pith. sign in

arxiv: 2605.17966 · v1 · pith:SV7WCECPnew · submitted 2026-05-18 · 🧮 math.OC

Control-Channel Informativity for Koopman EDMDc under Behavior-Policy Data

Yue Wu This is my paper

Pith reviewed 2026-05-20 09:41 UTC · model grok-4.3

classification 🧮 math.OC
keywords EDMDcKoopmancontrol-channel identifiabilitySchur complementbehavior policyintervention certificatefinite-sampleresidual covariance
0
0 comments X

The pith

The strict positivity of residual input covariance after state projection is necessary and sufficient for finite-sample identifiability of the lifted control-channel block in EDMDc.

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

When training extended dynamic mode decomposition with control on trajectories from a behavior policy, the data may match observed transitions accurately without determining how new inputs would affect the lifted state. The paper introduces a conditional intervention certificate equal to the residual input covariance after projecting out the lifted-state features, which is the Schur complement of the state block in the information matrix. It proves that strict positivity of this certificate is necessary and sufficient for the finite-sample identifiability of the control-channel block. If the certificate is zero, different models can fit the same data but disagree on counterfactual inputs. A bound and examples show quadratic growth of information with dither amplitude and complementary roles for coverage and excitation.

Core claim

The central claim is that the conditional intervention certificate, defined as the residual covariance of inputs after orthogonal projection away from the span of active lifted-state features and realized as the Schur complement of the lifted-state block in the EDMDc information matrix, must be strictly positive to guarantee that the lifted control-channel block is identifiable from finite samples. When this certificate vanishes, there exist distinct lifted models that agree on every collected transition yet produce different predictions under counterfactual inputs. The result is supported by a closed-loop statistical bound that uses predictable regressors and conditionally sub-Gaussian nose

What carries the argument

The conditional intervention certificate, which measures residual input covariance after projection onto the orthogonal complement of the active lifted-state feature span.

If this is right

  • Strict positivity of the certificate guarantees unique recovery of the lifted control-channel block from the given finite samples.
  • Vanishing certificate allows multiple models to agree on collected transitions while differing on counterfactual inputs.
  • Under scalar dithered feedback the residual intervention information grows quadratically with dither amplitude.
  • Control-channel estimation error scales inversely with the intervention signal-to-noise ratio.
  • State coverage, joint-regression conditioning, and intervention excitation function as complementary rather than interchangeable diagnostics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Behavior policies used for data collection could be deliberately augmented with small persistent excitation to keep the certificate positive and thereby support later control design.
  • The same residual-covariance diagnostic may be useful in other off-policy model-learning settings to decide when collected trajectories suffice for intervention effects.
  • Online monitoring of the certificate during data acquisition could trigger adaptive policy adjustments that restore identifiability without full re-collection.
  • In high-dimensional or partially observed systems the certificate could serve as a practical stopping criterion for data gathering before attempting control synthesis.

Load-bearing premise

Transition noise is conditionally sub-Gaussian and regressors are predictable.

What would settle it

A finite data set generated by a behavior policy in which the computed Schur complement is zero or negative, yet every pair of lifted models that fit the observed transitions produce identical outputs under new input sequences, would disprove necessity and sufficiency.

Figures

Figures reproduced from arXiv: 2605.17966 by Yue Wu.

Figure 1
Figure 1. Figure 1: EDMDc can fit behavior-policy transitions while leaving the response to new commands unidentified. The proposed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Scalar closed-loop scaling experiment. Residual intervention information grows quadratically with dither amplitude, and [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Nonlinear feedback/dither experiments on Duffing and Van der Pol systems. Dither creates residual intervention excitation [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Broad EDMDc acquisition grid. Mean Cint, input predictability, and counterfactual error separate different data-quality regimes; no single certificate is a universal leaderboard. VIII. PRACTICAL GUIDANCE AND LIMITATIONS The recommended workflow is simple. Choose the dictionary, remove inactive coordinates, and standardize lifted features and inputs on the identification data. Compute Creg to check whether … view at source ↗
read the original abstract

