Pith. sign in

REVIEW 5 minor 41 references

Machines that Predict Trajectories from Templates

T0 review · 0 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Stored output trajectories form behavioral spaces that predict futures without identifying the generating system.

desk verdict Solid theory paper that turns multi-system trajectory libraries into exact predictors, with clean spectral and interconnection extensions; classical assumptions, no load-bearing holes. read the letter →

arxiv 2607.11551 v1 pith:2SP25NXA submitted 2026-07-13 eess.SY cs.SY

Claudio De Persis , Pietro Tesi This is my paper · ORCID
classification eess.SYcs.SY
keywords trajectorypredictiontemplatelibrariesbehavioralspacescontinuationmapsspectralinclusiondata-drivenimmersioninterconnected
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 shows that a library of finite-length output windows, drawn from one or several linear systems, can be used as a prediction machine: given only the past of a new trajectory, a linear fit against the library past block recovers the future exactly whenever the past is long enough and the data are rich. Exactness holds even when the trajectory comes from a different system whose visible eigenvalues already appear in the library, and it continues to hold after the library is enlarged by feedback, series or parallel interconnection constraints that create new modes. The same mechanism extends to nonlinear systems whose outputs lie in a finite-dimensional linear observation space. The practical payoff is that prediction no longer requires first recovering a state-space model; the stored templates themselves act as the generative model.

What carries the argument

The continuation map of the aggregate (or interconnection-generated) library: the linear relation Hf = Lmix Hp that turns any feasible encoding of the past into an exact future decoding, independent of which particular coefficient vector is chosen.

What would settle it

Build a library from several diagonalizable linear systems whose visible eigenvalues are known; generate a new scalar-output trajectory whose spectrum is not contained in that set and check whether the least-squares predictor still recovers the future exactly for past length equal to the aggregate observability index.

Watch

Extended reading notes

Core claim

Libraries of stored output trajectories define behavioral spaces that serve as exact prediction machines: under data richness and a past window at least as long as the aggregate observability index, any coefficient vector that matches an observed past against the library yields the correct future for every trajectory generated by any system represented in the library, and for many systems outside it whose output-visible spectrum is covered by the library.

Load-bearing premise

The stored trajectories must be rich enough to span every mode of every template system, and the observed past must be at least as long as the observability index of the whole library.

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

0 major / 5 minor

Summary. The paper develops a theory of trajectory prediction from libraries of stored output templates, without identifying the generating state-space model. Libraries of finite-length output windows from one or more LTI systems define behavioral spaces; exact prediction is characterized via continuation maps (Lemma 3), mixed-library factorization (Lemma 4, Theorem 1), and, for out-of-library systems, spectral inclusion of output-visible eigenvalues (Proposition 2 / Theorem 5). Robustness to noisy observations and libraries is quantified by exact error identities (Theorems 3–4); interconnection constraints (feedback, series, parallel) generate new libraries with emergent modes (Lemmas 5–10, Theorem 6, Proposition 6); and the framework extends to nonlinear systems whose outputs are contained in or immersed into finite-dimensional linear behaviors (Proposition 7). Numerical examples illustrate noisy prediction, sparse modal recovery, feedback-generated modes, and nonlinear immersion.

Significance. If the results hold, the paper supplies a clean algebraic foundation for template-based prediction machines that generalize beyond stored trajectories and, via spectral inclusion or interconnection, beyond the generating systems themselves. The core contributions rest on standard linear-algebra and systems-theory facts (observability index, data richness, Vandermonde structure) with explicit hypotheses and exact error identities; the interconnection construction and the immersion route to nonlinear trajectories are particularly distinctive. Strengths include parameter-free exact statements under clearly scoped assumptions, explicit equivalence between spectral inclusion, output inclusion and linear immersion (Lemmas 11–12, Corollary 2), and concrete numerical illustrations. The work sits naturally at the intersection of behavioral data-driven control and sequence completion, and opens well-defined directions (library design, inputs, online recursion).

