Pith. sign in

REVIEW 5 minor 1 cited by

Finite-Sample Closed-Loop Stability of Model Predictive Path Integral Control for Linear Time-Invariant Systems

T0 review · 0 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Finite-sample MPPI on linear systems is a controllable perturbation of LQR, with explicit sample counts that guarantee practical exponential closed-loop stability.

desk verdict First explicit finite-sample closed-loop stability certificate for MPPI on LTI systems; scoped tightly, math holds, useful M* formula. read the letter →

arxiv 2607.04006 v1 pith:XOS25AFK submitted 2026-07-04 math.OC cs.ROcs.SYeess.SY

Hyung-Jin Yoon , Hunmin Kim This is my paper · ORCID
classification math.OCcs.ROcs.SYeess.SY MSC 93D1593E1549N10
keywords MPPImodelpredictivecontrolfinite-samplestabilityLQRLyapunovperturbationsampling-basedMPCDAREterminalcostpracticalexponential
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 gives the first finite-sample closed-loop stability guarantees for Model Predictive Path Integral (MPPI) control on discrete-time linear systems with additive Gaussian noise. The key observation is that, with the DARE terminal cost, the exact finite-horizon optimal first action equals the infinite-horizon LQR law for every planning horizon, so finite-sample MPPI can be treated as a stochastic perturbation of classical LQR. The authors prove that the control error splits into a Monte Carlo term that shrinks like one over square-root of the sample count and a temperature bias that vanishes as temperature goes to zero; once both are small enough relative to the LQR Lyapunov margin, the closed-loop state decays exponentially in expectation on high-probability paths that stay inside a compact sublevel set over any finite operating horizon. Three residual floors remain—process noise, MPPI approximation, and per-step sampling failure—and the sufficient sample threshold is written explicitly in terms of the DARE solution, the LQR margin, temperature, and horizon. A sympathetic reader cares because the result finally answers, with a computable formula, how many MPPI samples are enough for certified stability, and recovers the classical stochastic LQR bound in the infinite-sample, zero-temperature limit.

What carries the argument

The exact coincidence of the first finite-horizon optimal action with the infinite-horizon LQR gain for every planning horizon when the terminal cost is the DARE solution. That identity lets the authors view MPPI as a stochastic perturbation of LQR, decompose the error into Monte Carlo concentration plus closed-form temperature bias (via Gaussian completing-the-square), and absorb both into the classical LQR Lyapunov decrease through a stopped-process supermartingale argument.

What would settle it

On the double-integrator benchmark (or any LTI plant meeting the assumptions), compute the analytical M-star from the paper’s Corollary 1; if for every M at or above that threshold the empirical median Lyapunov ratio remains strictly larger than the certified decay rate while process noise and temperature stay at the paper’s values, the practical-stability claim is false.

Watch

Extended reading notes

Core claim

For unconstrained LTI systems with quadratic costs and DARE terminal cost, finite-sample MPPI approximates the LQR feedback with high probability: the error decomposes into a Monte Carlo term of order M to the minus one-half and an infinite-sample temperature bias that vanishes as temperature tends to zero. Under a small-gain condition on that bias, a Lyapunov perturbation argument yields practical exponential stability in expectation on sample paths that remain in a compact Lyapunov sublevel set over a finite horizon, with three explicit residual floors and an explicit sufficient sample threshold M-star computable from the DARE solution and LQR stability margin.

Load-bearing premise

The whole argument rests on the system being unconstrained, linear, and quadratic with the exact infinite-horizon cost-to-go used as the terminal cost, so that every finite-horizon plan has the same first move as classical LQR.

Editorial extensions

If this is right

  • With sample count above an explicit threshold built from the DARE solution and LQR margin, the closed loop is practically exponentially stable in expectation on high-probability finite-horizon paths.
  • In the joint limit of infinite samples and vanishing temperature the bound recovers the classical stochastic LQR stability certificate.
  • The sample threshold and residual floors are parametrized by planning horizon, temperature, and sampling covariance, so designers can trade samples against temperature and horizon.
  • The same bound admits a practical input-to-state stability reading with three explicit gains for process noise, MPPI approximation, and confidence loss.
  • On the double-integrator the analytical threshold is numerically computable and qualitatively marks the onset of certified decay rates.

