Pith. sign in

REVIEW 3 major objections 5 minor 21 references

On data usage and predictive behavior of data-driven predictive control with 1-norm regularization

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read With 1-norm regularization, data-driven predictive control uses only extreme points of the mirrored data hull and yields an odd piecewise-affine predictor.

desk verdict Theorem 2's proof is wrong but its conclusion is right; the paper's structural results on 1-norm DPC deserve a careful referee. read the letter →

arxiv 2505.22307 v2 pith:AKQTPTPR submitted 2025-05-28 eess.SY cs.SYmath.OC

Manuel Klädtke , Moritz Schulze Darup This is my paper
classification eess.SYcs.SYmath.OC
keywords data-drivenpredictivecontrol1-normregularizationatomicnormimplicitpredictorpiecewiseaffinecriticalregionsdatausageconvexhull
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 analyzes data-driven predictive control (DPC) when the regularization is the 1-norm of the combination weights. It establishes that some recorded trajectory samples are never used by the optimizer, so they can be removed offline without changing the optimal cost or the chosen inputs and outputs. It connects the per-trajectory regularization cost to the atomic norm of a preprocessed, mirrored data set, giving a geometric gauge interpretation. It shows the implicit predictor is a symmetric piecewise-affine function whose critical regions, which dictate local data usage, scale linearly with the regularization weight. Together these results let a designer prune data sets and check whether the scheme's induced structure matches the true system.

What carries the argument

The load-bearing object is the mirrored data matrix $D_{\pm}=[D,\,-D]$ and the convex hull of its columns, $\operatorname{conv}(D_{\pm})$. The 1-norm over combination weights becomes a conical combination with nonnegative coefficients over $D_{\pm}$, and the extreme points of $\operatorname{conv}(D_{\pm})$ form the reduced set $\bar{D}_{\pm}$ (with unmirrored counterpart $\bar{D}$). This reduction preserves the span and conical hull (Lemma 1) and underlies Theorem 2's data-removal result. The same set induces the atomic norm via its gauge, providing the connection in Theorem 3. The implicit predictor's critical regions are then polyhedral cells of the parameter space labeled by which data atoms are active, and their scaling law comes from a nullspace reparametrization of the multiparametric quadratic program.

What would settle it

For a deliberately rank-deficient data set (fewer than $(m+p)(N_p+N)$ independent rows), solve the trajectory-specific program (5) for several $w$ lying in $\operatorname{span}(D)$ and compare the result to the atomic norm $\lambda\|w\|_{D_{\pm}}$; if the two disagree or the infimum is not attained, Theorem 3's equivalence fails exactly when Assumption 1 is violated. Alternatively, for any full-rank $D$ and two values $\lambda$ and $\eta\lambda$, solve the mpQP (11) with a multiparametric solver and check whether the critical regions satisfy $C_{\eta,i}=\eta C_i$; a violation would refute Theorem 4.

Watch

Extended reading notes

Core claim

The central claim is that the 1-norm regularized DPC problem has a hidden geometric structure that can be exploited and audited. With $h(a)=\lambda\|a\|_1$, splitting the combination weights into positive and negative parts turns the trajectory synthesis into a conical combination over the mirrored data set $D_{\pm}$. Only the extreme points of $\operatorname{conv}(D_{\pm})$ matter: Theorem 2 shows that removing all other data columns leaves the optimal cost and the optimal $(u^*, y^*)$ unchanged. Under a full row-rank assumption, Theorem 3 identifies the trajectory-specific effect $h^*(w)$ with the atomic norm $\lambda\|w\|_{D_{\pm}}$, so the regularization cost is the gauge of the convex hull of the reduced mirrored data. For a quadratic output objective, the implicit predictor is a piecewise-affine function of the initial condition and input; Theorem 4 proves its critical regions scale as $C_{\eta,i}=\eta C_i$ under weight scaling $\lambda\mapsto\eta\lambda$, and Theorem 5 proves the predictor is odd whenever the output cost and regularization are even. The upshot is a concrete picture of which data is used where in the state-input space, and how that usage changes as $\lambda$ is tuned.