minor comments (5)
  1. The manuscript is long and dense; a short roadmap or table of main theorems (Theorem 1, Proposition 2, Theorem 6, Proposition 7) early in Section II would help readers navigate the progression from single-system to multi-system, out-of-library, interconnection and nonlinear cases.
  2. Notation for the continuation map switches among L*, Lmix and L; a single consistent symbol (with subscripts only when needed) would reduce cognitive load, especially in Sections V–VII.
  3. Figure 1 (noisy-data histogram) and Figure 3 (sparse modal prediction) would benefit from explicit axis units and a brief caption statement of the SNR / r / T values used, so that the plots are self-contained.
  4. The restriction of the spectral results (Propositions 2–4, 6–7) to scalar-output diagonalizable systems is clearly stated, but a short remark on the multi-output or non-diagonalizable obstacles (and whether they are merely technical) would help readers assess the scope.
  5. A few typographical slips remain (e.g., “visible the spectral inclusion” in the proof of Proposition 2; occasional missing spaces around math). A final copy-edit pass would polish the presentation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: exact-prediction claims follow from observability-index continuation maps, data-richness image factorizations, and spectral-to-output inclusion, all proved algebraically under explicit hypotheses.

full rationale

The derivation chain is self-contained linear algebra and classical systems theory. Lemma 3 constructs the continuation map L* from the definition of the observability index (Fact 1 / Definition 1) and shows that any feasible latent g (or any L solving the data equation) recovers the future exactly when rank(X)=n; Theorem 1 simply applies the same argument to the block-diagonal aggregate pair. Proposition 2 / Theorem 5 prove that visible spectral inclusion implies Im(OT(Anew,Cnew))subseteq Im(OT(Amix,Cmix)) by expanding diagonalizable trajectories in Vandermonde modes, then invoke data richness to place every such trajectory in Im(HY) so that the continuation identity applies; the converse necessity (Proposition 3) is likewise a linear-independence argument on Vandermonde columns. Interconnection libraries (Lemmas 5-10) are constructed by explicit kernel constraints on atomic input-output matrices and factorized by the closed-loop realization; the same continuation then yields Theorem 6 / Proposition 6. Nonlinear extension (Proposition 7) uses the standard immersion of a finite-dimensional observation space into a companion linear system, then reduces to the linear spectral case. No parameter is fitted to data and then re-used as a prediction; numerical examples merely illustrate the exact or bounded-error statements under prescribed SNRs and eigenvalues. Self-citations to the authors' earlier data-driven-control papers are background only and are not load-bearing for any uniqueness or existence claim. Consequently the central theorems do not reduce to their inputs by construction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 3 invented entities

Central claims rest on standard linear systems theory (observability index, Vandermonde structure of modal trajectories, Moore-Penrose least squares) plus domain assumptions of data richness and past length at least the observability index. No free parameters are fitted for the theorems; invented entities are definitional (continuation map, mixed library, visible spectrum) rather than new physical objects.

assumptions (4)
  • standard math Observability index s of (A,C) is the smallest integer such that rank Os = rank Os+1; for r >= s a linear continuation map exists (Fact 1 / Definition 1).
    Classical linear systems theory; used throughout Lemmas 3 and Theorems 1, 5, 6.
  • domain assumption Data richness: rank Xi = ni for every template system i (Definition 2).
    Ensures Im(HY) equals the full output behavior of the aggregate system; required for uniform exact prediction over all trajectories.
  • standard math For scalar-output diagonalizable systems, visible spectral inclusion implies output-trajectory inclusion (Corollary 2 / Proposition 2).
    Follows from Vandermonde spanning of modal trajectories; necessity shown for sufficiently long T.
  • domain assumption Nonlinear systems with finite-dimensional observation space can be immersed into a linear companion-form system whose spectrum is the observation-space characteristic roots (Definition 4, Proposition 7).
    Standard immersion / finite-dimensional observation-space theory (Fliess, Isidori); used to transfer linear prediction guarantees to nonlinear outputs.
invented entities (3)
  • Continuation map L* (or Lmix) independent evidence
    purpose: Linear operator that maps any admissible past window to its unique future window for a given linear behavior.
    Definitional object extracted from the observability matrix once r >= s; not a new physical entity.
  • Visible spectrum specC(A) independent evidence
    purpose: Subset of eigenvalues whose eigenspaces are not annihilated by C; used for spectral inclusion tests.
    Standard notion of output-visible modes; independent of the paper's prediction claims.
  • Interconnection-generated library (feedback/series/parallel via compatibility kernel Theta) independent evidence
    purpose: Construct new output libraries with emergent modes directly from atomic input-output data without new experiments.
    Constructive definition via ker(Q); the resulting (A,C) is the classical closed-loop realization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machines that Predict Trajectories from Templates." pith.science (2026). https://pith.science/paper/2SP25NXA

