pith. sign in

arxiv: 2604.05832 · v1 · submitted 2026-04-07 · 📡 eess.SY · cs.SY

Local Sensitivity Analysis for Kernel-Regularized ARX Predictors in Data-Driven Predictive Control

Aihui Liu , Magnus Jansson This is my paper

Pith reviewed 2026-05-10 18:22 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords local sensitivity analysiskernel regularizationARX predictorsdata-driven predictive controlJacobian linearizationprediction uncertaintyMPC
0
0 comments X

The pith

A first-order linearization of the implicit multi-step predictor yields a Jacobian that approximates control-relevant prediction uncertainty and shapes kernel regularization in ARX-based data-driven predictive control.

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

The paper derives a local first-order linearization of the implicit predictor map used in structured ARX-based data-driven predictive control. Although parameter estimation is linear, the lifted multi-step predictor depends on those parameters in a nonlinear way, which makes uncertainty propagation and task-aware regularization difficult. The resulting Jacobian supplies an approximate control-relevant prediction uncertainty term and a sensitivity metric that can be used to shape kernel regularization according to the control task. A reader would care because this approach targets robustness specifically in weak-excitation regimes where standard regularization already helps but further task-dependent refinement can add value.

Core claim

We derive a local first-order linearization of the implicit predictor map for structured ARX-based data-driven predictive control. The resulting Jacobian yields both an approximate control-relevant prediction uncertainty term and a task-dependent sensitivity metric for shaping kernel regularization. Numerical results show that the proposed analysis is most useful in weak-excitation regimes, where baseline regularization already provides substantial robustness gains and the proposed sensitivity shaping yields a further smaller improvement.

What carries the argument

The Jacobian of the first-order linearization of the implicit multi-step predictor map, which maps small ARX parameter perturbations to changes in control-relevant predictions for uncertainty estimation and regularization tuning.

Load-bearing premise

The first-order linearization remains accurate only for small changes in the estimated ARX parameters.

What would settle it

Perturb the estimated ARX parameters by a small amount in a simulation, recompute the actual multi-step predictions, and check whether the change matches the value predicted by the Jacobian within the expected error bound.

Figures

Figures reproduced from arXiv: 2604.05832 by Aihui Liu, Magnus Jansson.

Figure 1
Figure 1. Figure 1: Closed-loop cost J over 500 Monte Carlo runs in the informative￾data regime. Since the identification problem is already reasonably well￾conditioned, all ARX-based variants are competitive and only limited gains are obtained from additional kernel regularization. are A =  0.7326 −0.0861 0.1722 0.9909  , B =  0.0609 0.0064 , C = [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Mean and standard deviation of closed-loop trajectories in the weak-excitation regime, for the output (top row) and the input (bottom row). The [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Closed-loop cost J in the weak-excitation regime. Baseline SS regularization provides the main stabilization relative to OLS, while the proposed sensitivity shaping yields an additional improvement over SS. TABLE I AVERAGE POSTERIOR COVARIANCE tr(Σθ ) AFTER IDENTIFICATION OLS SS SS+W 0.0470± 0.0028 0.0113 ± 0.0044 0.0019 ± 0.0012 covariance tr(Σθ ) after identification in Table I. This reduc￾tion in the pa… view at source ↗
Figure 4
Figure 4. Figure 4: Heatmaps of (a) the empirical-Bayes tuned baseline SS kernel [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

We study local sensitivity of structured ARX-based data-driven predictive control. Although predictor estimation is linear in the ARX parameters, the lifted multi-step predictor used in MPC depends on them implicitly, which complicates both uncertainty propagation and task-aware regularization. We derive a local first-order linearization of this implicit predictor map. The resulting Jacobian yields both an approximate control-relevant prediction uncertainty term and a task-dependent sensitivity metric for shaping kernel regularization. Numerical results show that the proposed analysis is most useful in weak-excitation regimes, where baseline SS regularization already provides substantial robustness gains and the proposed sensitivity shaping yields a further smaller improvement.

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

1 major / 2 minor

Summary. The paper derives a local first-order linearization (via the Jacobian) of the implicit multi-step ARX predictor map arising in kernel-regularized data-driven predictive control. This Jacobian is used to obtain both an approximate control-relevant prediction uncertainty term and a task-dependent sensitivity metric that shapes the kernel regularization. Numerical experiments indicate the approach is most useful in weak-excitation regimes, where it yields modest additional robustness gains beyond standard sum-of-squares regularization.

Significance. If the first-order approximation is sufficiently accurate for the parameter perturbations encountered in practice, the work supplies a concrete, control-oriented way to propagate uncertainty through the lifted predictor and to perform task-aware regularization. This could be valuable in data-driven MPC settings with limited excitation, where standard regularization already helps but further tailoring may improve closed-loop performance. The derivation itself follows standard implicit differentiation and does not introduce new free parameters.

major comments (1)
  1. [§3 (derivation) and §4 (numerics)] The central claim that the Jacobian supplies a usable control-relevant uncertainty term and sensitivity metric rests on the first-order Taylor remainder being negligible. The manuscript provides neither analytic bounds on the remainder nor systematic numerical verification (finite-difference checks, Monte Carlo sampling of the nonlinear map, or comparison to higher-order terms) precisely in the weak-excitation regimes highlighted in the abstract and numerical results. This verification is load-bearing for the practical utility asserted in §4.
minor comments (2)
  1. [Abstract] The abstract states that the sensitivity shaping yields a 'further smaller improvement' but does not quantify the effect size relative to baseline SS regularization or report confidence intervals on the closed-loop metrics.
  2. [§2] Notation for the lifted predictor map and the implicit dependence on ARX parameters could be introduced earlier and used consistently when defining the Jacobian.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the contribution. We agree that additional verification of the first-order approximation is needed to support the claims in weak-excitation regimes and will incorporate it in the revision.

read point-by-point responses

  1. Referee: [§3 (derivation) and §4 (numerics)] The central claim that the Jacobian supplies a usable control-relevant uncertainty term and sensitivity metric rests on the first-order Taylor remainder being negligible. The manuscript provides neither analytic bounds on the remainder nor systematic numerical verification (finite-difference checks, Monte Carlo sampling of the nonlinear map, or comparison to higher-order terms) precisely in the weak-excitation regimes highlighted in the abstract and numerical results. This verification is load-bearing for the practical utility asserted in §4.

