pith. sign in

arxiv: 2604.17098 · v1 · submitted 2026-04-18 · 🧮 math.OC

Reference Condensation for Model Predictive Control with Preview

Daniel Arnstr\"om This is my paper

Pith reviewed 2026-05-10 06:09 UTC · model grok-4.3

classification 🧮 math.OC
keywords model predictive controlpreviewreference condensationleast-squares projectiontrackingexplicit MPC
0
0 comments X

The pith

A linear condensation map compresses future references into one setpoint for model predictive control, keeping the parameter dimension fixed regardless of preview length.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Model predictive control can use preview of upcoming reference signals to improve tracking, yet this normally makes the online problem larger as the horizon grows. The paper introduces reference condensation, a fixed linear map that reduces the entire preview sequence to a single equivalent setpoint. The map is obtained by least-squares projection for the unconstrained case and by a weighted variant that emphasizes the immediate control move in receding-horizon MPC. Numerical tests on a double integrator and an aircraft model show the weighted map recovers nearly the full benefit of preview while the decision-variable count stays constant. The result is that longer preview horizons no longer inflate computational or memory requirements.

Core claim

Reference condensation applies a precomputed linear map to any preview trajectory so that the resulting single setpoint produces tracking performance comparable to using the full preview sequence inside the MPC optimization; the map is derived offline from a least-squares fit (or weighted variant) and leaves the online parameter dimension independent of horizon length.

What carries the argument

The linear condensation map obtained by least-squares projection of the reference trajectory onto the system's predicted response.

Load-bearing premise

The offline linear map, built from a nominal projection, continues to produce good closed-loop behavior when the actual reference deviates from the assumed shape or when the plant differs from the model used to derive the map.

What would settle it

Apply both the condensed and full-preview controllers to a reference containing abrupt jumps or high-frequency content outside the projection basis and measure whether the integrated tracking error of the condensed version exceeds that of full preview by more than a few percent.

Figures

Figures reproduced from arXiv: 2604.17098 by Daniel Arnstr\"om.

Figure 1
Figure 1. Figure 1: Step responses of four MPC controllers. The preview [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Entries of the condensation matrix S (row vector for SISO). Near-term references receive the highest weight, with an approximately exponential decay. The weights sum to one (affine combination, Lemma 2). begin accelerating before t = 5 s, which gives much faster tracking than no preview. The average-reference baseline also anticipates the change, but tracks it less well. The reference￾condensation trajecto… view at source ↗
Figure 3
Figure 3. Figure 3: Sinusoidal reference tracking (r(t) = sin(0.5t), N = 50). Preview and reference condensation achieve nearly identical tracking, both dramatically outperforming no preview and the average-reference baseline. ISE values are reported in Table I. TABLE I: Integrated squared error (ISE) for step and sinu￾soidal references with horizon N = 50. Reference No preview Average ref. Ref. cond. Preview Step 1.114 0.628… view at source ↗
read the original abstract

In model predictive control (MPC), preview information can greatly improve tracking. Including preview information does, however, increase the parameter dimension linearly with the preview horizon, which increases online cost and, more importantly, the complexity of explicit MPC. We introduce reference condensation, a method that compresses a future reference trajectory into a single setpoint through a linear map. For the unconstrained tracking problem, the map follows from a least-squares projection. For receding-horizon MPC, we also study a weighted variant that prioritizes the first applied control. Numerical experiments on a double integrator and a higher-order aircraft example show that the weighted condensation closely matches full preview while keeping the parameter dimension independent of the preview horizon.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes reference condensation to compress a preview reference trajectory into a single setpoint for MPC via a linear map. For the unconstrained tracking problem the map is obtained from least-squares projection; a weighted variant prioritizing the first control input is studied for receding-horizon MPC. Numerical experiments on a double integrator and a higher-order aircraft model are used to show that the weighted condensation closely matches full-preview performance while keeping the parameter dimension independent of the preview horizon.