@misc{pith2026260711551,
  author       = {Pith},
  title        = {Pith review of: Machines that Predict Trajectories from Templates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SP25NXA}},
  note         = {Machine review of arXiv:2607.11551}
}
read the original abstract

We study trajectory prediction from libraries of stored output templates. Given the past of an unknown trajectory, the goal is to predict its future without identifying the state-space model that generated it. We show that libraries of trajectories generated by one or more dynamical systems define behavioral spaces that can be used as prediction mechanisms. For linear systems, we characterize exact prediction in terms of continuation maps, behavioral containment, and spectral conditions on output-visible eigenvalues. We also analyze robustness to noisy observations and noisy libraries, derive error bounds for out-of-library trajectories, and show how interconnection constraints can compose template libraries into new behavioral spaces with emergent modes. Finally, we extend the framework to nonlinear systems whose output trajectories are contained in, or immersed into, finite-dimensional linear behaviors. These results provide a theory of template-based prediction machines capable of generalizing beyond the stored trajectories and, in some cases, beyond the systems that generated them.

Figures

Figures reproduced from arXiv: 2607.11551 by the authors.

Figure 1
Figure 1. Histogram of the mean squared prediction error over [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Machine that performs out-of-library prediction from a mixed template library. The observed past [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Sparse modal prediction using five second-order sentinel systems. The [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The following lemma provides a state-space factorization of the feedback-interconnected library. Lemma 6 (Factorization of the feedback-interconnected li￾brary). Let Q and Θ be defined as in Lemma 5. Let A :=  A1 B1C2 B2C1 A2  , C := C1 0 [PITH_FULL_IMAGE:figures/fu…
Figure 4
Figure 4. Figure 4: Construction of a feedback-generated library from two atomic input-output libraries. The matrix [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Trajectory prediction by observing r = 3 samples (black dots). The feedback-generated library is able to predict exactly the trajectory even though its natural modes are not present in either atomic system. The eigenvalues of A are λ1,2 = 0.5 ± √ 4.01 2 which are diffe…
Figure 6
Figure 6. Figure 6: Prediction of a nonlinear output trajectory from a mismatched linear [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 2 linked inside Pith

  1. [1]

    Vapnik,The Nature of Statistical Learning Theory

    V . Vapnik,The Nature of Statistical Learning Theory. Springer-Verlag, New York, 1995

    1995

  2. [2]

    A tutorial on support vector machines for pattern recog- nition,

    C. Burges, “A tutorial on support vector machines for pattern recog- nition,”Data Mining and Knowledge Discovery, vol. 2, pp. 121–167, 1998

    1998

  3. [3]

    Instance-based learning algorithms,

    D. W. Aha, D. Kibler, and M. K. Albert, “Instance-based learning algorithms,”Machine Learning, vol. 6, pp. 37–66, 1991

    1991

  4. [4]

    A survey of fault diagnosis and fault-tolerant techniques–part i: Fault diagnosis with model-based and signal-based approaches,

    Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and fault-tolerant techniques–part i: Fault diagnosis with model-based and signal-based approaches,”IEEE Transactions on Industrial Electronics, vol. 62, no. 6, pp. 3757–3767, 2015

    2015

  5. [5]

    Linear predictors for nonlinear dynamical sys- tems: Koopman operator meets model predictive control,

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

    2018

  6. [6]

    Data-driven approximation of the Koopman generator: Model reduction, system identification, and control,

    S. Klus, F. Nuske, J. Peitz, C. Clementi, and C. Sch ¨utte, “Data-driven approximation of the Koopman generator: Model reduction, system identification, and control,”Physica D: Nonlinear Phenomena, vol. 406, p. 132416, 2020

    2020

  7. [7]

    Koopman operator theory: fundamentals, control, and applications,

    I. Mezi ´c, J. Cort´es, K. Worthmann, M. Lazar, and A. Lederer, “Koopman operator theory: fundamentals, control, and applications,”arXiv preprint arXiv:2607.01819, 2026

    arXiv 2026

  8. [8]

    Finite-dimensional observation-spaces for non-linear sys- tems,

    M. Fliess, “Finite-dimensional observation-spaces for non-linear sys- tems,” inFeedback Control of Linear and Nonlinear Systems.Proceed- ings of the Joint Workshop on Feedback and Synthesis of Linear and Nonlinear Systems Bielefeld/Roma. Springer, 2005, pp. 73–77