Extended dynamic mode decomposition with control (EDMDc) is often trained from trajectories generated by a behavior policy or a pre-existing feedback controller. Such data can predict the observed behavior accurately while failing to identify how new input commands change the lifted state. This paper studies that failure as a control-channel informativity problem. We introduce a conditional intervention certificate, defined as the residual input covariance after projecting the input data away from the active lifted-state feature span. The certificate is the Schur complement of the lifted-state block in the EDMDc information matrix. We prove that its strict positivity is necessary and sufficient for finite-sample sample- identifiability of the lifted control-channel block. If the certificate vanishes, distinct lifted models agree on every collected transition but disagree under counterfactual inputs. We then give a closed-loop statistical bound using predictable regressors, conditionally sub-Gaussian transition noise, and a regularized Schur complement. A scalar feedback example shows the unavoidable scaling: under dithered feedback, residual intervention information grows quadratically with the dither amplitude and the control-channel error decreases with the inverse intervention signal-to-noise scale. New experiments verify these scalings exactly in a linear system and diagnostically in controlled Duffing and Van der Pol benchmarks. A larger EDMDc acquisition grid further shows that state coverage, joint regression conditioning, and intervention excitation are complementary diagnostics rather than interchangeable performance score.

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 / 3 minor

Summary. The paper studies control-channel informativity in EDMDc when trajectories are generated under a behavior policy or existing feedback controller. It defines a conditional intervention certificate as the Schur complement of the lifted-state block within the EDMDc Gram matrix and proves that strict positivity of this certificate is necessary and sufficient for finite-sample identifiability of the lifted control-channel parameters. The work further derives a closed-loop statistical bound under predictable regressors and conditionally sub-Gaussian transition noise, illustrates unavoidable scaling with dither amplitude in a scalar feedback example, and validates the scalings on linear and nonlinear benchmark systems.

Significance. If the algebraic characterization and statistical bound hold, the paper supplies a concrete, computable diagnostic for whether behavior-policy data can identify control effects in Koopman models. This is valuable for data-driven control applications where open-loop excitation is unavailable. The direct link to the normal equations of joint least-squares and the explicit necessity construction via an alternative control matrix are clear strengths; the scaling result with dither amplitude offers practical guidance.

major comments (2)
  1. [Section 3] Section 3 (finite-sample identifiability): The necessity direction constructs an alternative control matrix that agrees on observed transitions when the certificate vanishes. Please confirm that this construction remains valid for arbitrary feature maps and does not implicitly require the lifted state to be a faithful representation of the original dynamics.
  2. [Section 4] Section 4 (closed-loop statistical bound): The bound is stated for the regularized Schur complement. Clarify whether the regularization parameter must be chosen independently of the data or can be adapted, and whether the resulting high-probability statement still implies practical identifiability when the unregularized certificate is only marginally positive.
minor comments (3)
  1. Notation: The term 'conditional intervention certificate' is introduced without an explicit symbol; introducing a compact notation (e.g., C_I) would improve readability when the quantity is referenced repeatedly in theorems and experiments.
  2. Experiments: The Duffing and Van der Pol results are described as 'diagnostic'; adding a quantitative table that reports the certificate value alongside the observed control-channel error for each acquisition grid would make the complementarity claim easier to verify.
  3. References: The connection to existing persistency-of-excitation conditions in adaptive control or subspace identification is mentioned only briefly; a short paragraph contrasting the Schur-complement certificate with classical rank conditions would help situate the contribution.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and the encouraging recommendation for minor revision. The comments help clarify important aspects of the identifiability results and statistical bounds. We address each major comment below.

read point-by-point responses

  1. Referee: [Section 3] Section 3 (finite-sample identifiability): The necessity direction constructs an alternative control matrix that agrees on observed transitions when the certificate vanishes. Please confirm that this construction remains valid for arbitrary feature maps and does not implicitly require the lifted state to be a faithful representation of the original dynamics.

