pith. sign in

arxiv: 2606.23957 · v1 · pith:5UFTXBAInew · submitted 2026-06-22 · 💻 cs.LG · cs.SY· eess.SY· q-bio.MN

Learning the Koopman Operator using Attention Free Transformers

Mohammed Nagdi , Evangelos-Marios Nikolados , Alexey Yermakov , Mars Gao , Nathan Kutz , Filippo Menolascina This is my paper

Pith reviewed 2026-06-26 08:47 UTC · model grok-4.3

classification 💻 cs.LG cs.SYeess.SYq-bio.MN
keywords Koopman operatorattention-free transformerlatent memorydynamic re-encodingnonlinear dynamical systemslong-horizon predictionautoencoderserror accumulation
0
0 comments X

The pith

Koopman autoencoders gain stability over long horizons by adding an attention-free latent memory block and change-point re-encoding.

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

The paper sets out to show that Koopman latent predictors can avoid drifting off the learned manifold during extended rollouts by inserting an attention-free block that corrects each new latent using a short history of prior ones before the linear update. This is paired with lightweight online detectors that sense when predictions have wandered and force a fresh projection back through the autoencoder. A reader would care because the usual Koopman approach accumulates phase and amplitude errors on systems that switch, have continuous spectra, or show strong transients, limiting reliable forecasting. The additions keep the model compact, linear-time, and faster at inference than matched multi-head attention while cutting error on the Duffing oscillator, Repressilator, and IRMA benchmarks.

Core claim

An attention-free latent memory block aggregates a short window of past latents to produce a corrected state before each Koopman step, and dynamic re-encoding via EWMA, CUSUM, or two-sample tests detects drift and resets predictions to the autoencoder manifold; together these yield lower accumulated error over horizons up to 1000 steps than plain Koopman autoencoders or capacity-matched multi-head attention models on three benchmark systems while preserving lower inference latency.

What carries the argument

The attention-free latent memory (AFT) block, which aggregates past latents in linear time to correct the input to each Koopman operator update.

If this is right

  • Error accumulation drops consistently across the three benchmark systems over long prediction horizons.
  • Inference latency stays lower than that of matched-capacity multi-head attention models.
  • Gains appear both with and without the optional re-encoding step.
  • Ablations confirm that different trigger policies still deliver the error reduction.

Where Pith is reading between the lines

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

  • The linear-time correction could let the same idea stabilize other latent-space linear predictors where attention cost would be prohibitive.
  • If the detectors work, hybrid linear-nonlinear forecasting pipelines become practical for real-time settings that need to stay on a learned manifold.
  • Applying the triggers to systems dominated by continuous spectra would test whether the current change-point logic generalizes beyond the reported benchmarks.

Load-bearing premise

Lightweight change-point detectors can reliably flag latent drift and trigger re-encoding without creating new phase or amplitude errors on the target systems.

What would settle it

If the 1000-step rollout error on the Duffing oscillator or Repressilator is not lower for the AFT model than for the plain Koopman autoencoder, the central claim does not hold.

Figures

Figures reproduced from arXiv: 2606.23957 by Alexey Yermakov, Evangelos-Marios Nikolados, Filippo Menolascina, Mars Gao, Mohammed Nagdi, Nathan Kutz.

