Pith. sign in

REVIEW 3 major objections 8 minor 20 references

A three-stage scheme makes fitting nonlinear state-space models faster and more reliable by first inferring latent signals, then learning the nonlinearity, then refining with multiple shooting.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-02 20:31 UTC pith:PL5SPNIY

load-bearing objection A practical three-stage initialization for NL-LFR identification that mostly delivers on its claims; the main gap is an ablation that isolates what the guided residual search actually contributes. the 3 major comments →

arxiv 2602.22964 v2 pith:PL5SPNIY submitted 2026-02-26 eess.SP

A guided residual search for nonlinear state-space identification

Merijn Floren , Jan Swevers This is my paper
classification eess.SP
keywords nonlinear state-space identificationNL-LFR modelsguided residual searchmultiple shootingdistribution shiftneural network initializationSilverbox benchmarkF-16 ground vibration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a difficult nonlinear state-space identification problem can be broken into three tractable stages: a guided residual search that infers the latent nonlinear input w and state x from a linear baseline model, a neural-network fit of the nonlinearity on those inferred signals, and a multiple-shooting refinement that corrects the distribution shift caused by the decoupled training. The authors argue that this decomposition lands the final optimization in a favorable attraction region, improving convergence over naive linear-only initialization and matching black-box methods on two experimental benchmarks. If true, it offers a principled initialization strategy for NL-LFR models, making them a practical choice for nonlinear system identification.

Core claim

The central claim is that the proposed sequential scheme—guided residual search to infer latent signals w and x and fit B_w, D_yw; parametric neural-network learning of the nonlinear residual; then joint multiple-shooting refinement—improves the reliability and efficiency of the estimation process. The paper shows that the guided residual search produces state and residual estimates that, despite a known distribution shift (formalized in Theorem 1), place the final multiple-shooting optimization in a parameter region where it converges to lower simulation NRMSE than a linear-only initialization. Experiments on Silverbox and F-16 ground vibration benchmarks demonstrate competitive performance

What carries the argument

The key machinery is the bilevel guided residual search: an inner convex problem that, for a fixed B_w and D_yw, estimates the latent input w over a sliding window via a closed-form least-squares solution, and an outer problem that updates B_w and D_yw to minimize the output mismatch. The inferred signals form a dataset D* used to train the neural network nonlinearity (with C_z and D_zu), and the resulting states seed a multiple-shooting optimization that enforces continuity across intervals. Theorem 1 provides the formal justification for the final multiple-shooting stage by showing that one-step-ahead training errors grow exponentially with rollout length.

Load-bearing premise

The guided residual search must return latent signals w and x that are meaningfully related to the true nonlinear residual; if the inferred residual is meaningless, the neural network and the subsequent initialization inherit the error, and the method loses its advantage over random initialization.

What would settle it

A simple counterexample would be a well-defined NL-LFR system with known parameters where the guided residual search, run with several random initializations of B_w and D_yw, consistently yields latent signals that cannot be fitted by the neural network (low validation accuracy) while the final multiple-shooting optimization converges to a worse solution than a random initialization. This could be tested on a simulated system with known ground truth.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • The guided residual search provides an initialization that consistently reduces the final simulation error compared to random or linear-only initialization, regardless of whether the final optimization uses single or multiple shooting.
  • The theoretical distribution-shift analysis explains why neural-network training loss decreases while simulation error worsens, and justifies the use of multiple shooting in the final stage.
  • The method does not require prior knowledge of the system's equations or periodic data, extending guided residual search to general input-output data.
  • On the Silverbox benchmark, the method achieves competitive RMSE compared to SUBNET and PNLSS, though with less extrapolation capability than the polynomial-based PNLSS.
  • On the F-16 benchmark, the method improves per-output simulation NRMSE over the BLA by 2-7 percentage points on unseen test data.