    Authors: We acknowledge that the manuscript lacks both analytic bounds on the Taylor remainder and dedicated numerical verification of the first-order approximation accuracy in the weak-excitation regimes. This is a valid observation. In the revised manuscript we will add, in §4, finite-difference checks comparing the Jacobian linearization to the true nonlinear predictor map, along with Monte Carlo sampling of parameter perturbations drawn from the low-excitation data regimes used in the numerical examples. These checks will quantify the remainder size and directly support the asserted practical utility. No changes are required to the derivation in §3. revision: yes

Circularity Check

0 steps flagged

No significant circularity; Jacobian derived via standard differentiation of implicit map

full rationale

The central derivation applies first-order Taylor linearization and Jacobian computation to the lifted multi-step ARX predictor, which is an implicit function of the ARX parameters. This is a direct application of implicit differentiation to the predictor equations and does not reduce any claimed prediction or sensitivity metric to a fitted quantity by construction. Minor self-citations appear for background on kernel regularization and data-driven MPC but are not load-bearing for the Jacobian step itself. The result remains independent of the target uncertainty or sensitivity outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract; the work relies on standard linearization and kernel methods from control theory.

pith-pipeline@v0.9.0 · 5395 in / 1024 out tokens · 25858 ms · 2026-05-10T18:22:04.127000+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

12 extracted references · 12 canonical work pages

  1. [1]

    Handbook of linear data- driven predictive control: Theory, implementation and design,

    P. Verheijen, V . Breschi, and M. Lazar, “Handbook of linear data- driven predictive control: Theory, implementation and design,”Annual Reviews in Control, vol. 56, p. 100914, 2023

    work page 2023

  2. [2]

    Bayesian kernel-based linear control design,

    A. Scampicchio, A. Chiuso, S. Formentin, and G. Pillonetto, “Bayesian kernel-based linear control design,” in2019 IEEE 58th Conference on Decision and Control (CDC). IEEE, Dec. 2019, p. 822–827

    work page 2019

  3. [3]

    Control-oriented regularization for linear system identification,

    S. Formentin and A. Chiuso, “Control-oriented regularization for linear system identification,”Automatica, vol. 127, p. 109539, May 2021

    work page 2021

  4. [4]

    A bias-variance perspective of data-driven control,

    K. Colin, Y . Ju, X. Bombois, C. R. Rojas, and H. Hjalmarsson, “A bias-variance perspective of data-driven control,”IFAC-PapersOnLine, vol. 58, no. 15, p. 85–90, 2024

    work page 2024

  5. [5]

    Harnessing uncertainty for a separation principle in direct data-driven predictive control,

    A. Chiuso, M. Fabris, V . Breschi, and S. Formentin, “Harnessing uncertainty for a separation principle in direct data-driven predictive control,”Automatica, vol. 173, p. 112070, Mar. 2025

    work page 2025

  6. [6]

    A new kernel-based approach for linear system identification,

    G. Pillonetto and G. De Nicolao, “A new kernel-based approach for linear system identification,”Automatica, vol. 46, no. 1, p. 81–93, Jan. 2010

    work page 2010

  7. [7]

    Ker- nel methods in system identification, machine learning and function estimation: A survey,

    G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and L. Ljung, “Ker- nel methods in system identification, machine learning and function estimation: A survey,”Automatica, vol. 50, no. 3, p. 657–682, Mar. 2014

    work page 2014

  8. [8]

    Data-driven output prediction and control of stochastic systems: An innovation-based approach,

    Y . Wang, K. You, D. Huang, and C. Shang, “Data-driven output prediction and control of stochastic systems: An innovation-based approach,”Automatica, vol. 171, p. 111897, Jan. 2025

    work page 2025

  9. [9]

    Data-driven predictive control in a stochastic setting: a unified framework,

    V . Breschi, A. Chiuso, and S. Formentin, “Data-driven predictive control in a stochastic setting: a unified framework,”Automatica, vol. 152, p. 110961, Jun. 2023

    work page 2023

  10. [10]

    SPC: Subspace predictive control,

    W. Favoreel, B. D. Moor, and M. Gevers, “SPC: Subspace predictive control,”IFAC Proceedings Volumes, vol. 32, no. 2, p. 4004–4009, Jul. 1999

    work page 1999

  11. [11]

    Data-driven predictive control using closed-loop data: An instrumental variable approach,

    Y . Wang, Y . Qiu, M. Sader, D. Huang, and C. Shang, “Data-driven predictive control using closed-loop data: An instrumental variable approach,”IEEE Control Systems Letters, vol. 7, p. 3639–3644, 2023

    work page 2023

  12. [12]

    Closed-loop consistent, causal data-driven predictive control via ssarx,

    A. Liu and M. Jansson, “Closed-loop consistent, causal data-driven predictive control via ssarx,” arXiv preprint arXiv:2512.14510, 2025

    work page arXiv 2025