Load-bearing premise

Assumption 1, that the data matrix $D$ has full row rank, is the load-bearing premise: without it the trajectory-specific effect may not be attained as a finite atomic norm, so the geometric interpretation and the reformulation of DPC as atomic-norm regularization can fail.

Editorial extensions

If this is right

  • Offline preprocessing can remove all data points outside the extreme points of $\operatorname{conv}(D_{\pm})$, reducing online computation without changing the control outcome.
  • The 1-norm penalty's trajectory-specific price is exactly the atomic norm on the reduced mirrored set, so regularization cost can be visualized as the gauge of $\operatorname{conv}(D_{\pm})$.
  • The implicit predictor's critical regions scale linearly with $\lambda$, meaning tuning $\lambda$ rescales the regions of local data usage rather than arbitrarily rearranging them.
  • The implicit predictor is odd if the objective and regularization are even, implying $\hat{y}_{\mathrm{DPC}}(0,0)=0$ and a systematic mismatch near the origin for systems without that symmetry.
  • The choice of data coordinates and exploration strategy changes $\operatorname{conv}(D_{\pm})$ and thus which samples survive pruning.

Reading between the lines

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

  • Because the removal criterion is purely geometric (extreme points of $\operatorname{conv}(D_{\pm})$), one could precompute the survival set for a given data set and use it as a diagnostic for whether the exploration strategy is well-matched to the operating regions of interest.
  • The symmetry result suggests a concrete design rule: for systems with a nonzero equilibrium or asymmetric dynamics, a designer could break the evenness artificially (e.g., by adding a linear term or using an affine data representation) to avoid systematic error near the origin.
  • The atomic-norm equivalence extends to other gauge-like regularizations only when the corresponding atom set remains centrally symmetric; testing this for quadratic regularization would delimit how special the 1-norm case is.
  • For noisy data, the full row-rank assumption may fail; quantifying how the atomic-norm equality degrades with rank deficiency would tell practitioners when the result is approximately usable.
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 analyzes data-driven predictive control (DPC) with 1-norm regularization, focusing on which trajectory data are used in predictions. It claims that certain data points can be removed offline without changing the optimal cost or optimizers, that the trajectory-specific cost of 1-norm regularization equals a scaled atomic norm over a reduced, mirrored data set, and that the resulting implicit predictor is a symmetric piecewise affine function whose critical regions scale linearly with the regularization weight. The analysis builds on the authors' earlier concepts of trajectory-specific effect and implicit predictors, and includes low-dimensional numerical examples and a link to publicly available code.

Significance. If the results are correct, they offer a concrete way to reduce data complexity in 1-norm-regularized DPC and provide structural insight into how regularization interacts with data usage and prediction behavior. The connection to the atomic norm is elegant and potentially useful for design and analysis. The paper uses established tools, is clearly written, and ships reproducible code. However, a central proof (Theorem 2) is invalid as written, which directly affects the advertised claim that some data is 'never utilized' and also undermines the derivation of the atomic-norm identity in Theorem 3. The core ideas are plausible and likely repairable, but the current manuscript does not rigorously establish its main claims.

