Pith. sign in

REVIEW 2 major objections 4 minor 37 references

Stochastic MPC with Online-optimized Policies and Closed-loop Guarantees

T0 review · 2 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Stochastic MPC optimizes feedback online without losing closed-loop safety guarantees

desk verdict Closes the gap between online-optimized feedback and closed-loop chance constraint guarantees in SMPC, with honest caveats about the offline terminal set computation. read the letter →

arxiv 2502.06469 v3 pith:ZH4KY6FZ submitted 2025-02-10 eess.SY cs.SYmath.OC

classification eess.SYcs.SYmath.OC MSC 93C5593E20
keywords stochasticmodelpredictivecontrolchanceconstraintsdisturbancefeedbacksystemlevelsynthesismaximaladmissiblesetreconditioningrecursivefeasibilityGaussiandisturbances
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 proposes a stochastic model predictive control (SMPC) method for linear time-invariant systems with additive Gaussian disturbances that optimizes affine disturbance-feedback policies online at every time step. The authors claim this is the first SMPC scheme to combine online-optimized feedback with recursive feasibility, closed-loop satisfaction of probabilistic (chance) constraints, and a bounded asymptotic average cost. The key enablers are a finitely determined maximal admissible set used as a terminal constraint, and a reconditioning step that updates the predicted chance constraints using the realized past disturbances each time the optimization is solved. A building temperature control example illustrates that the method operates close to the probabilistic constraint while meeting it in closed loop.

What carries the argument

The machinery is the system level parameterization (SLP), which expresses the controlled error trajectory as a convolution of past disturbances with the optimized feedback matrices and turns the chance constraints into second-order cone constraints. A finitely determined maximal admissible set $S_\mu = S_\infty$ for the probabilistic tail constraints serves as the terminal set at $k=0$; Algorithm 1 computes it using maximal admissible set theory. At $k>0$, a reconditioning update computes each required probability level $\alpha^i_{j|k}$ from the previous optimal policy shifted and conditioned on the latest disturbance $w_{k-1}$, and the terminal set is replaced by the fixed shifted tail (40), which trivially satisfies the tail constraints.

What would settle it

Simulate Algorithm 2 on a system satisfying Assumption 1 over many (e.g. $10^5$) i.i.d. Gaussian disturbance sequences and record, at every time step, whether the optimization (41) is infeasible or the empirical frequency of constraint violation falls below the required $p_j$; any such event contradicts the theorem. A cheaper offline falsifier is to run Algorithm 1 with the S-procedure sufficient conditions on a system satisfying Assumption 1 and observe non-termination, which would show the proposed construction is not actually computable.

Watch

Extended reading notes

Core claim

The central claim is Theorem 2: if Assumption 1 holds and the initial SOCP (19) is feasible, then the receding-horizon scheme Algorithm 2, which solves (19) at $k=0$ and (41) for $k>0$, keeps (41) recursively feasible for all $k>0$, satisfies the chance constraints (2) in closed loop, and bounds the asymptotic average cost by the performance of the fixed terminal feedback $K$. This closes a gap identified in the paper: prior methods with online-optimized disturbance feedback either assumed bounded disturbances, used recovery mechanisms, or forfeited closed-loop chance-constraint guarantees, while methods with such guarantees used fixed feedback policies. The proof works by reconditioning the probabilistic constraints at each step on the realized disturbance history, so the predicted distribution of the trajectory under the candidate shifted policy matches the distribution that the previous policy would have produced in hindsight.

Load-bearing premise

The scheme is only guaranteed to exist if the offline Algorithm 1 terminates and returns a non-empty terminal set $S_\mu$, which its guaranteed convergence requires Assumption 1 (bounded constraint set, observability of $(C_K,A_K)$, and the margin condition (17)); with the practical S-procedure checks, termination is not guaranteed.