Where Pith is reading between the lines

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

  • The sequential decomposition could be adapted to other model classes: any gray-box state-space model with a structured nonlinearity might benefit from inferring latent signals before fitting the nonlinear part.
  • The sensitivity to random initialization of B_w and D_yw (Remark 2) suggests that a deterministic or physics-informed initialization of these matrices could further improve reliability, especially for high-dimensional systems.
  • The distribution-shift theorem applies to any learned one-step predictor; it could serve as a general argument for using shooting methods in other identification problems with long horizons.
  • The computational cost breakdown (21 s for GRS plus neural network training on Silverbox) suggests the method is practical even on modest hardware, and the parallelism of multiple shooting could be exploited for further speedups.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. The paper proposes a three-stage identification procedure for nonlinear LFR state-space models. Starting from an initial linear model, the method (i) performs a guided residual search (GRS) that alternates between inferring latent residual signals (w) and states (x) and estimating the associated matrices B_w and D_yw; (ii) trains a neural-network parametrization of the static nonlinearity on the inferred latent variables; and (iii) refines all parameters with a multiple-shooting optimization initialized from the GRS states. The authors claim that this sequential scheme improves convergence over naive linear-only initialization, that multiple shooting mitigates the distribution shift introduced by one-step residual learning, and that the method is competitive with state-of-the-art black-box approaches on the Silverbox and F-16 ground-vibration benchmarks.

Significance. If the central claim holds, the paper offers a practical and relatively cheap initialization strategy for nonlinear state-space identification, a problem that is notoriously sensitive to initialization. The paper also formalizes, via a simple error-propagation bound (Theorem 1), why one-step residual training can lead to exponential simulation-error growth, and demonstrates empirically that multiple shooting can recover from this effect. Strengths include the use of two standard experimental benchmarks, a transparent description of hyperparameters, 25 random seeds for the Silverbox study, and comparisons against SUBNET and PNLSS. However, the experiments do not currently isolate the contribution of the GRS itself from other parts of the pipeline, and the F-16 demonstration rests on a single initialization, which is a concern given the paper's own Remark 2.

major comments (3)
  1. [Section IV.A, Fig. 2 and scenarios S1–S4] The comparison does not isolate the effect of the guided residual search. S1/S2 run the full GRS and then train the neural network on the inferred dataset D*, whereas S3/S4 skip both: they use a randomly initialized θ_nl and B_w,D_yw initialized from U(−1e−4,1e−4). The text attributes the improved initial NRMSE to 'the guided residual search' (Fig. 2 caption, Section IV.A.1), but the observed benefit could come from the NN training stage alone, from the different scale of B_w,D_yw, or from their interaction. Please add an ablation that keeps the NN training and initialization scale fixed and varies only the source of the latent residual (e.g., train the same NN on residuals from the BLA model without GRS, or on a deliberately corrupted/random w*). Without such an ablation, the central mechanism claim that GRS provides a favorable attraction region is underdetermined.
  2. [Section III.A, Remark 2 and Eq. (12)] Remark 2 concedes that the outer loss (12) has multiple effective minimizers, not all of which yield meaningful relationships between x and w. This is load-bearing because D*, produced by the GRS, is the only training signal for the neural network and the basis for the multiple-shooting initialization. The paper recommends multiple initializations but provides no diagnostic to detect or discard meaningless D*. Please quantify the behavior across the 25 Silverbox seeds: report the distribution of the GRS objective values and the correlation between GRS loss and final NRMSE, and for the F-16 case report sensitivity to GRS restarts. This would directly support the claim that the favorable attraction region is due to meaningful latent inference rather than to a lucky run or to the regularization making the fit trivially good.
  3. [Section IV.B, Fig. 5] The F-16 result is the key demonstration that the proposed method escapes a local minimum that traps linear-only initialization, despite a higher initial simulation error. However, this figure appears to be based on a single initialization of the GRS and the final optimization. Given the acknowledged sensitivity of GRS to initialization (Remark 2), a single trajectory is insufficient to establish the reliability of the claimed behavior. Please provide multiple GRS restarts and show the spread of optimization trajectories (e.g., median with percentiles), or at least a small seed study. This is particularly important because Table II reports only one final model with no uncertainty.
minor comments (8)
  1. [Section IV.A] The training/test split for the Silverbox data is not stated explicitly. Clarify which portion of the multisine segment is used for estimation and whether the arrowhead segment is used only for testing.
  2. [Section III.C, Eqs. (17b) and (17e)] The continuity constraint (17e) appears redundant with (17b) when n+d is also a shooting index in I. If there is a different intended role for (17e), please explain; otherwise remove it for clarity.
  3. [Section III.A.1, Eq. (9)] Consider writing the closed-form solution as S_w^* = G^{-1} T_w^T (S_y − O_x x* − T_u S_u) to avoid the double negative in the current expression.
  4. [Section III.B] The phrase 'by implicitly treating the samples in D* as independent and identically distributed' is confusing because the samples are not i.i.d. by construction. It is clearer to say that the temporal dependencies are ignored during this stage.
  5. [Section IV.A.1 and IV.B] The sentences 'for 25 Levenberg-Marquardt iterations' and 'for 50 Levenberg-Marquardt iterations' are ambiguous: they should specify whether these are inner-loop iterations, outer-loop iterations, or whole GRS restarts.
  6. [Fig. 4] The y-axis label contains a garbled '×10□6'; please fix the formatting.
  7. [Table I] The text says 'The best-performing model from scenario S1 is evaluated' but the table reports mean±standard deviation. Clarify whether the entries are averaged over the 25 models or are for the single best model.
  8. [Abstract and Section IV.A.2] The phrase 'competitive with state-of-the-art black-box methods' is somewhat strong given that PNLSS is about 4× better on the full Silverbox arrowhead test (0.26 mV vs. 1.13 mV). The paper acknowledges the polynomial basis explains this, but consider tempering the wording or emphasizing the multisine test where the proposed method is competitive.