Reading between the lines

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

  • The same first-action coincidence may serve as a template for sample-complexity certificates of other sampling-based MPC methods that still lack closed-loop guarantees.
  • Bounded or truncated process noise would lift the finite-horizon localization to infinite-horizon high-probability invariance, exactly as the paper’s own future-work section anticipates.
  • Because the certificate is deliberately conservative, tighter problem-specific concentration constants could convert the sufficient M-star into a sharper design knob without changing the controller architecture.
  • The closed-form temperature-bias gain suggests an adaptive schedule that lowers temperature once samples have concentrated, shrinking the residual floor at no 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

0 major / 5 minor

Summary. The paper establishes finite-sample closed-loop practical exponential stability for Model Predictive Path Integral (MPPI) control on discrete-time LTI systems with quadratic costs and additive Gaussian process noise. Under the DARE terminal cost, the exact finite-horizon MPC first action coincides with infinite-horizon LQR for every horizon N, so finite-sample MPPI is treated as a stochastic perturbation of LQR. Lemmas 1–3 give high-probability approximation of the LQR feedback by a Monte Carlo term O(M^{-1/2}) plus an infinite-sample temperature bias characterized in closed form via completing-the-square (Proposition 1). Lemma 5 establishes high-probability finite-horizon invariance of a Lyapunov sublevel set via a stopped supermartingale, resolving the compact-set circularity. Theorem 1 then yields an unconditional Lyapunov bound on the stopped process and, on paths that remain in Ω_R over [0,T] (probability ≥1−δ), the bound E[∥x_k∥ 1_{τ_R>T}] ≤ c ρ^k ∥x_0∥ + γ_w √tr(Σ_w) + γ_M e_M(η) + γ_η √η. Corollary 1 supplies an explicit sufficient sample threshold M* computable from the DARE solution, LQR margin, temperature, and horizon; the joint limit recovers the stochastic LQR certificate. Simulations on a double integrator illustrate the certificate’s conservatism and qualitative consistency.

Significance. Closed-loop stability of finite-sample MPPI under receding-horizon execution has been identified as an open problem; this manuscript supplies the first explicit finite-sample certificate for the unconstrained LTI/quadratic case. The reduction via DARE terminal cost is clean, the bias formula is closed-form, the sample threshold is computable rather than existential, and the three residual floors (process noise, MPPI approximation, confidence) are transparent. Simulation code is released and the conservatism of ρ and M* is documented rather than hidden. Within its stated scope the result is a solid foundation that recovers classical stochastic LQR in the appropriate limit and connects sampling-based MPC to the Mayne et al. Lyapunov framework. The restriction to unconstrained LTI/quadratic systems is a genuine limitation of scope, not a flaw in the argument as written.

minor comments (5)
  1. In the abstract and §I the phrase “finite-sample certificate is parametrized by the selected planning horizon” is accurate (Remark 5), but a short explicit pointer in Corollary 1 to the N-dependence of C_{X,U}, H_N and F_N would help readers who might otherwise expect horizon-independent constants.
  2. Table I lists ρ(certificate bound)=(1−α/2)^{1/2}=0.9429; the same quantity appears as ρ in (36). A single consistent symbol (or a parenthetical note that the square-root form is used for the state-norm bound) would avoid a momentary notational mismatch.
  3. Experiment 1 caption states that the plotted envelope omits the calibrated MPPI approximation floor. Adding a second curve that includes a representative γ_M e_M term (even if only for illustration) would make the comparison with Theorem 1 more direct.
  4. Assumption 6 (bounded MPPI update) is essential for the linear-in-η bad-event residual. A one-sentence remark that actuator saturation or truncated sampling already implements this bound in every practical MPPI code base would further reassure readers.
  5. The arXiv identifier of the companion nonlinear paper [15] is still “in preparation”; once available, a forward reference with the arXiv number would improve traceability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: LQR reduction, concentration, bias, and Lyapunov perturbation are derived from stated assumptions and standard tools, not by construction from fitted targets or load-bearing self-citations.