    Authors: The necessity construction is algebraic and operates solely in the space of lifted features. Given a vanishing Schur complement, we explicitly construct an alternative control matrix B' such that the lifted model with B' produces identical one-step predictions on all observed (lifted-state, input) pairs, yet differs under new inputs. This argument relies only on the linear algebra of the Gram matrix and the definition of the Schur complement; it holds for any feature map and makes no reference to whether the lifted coordinates faithfully embed the original state space. The result concerns identifiability of the lifted control parameters, which is the relevant quantity for subsequent control design. A short clarifying paragraph has been added to Section 3. revision: yes

  2. Referee: [Section 4] Section 4 (closed-loop statistical bound): The bound is stated for the regularized Schur complement. Clarify whether the regularization parameter must be chosen independently of the data or can be adapted, and whether the resulting high-probability statement still implies practical identifiability when the unregularized certificate is only marginally positive.

    Authors: The regularization parameter is a fixed positive constant selected independently of the data, typically on the order of the noise variance divided by the sample size or a similar a priori quantity. Adaptation to the data is not required for the high-probability statement and would complicate the analysis. When the unregularized certificate is only marginally positive, the bound on the regularized quantity remains valid and implies that the control-channel estimation error is controlled by the inverse of (certificate - λ); the resulting guarantee is correspondingly weaker, which accurately reflects the limited practical identifiability in that regime. We have inserted additional explanatory text after the main theorem in Section 4 to make these distinctions explicit. revision: yes

Circularity Check

1 steps flagged

Central identifiability claim reduces to algebraic definition of the certificate

specific steps
  1. self definitional [Abstract]
    "We introduce a conditional intervention certificate, defined as the residual input covariance after projecting the input data away from the active lifted-state feature span. The certificate is the Schur complement of the lifted-state block in the EDMDc information matrix. We prove that its strict positivity is necessary and sufficient for finite-sample sample-identifiability of the lifted control-channel block."

    The certificate is defined to be the Schur complement, which by construction is the precise algebraic condition (full column rank of the residual regressors) that guarantees a unique least-squares solution for the control coefficients in the joint normal equations. The claimed necessity-and-sufficiency proof therefore reduces directly to this definition, with no additional derivation or external content.

full rationale

The paper defines the conditional intervention certificate explicitly as the Schur complement of the lifted-state block in the EDMDc Gram matrix (i.e., residual input covariance after projection onto the state features). It then claims to prove that strict positivity of this quantity is necessary and sufficient for finite-sample identifiability of the control-channel block. This equivalence is exactly the condition for unique solvability of the normal equations in the joint least-squares problem, which holds by standard block-matrix linear algebra once the certificate is so defined. The necessity direction is shown by constructing an alternative control matrix that matches all observed transitions precisely when the residual covariance is singular. No external assumptions, self-citations, or fitted parameters are required for the equivalence; the result is therefore self-definitional rather than independently derived.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The work relies on standard statistical assumptions for dynamical systems and introduces one new mathematical object whose properties are proved from the data matrix.

axioms (2)
  • domain assumption Transition noise is conditionally sub-Gaussian
    Invoked to obtain the closed-loop statistical bound on the control-channel error.
  • domain assumption Regressors are predictable
    Required for the martingale concentration used in the finite-sample bound.
invented entities (1)
  • Conditional intervention certificate no independent evidence
    purpose: Quantifies residual input covariance after projection onto the lifted-state feature span
    Newly defined as the Schur complement of the lifted-state block in the EDMDc information matrix.

pith-pipeline@v0.9.0 · 5773 in / 1307 out tokens · 39438 ms · 2026-05-20T09:41:06.077334+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We introduce a conditional intervention certificate, defined as the residual input covariance after projecting the input data away from the active lifted-state feature span. The certificate is the Schur complement of the lifted-state block in the EDMDc information matrix. We prove that its strict positivity is necessary and sufficient for finite-sample sample-identifiability of the lifted control-channel block.

  • IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    If βN = 0, then there exist a ≠ 0 and h such that a⊤U = h⊤Z. For any c, define Ac = A⋆ − c h⊤, Bc = B⋆ + c a⊤. Then AcZ + BcU = A⋆Z + B⋆U on the collected data.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    A data-driven approximation of the koopman operator: Extending dynamic mode decomposition,

    M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, “A data-driven approximation of the koopman operator: Extending dynamic mode decomposition,”Journal of Nonlinear Science, vol. 25, pp. 1307–1346, 2015

    work page 2015

  2. [2]

    Dynamic mode decomposition with control,

    J. L. Proctor, S. L. Brunton, and J. N. Kutz, “Dynamic mode decomposition with control,”SIAM Journal on Applied Dynamical Systems, vol. 15, no. 1, pp. 142–161, 2016

    work page 2016

  3. [3]

    Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control,

    M. Korda and I. Mezi ´c, “Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control,”Automatica, vol. 93, pp. 149–160, 2018

    work page 2018

  4. [4]

    Modern koopman theory for dynamical systems,

    S. L. Brunton, M. Budiši ´c, E. Kaiser, and J. N. Kutz, “Modern koopman theory for dynamical systems,”SIAM Review, vol. 64, no. 2, pp. 229–340, 2022. 13

    work page 2022

  5. [5]

    Active learning of dynamics for data-driven control using koopman operators,

    I. Abraham and T. D. Murphey, “Active learning of dynamics for data-driven control using koopman operators,”IEEE Transactions on Robotics, vol. 35, no. 5, pp. 1071–1083, 2019

    work page 2019

  6. [6]

    Koopman operators for generalized persistence of excitation conditions for nonlinear systems,

    N. Boddupalli, A. Hasnain, S. P. Nandanoori, and E. Yeung, “Koopman operators for generalized persistence of excitation conditions for nonlinear systems,” inProceedings of the IEEE 58th Conference on Decision and Control (CDC). IEEE, 2019, pp. 8106–8111

    work page 2019

  7. [7]

    Willems’ fundamental lemma for nonlinear systems with koopman linear embedding,

    X. Shang, J. Cortés, and Y . Zheng, “Willems’ fundamental lemma for nonlinear systems with koopman linear embedding,”IEEE Control Systems Letters, 2024

    work page 2024

  8. [8]

    Data-driven mpc with stability guarantees using extended dynamic mode decomposition,

    L. Bold, L. Grüne, M. Schaller, and K. Worthmann, “Data-driven mpc with stability guarantees using extended dynamic mode decomposition,”IEEE Transactions on Automatic Control, vol. 70, no. 1, pp. 534–541, 2025

    work page 2025

  9. [9]

    Koopman-based feedback design with stability guarantees,

    R. Strässer, M. Schaller, K. Worthmann, J. Berberich, and F. Allgöwer, “Koopman-based feedback design with stability guarantees,”IEEE Transactions on Automatic Control, vol. 70, no. 1, pp. 355–370, 2025

    work page 2025

  10. [10]

    Error analysis of kernel edmd for prediction and control in the koopman framework,

    F. M. Philipp, M. Schaller, K. Worthmann, S. Peitz, and F. Nüske, “Error analysis of kernel edmd for prediction and control in the koopman framework,”Journal of Nonlinear Science, vol. 35, p. 92, 2025

    work page 2025

  11. [11]

    Kernel-based koopman approximants for control: Flexible sampling, error analysis, and stability,

    L. Bold, F. M. Philipp, M. Schaller, and K. Worthmann, “Kernel-based koopman approximants for control: Flexible sampling, error analysis, and stability,”SIAM Journal on Control and Optimization, vol. 63, no. 6, pp. 4044–4071, 2025

    work page 2025

  12. [12]

