Biological Cybernetics, 116, 307-325

  18. [25]

    Sun, J. & J. Yong. (2023). Stochastic linear-quadratic optimal control problems-Recent developments. Annual Reviews in Control , 56, 100899

  19. [26]

    & Benjamin, R

    Stephen, T. & Benjamin, R. (2019). The gap between model -based and model-free methods on the linear quadratic regulator: An asymptotic viewpoint . In Proceedings of the Confer- ence on Learning Theory , 3036-3083. 31

  20. [27]

    Vrabie, D., Pastravanu, O., Abu-Khalaf, M., & Lewis, F. L. (2009). Adaptive optimal control for continuous-time linear systems based on policy iteration. Automatica, 45(2), 477-484

  21. [28]

    Wang, Y., You, K., Huang, D., & Shang, C. (2025). Data-dr iven output prediction and control of stochastic systems: An innovation-based approa ch. Automatica, 171, 111897

  22. [29]

    J., & Dayan, P

    Watkins, C. J., & Dayan, P. (1992). Q-Learning. Machine Learning, 8, 279-292

  23. [30]

    Werbos, P. J. (1991). A Menu of Designs for Reinforcement Learning Over Time . Cam- bridge, MA, USA: MIT Press

  24. [31]

    Wonham, W. M. (1968). On a matrix Riccati equation of sto chastic control. SIAM Journal on Control, 6(4), 681-697

  25. [32]

    D., Zhang, S., & Zhou, X

    Yao, D. D., Zhang, S., & Zhou, X. Y. (2001). Stochastic linear-quadratic control via semidef- inite programming. SIAM Journal on Control and Optimization , 40(3), 801-823

  26. [33]

    Y., Nevena, L., & Csaba, S

    Yasin, A. Y., Nevena, L., & Csaba, S. (2019). Model-free linear quadratic control via reduction to expert prediction. In Proceedings of the International Conference on Artificial Intelligence and Statistics , 3108-3117

  27. [34]

    Yuan, K., Xu, W., & Ling, Q. (2020). Can primal methods outperform primal-dual methods in decentralized dynamic optimization?. IEEE Transactions on Signal Processing, 68, 4466- 4480

  28. [35]

    Zhang, H., & Ringh, A. (2023). Inverse linear-quadrati c discrete-time finite horizon opti- mal control for indistinguishable homogeneous agents: A co nvex optimization approach. Automatica, 148, 110758

  29. [36]

    Zhang, W., & Chen, B. S. (2004). On stabilizability and e xact observability of stochastic systems with their applications. Automatica, 40(1), 87-94

  30. [38]

    Zhang, W., Zhang, H., & Chen, B. S. (2008). Generalized L yapunov equation approach to state-dependent stochastic stabilization/detectabil ity criterion. IEEE Transactions on Automatic Control, 53(7), 1630-1642. 32

  31. [39]

    Zhang, W. (1998). Study on the algebraic Riccati equati on arising from infinite horizon stochastic LQ optimal control. Zhejiang University, PhD di ssertation

  32. [40]

    Zhang, W., Guo, J., & Jiang, X. (2025). Model-free H∞ control of Itˆ o stochastic system via off-policy reinforcement learning. Automatica, 174, 112144

  33. [41]

    Zhang, W., Yu, Z., & Jiang, X. (2024). Finite-time annul ar domain stability and asyn- chronous H∞ control for stochastic switching Markov jump systems. IEEE Transactions on Automatic Control , 699, 6277-6284

  34. [42]

    Zhao, B., & You, K. (2023). Survey of recent progress in d ata-driven policy optimization for controller design (in Chinese). Scientia Sinica Informationis , 53(6), 1027–1049. 33

Pith tools

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