Editorial extensions

If this is right

  • Recursive feasibility is maintained even though the Gaussian disturbance has unbounded support, without softening constraints or adding a recovery mechanism.
  • The controller can operate close to the probabilistic constraint, achieving lower expected cost than fixed-feedback SMPC or robust constraint-tightening approaches.
  • Closed-loop chance constraint satisfaction is guaranteed despite the feedback matrices being random variables, because the constraints are reconditioned on the realized disturbance history at each step.
  • The method extends the reconditioning idea from mission-wide joint chance constraints to a receding-horizon setting with causal disturbance-feedback policies.
  • The asymptotic average cost is no worse than that of the static terminal feedback $K$, providing a performance ceiling.

Reading between the lines

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

  • The practical construction of the terminal set relies on S-procedure sufficient conditions that, as the paper's Remark 4 states, may not terminate; should Algorithm 1 fail for a given instance, the scheme cannot be built even if the system satisfies Assumption 1.
  • The paper's numerical example actually violates the boundedness part of Assumption 1 and still converges, suggesting finite determination may hold under weaker conditions than assumed; this is a testable conjecture, not a paper claim.
  • A natural extension, implied rather than proven here, would replace Gaussian disturbances with subgaussian or bounded-variance distributions and keep the same reconditioning structure, though the SOC reformulations would need new tail bounds.
  • The terminal constraint (40) at $k>0$ is deliberately conservative, tying the tail to the previous optimal solution; relaxing it with a re-computed finite index, as the modified variant RC-mod does, seems to recover performance at extra online cost.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a stochastic model predictive control (SMPC) scheme for linear time-invariant systems with additive i.i.d. Gaussian disturbances, in which the disturbance-feedback matrices are optimized online via the system level parameterization. The central mechanism is a reconditioning framework that updates the predicted probabilistic constraints on the current disturbance history, combined with a finitely determined maximal admissible terminal set for the infinite-horizon tail under a fixed terminal controller. The authors prove that the receding-horizon implementation is recursively feasible, satisfies the original chance constraints in closed loop, and yields a bounded asymptotic average cost (Theorem 2). The method is demonstrated on a building temperature control example, with numerical comparisons against several existing SMPC schemes.

Significance. If the results hold, the paper fills a genuine gap: existing SMPC methods with online-optimized feedback either lack closed-loop chance-constraint guarantees or rely on bounded disturbances. The SOC reformulation of the probabilistic constraints, the finite-determination argument for the probabilistic terminal set, and the recursive-feasibility proof via shifted candidate policies are the main technical strengths. The paper provides detailed proofs, a convex formulation, and released code, which are all positive features. The principal caveat is the offline terminal-set construction: the implementable S-procedure checks are only sufficient and may not terminate, and the numerical example lies outside the assumptions of the main theorem. These issues are fixable but currently leave the practical scheme without a finite-time certificate in the demonstrated setting.

major comments (2)
  1. [Algorithm 1, Remarks 4–5, Appendices E–F] Theorem 2's closed-loop guarantees are conditional on the existence of the terminal set S_μ, and Proposition 1 proves termination of Algorithm 1 only under exact set-inclusion checks. The actual implementation proposed in Appendices E–F uses lossy S-procedure sufficient conditions, and Remark 4 explicitly concedes that these may not terminate. Thus Algorithm 2 has no finite-time certificate for constructing S_μ in the form that is implemented. This is load-bearing because without S_μ the online problem (19) and all subsequent guarantees are void. Please either provide conditions under which the S-procedure checks are exact or terminate, or restate the main theorem with an explicit computability assumption and discuss the resulting gap.
  2. [Section VII, constraint (44), Remark 5] The numerical example uses the one-sided chance constraint [1,0,0]x_k ≥ −0.5, which makes the set L unbounded and hence violates the boundedness part of Assumption 1. Consequently, Theorem 2 does not apply to the demonstrated scenario; the paper relies on Remark 5 and on Algorithm 1 having converged in that specific instance. Please modify the example so that Assumption 1 is satisfied (for instance, by adding an upper bound on the room temperature so that L is bounded), or provide a separate proof that the one-sided case still yields a finitely determined S_μ under the other conditions of Assumption 1.
minor comments (4)
  1. [Remark 1] The notation in Remark 1 contains a corrupted symbol where 'mathbfΣ' appears; it should read Σ_w^0.
  2. [Section II-A] The notation χ2(·) for the inverse cumulative distribution function of the chi-squared distribution with one degree of freedom is nonstandard and could be confused with the chi-squared random variable; a more explicit notation such as Φ^{-1}_{χ^2_1}(·) would improve clarity.
  3. [Appendix H, proof of Theorem 2 Part I] The proof of recursive feasibility states that the candidate satisfies the constraints 'with equality', but this is not literally true in the degenerate case (33)–(34) and in the α<0 case (38), where the candidate satisfies the relevant set definition by construction rather than by equality. A sentence clarifying the case distinction would make the argument easier to verify.
  4. [Section VII, Table I] The IF method's closed-loop cost is reported in parentheses with a satisfaction level of 0.0%; it would help to state explicitly that this cost is not comparable because the method violates the chance constraints.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular step found; the closed-loop guarantees are proven from the stated assumptions via a constructive feasibility argument, with the main caveats being completeness and assumption violations, not circularity.