major comments (3)
  1. [III.B, Theorem 2] The proof of Theorem 2 asserts that for every w± in D±\D± there exists a conical combination of extreme points with coefficient sum strictly less than 1. This is false for boundary points of conv(D±): for example, in conv({±e1, ±e2}) the point (0.5, 0.5) is a non-extreme boundary point with gauge 1, so every nonnegative representation by extreme points has coefficient sum at least 1, and equality is attainable. Substituting such a representation into an optimizer with a*±,j > 0 yields a feasible point with equal, not strictly smaller, regularization cost. Therefore the indirect proof does not force a*±,j = 0. The theorem's conclusion that replacing D with D does not alter optimal cost and (u*, y*) may still be true via an equal-cost argument, but the proof as written is invalid. Moreover, the strong interpretation that such data are 'never used' is not justified, since optimal solutions may exist that use boundary non-extreme points.
  2. [III.C, Theorem 3] The proof of Theorem 3 invokes Theorem 2 to replace D± by D±. Because the proof of Theorem 2 is flawed, the derivation of the atomic-norm identity is not valid as written. An equal-cost argument might preserve the identity, but the current manuscript does not supply it. In addition, Assumption 1 (full row-rank of D) is load-bearing here: without it, some trajectories w are not synthesizable, the gauge in (7b) can be infinite, and the claimed equality h*(w) = λ||w||_{D±} fails for those w. This assumption is stated as 'realistic' but is not guaranteed, so its role should be made explicit in the statement and proof.
  3. [III.D, Theorem 4] The proof of Theorem 4 contains notational and algebraic issues. The expression 'V^T H V^T' appears where 'V^T H V' is intended. More importantly, when eliminating the equality constraint, the cross term in the quadratic objective includes a factor 2 that is omitted in the displayed equations (12) and (13); the step 'α*(z) = β*(z̃)' is not justified by the equations as written. The scaling property C_{η,i} = η C_i is plausible and may be correct, but the proof needs to be rewritten with correct coefficients, or it must show that the omitted factor does not affect the comparison.
minor comments (5)
  1. [III.D, Theorem 4] The definitions of the matrices G and E in the proof of Theorem 4 are ambiguous; the block structure should be spelled out explicitly, for example by writing G and E as block matrices with clear dimensions.
  2. [II.B, Assumption 1] The paper states that full row-rank of D is 'realistic' for noisy or nonlinear systems, but it would be helpful to discuss what happens when this assumption fails, since Theorem 3 and the atomic-norm interpretation rely on it.
  3. [I and IV] The phrases 'some data is never utilized' (Introduction) and 'certain trajectory samples are never used for predictions' (Conclusions) are stronger than what is established. In light of the flaw in Theorem 2's proof, these statements should be softened to indicate that such data are not needed for optimal cost, or that there exists an optimal solution not using them.
  4. [III.D, Figure 2] The caption of Figure 2 refers to 'bottom left and top right' in a way that is difficult to follow; the layout of the figure should be described more clearly.
  5. [III.B, Lemma 1] The proof of Lemma 1 is terse: the statement 'D±\D± ⊆ coni(D±) as per the construction procedure' would benefit from a one-sentence justification that D±\D± ⊆ conv(D±) ⊆ coni(D±).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: new results follow from the stated OCP, convex geometry, and mpQP theory; the proof gap in Theorem 2 is a correctness issue, not a circular one.

full rationale

The paper does not fit parameters and does not present any quantity as a prediction that was used to construct it. Theorems 2, 4, and 5 are derived from the OCP (4)/(8), convex hull geometry, and standard multiparametric programming arguments; none of the conclusions is assumed in its premises. Definitions 1 and 2 are attributed to the authors' prior work [10,11], but they are restated as definitions in Section II.C and are not used as unverified external facts; no theorem's conclusion reduces to them by construction. Theorem 3 equates h*(w) with the atomic norm ||w||_{D±}; this is a direct reformulation of the l1 trajectory-specific effect as the gauge of conv(D±), and the paper transparently presents it as a connection to the independent atomic-norm framework [12], not as an empirical derivation. A proof gap does exist in Theorem 2 (Section III.B): the claim that every removed point has a conical combination of extreme points with coefficient sum <1 fails for non-extreme boundary points, whose gauge is 1; an equal-cost replacement argument is needed. This is a correctness/rigor concern, not circularity. Theorems 4 and 5 are algebraic and symmetry-based, with no dependence on fitted values or on the data itself beyond the fixed data matrix. Accordingly, no circular step meets the quoted-evidence bar.

