Pith. sign in

REVIEW 3 major objections 5 minor 13 references

Fast Stochastic MPC using Affine Disturbance Feedback Gains Learned Offline

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper claims that stochastic MPC can solve online over a handful of offline-learned feature feedback gains instead of the full disturbance-feedback policy, cutting computation about tenfold while matching region of attraction and…

desk verdict The offline SVD-plus-scaling pipeline is a genuinely useful idea, but the implemented small-singular-value truncation falls outside the paper's own guarantee, so the confidence claims need work before this is trustworthy. read the letter →

arxiv 2411.13935 v1 pith:PRIA27WI submitted 2024-11-21 eess.SY cs.SY

Hotae Lee , Francesco Borrelli This is my paper · ORCID
classification eess.SYcs.SY
keywords stochasticmodelpredictivecontrolaffinedisturbancefeedbackchanceconstraintsscenarioapproachsingularvaluedecompositionfeatureextractionprobabilisticscalingregionofattraction
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

Stochastic model predictive control (SMPC) for uncertain linear systems with chance constraints normally optimizes over a high-dimensional affine disturbance feedback policy—a matrix of gains mapping past disturbances to future inputs—enforced against many sampled disturbance scenarios. This paper tries to make that problem fast enough for real-time use by learning the policy's structure offline. The authors stack offline disturbance samples with the dynamics and constraint matrices, apply a singular value decomposition, and keep only the feedback-gain directions that correspond to nonzero singular values; they then use a probabilistic scaling method to replace the many sampled constraints with one small, guaranteed-confidence feasible set. Online, the MPC solves a quadratic program over the nominal inputs and the reduced set of feature gains. In the lane-keeping simulation, the proposed controller matches the full feedback policy's region of attraction and closed-loop costs while solving at least ten times faster, about 0.01 seconds per update.

What carries the argument

The load-bearing object is the stacked constraint-dynamics matrix $\bar{A}$ assembled from $N_s$ offline disturbance samples and the system matrices, together with its structured singular value decomposition $\bar{A} = (U_W \otimes U_B)(\Sigma_W \otimes \Sigma_B)(V_W^\top \otimes V_B^\top)$. The diagonal middle factor shows which rotated components of the vectorized affine feedback gains $y_t$ can affect the sampled constraints; components multiplied by zero singular values are redundant. The truncation $y_t^{\mathrm{trun}} = P y_t$ removes those redundant components, and the reconstruction map $z_t^{\mathrm{recon}} = P_{\mathrm{re}} z_t^{\mathrm{trun}}$ recovers the original policy variables, preserving time-causal zero structure when needed. The second mechanism is the adjusted probabilistic scaling construction $S(\gamma^\star) = z_c \oplus \gamma^\star H B_\infty^n$, which compresses the remaining large set of sampled inequalities into one infinity-norm ball inclusion. Proposition 6 is the result that carries the confidence guarantee: if the same sample set is used, the reconstruction of any point in $S(\gamma^\star)$ lies in the chance-constrained set with confidence $1-\delta$, and the required number of scaling samples $N_\gamma \ge \frac{7.47}{\epsilon} \ln\frac{1}{\delta}$ is unchanged.

What would settle it

A direct test: use fresh disturbance realizations that were not part of the offline construction, run the controller from initial states near the boundary of the reported region of attraction, and count constraint violations across many trials; if the empirical violation frequency exceeds $\epsilon$ more often than the claimed $1-\delta$ confidence allows, then the small-singular-value truncation or the sample reuse has broken the guarantee. A sharper version tunes the smallest nonzero singular value of the sample matrix to be small but nonzero; Proposition 6 predicts no change in feasibility only when that value is exactly zero.

