Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Offset-free model predictive control: stability under plant-model mismatch

T0 review · 2 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper proves that a particular nonlinear offset-free model predictive control design is robustly exponentially stable with respect to setpoint tracking error, despite plant-model mismatch and persistent disturbances, provided the…

desk verdict A solid conditional theorem for nonlinear offset-free MPC, but the key estimator assumption is not demonstrated for any practical estimator; deserves review but needs major revision. read the letter →

arxiv 2412.08104 v2 pith:QTWZHBEB submitted 2024-12-11 eess.SY cs.SYmath.OC

Steven J. Kuntz , James B. Rawlings This is my paper · ORCID
classification eess.SYcs.SYmath.OC MSC 93D0993D3093C1093C55
keywords offset-freemodelpredictivecontrolnonlinearMPCrobuststabilityplant-modelmismatchdisturbanceestimationsetpointtrackingLyapunovmovinghorizon
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

The paper supplies the first general stability theory for nonlinear offset-free model predictive control. It shows that a specific design, built from a steady-state target problem, a finite-horizon regulator, and a joint state-and-disturbance estimator, drives the setpoint tracking error to zero when setpoint and disturbance increments vanish, even under plant-model mismatch. The central result, Theorem 6, states that the closed-loop system is regionally robustly exponentially stable with respect to the tracking error $\delta_r := r - r_{sp}$ under Assumptions 1 to 9. A sympathetic reader cares because offset-free MPC has been used for decades without a stability guarantee that covers mismatch and persistent disturbances.

What carries the argument

The load-bearing structure is the three-part offset-free MPC design: a steady-state target problem (SSTP) that selects a feasible target pair $(x_s,u_s)$ meeting the setpoint, a finite-horizon optimal control regulator driving the state to that target, and a joint state-and-disturbance estimator providing integral action. Stability is carried by a joint Lyapunov theorem (Theorem 3) that combines the regulator value function $V_N^0$ and an estimator Lyapunov function $V_e$ into a single contractive quantity, with plant-model mismatch handled by steady-state correction functions $(\Delta x_s, d_s)$ that align plant and model steady states. Assumption 6, requiring the estimator to admit a global quadratic Lyapunov function, is what makes the interconnection of controller and estimator tractable.

What would settle it

Take the pendulum or CSTR example with the proposed moving-horizon estimator and a small, asymptotically constant mismatch; Theorem 6 predicts the tracking error converges to zero only if Assumption 6 holds. A single simulation where the mismatch increments vanish but the output offset remains bounded away from zero, with the theorem's other assumptions visibly satisfied, would falsify the claim's applicability, as would an explicit proof that the moving-horizon estimator cannot admit any global quadratic Lyapunov function satisfying (19).

Watch

Extended reading notes

Core claim

The paper's central claim is that offset-free performance can be guaranteed, not assumed, for a nonlinear MPC design. Theorem 6 establishes that there exist constants $\tau, \delta_w, \delta_\alpha > 0$ such that the closed-loop offset-free MPC system is regionally robustly exponentially stable with respect to the setpoint tracking error $\delta_r := r - r_{sp}$, and equivalently that as setpoint and disturbance increments vanish, both tracking error and estimator error converge to zero. The proof proceeds in three stages: nominal stability and offset-free performance, robustness to estimate errors and setpoint or disturbance changes, and finally robustness to sufficiently small plant-model mismatch. The result holds under quadratic costs, differentiability of plant and model functions, constraint backoffs at steady state, and a robustly stable state and disturbance estimator.

Load-bearing premise

The whole guarantee rests on the estimator having a global quadratic Lyapunov function (Assumption 6), which the paper notes is not known to hold for the moving-horizon estimators used in its own examples.

Editorial extensions

If this is right

  • Offset-free performance no longer needs to be assumed: under the stated assumptions it follows as a stability property of the closed-loop design.
  • The design applies to unstable nonlinear plants, as demonstrated on a continuously stirred-tank reactor operating near a Hopf bifurcation.
  • The theorem implies that tracking error and estimator error converge to zero whenever the setpoint and disturbance increments vanish, not only when signals are asymptotically constant.
  • For linearized systems, the rank and invertibility conditions on the matrices $M_1$ and $M_2$ recover classical linear offset-free MPC conditions, connecting the nonlinear result to existing practice.
  • Any estimator that satisfies the global quadratic Lyapunov condition can be plugged into the design and inherit the stability guarantee; the paper notes that moving-horizon estimators are not yet known to satisfy that condition.