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

The paper introduces no free parameters beyond the design weight lambda (not fitted) and no invented entities. All axioms are either standard DPC setup assumptions, the explicit full row-rank assumption, or specific restrictions on the objective stated in the theorems. The most load-bearing is Assumption 1, since it underpins the atomic norm result.

assumptions (5)
  • domain assumption Assumption 1: The data matrix D has full row rank.
    Stated in Section II.B. Needed for Theorem 3 so that the trajectory-specific effect is finite on the whole space and the atomic norm is a norm. Without it the atomic norm equivalence breaks.
  • domain assumption Data generated by an LTI system and the persistency of excitation condition (2) hold for the ideal setting.
    Section II.A, standard DPC setup. The image representation (3) and the discussion of exact predictions rely on this. The paper treats noisy/nonlinear data as deviations from this baseline.
  • domain assumption For Theorem 4: the control objective is quadratic output tracking with zero reference and no output constraints (Y = R^{pN}).
    Section III.D, equation (10) and following text. The mpQP structure and the scaling property are derived for this specific objective; the paper notes extensions require future work.
  • domain assumption For Theorem 5: Jy(z,y) and h(a) are even functions, and the parametric optimizer is unique for all feasible z.
    Section III.D, stated in the theorem. The oddness conclusion depends on these symmetry and uniqueness conditions.
  • standard math The atomic norm preconditions: the atom set is centrally symmetric and the origin is the centroid of its convex hull.
    Section II.D. These are standard conditions from [12]; the paper notes they are satisfied by the mirrored data set D± by construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On data usage and predictive behavior of data-driven predictive control with 1-norm regularization." pith.science (2026). https://pith.science/paper/AKQTPTPR

@misc{pith2026250522307,
  author       = {Pith},
  title        = {Pith review of: On data usage and predictive behavior of data-driven predictive control with 1-norm regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKQTPTPR}},
  note         = {Machine review of arXiv:2505.22307}
}
read the original abstract

We investigate the data usage and predictive behavior of data-driven predictive control (DPC) with 1-norm regularization. Our analysis enables the offline removal of unused data and facilitates a comparison between the identified symmetric structure and data usage against prior knowledge of the true system. This comparison helps assess the suitability of the DPC scheme for effective control.

Figures

Figures reproduced from arXiv: 2505.22307 by the authors.

Figure 1
Figure 1. Visualization of data sets with elements in D drawn from N (0, I). (a) Original sets D in green (×), and D± in green and orange (×, ◦). The convex hull conv (D±) is also highlighted. (b) Reduced sets D in green (×), and D± in green and orange (×, ◦). Theorem 2. Consider the DPC problem (4) with 1-norm regu￾larization h(a) = λ∥a∥1. Replacing the data matrix D with a data matrix D whose columns are the elements of D d… view at source ↗
Figure 2
Figure 2. Importantly, since every CR is associated with a subset of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of true dynamics and implicit predictor. (a) System dynamics f(x, u) = 2x 2 + 2u 2 − 1 (green) with 20 data samples D generated from x (i) 0 , u (i) ∼ U[−1,1] (blue). (b) Implicit predictor ˆxDPC(x0, u) (orange) for λ = 100, Q = 1 resulting from the 8 relevant data samples D (blue). IV. CONCLUSIONS AND OUTLOOK We analyzed the data usage and predictive behavior of DPC with 1-norm regularization using the c… view at source ↗
Figures from the paper (1 more)
Figure 2
Figure 2. Figure 2: CRs of the implicit predictor ˆxDPC(x0, u) for Q = 1 and the same data D (blue) as in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Yang and S

    H. Yang and S. Li. A new method of direct data-driven predictive con- troller design. 2013 9th Asian Control Conference, pp. 1–6, 2013

    work page 2013

  2. [2]

    Coulson, J

    J. Coulson, J. Lygeros, and F. D ¨orfler. Data-enabled predictive control: In the shallows of the DeePC. 18th European Control Conference , pp. 307–312, 2019

    work page 2019

  3. [3]

    J. C. Willems, P. Rapisarda, I. Markovsky, and B. L.M. De Moor. A note on persistency of excitation. Systems & Control Letters, 54(4):325–329, 2005

    work page 2005

  4. [4]