Figure 1
Figure 1. Figure 1: Workflow of the Koopman autoencoder with AFT and Dynamic Re-encoding. (a) Sampled trajectories from a Duffing Oscillator serve as input. (b) The core Koopman autoen￾coder learns a linear latent representation by minimizing reconstruction, linearization, and predic￾tion losses. (c) The prediction process uses a Dynamic Re-encoding module with AFT attention to refine the latent state (zt → z˜t), which is the… view at source ↗
Figure 2
Figure 2. Figure 2: Multi-trajectory rollouts on dynamical systems. AFT reduces phase drift across initial conditions and enables accurate detection of switching dynamics in bistable systems. Duffing Oscillator. AFT achieves the lowest error by a wide margin (MSE 0.0124 vs. 0.0957 / 0.1137 for 10/4-head MHA; ∼ 8–9× lower), and flattens error growth (MCAE 10.95 vs. 49.09 / 52.98; ∼ 4.5–4.8× lower). This matches the intuition t… view at source ↗
Figure 3
Figure 3. Figure 3: AFT vs. MHA (4 and 10 heads) on the three primary systems. The top row shows sampled trajectories, and the bottom row shows the MCAE curves. AFT reduces error growth and outperforms matched-capacity MHA. 4.3 EFFECT OF DYNAMIC RE-ENCODING (STREAMING TRIGGERS) We study the dynamic re-encoding with streaming triggers on the Duffing oscillator (§2.3). Ta￾ble 3 shows that the sequential two-sample detector atta… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of re-encoding methods. (a) Cumulative MAE per timestep comparison [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: AFT vs. plain KAE on the Repressilator over a 1000-step rollout (two representative trajectories, first observable). AFT (red) stays phase-locked to the reference (blue) across the full horizon, whereas the plain KAE (green) accumulates phase and amplitude drift, illustrating the long￾horizon stability gained from the latent memory. (a) Repressilator — 3D prediction (b) IRMA — 3D prediction (c) Duffing — m… view at source ↗
Figure 6
Figure 6. Figure 6: Phase Plane Visualization of the systems. Dynamic re-encoding prevents rare-but￾catastrophic divergence on long rollouts and provides robust trajectory prediction across different initial conditions. the AFT-only curve remains lowest and most stable, consistent with [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean cumulative absolute error (MCAE) results for our three dynamical systems, comple [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional Dynamical systems. AFT (and AFT+Re-enc where helpful) improves or matches the baseline across diverse regimes. We did not perform additional, extensive per-system tuning. 0 200 400 600 800 1000 Timestep 10 2 10 1 10 0 10 1 10 2 Average Cumulative Absolute Error (Log Scale) Prediction Performance Across Models Koopman AFT 10 Koopman AFT 20 Koopman AFT 30 Koopman AFT 50 [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 9
Figure 9. Figure 9: Ablation studies on Koopman and AFT parameters. Left: AFT robustness vs. Koop￾man with different operator sizes on Repressilator. Dense K achieves the best accuracy; constrained forms need larger widths for parity. Right: AFT with different context lengths. Small context length enable learning temporal changes while longer context might lead to noise updates. and we use full-horizon weighting in equation 2… view at source ↗
Figure 10
Figure 10. Figure 10: Trajectory segmentation for training For complex dynamical systems exhibiting chaotic behavior, switching dynamics, or continuous spectra, we employ shorter prediction lengths during training, as this approach yields better perfor￾mance and more stable training dynamics. The model unrolls predictions from the initial condition x0 across the specified prediction horizon, computing both latent space predict… view at source ↗
read the original abstract

Learning Koopman operators with autoencoders enables linear prediction in a latent space, but long-horizon rollouts often drift off the learned manifold, leading to phase and amplitude errors on systems with switching, continuous spectra, or strong transients. We introduce two complementary components that make Koopman predictors more robust. First, we add an attention-free latent memory (AFT) block that aggregates a short window of past latents to produce a corrected latent before each Koopman update. Unlike multi-head attention, AFT operates in linear time and adds only $\approx$30k parameters ($3d^2 + T^2$, fewer than matched multi-head attention), yet captures the local temporal context needed to suppress error divergence. Second, we propose dynamic re-encoding: lightweight, online change-point triggers (EWMA, CUSUM, and sequential two-sample tests) that detect latent drift and project predictions back onto the autoencoder manifold. Across three benchmark systems -- Duffing oscillator, Repressilator, IRMA -- our model consistently reduces error accumulation compared to a Koopman autoencoder and matched-capacity multi-head attention. We also compare against GRU and Transformer autoencoders, evaluated both from initial conditions and with a 50-step context, and find that Koopman+AFT (with optional re-encoding) attains markedly lower long-horizon error while maintaining lower inference latency. We report improvements over horizons up to 1000 steps, together with ablations over trigger policies. The result is a fast, compact predictor that stays on the learned manifold over long horizons.