full rationale

The derivation is self-contained against external benchmarks. Theorem 2's recursive-feasibility claim is established by constructing the shifted previous solution (30)-(32) and verifying that the reconditioned constraints (41d)-(41e) hold for it with equality; this is a proof-by-construction, not a fitted quantity renamed as a prediction. Closed-loop chance constraint satisfaction follows from the inductive probability recursion in (26)-(27) and the argument in Part II of the proof of Proposition 3, which starts from the exact Gaussian reformulation in Lemma 1. The finite determination of the terminal set is imported from the external maximal admissible set theory of Gilbert and Tan [29], and the S-procedure implementation in Appendices E-F is an independently verifiable sufficient condition. The limitations flagged in Remarks 4 and 5 (possible non-termination of Algorithm 1 and violation of the boundedness part of Assumption 1 in the numerical example) are completeness or assumption-satisfaction issues, not circular reductions: the main theorem is conditional on Assumption 1, and the example relies on the explicitly stated conditional statement that convergence of Algorithm 1 still yields S_infinity = S_mu. Self-citations such as [13] and [18] are used as standard building blocks for Gaussian chance-constraint reformulation and system level parameterization, and they are not load-bearing in a way that forces the paper's conclusions.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No parameters are fitted to data. The horizon N, probability levels p_j, and terminal gain K are standard design choices; K is synthesized via the SDP in Appendix B. All assumptions are stated in Assumption 1 and the problem setup. The analysis relies on standard machinery (maximal admissible sets, S-procedure, system level synthesis) from the cited literature.

assumptions (6)
  • domain assumption Assumption 1: L = {y | L y ≤ b} is bounded, (C_K, A_K) observable, and b_j - sqrt(p̃_j) ||(Σ_x^∞)^{1/2} C_K^T L_j^T|| > 0 for all j
    Invoked in Theorem 1 and Proposition 1 to guarantee finite determination of the terminal set. The numerical example violates the boundedness part, so the guarantee is not strictly applicable there, though Algorithm 1 converges in practice.
  • domain assumption There exists a terminal feedback gain K such that A+BK is Schur and constraint (17) holds
    Needed for the terminal set to be non-empty and for the Lyapunov-based cost bound. If not, p_j must be reduced (Appendix B provides an SDP to synthesize K).
  • domain assumption Initial feasibility of Problem (19) at time step k=0
    Theorem 2 and Proposition 3 condition all recursive feasibility and closed-loop guarantees on this. If (19) is infeasible, the method cannot start.
  • domain assumption Gaussian i.i.d. disturbances with known covariance Σ_w and direct state measurement
    Used in Lemma 1 for the exact SOC reformulation of chance constraints, and throughout the reconditioning derivation for the Gaussian conditional distributions.
  • standard math Gilbert-Tan maximal admissible set theory and the S-procedure for containment checking
    Foundation of Theorem 1 (finite determination) and Algorithm 1 (Steps 1-2 with LMIs), cited from [29] and [30].
  • standard math System level parameterization constraint (6)
    The linear constraint (6) defines the relationship between state and input response matrices; it is a known result from [21], used to keep the optimization convex.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stochastic MPC with Online-optimized Policies and Closed-loop Guarantees." pith.science (2026). https://pith.science/paper/ZH4KY6FZ