full rationale

The central chain (Assumption 3 DARE terminal cost implies exact first-action coincidence of finite-horizon MPC with infinite-horizon LQR for every N; Lemmas 1–3 give high-probability Monte Carlo + temperature-bias approximation of LQR via Hoeffding/sub-Gaussian concentration and completing-the-square; Lemma 5 high-probability finite-horizon invariance of a Lyapunov sublevel via stopped supermartingale; Theorem 1 practical exponential stability under the small-gain condition Φ(β∞) ≤ αP/2; Corollary 1 explicit M*) is self-contained. Constants are expressed from the DARE solution, system matrices, λ, Σϵ, N and concentration parameters; they are not fitted to the stability claim itself. Self-citations ([5] open-loop sampling complexity; [15] companion nonlinear paper in preparation) supply background or future work and are not used as uniqueness theorems or load-bearing premises for the closed-loop certificate. Simulation constant C1 = 0.22 calibrates only a numerical M* illustration and does not enter the theorems. Finite-horizon localization under unbounded Gaussian noise is acknowledged (Remark 10) rather than smuggled. No equation reduces to its own input by definition, and no prediction is a renamed fit. Minor self-citation presence without load-bearing role yields score 1 rather than 0.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The certificate rests on standard LQR/DARE theory, classical concentration inequalities, and six modeling assumptions that restrict the setting to unconstrained LTI/quadratic systems with bounded warm-starts and controls. No new physical entities are postulated. Free parameters are the usual MPPI design knobs (M, λ, N, Σϵ, η, δ, R) and one simulation calibration constant C1; the theorem’s M* is expressed in terms of system and design quantities rather than fitted to closed-loop data.

free parameters (5)
  • MPPI temperature λ
    Design parameter that must be small enough (λ ≤ λ*) for the small-gain condition; chosen by the user, not fitted to stability data.
  • Sample count M and failure probability η
    User-chosen accuracy/confidence parameters that enter εM(η) and M*; not fitted from closed-loop outcomes in the theorem.
  • Planning horizon N and sampling covariance Σϵ
    Horizon-dependent stacked cost matrices and sampling law enter CX,U, κλ, and β coefficients; design choices that parametrize the certificate.
  • Simulation concentration constant C1 = 0.22
    Used only to numerically evaluate M* = 153 for the double-integrator benchmark; not part of the general theorem statement.
  • Invariance level R and failure probability δ
    Chosen via Lemma 5 from V(x0), Δ, and α to enforce P(τR ≤ T) ≤ δ; free design of the localization set.
assumptions (7)
  • domain assumption Pair (A,B) stabilizable and (A,Q^{1/2}) detectable so the DARE has a unique P ≻ 0 (Assumptions 1–2).
    Standard LQR existence conditions; invoked to guarantee the CLF decrease (7).
  • domain assumption Terminal cost P is the DARE solution, so the finite-horizon first action equals LQR for every N (Assumption 3).
    Load-bearing structural assumption that removes horizon-truncation error and enables the LQR-perturbation analysis.
  • domain assumption MPPI draws i.i.d. Gaussian perturbations independent across time (Assumption 4).
    Enables concentration and independence of sampling noise from process noise.
  • domain assumption Nominal sequence stays in a compact warm-start set UN of finite diameter DU (Assumption 5).
    Needed for uniform lower bound Z on E[w] and for bias coefficients β0, Cbad.
  • domain assumption Implemented MPPI control is almost surely bounded (Assumption 6).
    Makes the bad-event Lyapunov residual linear in η rather than √η; matches practical saturation.
  • standard math Hoeffding and sub-Gaussian concentration for bounded weights and Gaussian perturbations (Lemma 2).
    Standard concentration tools used to obtain εM(η) = O(√(log(m/η)/M)).
  • domain assumption Process noise wk ~ N(0, Σw) i.i.d., unbounded (system model).
    Forces finite-horizon localization rather than infinite-horizon almost-sure invariance (Remark 10).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Finite-Sample Closed-Loop Stability of Model Predictive Path Integral Control for Linear Time-Invariant Systems." pith.science (2026). https://pith.science/paper/XOS25AFK