Significance. If the fixed linear map generalizes, the method would reduce online computational cost and the complexity of explicit MPC formulations that incorporate preview, without substantial degradation in tracking. The reported numerical results on two systems provide concrete support for the performance claim under the tested conditions, and the parameter-free nature of the least-squares derivation (for the unconstrained case) is a positive feature.

major comments (2)
  1. [Method / Derivation of the condensation map] The central performance claim rests on the fixed linear condensation map producing a setpoint whose closed-loop effect nearly equals full preview. The manuscript should supply explicit conditions or error bounds showing when this holds (e.g., invariance to reference class or plant perturbations), as the least-squares projection is derived offline from an assumed trajectory ensemble.
  2. [Numerical Experiments] Numerical experiments (double integrator and aircraft) demonstrate close match under specific conditions, but no sensitivity analysis is provided for references containing steps, ramps, or frequencies outside the implicit span of the projection basis, nor for unmodeled dynamics. These tests are load-bearing for the claim that the map remains effective in receding-horizon use.
minor comments (2)
  1. [Weighted variant] Clarify the exact weighting factor used in the receding-horizon variant and whether it is the only free parameter.
  2. [Introduction / Abstract] The abstract states the parameter dimension is independent of the preview horizon; confirm this is preserved after condensation in the explicit MPC setting.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments and the recommendation for major revision. We address each point below, clarifying the scope of our contribution and outlining targeted revisions to the manuscript.

read point-by-point responses

  1. Referee: [Method / Derivation of the condensation map] The central performance claim rests on the fixed linear condensation map producing a setpoint whose closed-loop effect nearly equals full preview. The manuscript should supply explicit conditions or error bounds showing when this holds (e.g., invariance to reference class or plant perturbations), as the least-squares projection is derived offline from an assumed trajectory ensemble.

    Authors: The condensation map is constructed offline by solving a least-squares projection of full preview trajectories onto a single setpoint, using an ensemble of representative references; this yields a fixed linear map whose performance is therefore tied to the statistical properties of that ensemble. We do not claim invariance to arbitrary reference classes or plant perturbations. In the revision we will add a dedicated paragraph in the derivation section that explicitly states the ensemble assumptions, shows the projection residual for the unconstrained LQ case, and provides heuristic guidance on when the approximation is expected to remain accurate (i.e., for references whose frequency content lies within the span of the training data). Deriving rigorous, a-priori error bounds that hold uniformly for all possible references and under model mismatch would require a separate worst-case analysis that lies outside the present contribution, which centers on practical dimension reduction for preview MPC. revision: partial

  2. Referee: [Numerical Experiments] Numerical experiments (double integrator and aircraft) demonstrate close match under specific conditions, but no sensitivity analysis is provided for references containing steps, ramps, or frequencies outside the implicit span of the projection basis, nor for unmodeled dynamics. These tests are load-bearing for the claim that the map remains effective in receding-horizon use.

    Authors: We agree that the current numerical section would be strengthened by broader sensitivity tests. In the revised manuscript we will augment the experiments with additional closed-loop simulations on both the double integrator and the aircraft model, using step references, linear ramps, and sinusoidal trajectories whose frequencies lie outside the original training ensemble. We will also report results under modest plant-parameter perturbations to illustrate behavior under unmodeled dynamics. Quantitative tracking-error and control-effort metrics comparing weighted condensation against full-preview MPC will be included, thereby providing direct evidence for the receding-horizon applicability of the method. revision: yes

standing simulated objections not resolved
  • Deriving explicit general error bounds or invariance conditions that hold for arbitrary reference classes and plant perturbations

Circularity Check

0 steps flagged

No significant circularity; linear map derived from independent least-squares projection and validated by separate numerical experiments.

full rationale

The paper defines the condensation map explicitly as the solution to a least-squares projection problem for the unconstrained tracking case (and a weighted variant for receding-horizon MPC), which is a standard, self-contained optimization step whose output is not presupposed by the performance claim. The assertion that the condensed setpoint 'closely matches full preview' is supported only by numerical experiments on a double integrator and aircraft model, not by any algebraic identity or self-referential definition that would force equivalence. No self-citations, uniqueness theorems, or ansatzes imported from prior author work are invoked to justify the central result. The derivation chain therefore remains independent of its empirical validation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The claim rests on the assumption that the system is linear and that an offline least-squares map computed on a nominal reference remains useful online. No new physical entities are introduced. One tunable weight appears in the weighted variant.

free parameters (1)
  • weighting factor for first input
    The weighted condensation variant introduces a scalar that prioritizes the immediate control move; its value is chosen to balance tracking and stability.
axioms (1)
  • domain assumption The plant is linear time-invariant
    The tracking problem and least-squares projection are formulated for linear systems; the aircraft example is treated as linear.

pith-pipeline@v0.9.0 · 5400 in / 1223 out tokens · 42451 ms · 2026-05-10T06:09:00.425992+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Optimal preview control: A review,