@misc{pith2026250206469,
  author       = {Pith},
  title        = {Pith review of: Stochastic MPC with Online-optimized Policies and Closed-loop Guarantees},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZH4KY6FZ}},
  note         = {Machine review of arXiv:2502.06469}
}
read the original abstract

This paper proposes a stochastic model predictive control method for linear systems affected by additive Gaussian disturbances that optimizes over disturbance feedback matrices online. Closed-loop satisfaction of probabilistic constraints and recursive feasibility of the underlying convex optimization problem is guaranteed. Optimization over feedback policies online increases performance and reduces conservatism compared to fixed-feedback approaches. The central mechanism is a finitely determined maximal admissible set for probabilistic constraints, together with the reconditioning of the predicted probabilistic constraints on the current knowledge at every time step. The proposed method's applicability is demonstrated on a building temperature control example.

Figures

Figures reproduced from arXiv: 2502.06469 by the authors.

Figure 1
Figure 1. Closed-loop evolution of the room temperature using the two variations [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 25 canonical work pages

  1. [17]

    A tractable approximation of chance constrained stochastic MPC based on affine disturbance feedback,

    F. Oldewurtel, C. N. Jones, and M. Morari, “A tractable approximation of chance constrained stochastic MPC based on affine disturbance feedback,” inProc. 47th IEEE Conference on Decision and Control (CDC). IEEE, 2008, pp. 4731–4736

  2. [13]

    Recursively feasible stochastic model predictive control using indirect feedback,

    L. Hewing, K. P. Wabersich, and M. N. Zeilinger, “Recursively feasible stochastic model predictive control using indirect feedback,”Automatica, vol. 119, p. 109095, 2020

  3. [21]

    System level synthesis,

    J. Anderson, J. C. Doyle, S. H. Low, and N. Matni, “System level synthesis,”Annual Reviews in Control, vol. 47, pp. 364–393, 2019

  4. [1]

    J. B. Rawlings, D. Q. Mayne, M. Diehlet al.,Model predictive control: theory, computation, and design. Nob Hill Publishing Madison, WI, 2017, vol. 2

  5. [2]

    Model predictive control,

    B. Kouvaritakis and M. Cannon, “Model predictive control,”Switzer- land: Springer International Publishing, vol. 38, pp. 13–56, 2016

  6. [3]

    Stochastic linear model predictive control with chance constraints–a review,

    M. Farina, L. Giulioni, and R. Scattolini, “Stochastic linear model predictive control with chance constraints–a review,”Journal of Process Control, vol. 44, pp. 53–67, 2016

  7. [4]

    Stochastic model predictive control: An overview and perspectives for future research,

    A. Mesbah, “Stochastic model predictive control: An overview and perspectives for future research,”IEEE Control Systems Magazine, vol. 36, no. 6, pp. 30–44, 2016

  8. [5]

    Stochastic tubes in model predictive control with probabilistic constraints,

    M. Cannon, B. Kouvaritakis, S. V . Rakovi ´c, and Q. Cheng, “Stochastic tubes in model predictive control with probabilistic constraints,”IEEE Transactions on Automatic Control, vol. 56, no. 1, pp. 194–200, 2010

Show all 37 references
  1. [6]

    Constraint- tightening and stability in stochastic model predictive control,

    M. Lorenzen, F. Dabbene, R. Tempo, and F. Allg ¨ower, “Constraint- tightening and stability in stochastic model predictive control,”IEEE Transactions on Automatic Control, vol. 62, no. 7, pp. 3165–3177, 2016

  2. [7]

    Joint chance-constrained model predictive control with prob- abilistic resolvability,

    M. Ono, “Joint chance-constrained model predictive control with prob- abilistic resolvability,” inProc. American Control Conference (ACC). IEEE, 2012, pp. 435–441