Reading between the lines

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

  • The paper leaves explicit bounds on 'sufficiently small' mismatch unquantified; a natural extension would turn the constants in Proposition 7 into computable margins, as has been done for linear systems.
  • Assumption 8, requiring known steady-state mismatch corrections, is strong; the theorem does not cover online identification of those corrections, so a practical extension would investigate simultaneous learning and offset-free control.
  • The proof suggests a modular design philosophy: pair any estimator with a quadratic Lyapunov function with any regulator satisfying the backoff terminal condition to obtain offset-free tracking, decoupling estimation design from controller design.
  • One testable extension is to seek estimators satisfying Assumption 6, for instance via N-step Lyapunov constructions; success would immediately make the theorem applicable to moving-horizon estimation.
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 / 6 minor

Summary. The manuscript develops an offset-free model predictive control architecture for nonlinear discrete-time systems with plant-model mismatch, consisting of a steady-state target problem, a finite-horizon regulator with terminal ingredients constructed by linearization, and a joint state-and-disturbance estimator. It first proves nominal exponential stability with respect to target and setpoint tracking errors (Theorem 4), then robust stability with respect to estimate errors and setpoint/disturbance changes in the absence of mismatch (Theorem 5), and finally robust exponential stability of the joint controller-estimator loop for sufficiently small plant-model mismatch (Theorem 6) under Assumptions 1 to 9. The proof framework introduces an ISS/Lyapunov theory with respect to two measurement functions (Theorems 2 and 3) and is applied to the offset-free MPC loop. Numerical experiments on a pendulum and a CSTR compare the proposed offset-free MPC with a tracking MPC.

Significance. If the results are taken together with a constructive estimator satisfying Assumption 6, this would be a substantial contribution, filling a long-standing gap in the nonlinear offset-free MPC literature. The proof structure is careful and detailed, with complete appendices, and Section 7 provides verifiable local conditions for the steady-state target problem through rank conditions on linearizations. The main weakness is that the paper's key estimator assumption, Assumption 6, is not instantiated by any estimator used in the examples, so the central theorem currently lacks a demonstrated domain of application in the nonlinear setting claimed by the abstract.

major comments (2)
  1. [§8.1, Assumption 6, Theorem 6] Assumption 6, stated in Section 2.2.3 as requiring a global quadratic ISS Lyapunov function satisfying (19a)-(19b), is the most restrictive hypothesis of the central result. The paper itself states in Section 8.1 that the MHE estimators used in the numerical examples 'should be RGES' but 'it is not known if they satisfy Assumption 6'. Consequently, none of the simulations instantiates Theorem 6, and the theorem's domain of application is not demonstrated. The cited stability results for MHE (Allan and Rawlings 2021; Schiller et al. 2023) provide a Q-function and an N-step Lyapunov function, respectively, not the one-step global quadratic Lyapunov function required here, and no conversion argument is supplied. Because the abstract claims 'a nonlinear offset-free MPC design that is robustly stable', the authors need either to construct estimators that provably satisfy Assumption 6 and use them in the examples, or to relax Assumption 6 to a Q-function/N-step Lyapunov condition and adapt Theorems 3 and 6 accordingly. As written, the central nonlinear stability claim is conditional on an assumption that is not shown to be satisfiable by any practical estimator.
  2. [Abstract, Section 9] The abstract and conclusions present the result as an unconditional 'first general stability results' statement, but Theorem 6 is a conditional statement relying on Assumption 6, which the paper's own Section 8.1 admits is unverified for the MHE estimators actually used. Section 9 also lists 'the requirement of a Lyapunov function for the estimator (Assumption 6)' as an open direction for future work. The claims should be rephrased to state explicitly that the stability guarantee holds for estimators satisfying Assumption 6, and that no such estimator is constructed or validated in the paper. This is not merely a wording issue: it determines whether the paper delivers a design or only a conditional theorem.
minor comments (6)
  1. [§8.1] The default simulation parameters in Section 8.1 appear to swap the roles of (wP)4 and (wP)5: the text says 'discretization parameter (wP)4 = 1' and 'no measurement noise (wP)5 = 0', but in equations (50)-(51) (wP)4 is the measurement offset and (wP)5 scales the discretization error; as written, (wP)4 = 1 violates the stated bound (wP)4 ∈ [-0.05, 0.05] and contradicts the claim of no measurement noise.
  2. [§8.1] In the third pendulum experiment, the text says 'we have measurement noise (wP)5 ∼ N(0,10^-4)'; given (50b), the measurement noise is (wP)4, not (wP)5. The same subsection also writes 'where (wP)3 ∼ N(0,10^-2)' for the integrating disturbance, which should presumably be the increment (ΔwP)3.
  3. [§8.1 and §8.2] The phrase 'Runga-Kutta' should be 'Runge-Kutta' in both example sections.
  4. [§1] The sentence 'We also consider we softened regulator output constraints' contains an extra 'we' and should read 'We also consider softened regulator output constraints'.
  5. [Theorem 6(c), Definition 7] The statement 'RES w.r.t. (δr, δx̂)' is ambiguous because Definition 7 bounds |(ζ1, ε)| rather than |(ζ1, ζ2)|, and the proof of Theorem 6(c) indeed bounds |(δr, e)|. The authors should state explicitly that the conclusion includes the estimator error e, consistent with Definition 7.
  6. [§8.2] The CSTR terminal region is reported as cf ≈ 6.5 × 10^-16, which is close to floating-point precision. Since the simulations are intended to illustrate the theory, the authors should report how the terminal constraint Xf is verified numerically and whether the optimizer reliably returns states inside this extremely small terminal region.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 6 is a conditional result proved from stated assumptions, and the main weakness (Assumption 6 unverified for the demonstrated MHE) is an applicability gap, not a circular reduction.