Circularity Check

0 steps flagged

No significant circularity: the derivation is self-contained, evaluated on independent benchmarks, and its flagged limitation (Remark 2) weakens assurance without being circular.

full rationale

The paper does not reduce any predicted result to a fitted input. Theorem 1 is a standard error-propagation bound proved in the text from Lipschitz continuity and a bounded one-step error; it does not reuse the estimated parameters. The GRS inner problem (Eqs. 6-9) and outer update (Eq. 12) form an alternating heuristic, not a definitional identity: the inferred w* is optimized from data and then used to train the NN (Eq. 13), and the final multiple-shooting stage (Eq. 17) is a separate joint optimization. The reported improvements are measured against independent baselines (BLA, SUBNET, PNLSS) on external benchmark data, so the central 'improved convergence' claim is empirically falsifiable rather than forced by construction. The authors candidly flag the main limitation in Remark 2: 'This dependence introduces multiple effective minimizers of the outer loss (12), not all yielding meaningful relationships between x and w.' This is a real weakness in assurance, but it is not circularity; it merely says the heuristic can fail and should be run with multiple seeds. Self-citations [6]-[9] motivate the GRS idea and are used for the BLA parametrization procedure, but the load-bearing claims of this paper—decomposition for arbitrary data, distribution-shift analysis, and benchmark performance—do not reduce to those citations. No uniqueness theorem is imported, no fitted quantity is renamed as a prediction, and no known result is repackaged. Score 0.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The method's claims rest on standard identifiability assumptions (noise, excitation, output controllability), a usable initial linear model, and the NL-LFR model class. The most fragile non-empirical premise is that the guided residual search returns latent signals w*,x* that are meaningful; the authors explicitly note multiple minimizers and recommend multiple initializations. All numerical hyperparameters (λ, H, d, architectures, model orders) are tuned per dataset and count as free parameters. No new physical entities are introduced.