Watch

Extended reading notes

Core claim

The paper's central claim is that affine disturbance feedback policies for chance-constrained SMPC contain a low-dimensional feature subspace that can be identified offline without changing the sampled feasible set, and that optimizing in this subspace reproduces the behavior of the full policy. The constraint matrix built from $N_s$ disturbance samples factors through a structured SVD, $\bar{A} = (U_W \otimes U_B)(\Sigma_W \otimes \Sigma_B)(V_W^\top \otimes V_B^\top)$, so the rotated gain vector $V^\top y_t$ has components that are multiplied by singular values $\sigma_i$. Setting to zero every component corresponding to $\sigma_i = 0$ leaves the sampled constraints untouched, giving a truncated feature policy $y_t^{\mathrm{trun}} = P y_t$ from which the original policy can be reconstructed. The paper then wraps this reduced variable in the probabilistically scaled set $S(\gamma^\star) = z_c \oplus \gamma^\star H B_\infty^n$, and proves in Proposition 6 that the confidence level $1-\delta$ survives the truncation when the truncation removes only exact zeros. On the lane-keeping example this yields a QP with a simple norm-ball constraint set that solves in less than 0.01 seconds, compared with 0.1--0.2 seconds for the full feedback policy, with comparable region of attraction and closed-loop cost.

Load-bearing premise

The load-bearing premise is that the offline sample set can be used twice—once to learn which feedback directions matter and once to size the approximate feasible set—and that truncating small singular values leaves the $1-\delta$ confidence intact; Proposition 6 proves only the exact-zero truncation case, and Remark 2 concedes the practical case is unproven.

Editorial extensions

If this is right

  • The reported speedup means stochastic MPC with affine disturbance feedback can be used where the full feedback policy is too slow: the online solve took under 0.01 seconds in the example, independent of whether 10, 100, or 1000 scenarios were used offline.
  • The offline computation is amortized over the closed loop: the approximate feasible set is built once for a given constraint set and does not need to be regenerated when the control cost changes.
  • The approach inherits the scenario method's distribution-free property: the controller needs only samples of the disturbance support, not a known probability distribution, so it applies to non-Gaussian uncertainties such as the uniform samples in the example.
  • Because the remaining decision variables include a reduced set of feedback gains as well as nominal inputs, the controller still adapts online to the current state and disturbance history, rather than committing to a fixed state-feedback law.
  • If the small-singular-value truncation is admitted as safe, the feasible region and cost performance are close to the full feedback policy; the paper reports comparable regions of attraction and average closed-loop costs of about 56.85 versus 56.02 in the main scenario.

Reading between the lines

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

  • Beyond the paper: if the feature extraction is valid, the same compression applies to any policy class that enters the constraints linearly, so state-feedback variants or longer horizons with the same constraint matrices should see similar reductions.
  • Beyond the paper: because the online problem's dimension depends on the number of retained singular directions rather than the raw sample count, the computational advantage should grow with the number of scenarios and the horizon, a prediction that could be tested by scaling $N_s$ and $N$.
  • Beyond the paper: the unproven small-singular-value truncation and the double use of the same samples for feature selection and scaling calibration suggest a natural follow-up—split the samples or add a union bound and derive a violation bound that depends on the truncation threshold; until then, the strict $1-\delta$ statement covers only exact-zero truncation.
  • Beyond the paper: a practical tuning rule for the truncation threshold could be obtained by measuring the sensitivity of the computed region of attraction to the smallest retained singular value; if the ROA is stable over a range of thresholds, the method becomes robust to the approximation.
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 an offline feature-extraction method for stochastic MPC with affine disturbance feedback policies. Scenario constraints are built from offline disturbance samples, and an SVD of the stacked constraint matrix is used to reduce the dimension of the feedback-gain decision variables. A probabilistic scaling approach then approximates the feasible set of the truncated variables by a norm ball, yielding a small online QP. Numerical simulations of a lane-keeping example show closed-loop costs and regions of attraction comparable to a full affine-feedback scenario MPC, with roughly an order-of-magnitude online speedup. The central claim is that the offline learning preserves a specified chance-constraint confidence level.