    N. Birla and A. Swarup, “Optimal preview control: A review,”Optimal Control Applications and Methods, vol. 36, no. 2, pp. 241–268, 2015

    work page 2015

  2. [2]

    Optimal continuous finite preview problem,

    M. Tomizuka, “Optimal continuous finite preview problem,”IEEE Transactions on Automatic Control, vol. 20, no. 3, pp. 362–365, 1975

    work page 1975

  3. [3]

    J. B. Rawlings, D. Q. Mayne, and M. Diehl,Model Predictive Control: Theory, Computation, and Design, 2nd ed. Nob Hill Publishing, 2017

    work page 2017

  4. [4]

    Borrelli, A

    F. Borrelli, A. Bemporad, and M. Morari,Predictive Control for Linear and Hybrid Systems. Cambridge University Press, 2017

    work page 2017

  5. [5]

    The explicit linear quadratic regulator for constrained systems,

    A. Bemporad, M. Morari, V . Dua, and E. N. Pistikopoulos, “The explicit linear quadratic regulator for constrained systems,”Automatica, vol. 38, no. 1, pp. 3–20, 2002

    work page 2002

  6. [6]

    A survey on explicit model predictive control,

    A. Alessio and A. Bemporad, “A survey on explicit model predictive control,” inNonlinear Model Predictive Control. Springer, 2009, pp. 345–369

    work page 2009

  7. [7]

    Reference and com- mand governors for systems with constraints: A survey on theory and applications,

    E. Garone, S. Di Cairano, and I. Kolmanovsky, “Reference and com- mand governors for systems with constraints: A survey on theory and applications,”Automatica, vol. 75, pp. 306–328, 2017

    work page 2017

  8. [8]

    MPC for tracking piecewise constant references for constrained linear systems,

    D. Limon, I. Alvarado, T. Alamo, and E. F. Camacho, “MPC for tracking piecewise constant references for constrained linear systems,” Automatica, vol. 44, no. 9, pp. 2382–2387, 2008

    work page 2008

  9. [9]

    Model predictive control for tracking using artificial references: Fundamentals, recent results and practical imple- mentation,

    P. Krupa, J. K ¨ohler, A. Ferramosca, I. Alvarado, M. N. Zeilinger, T. Alamo, and D. Limon, “Model predictive control for tracking using artificial references: Fundamentals, recent results and practical imple- mentation,” in2024 IEEE 63rd Conference on Decision and Control (CDC). IEEE, 2024, pp. 2977–2991

    work page 2024

  10. [10]

    B. D. O. Anderson and J. B. Moore,Optimal Control: Linear Quadratic Methods. Courier Corporation, 2007

    work page 2007

  11. [11]

    G. H. Golub and C. F. Van Loan,Matrix Computations, 4th ed. Johns Hopkins University Press, 2013

    work page 2013

  12. [12]

    Evaluation of piecewise affine control via binary search tree,

    P. Tøndel, T. A. Johansen, and A. Bemporad, “Evaluation of piecewise affine control via binary search tree,”Automatica, vol. 39, no. 5, pp. 945–950, 2003

    work page 2003

  13. [13]

    A dual active-set solver for embedded quadratic programming using recursive LDL T updates,

    D. Arnstr ¨om, A. Bemporad, and D. Axehill, “A dual active-set solver for embedded quadratic programming using recursive LDL T updates,” IEEE Transactions on Automatic Control, vol. 67, no. 8, pp. 4362–4369, 2022

    work page 2022

  14. [14]

    Design of feedback control systems for stable plants with saturating actuators,

    P. Kapasouris, M. Athans, and G. Stein, “Design of feedback control systems for stable plants with saturating actuators,” inProceedings of the 27th IEEE Conference on Decision and Control, 1988, pp. 469–479 vol.1

    work page 1988

  15. [15]

    A high-performant multi-parametric quadratic programming solver,

    D. Arnstr ¨om and D. Axehill, “A high-performant multi-parametric quadratic programming solver,” in2024 IEEE 63rd Conference on Decision and Control (CDC). IEEE, 2024, pp. 303–308. APPENDIXA PROOF OFLEMMA1 We write the tracking problem (3) in batch form. Rolling out the dynamics (1) over the horizon gives   x1 x2 ... xN   | {z } ≜ˆx =   A A2 ...

    work page 2024