@misc{pith2026260704006,
  author       = {Pith},
  title        = {Pith review of: Finite-Sample Closed-Loop Stability of Model Predictive Path Integral Control for Linear Time-Invariant Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XOS25AFK}},
  note         = {Machine review of arXiv:2607.04006}
}
read the original abstract

We establish finite-sample closed-loop stability guarantees for Model Predictive Path Integral (MPPI) control applied to discrete-time Linear Time-Invariant (LTI) systems with additive Gaussian process disturbances. The key observation is that, for unconstrained LTI/quadratic systems with the DARE terminal cost, the exact finite-horizon MPC law has the same first control action as the infinite-horizon LQR law for every planning horizon. Thus, finite-sample MPPI can be analyzed as a stochastic perturbation of LQR. First, we show that the MPPI control law approximates the LQR feedback with high probability. The approximation error decomposes into a Monte Carlo term that decreases with the sample count and an infinite-sample temperature bias that persists at finite temperature but vanishes as the temperature is reduced. The resulting constants are written in terms of the horizon-dependent stacked cost matrices, making explicit that the finite-sample certificate is parametrized by the selected planning horizon. Second, we use a Lyapunov perturbation argument to prove practical exponential stability in expectation. On sample paths that remain in a compact Lyapunov sublevel set over a finite operating horizon, the expected state norm decays exponentially up to three residual floors: a process-noise floor, an MPPI approximation floor, and a confidence floor from the per-step sampling failure probability. The sufficient sample threshold is explicit and computable from the DARE solution, LQR stability margin, MPPI sampling parameters, temperature, and planning horizon. In the joint limit of infinite samples and vanishing temperature bias, the result recovers the stochastic LQR stability bound.

Figures

Figures reproduced from arXiv: 2607.04006 by the authors.

Figure 1
Figure 1. Experiment 1: Empirical E[∥xk∥] for M ∈ {50, 200, 1000} compared with the nominal certificate decay term and the LQR reference decay. Shaded bands show ±1σ across 300 Monte Carlo trials. Certified trajectories (M ≥ M∗ = 153) remain bounded and decay toward a practical steady-state level. The displayed envelope is qualitative because the explicitly calibrated finite-sample MPPI approximation floor is not included in … view at source ↗
Figure 2
Figure 2. Experiment 2: Empirical decay rate ρˆ(M) vs. sample count on a log scale. Shaded regions mark the unstable zone (ρ >ˆ 1, red), the uncertified-but-stable zone (ρ < ρˆ ≤ 1, orange), and the certified zone (ρˆ ≤ ρ = 0.943, green). The vertical dashed line marks M∗ = 153. Annotated values are ρˆ at each tested M. 10 5 0 5 10 x1 (position) 8 6 4 2 0 2 4 6 8 x2 (v elo city) M = 50 < M * = 153 (uncertified) LQR 2 ellipse … view at source ↗
Figure 3
Figure 3. Experiment 3: Phase portraits for M = 50 (left, red, uncertified) and M = 500 (right, blue, certified). Circles mark trajectory starts; squares mark ends. The dashed ellipse is the LQR 2σ steady-state set. Certified trajectories concentrate visibly closer to the origin (mean terminal norm 0.92 vs. 1.65); no divergence occurs in either case. F. Experiment 4: Analytical vs. Empirical M∗ Table II and [PITH_FULL_IMAGE:… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Experiment 5: Normalized ESSk/M (top) and ∥xk∥ (bottom) over T = 200 steps at M = 500. Orange shading marks steps where ESS/M falls below the empirical threshold Mˆ ∗/M = 0.060. Despite ESS/M ≈ 0.003 throughout, the trajectory converges stably, confirming that low ESS …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Stochastic Stability of Nonlinear MPPI via Contraction Theory and Control Lyapunov Functions

    eess.SY 2026-07 conditional novelty 6.0 of 10

    Finite-sample MPPI inherits the contraction-based stability of a nominal nonlinear MPC policy under an explicit small-gain condition on the approximation error, yielding finite-horizon high-probability localized mean ...