    2005

Show all 41 references
  1. [9]

    Exact finite Koopman embedding of block-oriented polynomial systems,

    L. Iacob, R. T ´oth, and M. Schoukens, “Exact finite Koopman embedding of block-oriented polynomial systems,”arXiv preprint arXiv:2507.15093, 2025

    2025 arXiv

  2. [10]

    Great: Grassmannian recursive algorithm for tracking & online system identification,

    A. Sasfi, A. Padoan, I. Markovsky, and F. D ¨orfler, “Great: Grassmannian recursive algorithm for tracking & online system identification,”IEEE Transactions on Automatic Control, 2025

    2025

  3. [11]

    Data-driven unknown-input ob- servers and state estimation,

    M. S. Turan and G. Ferrari-Trecate, “Data-driven unknown-input ob- servers and state estimation,”IEEE Control Systems Letters, vol. 6, pp. 1424–1429, 2021

    2021

  4. [12]

    Data-driven input reconstruction and experimental validation,

    J. Shi, Y . Lian, and C. N. Jones, “Data-driven input reconstruction and experimental validation,”IEEE Control Systems Letters, vol. 6, pp. 3259–3264, 2022

    2022

  5. [13]

    Data-driven inverse of linear systems and application to disturbance observers,

    Y . Eun, J. Lee, and H. Shim, “Data-driven inverse of linear systems and application to disturbance observers,” in2023 American Control Conference (ACC). IEEE, 2023, pp. 2806–2811

    2023

  6. [14]

    On the equivalence of model-based and data-driven approaches to the design of unknown-input observers,

    G. Disar `o and M. E. Valcher, “On the equivalence of model-based and data-driven approaches to the design of unknown-input observers,”IEEE Transactions on Automatic Control, vol. 70, no. 3, pp. 2074–2081, 2024

  7. [15]

    Data-enabled predictive control: In the shallows of the DeePC,

    J. Coulson, J. Lygeros, and F. D ¨orfler, “Data-enabled predictive control: In the shallows of the DeePC,” in18th European Control Conference, 2019

    2019

  8. [16]

    Linear tracking MPC for nonlinear systems–Part ii: The data-driven case,

    J. Berberich, K. K ¨ohler, M. M ¨uller, and F. Allg ¨ower, “Linear tracking MPC for nonlinear systems–Part ii: The data-driven case,”IEEE Trans- actions on Automatic Control, vol. 67, no. 9, pp. 4406–4421, 2022

    2022

  9. [17]

    Learning- based model predictive control: Toward safe learning in control,

    L. Hewing, K. P. Wabersich, M. Menner, and M. N. Zeilinger, “Learning- based model predictive control: Toward safe learning in control,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 3, pp. 269– 296, 2020

    2020

  10. [18]

    On direct vs indirect data-driven predictive control,

    V . Krishnan and F. Pasqualetti, “On direct vs indirect data-driven predictive control,” in60th IEEE Conference on Decision and Control (CDC), 2021, pp. 736–741

    2021

  11. [19]

    From system models to class models: An in-context learning paradigm,

    M. Forgione, F. Pura, and D. Piga, “From system models to class models: An in-context learning paradigm,”IEEE Control Systems Letters, vol. 7, pp. 3513–3518, 2023

    2023

  12. [20]

    The asymptotic behavior of attention in transformers,

    ´A. R. Abella, J. P. Silvestre, and P. Tabuada, “The asymptotic behavior of attention in transformers,”arXiv preprint arXiv:2412.02682, 2024

    2024

  13. [21]

    Multistability of self-attention dynamics in transformers,

    C. Altafini, “Multistability of self-attention dynamics in transformers,” IEEE Transactions on Automatic Control, 2026

    2026

  14. [22]

    Generalization error analysis for selective state-space models through the lens of attention,

    A. Honarpisheh, M. Bozdag, O. Camps, and M. Sznaier, “Generalization error analysis for selective state-space models through the lens of attention,”Advances in Neural Information Processing Systems, vol. 38, pp. 170 805–170 839, 2026

    2026

  15. [23]

    Selection mechanisms for sequence modeling using linear state space models,

    U. Casti, S. Zampieri, and F. Pasqualetti, “Selection mechanisms for sequence modeling using linear state space models,”arXiv preprint arXiv:2505.17932, 2025

    2025

  16. [24]

    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

    2005

  17. [25]

    Willems’ fundamental lemma for state-space systems and its extension to multiple datasets,