    Persistent excitation in adaptive systems,

    K. S. Narendra and A. M. Annaswamy, “Persistent excitation in adaptive systems,”International Journal of Control, vol. 45, no. 1, pp. 127–160, 1987

    work page 1987

  13. [13]

    A note on persistency of excitation,

    J. C. Willems, P. Rapisarda, I. Markovsky, and B. L. De Moor, “A note on persistency of excitation,” Systems & Control Letters, vol. 54, no. 4, pp. 325–329, 2005

    work page 2005

  14. [14]

    From experiment design to closed-loop control,

    H. Hjalmarsson, “From experiment design to closed-loop control,”Automatica, vol. 41, no. 3, pp. 393–438, 2005

    work page 2005

  15. [15]

    D-optimal input design for nonlinear fir-type systems: A dispersion-based approach,

    A. De Cock, M. Gevers, and J. Schoukens, “D-optimal input design for nonlinear fir-type systems: A dispersion-based approach,”Automatica, vol. 73, pp. 88–100, 2016

    work page 2016

  16. [16]

    Trajectory synthesis for fisher information maximization,

    A. D. Wilson, J. A. Schultz, and T. D. Murphey, “Trajectory synthesis for fisher information maximization,”IEEE Transactions on Robotics, vol. 30, no. 6, pp. 1358–1370, 2014

    work page 2014

  17. [17]

    Data informativity: A new perspective on data-driven analysis and control,

    H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel, “Data informativity: A new perspective on data-driven analysis and control,”IEEE Transactions on Automatic Control, vol. 65, no. 11, pp. 4753–4768, 2020

    work page 2020

  18. [18]

    Optimal excitation trajectories for mechanical systems 14 identification,

    T. Lee, B. D. Lee, and F. C. Park, “Optimal excitation trajectories for mechanical systems 14 identification,”Automatica, vol. 131, p. 109773, 2021

    work page 2021

  19. [19]

    Space-filling input design for nonlinear state-space identification,

    M. Kiss, R. Tóth, and M. Schoukens, “Space-filling input design for nonlinear state-space identification,”IFAC-PapersOnLine, vol. 58, no. 15, pp. 562–567, 2024

    work page 2024

  20. [20]

    Space-filling optimized excitation signals for nonlinear system identification of dynamic processes of a diesel engine,

    V . Smits and O. Nelles, “Space-filling optimized excitation signals for nonlinear system identification of dynamic processes of a diesel engine,”Control Engineering Practice, vol. 144, p. 105821, 2024

    work page 2024

  21. [21]

    On space-filling input design for nonlinear dynamic model learning: A gaussian process approach,

    Y . Liu, M. Kiss, R. Tóth, and M. Schoukens, “On space-filling input design for nonlinear dynamic model learning: A gaussian process approach,”IEEE Control Systems Letters, vol. 9, pp. 1868–1873, 2025

    work page 2025

  22. [22]

    Active learning-based model predictive coverage control,

    R. Rickenbach, J. Köhler, A. Scampicchio, M. N. Zeilinger, and A. Carron, “Active learning-based model predictive coverage control,”IEEE Transactions on Automatic Control, vol. 69, no. 9, pp. 5931–5946, 2024

    work page 2024

  23. [23]

    Improved algorithms for linear stochastic bandits,

    Y . Abbasi-Yadkori, D. Pál, and C. Szepesvári, “Improved algorithms for linear stochastic bandits,” inAdvances in Neural Information Processing Systems, vol. 24, 2011, pp. 2312–2320

    work page 2011

  24. [24]

    Learning without mixing: Towards a sharp analysis of linear system identification,

    M. Simchowitz, H. Mania, S. Tu, M. I. Jordan, and B. Recht, “Learning without mixing: Towards a sharp analysis of linear system identification,” inProceedings of the 31st Conference on Learning Theory, ser. Proceedings of Machine Learning Research, vol. 75. PMLR, 2018, pp. 439–473

    work page 2018