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 / 2 minor

Summary. The paper proposes two additions to Koopman autoencoders for improved long-horizon prediction: (1) an Attention-Free Transformer (AFT) block that aggregates a short window of past latents in linear time with ~30k parameters to correct the latent state before each Koopman step, and (2) optional dynamic re-encoding triggered by lightweight online change-point detectors (EWMA, CUSUM, sequential two-sample tests) to project predictions back onto the learned manifold. Experiments on the Duffing oscillator, Repressilator, and IRMA systems report lower error accumulation than a baseline Koopman autoencoder, matched-capacity multi-head attention, GRU, and Transformer autoencoders, with ablations on trigger policies and evaluations from both initial conditions and 50-step context, up to 1000-step horizons.

Significance. If the reported error reductions and latency advantages are confirmed by detailed quantitative results, the work offers a compact, linear-time mechanism for mitigating manifold drift in Koopman predictors. The explicit parameter count (3d² + T²) and optional re-encoding design are falsifiable strengths that could make the method attractive for resource-constrained forecasting or control tasks involving switching or transient dynamics.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'consistent' and 'markedly lower' error reduction across three benchmarks is stated without any numerical values, tables, error bars, or statistical tests. The manuscript must supply these quantitative results (including per-system, per-horizon metrics and ablation tables) to support the empirical contribution.
  2. [Method (dynamic re-encoding)] The reliability of the change-point detectors for triggering re-encoding without introducing phase/amplitude errors is load-bearing for the second component. The manuscript should provide explicit analysis or ablation results demonstrating that EWMA/CUSUM/sequential tests do not degrade prediction quality on the target systems when triggered.
minor comments (2)
  1. Clarify whether the reported ~30k parameter count for AFT includes the full encoder/decoder or only the AFT block, and confirm the matched-capacity multi-head attention baseline uses identical total parameters.
  2. The abstract mentions comparisons 'evaluated both from initial conditions and with a 50-step context'; the manuscript should explicitly state the context length used for all baselines to ensure fair comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to incorporate the requested quantitative details and expanded analysis.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'consistent' and 'markedly lower' error reduction across three benchmarks is stated without any numerical values, tables, error bars, or statistical tests. The manuscript must supply these quantitative results (including per-system, per-horizon metrics and ablation tables) to support the empirical contribution.

    Authors: We agree that the abstract would be strengthened by including concrete numerical support for the claims. In the revised manuscript we will add key quantitative results (per-system and per-horizon error reductions with error bars) and explicit references to the ablation tables already present in the experiments section. revision: yes

  2. Referee: [Method (dynamic re-encoding)] The reliability of the change-point detectors for triggering re-encoding without introducing phase/amplitude errors is load-bearing for the second component. The manuscript should provide explicit analysis or ablation results demonstrating that EWMA/CUSUM/sequential tests do not degrade prediction quality on the target systems when triggered.

    Authors: The current manuscript already reports ablations over trigger policies. To directly address the concern, we will expand the analysis in the revision with explicit side-by-side comparisons of long-horizon prediction error (including phase and amplitude metrics) on Duffing, Repressilator, and IRMA when the detectors are active versus inactive, confirming that triggering does not introduce degradation. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper introduces an empirical architecture (Koopman autoencoder augmented with an AFT block for latent aggregation and optional online change-point triggers for re-encoding) and evaluates it on three benchmark dynamical systems against matched baselines. No derivation chain, equation, or central claim reduces by construction to fitted inputs, self-citations, or renamed known results; performance improvements are presented as falsifiable experimental outcomes with explicit ablations and latency comparisons. The method description remains independent of the reported error metrics.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities beyond the standard autoencoder and Koopman assumptions already present in the cited literature.

pith-pipeline@v0.9.1-grok · 5839 in / 1134 out tokens · 17185 ms · 2026-06-26T08:47:19.785707+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

70 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    , title =

    Koopman, Bernard O. , title =. Proceedings of the National Academy of Sciences , year =. doi:10.1073/pnas.17.5.315 , url =

    work page doi:10.1073/pnas.17.5.315

  2. [2]

    and Mezi

    Rowley, Clarence W. and Mezi. Spectral analysis of nonlinear flows , journal =. 2009 , volume =. doi:10.1017/S0022112009992059 , url =

    work page doi:10.1017/s0022112009992059 2009

  3. [3]