full rationale

The derivation chain is self-contained. Nominal stability (Theorem 4) is proved from Assumptions 1–5 via standard MPC value-function arguments. Robust stability without mismatch (Theorem 5) is proved in Appendix B.2 from Propositions 4–5, which are derived from the assumptions. The mismatch result (Theorem 6) is proved in Appendix B.3: Proposition 6 derives the mismatch-noise bound (83) using Taylor's theorem and Assumptions 8–9; Proposition 7 combines this bound with the regulator cost decrease (90) and the estimator Lyapunov decrease (91); Theorem 3 then supplies a small-gain argument. None of these steps fits a parameter to data or renames a conclusion as an input. The cited companion paper (Kuntz and Rawlings 2024) is used for motivation ('As in Kuntz and Rawlings (2024), we use quadratic costs') and for a parallel error-bound comment (Remark 22), but the actual bounds used in the proof are derived inside this paper, so the citation is not load-bearing. The genuinely load-bearing assumption is Assumption 6, which requires a global quadratic ISS Lyapunov function for the estimator. The paper honestly states in Section 8.1 that the MHE estimators used in the examples 'should be RGES (Allan and Rawlings, 2021)' but 'it is not known if they satisfy Assumption 6'; this means the theorem has not been instantiated for the demonstrated estimators. That is a completeness or correctness risk, not circularity: the theorem does not assume its own conclusion, and the estimator Lyapunov condition (19) is strictly stronger than, and logically prior to, the RGES property derived in Theorem 1. Similarly, the admitted lack of quantification of 'sufficiently small mismatch' in the Conclusions is a limitation, not a circular step.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

No parameters fitted to data; Q,R and back-offs are design choices and the theorem holds for any positive definite Q,R; tau, delta_w, delta_alpha are existential. The integrating disturbance and corrected model state are standard modeling constructs, not new entities.

assumptions (8)
  • domain assumption Assumption 1: continuity of plant, model, and reference functions plus nominal consistency at d=0.
    Standard smoothness and consistency needed to compare model and plant; invoked throughout.
  • domain assumption Assumption 2: X,Y closed, U,W,D compact, constraint back-offs b>0.
    Ensures SSTP feasibility and terminal set construction; back-offs are a design choice.
  • domain assumption Assumption 3: SSTP existence/coercivity.
    Guarantees steady-state target problem has solutions.
  • ad hoc to paper Assumptions 4 and 5: terminal control law and quadratic stage/terminal costs.
    The design is restricted to quadratic costs; necessary for the Lyapunov arguments.
  • ad hoc to paper Assumption 6: estimator admits a robust Lyapunov function satisfying (19a)-(19b).
    Most fragile; not known for MHE/FIE used in examples (Section 8.1).
  • domain assumption Assumption 7: SSTP continuity and robust feasibility under estimator errors.
    Needed for recursive feasibility of the target problem.
  • ad hoc to paper Assumption 8: existence, uniqueness, and continuity of mismatch corrections (x_P,s, d_s).
    Strong structural condition; verified in examples by explicit formulas.
  • domain assumption Assumption 9: differentiability of plant and model functions.
    Used in Taylor expansion to bound mismatch noise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Offset-free model predictive control: stability under plant-model mismatch." pith.science (2026). https://pith.science/paper/QTWZHBEB