    H. van Waarde, C. De Persis, K. Camlibel, and P. Tesi, “Willems’ fundamental lemma for state-space systems and its extension to multiple datasets,”IEEE Control Systems Letters, vol. 4, no. 3, pp. 602–607, 2020

    2020

  18. [26]

    On the design of persistently exciting inputs for data-driven control of linear and nonlinear systems,

    M. Alsalti, V . G. Lopez, and M. A. M¨uller, “On the design of persistently exciting inputs for data-driven control of linear and nonlinear systems,” IEEE Control Systems Letters, vol. 7, pp. 2629–2634, 2023

    2023

  19. [27]

    On controllability and persistency of excitation in data-driven control: Extensions of willems’ fundamental lemma,

    Y . Yu, S. Talebi, H. van Waarde, U. Topcu, M. M, and B. Ac ¸ikmes ,e, “On controllability and persistency of excitation in data-driven control: Extensions of willems’ fundamental lemma,” in2021 60th IEEE Con- ference on Decision and Control (CDC), 2021, pp. 14–17

    2021

  20. [28]

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

    X. Shang, J. Cort ´es, and Y . Zheng, “Willems’ fundamental lemma for nonlinear systems with koopman linear embedding,”IEEE Control Systems Letters, vol. 8, pp. 3135–3140, 2024

    2024

  21. [29]

    From product hilbert spaces to the generalized koop- man operator and the nonlinear fundamental lemma,

    M. Lazar, “From product hilbert spaces to the generalized koop- man operator and the nonlinear fundamental lemma,”arXiv preprint arXiv:2508.07494, 2025

    2025

  22. [30]

    Data-driven simulation and control,

    I. Markovsky and P. Rapisarda, “Data-driven simulation and control,” International Journal of Control, vol. 81, no. 12, pp. 1946–1959, 2008

    1946

  23. [31]

    Formulas for data-driven control: Stabilization, optimality, and robustness,

    C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization, optimality, and robustness,”IEEE Transactions on Automatic Control, vol. 65, no. 3, pp. 909–924, 2020

    2020

  24. [32]

    Learning controllers for nonlinear systems from data,

    C. De Persis and P. Tesi, “Learning controllers for nonlinear systems from data,”Annual Reviews in Control, vol. 56, p. 100915, 2023

    2023

  25. [33]

    Data-driven control of large- scale networks with formal guarantees: A small-gain-free approach,

    D. Ajeleye, A. Lavaei, and M. Zamani, “Data-driven control of large- scale networks with formal guarantees: A small-gain-free approach,” IEEE Control Systems Letters, vol. 7, pp. 3453–3458, 2023

    2023

  26. [34]

    Bridging direct and indirect data-driven control formulations via regularizations and relaxations,

    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, vol. 68, no. 2, pp. 883–897, 2023

    2023

  27. [35]

    Verhaegen and V

    M. Verhaegen and V . Verdult,Filtering and System Identification: A Least Squares Approach. Cambridge University Press, 2007

    2007

  28. [36]

    Subspace angles between ARMA models,

    K. De Cock and B. De Moor, “Subspace angles between ARMA models,”Systems & Control Letters, vol. 46, pp. 265–270, 2002

    2002

  29. [37]

    Behavioral uncertainty quantification for data-driven control,

    A. Padoan, J. Coulson, H. J. Van Waarde, J. Lygeros, and F. D ¨orfler, “Behavioral uncertainty quantification for data-driven control,” in2022 IEEE 61st Conference on Decision and Control (CDC). IEEE, 2022, pp. 4726–4731

    2022

  30. [38]

    G. H. Golub and C. F. Van Loan,Matrix computations. John Hopkins University Press, 2013

    2013

  31. [39]

    J. R. Partington,Linear operators and linear systems: an analytical approach to control theory. Cambridge University Press, 2004, no. 60

    2004

  32. [40]

    Isidori,Nonlinear control systems

    A. Isidori,Nonlinear control systems. Springer, London, Third Edition, 1995

    1995

  33. [41]

    Linear observer synthesis for nonlinear systems linear observer synthesis for nonlinear systems using Koopman operator framework,

    A. Surana and A. Banaszuk, “Linear observer synthesis for nonlinear systems linear observer synthesis for nonlinear systems using Koopman operator framework,” inIFAC-PapersOnLine 49-18, 2016, p. 716.723

    2016

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.