    Journal of Fluid Mechanics656, 5–28 (2010)

    Schmid, Peter J. , title =. Journal of Fluid Mechanics , year =. doi:10.1017/S0022112010001217 , url =

    work page doi:10.1017/s0022112010001217

  4. [4]

    Applied Koopmanism , journal =

    Budi. Applied Koopmanism , journal =. 2012 , volume =. doi:10.1063/1.4772195 , url =

    work page doi:10.1063/1.4772195 2012

  5. [5]

    Analysis of Fluid Flows via Spectral Properties of the Koopman Operator , journal =

    Mezi. Analysis of Fluid Flows via Spectral Properties of the Koopman Operator , journal =. 2013 , volume =. doi:10.1146/annurev-fluid-120710-101204 , url =

    work page doi:10.1146/annurev-fluid-120710-101204 2013

  6. [6]

    and Brunton, Steven L

    Proctor, Joshua L. and Brunton, Steven L. and Kutz, J. Nathan , title =. SIAM Journal on Applied Dynamical Systems , year =. doi:10.1137/15M1013857 , url =

    work page doi:10.1137/15m1013857

  7. [7]

    Linear Predictors for Nonlinear Dynamical Systems: Koopman Operator Meets Model Predictive Control , journal =

    Korda, Milan and Mezi. Linear Predictors for Nonlinear Dynamical Systems: Koopman Operator Meets Model Predictive Control , journal =. 2018 , volume =. doi:10.1016/j.automatica.2018.03.046 , url =

    work page doi:10.1016/j.automatica.2018.03.046 2018

  8. [8]

    Global Stability Analysis Using the Eigenfunctions of the Koopman Operator , journal =

    Mauroy, Alexandre and Mezi. Global Stability Analysis Using the Eigenfunctions of the Koopman Operator , journal =. 2016 , volume =. doi:10.1109/TAC.2016.2518918 , url =

    work page doi:10.1109/tac.2016.2518918 2016

  9. [9]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  10. [10]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  11. [11]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  12. [12]

    and Kevrekidis, Ioannis G

    Dietrich, Felix and Thiem, Thomas N. and Kevrekidis, Ioannis G. On the Koopman Operator of Algorithms. SIAM Journal on Applied Dynamical Systems. 2020. doi:10.1137/19m1277059

    work page doi:10.1137/19m1277059 2020

  13. [13]

    and Potomkin, Kostiantyn

    Bak, Stanley and Bogomolov, Sergiy and Duggirala, Parasara Sridhar and Gerlach, Adam R. and Potomkin, Kostiantyn. Reachability of Black-Box Nonlinear Systems after Koopman Operator Linearization. ArXiv. 2021. doi:10.48550/arxiv.2105.00886

    work page doi:10.48550/arxiv.2105.00886 2021

  14. [14]

    and Townsend, Alex

    Colbrook, Matthew J. and Townsend, Alex. Rigorous data‐driven computation of spectral properties of Koopman operators for dynamical systems. Communications on Pure and Applied Mathematics. 2024. doi:10.1002/cpa.22125

    work page doi:10.1002/cpa.22125 2024

  15. [15]

    Journal of Nonlinear Science , author =

    Korda, Milan and Mezić, Igor. On Convergence of Extended Dynamic Mode Decomposition to the Koopman Operator. Journal of Nonlinear Science. 2018. doi:10.1007/s00332-017-9423-0

    work page doi:10.1007/s00332-017-9423-0 2018

  16. [18]

    Koopman Operators in Robot Learning

    Shi, Lu and Haseli, Masih and Mamakoukas, Giorgos and Bruder, Daniel and Abraham, Ian and Murphey, Todd and Cortes, Jorge and Karydis, Konstantinos. Koopman Operators in Robot Learning. ArXiv. 2024. doi:10.48550/arxiv.2408.04200

    work page doi:10.48550/arxiv.2408.04200 2024

  17. [19]

    and Kutz, J

    Kamb, Mason and Kaiser, Eurika and Brunton, Steven L. and Kutz, J. Nathan. Time-Delay Observables for Koopman: Theory and Applications. SIAM Journal on Applied Dynamical Systems. 2020. doi:10.1137/18m1216572

    work page doi:10.1137/18m1216572 2020

  18. [20]