Reference graph

Works this paper leans on

15 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Model predictive path integral control: From theory to parallel computation,

    G. Williams, A. Aldrich, and E. A. Theodorou, “Model predictive path integral control: From theory to parallel computation,”Journal of Guidance, Control, and Dynamics, vol. 40, no. 2, pp. 344–357, 2017

    2017

  2. [2]

    Information-theoretic model predictive control: Theory and applications to autonomous driving,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Information-theoretic model predictive control: Theory and applications to autonomous driving,”IEEE Transactions on Robotics, vol. 34, no. 6, pp. 1603–1622, 2018

    2018

  3. [3]

    Aggressive driving with model predictive path integral control,

    ——, “Aggressive driving with model predictive path integral control,” inProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2016, pp. 1433–1440

    2016

  4. [4]

    An online learning approach to model predictive control,

    N. Wagener, C.-A. Cheng, J. Sacks, and B. Boots, “An online learning approach to model predictive control,” inProceedings of Robotics: Science and Systems (RSS), 2019

    2019

  5. [5]

    Sampling complexity of path integral methods for trajectory optimization,

    H.-J. Yoon, C. Tao, H. Kim, N. Hovakimyan, and P. V oulgaris, “Sampling complexity of path integral methods for trajectory optimization,” inProceedings of the American Control Conference (ACC), 2022, pp. 3482–3487, arXiv:2203.10067

    arXiv 2022

  6. [6]

    CoVO-MPC: Theoretical analysis of sampling-based MPC and optimal covariance design,

    Z. Yi, C. Pan, G. He, G. Qu, and G. Shi, “CoVO-MPC: Theoretical analysis of sampling-based MPC and optimal covariance design,” in Proceedings of the 6th Annual Learning for Dynamics and Control Conference (L4DC), ser. Proceedings of Machine Learning Research, vol. 242, 2024, pp. 1122–1135

    2024

  7. [7]

    Optimality and suboptimality of MPPI control in stochastic and deterministic settings,

    H. Homburger, F. Messerer, M. Diehl, and J. Reuter, “Optimality and suboptimality of MPPI control in stochastic and deterministic settings,” IEEE Control Systems Letters, 2025

    2025

  8. [8]

    Model predictive path integral control as preconditioned gradient descent,

    M. Fazlyab, S. Sharifi, and J. Wang, “Model predictive path integral control as preconditioned gradient descent,”arXiv preprint arXiv:2603.24489, 2026

    arXiv 2026

Show all 15 references
  1. [9]

    Robust model predictive path integral control: Analysis and performance guarantees,

    M. S. Gandhi, B. Vlahov, J. Gibson, G. Williams, and E. A. Theodorou, “Robust model predictive path integral control: Analysis and performance guarantees,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3653–3660, 2021

    2021

  2. [10]

    On contraction analysis for nonlinear systems,

    W. Lohmiller and J.-J. E. Slotine, “On contraction analysis for nonlinear systems,”Automatica, vol. 34, no. 6, pp. 683–696, 1998

    1998

  3. [11]

    Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design,

    I. R. Manchester and J.-J. E. Slotine, “Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design,”IEEE Transactions on Automatic Control, vol. 62, no. 6, pp. 3046–3053, 2017

    2017

  4. [12]

    Constrained model predictive control: Stability and optimality,

    D. Q. Mayne, J. B. Rawlings, C. V . Rao, and P. O. M. Scokaert, “Constrained model predictive control: Stability and optimality,” Automatica, vol. 36, no. 6, pp. 789–814, 2000

    2000

  5. [13]

    Model predictive control via probabilistic inference: A tutorial and survey,

    K. Honda, “Model predictive control via probabilistic inference: A tutorial and survey,”Annual Reviews in Control, vol. 61, p. 101052, 2026

    2026

  6. [14]

    B. D. O. Anderson and J. B. Moore,Optimal Control: Linear Quadratic Methods. Englewood Cliffs, NJ: Prentice-Hall, 1990

    1990

  7. [15]

    Global stability of nonlinear MPPI control via contraction theory and control Lyapunov functions,

    H.-J. Yoon, “Global stability of nonlinear MPPI control via contraction theory and control Lyapunov functions,” 2026, companion paper, in preparation

    2026

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.