Significance. If the theoretical guarantees were established, the paper would make a useful contribution to real-time stochastic MPC: moving the expensive scenario-constraint construction and reduction offline is a principled idea, and the numerical results are encouraging. The paper is clearly written and the SVD--Kronecker factorization in Section 3.2 is elegant. However, the confidence guarantee is proven only for exact-zero singular-value truncation, while the implemented method relies on truncating small nonzero singular values; the paper also uses the same offline samples both to select the SVD basis and to calibrate the probabilistic scaling factor. These gaps are acknowledged in Remark 2 but are not resolved, and they directly affect the abstract's claim of a 'specified confidence level.' The paper's value is therefore conditional on closing or explicitly narrowing these gaps.

major comments (3)
  1. [Section 3.4, Remark 2, Section 6] The 1−δ confidence statement in Section 3.4 is proven only for truncating singular values that are exactly zero. In the numerical setup (Ns=10^4 and n(N+1)=24), the sample matrix W is full column rank with probability one, so Σ_W has no zero diagonal entries; unless Σ_B is rank-deficient, Σ_W⊗Σ_B has no zero singular values at all. The reported dimension reduction and speedup therefore necessarily come from the practical truncation of small nonzero singular values described in Remark 2, which explicitly states that this 'will affect chance constraints guarantees' and defers the needed analysis. The paper does not provide the promised new probability-satisfaction bounds as a function of the truncation parameters, so the implemented algorithm's confidence guarantee is not established.
  2. [Section 4.2, Proposition 6] The proof of Proposition 6 assumes that for every sample w in W, the constraints evaluated at z_trun are identical to those evaluated at the reconstructed z_recon. This identity holds exactly only when the truncation removes zero singular values and the reconstruction is exact. When small nonzero singular values are truncated, the reconstruction map is not an exact inverse, and a z_trun that satisfies the sampled constraints in the truncated space need not correspond to a z_recon satisfying the original constraints for all w∈W. Consequently, the proof does not cover the implemented algorithm, and the claim that 'the required number of samples in the adjusted probabilistic scaling approach ... remains unchanged' is unsupported.
  3. [Sections 3.4, 4.2, Eq. (19), Proposition 6] The same offline samples are used both to select the SVD feature space (the matrix P and the truncated basis) and to compute the scaling factor γ in Eq. (19) and Proposition 6. This double use undermines the i.i.d. assumptions behind the scenario bound in Section 3.4 and the probabilistic scaling sample requirement (20), because the sets Z_trun(w) and the reconstruction map z_recon = P_re z_trun depend on the same data that is used to certify the confidence level. The paper does not analyze this dependence; it should either use independent sample sets for feature selection and calibration, or provide a rigorous treatment of the dependent case.
minor comments (5)
  1. [Eq. (7)] The block structure of A^j_{0:N} is indicated with ellipses but the exact dimensions and the placement of the Hu blocks are hard to verify; a formal block expression would improve readability.
  2. [Section 3.2] The symbol W is used both for the stacked sample matrix and for the disturbance support set W of the problem formulation; please disambiguate these two uses.
  3. [Algorithm 1] The input line says the samples are indexed j=0,...,Ns, but the scenario sets in Eq. (6) use j=1,...,Ns; the indexing should be made consistent.
  4. [Section 6.2, Table 5] The claim of being 'at least 10X faster' is stronger than what Table 5 shows for the 10-sample case, where the speedup over full feedback is roughly 3x to 16x; the statement should be qualified to the regimes where it holds.
  5. [Section 3.3] The treatment of time-causality constraints via the matrix P̄ and the equality constraints in (15) is difficult to follow; a small example or a more explicit definition of P̄ and P̄_basis would clarify the reconstruction step.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity found; the confidence-guarantee gaps are non-circular rigor issues.