  3. [8]

    A probabilistic approach to model predictive control,

    M. Farina, L. Giulioni, L. Magni, and R. Scattolini, “A probabilistic approach to model predictive control,” inProc. 52nd IEEE conference on decision and control. IEEE, 2013, pp. 7734–7739

  4. [9]

    Stochastic model predictive control with joint chance constraints,

    J. A. Paulson, E. A. Buehler, R. D. Braatz, and A. Mesbah, “Stochastic model predictive control with joint chance constraints,”International Journal of Control, vol. 93, no. 1, pp. 126–139, 2020

  5. [10]

    Stochastic model predictive control for linear systems using probabilistic reachable sets,

    L. Hewing and M. N. Zeilinger, “Stochastic model predictive control for linear systems using probabilistic reachable sets,” inProc. IEEE Conference on Decision and Control (CDC). IEEE, 2018, pp. 5182– 5188

  6. [11]

    Recursively feasible stochastic predictive control using an interpolating initial state constraint,

    J. K ¨ohler and M. N. Zeilinger, “Recursively feasible stochastic predictive control using an interpolating initial state constraint,”IEEE Control Systems Letters, vol. 6, pp. 2743–2748, 2022

  7. [12]

    Stochastic model predictive control using initial state optimization,

    H. Schl ¨uter and F. Allg ¨ower, “Stochastic model predictive control using initial state optimization,”IFAC-PapersOnLine, vol. 55, no. 30, pp. 454– 459, 2022

  8. [14]

    Stochastic MPC with robustness to bounded parametric uncertainty,

    E. Arcari, A. Iannelli, A. Carron, and M. N. Zeilinger, “Stochastic MPC with robustness to bounded parametric uncertainty,”IEEE Transactions on Automatic Control, vol. 68, no. 12, pp. 7601–7615, 2023

  9. [15]

    Stochastic MPC for linear systems with unbounded multiplicative noise guaranteeing closed- loop chance constraints satisfaction,

    C. Mark, D. Ravasio, M. Farina, and D. G ¨orges, “Stochastic MPC for linear systems with unbounded multiplicative noise guaranteeing closed- loop chance constraints satisfaction,”IEEE Control Systems Letters, vol. 8, pp. 2081–2086, 2024

  10. [16]

    Approximations of closed-loop minimax mpc,

    J. L ¨ofberg, “Approximations of closed-loop minimax mpc,” in42nd IEEE International Conference on Decision and Control (IEEE Cat. No. 03CH37475), vol. 2. IEEE, 2003, pp. 1438–1442

  11. [18]

    A system level approach to tube-based model predictive control,

    J. Sieber, S. Bennani, and M. N. Zeilinger, “A system level approach to tube-based model predictive control,”IEEE Control Systems Letters, vol. 6, pp. 776–781, 2021

  12. [19]

    Robust nonlinear optimal control via system level synthesis,

    A. P. Leeman, J. K ¨ohler, A. Zanelli, S. Bennani, and M. N. Zeilinger, “Robust nonlinear optimal control via system level synthesis,”IEEE Transactions on Automatic Control, 2025

  13. [20]

    Optimization over state feedback policies for robust control with constraints,

    P. J. Goulart, E. C. Kerrigan, and J. M. Maciejowski, “Optimization over state feedback policies for robust control with constraints,”Automatica, vol. 42, no. 4, pp. 523–533, 2006

  14. [22]

    A randomized approach to stochastic model predictive control,

    M. Prandini, S. Garatti, and J. Lygeros, “A randomized approach to stochastic model predictive control,” inProc. 51st IEEE Conference on Decision and Control (CDC). IEEE, 2012, pp. 7315–7320

  15. [23]

    Recursively feasible data-driven distributionally robust model predictive control with additive disturbances,

    C. Mark and S. Liu, “Recursively feasible data-driven distributionally robust model predictive control with additive disturbances,”IEEE Con- trol Systems Letters, vol. 7, pp. 526–531, 2022

  16. [24]

    Data-driven stochastic output- feedback predictive control: Recursive feasibility through interpolated initial conditions,