    Berberich and F

    J. Berberich and F. Allg ¨ower. A trajectory-based framework for data- driven system analysis and control. 2020 European Control Conference, pp. 1365–1370, 2020. 2475-1456 ©2025 IEEE

    work page 2020

  5. [5]

    Coulson, J

    J. Coulson, J. Lygeros, and F. D ¨orfler. Regularized and distributionally robust data-enabled predictive control. 2019 IEEE 58th Conference on Decision and Control, pp. 2696–2701, 2019

    work page 2019

  6. [6]

    D ¨orfler, J

    F. D ¨orfler, J. Coulson, and I. Markovsky. Bridging direct and indirect data-driven control formulations via regularizations and relaxations.IEEE Transactions on Automatic Control, 68(2):883–897, 2023

    work page 2023

  7. [7]

    Berberich, J

    J. Berberich, J. K ¨ohler, M. M ¨uller, and F. Allg ¨ower. Data-driven model predictive control with stability and robustness guarantees. IEEE Trans- actions on Automatic Control, 66(4):1702–1717, 2021

    work page 2021

  8. [8]

    Markovsky

    I. Markovsky. Structured low-rank approximation and its applications. Automatica, 44(4):891–909, 2008

    work page 2008

Show all 21 references
  1. [9]

    Mattsson and T

    P. Mattsson and T. B. Sch ¨on. On the regularization in DeePC. IFAC- PapersOnLine, 56(2):625–631, 2023. 22nd IFAC World Congress

    2023

  2. [10]

    Kl ¨adtke and M

    M. Kl ¨adtke and M. Schulze Darup. Implicit predictors in regularized data- driven predictive control. IEEE Control Systems Letters , 7:2479–2484, 2023

    2023

  3. [11]

    Kl ¨adtke and M

    M. Kl ¨adtke and M. Schulze Darup. Towards explainable data-driven pre- dictive control with regularizations. at - Automatisierungstechnik, 2025. accepted January 17, 2025

    2025

  4. [12]

    Chandrasekaran, B

    V . Chandrasekaran, B. Recht, P. A. Parrilo, and A. S. Willsky. The con- vex geometry of linear inverse problems. Foundations of Computational Mathematics, 12:805–849, 2012

    2012

  5. [13]

    Markovsky and F

    I. Markovsky and F. D ¨orfler. Identifiability in the behavioral setting.IEEE Transactions on Automatic Control, 68(3):1667–1677, 2023

    2023

  6. [14]

    De Persis and P

    C. De Persis and P. Tesi. Formulas for data-driven control: Stabiliza- tion, optimality, and robustness.IEEE Transactions on Automatic Control, 65(3):909–924, 2020

    2020

  7. [15]

    Breschi, A

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

    2023

  8. [16]

    Mattsson, F

    P. Mattsson, F. Bonassi, V . Breschi, and T. B. Sch¨on. On the equivalence of direct and indirect data-driven predictive control approaches. IEEE Control Systems Letters, 8:796–801, 2024

    2024

  9. [17]

    Herceg, M

    M. Herceg, M. Kvasnica, C.N. Jones, and M. Morari. Multi-Parametric Toolbox 3.0. In Proc. of the European Control Conference, pp. 502–510, 2013

    2013

  10. [18]

    Boyd and L

    S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge Univer- sity Press, 2004

    2004

  11. [19]

    C. B. Barber, D P. Dobkin, and H. Huhdanpaa. The quickhull algorithm for convex hulls. ACM Trans. Math. Softw., 22(4):469–483, 1996

    1996

  12. [20]

    Borrelli, A

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

    2017

  13. [21]

    Padoan, F

    A. Padoan, F. D ¨orfler, and J. Lygeros. Data-driven representations of con- ical, convex, and affine behaviors. In 2023 62nd IEEE Conference on Decision and Control, pp. 596–601, 2023. 2475-1456 ©2025 IEEE

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.