free parameters (8)
  • regularization λ = 1 (Silverbox), 100 (F-16)
    Balances output fit against solution variability in the inner problem; tuned per dataset.
  • prediction window length H = 10 (both benchmarks)
    Sliding-window horizon in the guided residual search; chosen without sensitivity study.
  • transient discard N0 = 100 (Silverbox), 500 (F-16)
    First samples of each realization are discarded to remove zero-initial-state transient effects; tuned per dataset.
  • multiple-shooting interval length d = 1 (Silverbox), 200 (F-16)
    Interval length for the final multiple-shooting stage; F-16 value chosen as a memory/performance compromise.
  • neural-network architecture = 1 hidden layer, 15 ReLU (Silverbox); 1 hidden layer, 32 tanh (F-16)
    Hand-chosen capacity and activation; affects expressiveness and Lipschitz constant.
  • NN training iterations and learning rate = 100 Adam, lr=1e-3 (Silverbox); 1000 Adam, lr=4e-3 (F-16)
    Training schedule for the residual neural network; chosen by hand.
  • Levenberg-Marquardt iterations in GRS = 25 (Silverbox), 50 (F-16)
    Number of iterations of the bilevel guided residual search; selected per problem.
  • model orders (nx, nw, nz) = (2,1,1) Silverbox; (15,2,6) F-16
    Selected “based on empirical tuning” in Section IV-B; controls identifiability and expressiveness.
axioms (6)
  • domain assumption Assumption 1: measured output is corrupted by additive, zero-mean, stationary noise uncorrelated with the exactly known input.
    Invoked throughout the estimation; if violated (e.g., noise correlated with input), the residual estimates are biased.
  • domain assumption Assumption 2: the input signal is persistently exciting of sufficiently high order.
    Required for identifiability of the linear dynamics and of the nonlinear residual contribution.
  • domain assumption Assumption 3: the submodel M1 is output controllable with respect to its exogenous input w.
    Needed so the inner problem can infer w from output data; the authors note in Remark 2 that typical random initializations render it satisfied.
  • domain assumption An initial linear model M0 is available, stable, and provides a rough approximation of the linear input-output behavior.
    The method starts from M0; in the experiments it is obtained via the best linear approximation. If no such baseline exists, the guided residual search has no starting point.
  • domain assumption The true system belongs to the NL-LFR model class: LTI dynamics with a static neural-network nonlinearity in feedback.
    The method only fits parameters within this structure; no structural validation against other nonlinear model classes is performed.
  • domain assumption Theorem 1 assumes f is uniformly Lipschitz continuous in x with constant L>1 and one-step prediction error uniformly bounded by ε.
    Used to justify the distribution-shift concern; the authors argue L>1 is typical for neural networks (citing [11], [12]) but do not measure L in their trained models.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A guided residual search for nonlinear state-space identification." pith.science (2026). https://pith.science/paper/PL5SPNIY

@misc{pith2026260222964,
  author       = {Pith},
  title        = {Pith review of: A guided residual search for nonlinear state-space identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PL5SPNIY}},
  note         = {Machine review of arXiv:2602.22964}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Identifying the parameters of nonlinear state-space models from input-output data typically requires solving a highly non-convex optimization problem, which is prone to slow convergence and suboptimal local solutions. This work improves the reliability and efficiency of the estimation process by decomposing the overall optimization problem into a sequence of tractable subproblems. Starting from a linear baseline model, nonlinear residual dynamics are first estimated using a guided residual search (GRS) and subsequently refined through multiple-shooting optimization. Experiments on two benchmarks show competitive performance with state-of-the-art black-box methods and improved convergence over naive initialization.

Figures

Figures reproduced from arXiv: 2602.22964 by Jan Swevers, Merijn Floren.

Figure 1
Figure 1. Figure 1: Schematic overview of the NL-LFR structure. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: NL-LFR simulation NRMSEs over the iterations of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Best-performing NL-LFR model from S1 on the Silverbox arrowhead (left) and multisine (right) test data, compared to the BLA model. The dashed vertical line marks the start of the extrapolation region. TABLE I: Simulation root mean square errors (RMSEs) on the Silverbox test data (mean ± standard deviation in mV) for the NL-LFR models of S1, compared with state-of-the￾art black-box methods. proposed SUBNET … view at source ↗
Figure 4
Figure 4. Figure 4: Distribution shift over the F-16 neural network train [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: F-16 simulation NRMSEs during the final optimiza [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

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

20 extracted references · 2 linked inside Pith

  1. [1]

    Identification of nonlinear systems using polynomial nonlinear state space models,

    J. Paduart, L. Lauwers, J. Swevers, K. Smolders, J. Schoukens, and R. Pintelon, “Identification of nonlinear systems using polynomial nonlinear state space models,”Automatica, vol. 46, no. 4, 2010

    2010

  2. [2]

    On the initialization of nonlinear LFR model identification with the best linear approximation,

    M. Schoukens and R. T ´oth, “On the initialization of nonlinear LFR model identification with the best linear approximation,”IFAC- PapersOnLine, vol. 53, no. 2, pp. 310–315, 2020

    2020

  3. [3]

    On the smoothness of nonlinear system identification,

    A. H. Ribeiro, K. Tiels, J. Umenberger, T. B. Sch ¨on, and L. A. Aguirre, “On the smoothness of nonlinear system identification,”Automatica, vol. 121, p. 109158, 2020

    2020

  4. [4]

    Shooting methods for identification of nonlinear state-space grey-box models,

    A. Retzler, J. Swevers, J. Gillis, and Z. Koll ´ar, “Shooting methods for identification of nonlinear state-space grey-box models,” in2022 IEEE 17th International Conference on Advanced Motion Control (AMC). IEEE, 2022, pp. 207–212