    On Numerical Approximations of the Koopman Operator

    Mezić, Igor. On Numerical Approximations of the Koopman Operator. Mathematics. 2022. doi:10.3390/math10071180

    work page doi:10.3390/math10071180 2022

  19. [21]

    Consistent spectral approximation of Koopman operators using resolvent compactification

    Giannakis, Dimitrios and Valva, Claire. Consistent spectral approximation of Koopman operators using resolvent compactification. Nonlinearity. 2024. doi:10.1088/1361-6544/ad4ade

    work page doi:10.1088/1361-6544/ad4ade 2024

  20. [22]

    Data-Driven Nonlinear Stabilization Using Koopman Operator

    Huang, Bowen and Ma, Xu and Vaidya, Umesh. Data-Driven Nonlinear Stabilization Using Koopman Operator. Lecture Notes in Control and Information Sciences. 2020. doi:10.1007/978-3-030-35713-9\_12

    work page doi:10.1007/978-3-030-35713-9 2020

  21. [23]

    and Brunton, Steven L

    Proctor, Joshua L. and Brunton, Steven L. and Kutz, J. Nathan. Generalizing Koopman Theory to Allow for Inputs and Control. SIAM Journal on Applied Dynamical Systems. 2018. doi:10.1137/16m1062296

    work page doi:10.1137/16m1062296 2018

  22. [24]

    Learning Deep Neural Network Representations for Koopman Operators of Nonlinear Dynamical Systems

    Yeung, Enoch and Kundu, Soumya and Hodas, Nathan Oken. Learning Deep Neural Network Representations for Koopman Operators of Nonlinear Dynamical Systems. 2019 American Control Conference (ACC). 2019. doi:10.23919/acc.2019.8815339

    work page doi:10.23919/acc.2019.8815339 2019

  23. [25]

    Estimating Koopman operators for nonlinear dynamical systems: a nonparametric approach

    Zanini, Francesco and Chiuso, A. Estimating Koopman operators for nonlinear dynamical systems: a nonparametric approach. ArXiv. 2021. doi:10.48550/arxiv.2103.13752

    work page doi:10.48550/arxiv.2103.13752 2021

  24. [26]

    Koopman Operator Inspired Nonlinear System Identification

    Wilson, Dan. Koopman Operator Inspired Nonlinear System Identification. SIAM Journal on Applied Dynamical Systems. 2023. doi:10.1137/22m1512272

    work page doi:10.1137/22m1512272 2023

  25. [27]

    and de Silva, Brian M

    Pan, Shaowu and Kaiser, E. and de Silva, Brian M. and Kutz, J. and Brunton, S. PyKoopman: A Python Package for Data-Driven Approximation of the Koopman Operator. ArXiv. 2023. doi:10.48550/arxiv.2306.12962

    work page doi:10.48550/arxiv.2306.12962 2023

  26. [28]

    On Computation of Koopman Operator from Sparse Data

    Sinha, S. and Vaidya, U. and Yeung, Enoch. On Computation of Koopman Operator from Sparse Data. 2019 American Control Conference (ACC). 2019. doi:10.48550/arxiv.1901.03024

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1901.03024 2019

  27. [29]

    and D’Souza, Raissa M

    Salova, Anastasiya and Emenheiser, Jeffrey and Rupe, Adam and Crutchfield, James P. and D’Souza, Raissa M. Koopman operator and its approximations for systems with symmetries. Chaos: An Interdisciplinary Journal of Nonlinear Science. 2019. doi:10.1063/1.5099091

    work page doi:10.1063/1.5099091 2019

  28. [30]

    Nature , volume=

    A synthetic oscillatory network of transcriptional regulators , author=. Nature , volume=. 2000 , publisher=

    2000

  29. [31]

    2019 IEEE 58th Conference on Decision and Control (CDC) , pages=

    Koopman operators for generalized persistence of excitation conditions for nonlinear systems , author=. 2019 IEEE 58th Conference on Decision and Control (CDC) , pages=. 2019 , organization=

    2019

  30. [32]