    G. Pan, R. Ou, and T. Faulwasser, “Data-driven stochastic output- feedback predictive control: Recursive feasibility through interpolated initial conditions,” inProc. Learning for Dynamics and Control Confer- ence. PMLR, 2023, pp. 980–992

  17. [25]

    Distributionally robust stochastic data-driven predictive control with optimized feedback gain,

    R. Li, J. W. Simpson-Porco, and S. L. Smith, “Distributionally robust stochastic data-driven predictive control with optimized feedback gain,” arXiv preprint arXiv:2409.05727, 2024

  18. [26]

    Recursively feasible stochastic model predictive control for time-varying linear systems subject to unbounded disturbances,

    J. W. Knaup and P. Tsiotras, “Recursively feasible stochastic model predictive control for time-varying linear systems subject to unbounded disturbances,”arXiv preprint arXiv:2410.11107, 2024

  19. [27]

    Recursive feasibility of stochastic model predic- tive control with mission-wide probabilistic constraints,

    K. Wang and S. Gros, “Recursive feasibility of stochastic model predic- tive control with mission-wide probabilistic constraints,” inProc. 60th IEEE Conference on Decision and Control (CDC). IEEE, 2021, pp. 2312–2317

  20. [28]

    Chance-constrained controller state and reference governor,

    N. Li, A. Girard, and I. Kolmanovsky, “Chance-constrained controller state and reference governor,”Automatica, vol. 133, p. 109864, 2021

  21. [29]

    Linear systems with state and control constraints: The theory and application of maximal output admissible sets,

    E. G. Gilbert and K. T. Tan, “Linear systems with state and control constraints: The theory and application of maximal output admissible sets,”IEEE Transactions on Automatic control, vol. 36, no. 9, pp. 1008– 1020, 1991

  22. [30]

    S. Boyd, L. El Ghaoui, E. Feron, and V . Balakrishnan,Linear matrix inequalities in system and control theory. SIAM, 1994

  23. [31]

    Economic optimization using model predictive control with a terminal cost,

    R. Amrit, J. B. Rawlings, and D. Angeli, “Economic optimization using model predictive control with a terminal cost,”Annual Reviews in Control, vol. 35, no. 2, pp. 178–186, 2011

  24. [32]

    ApS,The MOSEK Python Fusion API manual

    M. ApS,The MOSEK Python Fusion API manual. Version 11.0.,

  25. [33]

    OSQP: an operator splitting solver for quadratic programs,

    B. Stellato, G. Banjac, P. Goulart, A. Bemporad, and S. Boyd, “OSQP: an operator splitting solver for quadratic programs,”Mathematical Programming Computation, vol. 12, no. 4, pp. 637–672, 2020

  26. [34]

    CVXPY: A Python-embedded modeling lan- guage for convex optimization,

    S. Diamond and S. Boyd, “CVXPY: A Python-embedded modeling lan- guage for convex optimization,”Journal of Machine Learning Research, vol. 17, no. 83, pp. 1–5, 2016

  27. [35]

    The vec-permutation matrix, the vec operator and kronecker products: A review,

    H. V . Henderson and S. R. Searle, “The vec-permutation matrix, the vec operator and kronecker products: A review,”Linear and multilinear algebra, vol. 9, no. 4, pp. 271–288, 1981. Marcell Bartosis a Ph.D. student under the supervision of Prof. Melanie Zeilinger and Prof. Flor...

  28. [2014]

    She was awarded the ETH medal for her PhD thesis, an SNF Professorship, the ETH Golden Owl for exceptional teaching in 2022 and the European Control Award in 2023

    From 2018 to 2019 she was a professor at the University of Freiburg, Germany. She was awarded the ETH medal for her PhD thesis, an SNF Professorship, the ETH Golden Owl for exceptional teaching in 2022 and the European Control Award in 2023. Her research interests include lear...

  29. [2025]

    Available: https://docs.mosek.com/latest/pythonfusion/ index.html

    [Online]. Available: https://docs.mosek.com/latest/pythonfusion/ index.html

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.