    2022

  5. [5]

    Nonlinear state-space identification using deep encoder networks,

    G. Beintema, R. T ´oth, and M. Schoukens, “Nonlinear state-space identification using deep encoder networks,” inLearning for Dynamics and Control. PMLR, 2021, pp. 241–250

    2021

  6. [6]

    Inference and learning of nonlinear LFR state-space models,

    M. Floren, J.-P. No ¨el, and J. Swevers, “Inference and learning of nonlinear LFR state-space models,”IEEE Control Systems Letters, vol. 9, pp. 1345 – 1350, 2025

    2025

  7. [7]

    A sliding-window approach for latent restoring force modeling,

    M. Floren and J. Swevers, “A sliding-window approach for latent restoring force modeling,”arXiv:2602.21918, 2026

    arXiv 2026

  8. [8]

    Nonlinear restoring force modelling using Gaussian processes and model predictive control,

    M. Floren and J.-P. No ¨el, “Nonlinear restoring force modelling using Gaussian processes and model predictive control,” inConference Proceedings of ISMA2022-USD2022, 2022, pp. 2493–2498

    2022

  9. [9]

    Identification of deformable linear object dynamics from input-output measurements in 3D space,

    M. Floren, S. Mamedov, J.-P. No ¨el, and J. Swevers, “Identification of deformable linear object dynamics from input-output measurements in 3D space,”IFAC-PapersOnLine, vol. 58, no. 15, pp. 468–473, 2024

    2024

  10. [10]

    Improving multi-step prediction of learned time series models,

    A. Venkatraman, M. Hebert, and J. Bagnell, “Improving multi-step prediction of learned time series models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 29, no. 1, 2015

    2015

  11. [11]

    Regularisation of neural networks by enforcing Lipschitz continuity,

    H. Gouk, E. Frank, B. Pfahringer, and M. J. Cree, “Regularisation of neural networks by enforcing Lipschitz continuity,”Machine Learning, vol. 110, no. 2, pp. 393–416, 2021

    2021

  12. [12]

    Some fundamental aspects about Lipschitz continuity of neural networks,

    G. Khromov and S. P. Singh, “Some fundamental aspects about Lipschitz continuity of neural networks,”arXiv:2302.10886, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    A multiple shooting algorithm for direct solution of optimal control problems,

    H. G. Bock and K.-J. Plitt, “A multiple shooting algorithm for direct solution of optimal control problems,”IFAC Proceedings Volumes, vol. 17, no. 2, pp. 1603–1608, 1984

    1984

  14. [14]

    Understanding the difficulty of training deep feedforward neural networks,

    X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” inProceedings of the Thirteenth Inter- national Conference on Artificial Intelligence and Statistics. JMLR Workshop and Conference Proceedings, 2010, pp. 249–256

    2010

  15. [15]

    Three free data sets for development and benchmarking in nonlinear system identification,

    T. Wigren and J. Schoukens, “Three free data sets for development and benchmarking in nonlinear system identification,” in2013 European Control Conference (ECC). IEEE, 2013, pp. 2933–2938

    2013

  16. [16]

    F-16 aircraft benchmark based on ground vibration test data,

    J.-P. No ¨el and M. Schoukens, “F-16 aircraft benchmark based on ground vibration test data,” 4TU.ResearchData,”Dataset, doi: 10.4121/12954911

    doi:10.4121/12954911

  17. [17]

    JAX: composable transformations of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman- Milne, and Q. Zhang, “JAX: composable transformations of Python+NumPy programs,” 2018, https://github.com/jax-ml/jax

    2018

  18. [18]

    Optax: composable gradient transformation and optimisation, in JAX!

    M. Hessel, D. Budden, F. Viola, M. Rosca, E. Sezener, and T. Hen- nigan, “Optax: composable gradient transformation and optimisation, in JAX!” 2020, https://github.com/deepmind/optax

    2020

  19. [19]

    Optimistix: modular optimisation in JAX and Equinox,

    J. Rader, T. Lyons, and P. Kidger, “Optimistix: modular optimisation in JAX and Equinox,”arXiv:2402.09983, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    CasADi: a software framework for nonlinear optimization and opti- mal control,

    J. A. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl, “CasADi: a software framework for nonlinear optimization and opti- mal control,”Mathematical Programming Computation, vol. 11, pp. 1–36, 2019

    2019

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.