    Automatica , volume=

    Optimal control formulation of pulse-based control using Koopman operator , author=. Automatica , volume=. 2018 , publisher=

    2018

  31. [33]

    arXiv preprint arXiv:2210.09343 , year=

    Data-driven observability decomposition with koopman operators for optimization of output functions of nonlinear systems , author=. arXiv preprint arXiv:2210.09343 , year=

    arXiv

  32. [34]

    Cell systems , volume=

    Combining a toggle switch and a repressilator within the AC-DC circuit generates distinct dynamical behaviors , author=. Cell systems , volume=. 2018 , publisher=

    2018

  33. [35]

    SIAM journal on applied dynamical systems , volume=

    -error Bounds for Approximations of the Koopman Operator by Kernel Extended Dynamic Mode Decomposition , author=. SIAM journal on applied dynamical systems , volume=. 2025 , publisher=

    2025

  34. [36]

    SIAM Journal on Applied Dynamical Systems , volume=

    Physics-informed probabilistic learning of linear embeddings of nonlinear dynamics with guaranteed stability , author=. SIAM Journal on Applied Dynamical Systems , volume=. 2020 , publisher=

    2020

  35. [37]

    Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=

    Deep learning enhanced dynamic mode decomposition , author=. Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=. 2022 , publisher=

    2022

  36. [38]

    SIAM Journal on Applied Dynamical Systems , volume=

    Linearly recurrent autoencoder networks for learning dynamics , author=. SIAM Journal on Applied Dynamical Systems , volume=. 2019 , publisher=

    2019

  37. [39]

    Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=

    Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator , author=. Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=. 2017 , publisher=

    2017

  38. [40]

    PLoS computational biology , volume=

    In-vivo real-time control of protein expression from endogenous and synthetic gene networks , author=. PLoS computational biology , volume=. 2014 , publisher=

    2014

  39. [41]

    Synthetic Gene Networks: Methods and Protocols , pages=

    Predicting synthetic gene networks , author=. Synthetic Gene Networks: Methods and Protocols , pages=. 2011 , publisher=

    2011

  40. [42]

    PloS one , volume=

    How to turn a genetic circuit into a synthetic tunable oscillator, or a bistable switch , author=. PloS one , volume=. 2009 , publisher=

    2009

  41. [43]

    Cell , volume=

    A yeast synthetic network for in vivo assessment of reverse-engineering and modeling approaches , author=. Cell , volume=. 2009 , publisher=

    2009

  42. [44]

    2019 IEEE Biomedical Circuits and Systems Conference (BioCAS) , pages=

    A data-driven method for quantifying the impact of a genetic circuit on its host , author=. 2019 IEEE Biomedical Circuits and Systems Conference (BioCAS) , pages=. 2019 , organization=

    2019

  43. [45]

    Nature communications , volume=

    Deep learning for universal linear embeddings of nonlinear dynamics , author=. Nature communications , volume=. 2018 , publisher=

    2018

  44. [46]

    Advances in enzyme regulation , volume=

    Oscillatory behavior in enzymatic control processes , author=. Advances in enzyme regulation , volume=. 1965 , publisher=

    1965

  45. [47]

    Journal of Fluid Mechanics , volume=

    A hierarchy of low-dimensional models for the transient and post-transient cylinder wake , author=. Journal of Fluid Mechanics , volume=. 2003 , publisher=

    2003

  46. [48]

    Neural computation , volume=

    Long short-term memory , author=. Neural computation , volume=. 1997 , publisher=

    1997

  47. [49]

    Cognitive science , volume=

    Finding structure in time , author=. Cognitive science , volume=. 1990 , publisher=

    1990

  48. [50]

    arXiv preprint arXiv:1406.1078 , year=

    Learning phrase representations using RNN encoder-decoder for statistical machine translation , author=. arXiv preprint arXiv:1406.1078 , year=

    Pith/arXiv arXiv

  49. [51]

    the Annals of Statistics , volume=

    Optimal stopping times for detecting changes in distributions , author=. the Annals of Statistics , volume=. 1986 , publisher=

    1986

  50. [52]

    Technometrics , volume=

    Control chart tests based on geometric moving averages , author=. Technometrics , volume=. 2000 , publisher=