full rationale

The claimed derivation chain is: affine disturbance feedback parametrization (4); scenario reformulation (5) via Calafiore-Campi; exact-zero SVD truncation (Sec. 3.2-3.3); and probabilistic scaling set approximation (Sec. 4) via Mammarella et al. The truncation step is an algebraic identity on the sampled constraint map when only exact-zero singular values are removed, so it is not a fitted parameter renamed as a prediction. The scaling step imports an external theorem; Proposition 6 attempts to extend it to truncated variables, but its proof assumes 'the same sample set from the truncation process' is reused for the scaling factor and simply affirms that the sample requirement 'remains unchanged.' That is an unsupported in-sample reuse, not a derivation that reduces the conclusion to the input by construction. Similarly, Remark 2 explicitly states that removing small nonzero singular values 'will affect chance constraints guarantees' and defers the analysis, so the implemented speedup is not covered by the paper's own 1-delta statement. These are correctness and rigor gaps in the headline guarantee, but they do not make any equation equal to its input by definition or make the numerical speed comparison circular. No self-citation is load-bearing: Lorenzen et al. and Mammarella et al. are external, and the authors' own Lee et al. 2023 citation is not the basis of the main theorem.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The method's theoretical guarantees are largely imported from external scenario and probabilistic scaling results. The paper's own assumptions (zero-only truncation, sample reuse) are unproven and affect the validity of the confidence guarantee.

free parameters (3)
  • SVD truncation threshold (number of retained features) = not reported
    The online speedup comes from truncating small nonzero singular values in (13), but the threshold is not specified. The theoretical guarantee only holds for exact zero singular values.
  • Initial candidate set center z_c and shape H = not reported
    Computed from Nini offline samples via (18); the choice of Nini and the algorithm for H are not given, and this design affects conservatism and feasibility of the online QP.
  • Probability parameters epsilon and delta = not reported
    The sample counts N_s=10^4 and N_gamma=2000 are given, but the corresponding epsilon and delta implied by the scenario bounds are not stated, so the confidence guarantee cannot be checked.
assumptions (6)
  • domain assumption The plant is LTI with known A,B and additive disturbance with compact support W.
    Standard setup from Section 2, Eq. (1).
  • domain assumption Affine disturbance feedback parametrization (4) is sufficient for the problem.
    Adopted from prior SMPC literature; restricts policy space but is the standard approach.
  • standard math The scenario approach sample bound N_s >= (5/epsilon)(ln(4/delta)+d ln(40/epsilon)) applies.
    Used in Section 3.4 to claim confidence 1-delta; from Calafiore and Campi (2006).
  • standard math The probabilistic scaling theorem from Mammarella et al. (2022) applies to the truncated variables.
    Used in Section 4.1 and Prop. 6 to justify the approximate set S(gamma*).
  • ad hoc to paper Truncating small nonzero singular values does not materially affect chance constraint satisfaction.
    Remark 2 admits this affects guarantees and defers derivations; the practical method depends on it.
  • ad hoc to paper The same offline samples can be used for feature extraction and for computing the scaling factor without invalidating the confidence guarantee.
    Assumed in Proposition 6; no analysis of double use of data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast Stochastic MPC using Affine Disturbance Feedback Gains Learned Offline." pith.science (2026). https://pith.science/paper/PRIA27WI

@misc{pith2026241113935,
  author       = {Pith},
  title        = {Pith review of: Fast Stochastic MPC using Affine Disturbance Feedback Gains Learned Offline},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PRIA27WI}},
  note         = {Machine review of arXiv:2411.13935}
}
read the original abstract

We propose a novel Stochastic Model Predictive Control (MPC) for uncertain linear systems subject to probabilistic constraints. The proposed approach leverages offline learning to extract key features of affine disturbance feedback policies, significantly reducing the computational burden of online optimization. Specifically, we employ offline data-driven sampling to learn feature components of feedback gains and approximate the chance-constrained feasible set with a specified confidence level. By utilizing this learned information, the online MPC problem is simplified to optimization over nominal inputs and a reduced set of learned feedback gains, ensuring computational efficiency. In a numerical example, the proposed MPC approach achieves comparable control performance in terms of Region of Attraction (ROA) and average closed-loop costs to classical MPC optimizing over disturbance feedback policies, while delivering a 10-fold improvement in computational speed.

Figures

Figures reproduced from arXiv: 2411.13935 by the authors.