@misc{pith2026241208104,
  author       = {Pith},
  title        = {Pith review of: Offset-free model predictive control: stability under plant-model mismatch},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QTWZHBEB}},
  note         = {Machine review of arXiv:2412.08104}
}
read the original abstract

We present the first general stability results for nonlinear offset-free model predictive control (MPC). Despite over twenty years of active research, the offset-free MPC literature has not shaken the assumption of closed-loop stability for establishing offset-free performance. In this paper, we present a nonlinear offset-free MPC design that is robustly stable with respect to the tracking errors, and thus achieves offset-free performance, despite plant-model mismatch and persistent disturbances. Key features and assumptions of this design include quadratic costs, differentiability of the plant and model functions, constraint backoffs at steady state, and a robustly stable state and disturbance estimator. We first establish nominal stability and offset-free performance. Then, robustness to state and disturbance estimate errors and setpoint and disturbance changes is demonstrated. Finally, the results are extended to sufficiently small plant-model mismatch. The results are illustrated by numerical examples.

Figures

Figures reproduced from arXiv: 2412.08104 by the authors.

Figure 1
Figure 1. Example systems. Remark 21. Invertibility of M2 is a key assumption in linear offset-free MPC (Muske and Badgwell, 2002; Pannocchia and Rawlings, 2003). In fact, it is known that the sys￾tem (45a)–(45c) is detectable if and only if M2 is full column rank and (A, C) is de￾tectable (Pannocchia and Rawlings, 2003, Lem. 1). Moreover, M2 full row rank can be interpreted as a steady-state observability condition: at stead… view at source ↗
Figure 2
Figure 2. Simulated closed-loop trajectories for the offset-free MPC and tracking MPC of [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Nominal steady states for the CSTR (54). [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Simulated closed-loop trajectories for the offset-free MPC and tracking MPC of [PITH_FULL_IMAGE:figures/full_fig_p029_4.png]

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. Comparing Model-based Control Strategies for a Quadruple Tank System: Decentralized PID, LMPC, and NMPC

    math.OC 2025-09 conditional novelty 4.0 of 10

    On a quadruple tank system, MPCs beat a SIMC-tuned decentralized PID for setpoint tracking with future setpoint preview, but the PID is competitive for disturbance rejection and without preview.

Reference graph

Works this paper leans on

19 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [9]

    Zeilinger

    TWCCC Technical Report 2024-04 55 Simon Muntwiler, Johannes K¨ ohler, and Melanie N. Zeilinger. MHE under parametric uncertainty – Robust state estimation without informative data,

    work page 2024

  2. [10]

    arXiv:2312.14049 [cs, eess]

    URL http: //arxiv.org/abs/2312.14049. arXiv:2312.14049 [cs, eess]. Kenneth R. Muske and Thomas A. Badgwell. Disturbance modeling for offset-free linear model predictive control. J. Proc. Cont. , 12(5):617–632,

    work page arXiv

  3. [11]

    doi: 10.1016/j.ifacol

    ISSN 2405-8963. doi: 10.1016/j.ifacol. 2018.09.332. Gabriele Pannocchia and James B. Rawlings. Disturbance models for offset-free MPC control. AIChE J. , 49(2):426–437,

    doi:10.1016/j.ifacol 2018

  4. [12]

    doi: 10.1016/j.ifacol.2015.11.304

    ISSN 2405-8963. doi: 10.1016/j.ifacol.2015.11.304. James B. Rawlings and Luo Ji. Optimization-based state estimation: Current status and some new results. J. Proc. Cont. , 22:1439–1444,

    work page doi:10.1016/j.ifacol.2015.11.304 2015

  5. [14]

    doi: 10.1109/TAC.2023.3280344

    ISSN 1558-2523. doi: 10.1109/TAC.2023.3280344. E. D. Sontag and Y. Wang. On the characterization of the input to state stability property. Sys. Cont. Let. , 24:351–359,

    work page arXiv 2023

  6. [18]