    2000

  51. [53]

    Journal of Quality Technology , volume=

    Two nonparametric control charts for detecting arbitrary distribution changes , author=. Journal of Quality Technology , volume=. 2012 , publisher=

    2012

  52. [54]

    2019 American Control Conference (ACC) , pages=

    Learning deep neural network representations for Koopman operators of nonlinear dynamical systems , author=. 2019 American Control Conference (ACC) , pages=. 2019 , organization=

    2019

  53. [55]

    Journal of Nonlinear Science , volume=

    A data--driven approximation of the koopman operator: Extending dynamic mode decomposition , author=. Journal of Nonlinear Science , volume=. 2015 , publisher=

    2015

  54. [56]

    arXiv preprint arXiv:2105.14103 , year=

    An attention free transformer , author=. arXiv preprint arXiv:2105.14103 , year=

    arXiv

  55. [57]

    arXiv preprint arXiv:2412.04578 , year=

    Loss Terms and Operator Forms of Koopman Autoencoders , author=. arXiv preprint arXiv:2412.04578 , year=

    arXiv

  56. [58]

    arXiv preprint arXiv:2310.15386 , year=

    Course correcting Koopman representations , author=. arXiv preprint arXiv:2310.15386 , year=

    arXiv

  57. [59]

    Physics Letters A , volume=

    An equation for continuous chaos , author=. Physics Letters A , volume=. 1976 , publisher=

    1976

  58. [60]

    bioRxiv , pages=

    DeepMapper: attention-based autoencoder for system identification in wound healing and stage prediction , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  59. [61]

    Proceedings of the Royal Society A , volume=

    Mori--Zwanzig latent space Koopman closure for nonlinear autoencoder , author=. Proceedings of the Royal Society A , volume=. 2025 , publisher=

    2025

  60. [62]

    arXiv preprint arXiv:2210.03675 , year=

    Koopman neural forecaster for time series with temporal distribution shifts , author=. arXiv preprint arXiv:2210.03675 , year=

    arXiv

  61. [63]

    arXiv preprint arXiv:2406.12062 , year=

    Entropic Regression DMD (ERDMD) Discovers Informative Sparse and Nonuniformly Time Delayed Models , author=. arXiv preprint arXiv:2406.12062 , year=

    arXiv

  62. [64]

    SIAM Journal on Applied Dynamical Systems , volume=

    Ergodic theory, dynamic mode decomposition, and computation of spectral properties of the Koopman operator , author=. SIAM Journal on Applied Dynamical Systems , volume=. 2017 , publisher=

    2017

  63. [65]

    Nature communications , volume=

    Chaos as an intermittently forced linear system , author=. Nature communications , volume=. 2017 , publisher=

    2017

  64. [66]

    Scientific Reports , volume=

    Temporally-consistent koopman autoencoders for forecasting dynamical systems , author=. Scientific Reports , volume=. 2025 , publisher=

    2025

  65. [67]

    arXiv preprint arXiv:2503.12930 , year=

    Augmented Invertible Koopman Autoencoder for long-term time series forecasting , author=. arXiv preprint arXiv:2503.12930 , year=

    arXiv

  66. [68]

    Measurement , volume=

    Output-only modal identification with recursive dynamic mode decomposition for time-varying systems , author=. Measurement , volume=. 2024 , publisher=

    2024

  67. [69]

    arXiv preprint arXiv:1511.06876 , year=

    Recursive dynamic mode decomposition of a transient cylinder wake , author=. arXiv preprint arXiv:1511.06876 , year=

    Pith/arXiv arXiv

  68. [70]

    Physical Review E , volume=

    Dynamic mode decomposition for multiscale nonlinear physics , author=. Physical Review E , volume=. 2019 , publisher=

    2019

  69. [71]

    ACS synthetic biology , volume=

    Growth defects and loss-of-function in synthetic gene circuits , author=. ACS synthetic biology , volume=. 2019 , publisher=

    2019

  70. [72]

    Proceedings of the National Academy of Sciences , volume=

    Mechanistic links between cellular trade-offs, gene expression, and growth , author=. Proceedings of the National Academy of Sciences , volume=. 2015 , publisher=

    2015