Figure 1
Figure 1. Comparisons of ROA with three control policies [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Comparisons of ROA with three control policies for the less tight scenarios with [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. (a) Comparison of the closed-loop costs over 50 trials, (b) Gap distance graph during 20 [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a) Comparison of the closed-loop costs over 50 trials, (b) Gap distance graph during 20 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

  1. [1]

    CasADi -- A software framework for nonlinear optimization and optimal control

    Joel A E Andersson, Joris Gillis, Greg Horn, James B Rawlings, and Moritz Diehl. CasADi -- A software framework for nonlinear optimization and optimal control. Mathematical Programming Computation, In Press, 2018

    2018

  2. [2]

    The scenario approach to robust control design

    Giuseppe Carlo Calafiore and Marco C Campi. The scenario approach to robust control design. IEEE Transactions on automatic control, 51 0 (5): 0 742--753, 2006

    work page 2006

  3. [3]

    Stochastic tube mpc with state estimation

    Mark Cannon, Qifeng Cheng, Basil Kouvaritakis, and Sa s a V Rakovi \'c . Stochastic tube mpc with state estimation. Automatica, 48 0 (3): 0 536--541, 2012

    work page 2012

  4. [4]

    Scenario-based probabilistic reachable sets for recursively feasible stochastic model predictive control

    Lukas Hewing and Melanie N Zeilinger. Scenario-based probabilistic reachable sets for recursively feasible stochastic model predictive control. IEEE Control Systems Letters, 4 0 (2): 0 450--455, 2019

    work page 2019

  5. [5]

    Recursively feasible stochastic model predictive control using indirect feedback

    Lukas Hewing, Kim P Wabersich, and Melanie N Zeilinger. Recursively feasible stochastic model predictive control using indirect feedback. Automatica, 119: 0 109095, 2020

    work page 2020

  6. [6]

    Strongly feasible stochastic model predictive control

    Milan Korda, Ravi Gondhalekar, Ji r \' Cigler, and Frauke Oldewurtel. Strongly feasible stochastic model predictive control. In 2011 50th IEEE Conference on Decision and Control and European Control Conference, pages 1245--1251. IEEE, 2011

    work page 2011

  7. [7]

    Stochastic mpc with realization-adaptive constraint tightening

    Hotae Lee, Monimoy Bujarbaruah, and Francesco Borrelli. Stochastic mpc with realization-adaptive constraint tightening. In 2023 American Control Conference (ACC), pages 1838--1843. IEEE, 2023

    work page 2023

  8. [8]

    Stochastic mpc with offline uncertainty sampling

    Matthias Lorenzen, Fabrizio Dabbene, Roberto Tempo, and Frank Allg \"o wer. Stochastic mpc with offline uncertainty sampling. Automatica, 81: 0 176--183, 2017

    work page 2017

Show all 13 references
  1. [9]

    Chance-constrained sets approximation: A probabilistic scaling approach

    Martina Mammarella, Victor Mirasierra, Matthias Lorenzen, Teodoro Alamo, and Fabrizio Dabbene. Chance-constrained sets approximation: A probabilistic scaling approach. Automatica, 137: 0 110108, 2022

    2022

  2. [10]

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

    Ali Mesbah. Stochastic model predictive control: An overview and perspectives for future research. IEEE Control Systems Magazine, 36 0 (6): 0 30--44, 2016

    2016

  3. [11]

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

    Frauke Oldewurtel, Colin N Jones, and Manfred Morari. A tractable approximation of chance constrained stochastic mpc based on affine disturbance feedback. In 2008 47th IEEE conference on decision and control, pages 4731--4736. IEEE, 2008

    2008

  4. [12]

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming

    Andreas W \"a chter and Lorenz T Biegler. On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming. Mathematical programming, 106 0 (1): 0 25--57, 2006

    2006

  5. [13]

    Stochastic model predictive control using a combination of randomized and robust optimization

    Xiaojing Zhang, Kostas Margellos, Paul Goulart, and John Lygeros. Stochastic model predictive control using a combination of randomized and robust optimization. In 52nd IEEE conference on decision and control, pages 7740--7745. IEEE, 2013

    2013

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.