    Marco Vaccari and Gabriele Pannocchia

    doi: 10.1109/CDC.2015.7402474. Marco Vaccari and Gabriele Pannocchia. A Modifier-Adaptation Strategy Towards Offset- Free Economic MPC. Processes, 5(1):2,

    arXiv 2015

  7. [19]

    doi: 10.3390/pr5010002. Megan A. Zagrobelny. MPC Performance Monitoring and Disturbance Model Identifi- cation. PhD thesis, University of Wisconsin–Madison, December

    work page doi:10.3390/pr5010002

  8. [1998]

    doi: 10.1137/S0895479895291303

    ISSN 0895-4798. doi: 10.1137/S0895479895291303. Duc N. Tran, Christopher M. Kellett, and Peter M. Dower. Input-to-state stabil- ity with respect to two measurement functions: Discrete-time systems. In 2015 54th IEEE Conference on Decision and Control (CDC) , pages 1817–1822,

    doi:10.1137/s0895479895291303 2015

Show all 19 references
  1. [1999]

    doi: 10.1016/S0167-6911(99)00070-5

    ISSN 01676911. doi: 10.1016/S0167-6911(99)00070-5. Ji-Guang Sun. Perturbation Theory for Algebraic Riccati Equations. SIAM J. Matrix Anal. and Appl. , 19(1):39–65,

    doi

  2. [2000]

    doi: 10.1137/ S0363012999350213

    ISSN 0363-0129. doi: 10.1137/ S0363012999350213. TWCCC Technical Report 2024-04 56 Eduardo D. Sontag and Yuan Wang. Notions of input to output stability. Sys. Cont. Let. , 38(4-5):235–248,

    2024

  3. [2008]

    doi: 10.1016/j.jprocont.2007.10.004

    ISSN 0959-1524. doi: 10.1016/j.jprocont.2007.10.004. Julian D. Schiller and Matthias A. M¨ uller. A moving horizon state and parameter esti- mation scheme with guaranteed robust convergence. IF AC–P. Online, 56(2):6759–6764,

    2007 doi

  4. [2010]

    doi: 10.1016/j.automatica

    ISSN 00051098. doi: 10.1016/j.automatica. 2010.05.023. Manfred Morari and Urban Maeder. Nonlinear offset-free model predictive control. Auto- matica, 48(9):2059–2067,

    2010 doi

  5. [2012]

    doi: 10.1007/978-1-4419-9982-5

    ISBN 978-1-4419-9981-8 978-1-4419-9982-5. doi: 10.1007/978-1-4419-9982-5. Daniel Limon, Antonio Ferramosca, Ignacio Alvarado, and Teodoro Alamo. Nonlinear MPC for Tracking Piece-Wise Constant Reference Signals. IEEE Trans. Auto. Cont., 63 (11):3735–3750,

    doi

  6. [2015]

    doi: 10.1080/00207179.2014.972464

    ISSN 0020-7179. doi: 10.1080/00207179.2014.972464. Timm Faulwasser and Gabriele Pannocchia. Toward a Unifying Framework Blending Real- Time Optimization and Economic Model Predictive Control. Ind. Eng. Chem. Res. , 58 (30):13583–13598,

    2014

  7. [2017]

    doi: 10.1016/j.sysconle.2017.03.005

    ISSN 0167-6911. doi: 10.1016/j.sysconle.2017.03.005. TWCCC Technical Report 2024-04 54 T. M. Apostol. Mathematical analysis. Addison-Wesley,

    2017 doi

  8. [2018]

    doi: 10.1109/TAC.2018

    ISSN 0018-9286, 1558-2523, 2334-3303. doi: 10.1109/TAC.2018. 2798803. Urban Maeder and Manfred Morari. Offset-free reference tracking with model predictive control. Automatica, 46(9):1469–1476,

    2018 doi

  9. [2019]

    doi: 10.1021/acs.iecr.9b00782

    ISSN 0888-5885. doi: 10.1021/acs.iecr.9b00782. Giacomo Galuppini, Lalo Magni, and Antonio Ferramosca. Nonlinear MPC for Tracking Piecewise-Constant Reference Signals: the Positive Semidefinite Stage Cost Case. IF AC– P. Online , 56(1):210–215,

    doi

  10. [2023]

    doi: 10.1016/j.ifacol.2023.02.036

    ISSN 2405-8963. doi: 10.1016/j.ifacol.2023.02.036. Jack Hale. Ordinary Differential Equations. Robert E. Krieger Publishing Company, second edition,

    2023 doi

  11. [2024]

    URL https://arxiv.org/abs/2411. 15452. arXiv:2411.15452 [eecs, math]. John M. Lee. Introduction to Smooth Manifolds , volume 218 of Graduate Texts in Mathe- matics. Springer, New York, NY,

    arXiv

Pith tools

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