Pith. sign in

REVIEW 3 major objections 4 minor 24 references

Efficient streaming dynamic mode decomposition

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper argues that the streaming dynamic mode decomposition can run on a single orthonormal basis instead of two, cutting computation and storage by a constant factor without sacrificing accuracy.

desk verdict The main theorem is false as stated and the pseudocode has a missing assignment, but the single-basis idea is salvageable with a rewrite. read the letter →

arxiv 2507.03770 v1 pith:U7XLJ2TA submitted 2025-07-04 eess.SY cs.SY

Aditya Kale , Marcos Netto , Xinyang Zhou This is my paper · ORCID
classification eess.SYcs.SY
keywords dynamicmodedecompositionKoopmanoperatormodalanalysisnonlineardynamicalsystemsstreamprocessingorthonormalbasisGrammatrixproperorthogonal
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 argues that the streaming dynamic mode decomposition (sDMD) can be reformulated so that only one orthonormal basis is updated online instead of two. The key observation is that in a streaming trajectory each new snapshot pair can be written as (yi-1, yi), so every new direction in the data enters through the y sequence and a single evolving basis can track both snapshot matrices. The authors prove a theorem stating that a basis Qk exists with span(Qk)=span(Xk)=span(Yk) at every step, and use it to build esDMD, which keeps the same Gram-matrix machinery but halves the per-step basis bookkeeping. Numerical tests on a nonlinear oscillator and a hundred-oscillator Kuramoto network show the same dominant eigenvalues and frequencies as sDMD and batch DMD at a lower measured execution time. If correct, this means online modal analysis can run at a constant-factor lower cost with no loss of fidelity.

What carries the argument

The load-bearing object is the single shared orthonormal basis Qk, initialized as QR([x1 y1]) and updated by appending the normalized leftover ek+1 = yk+1 - Qk Qk^T yk+1 when its size exceeds tolerance, then rotated by the top r eigenvectors of the padded Gram matrix G'_{Yk}. The identity that carries the argument is the shift relation yi = xi+1 (equivalently xi = yi-1), which makes every new column of Xk an old column of Yk and thereby lets one basis span both data matrices. The low-dimensional operator takes the form \tilde{A}_k = C_k G_{X_k}^+, with Ck and GXk accumulated from projected snapshot coordinates, so no explicit pseudoinverse of the full data matrix is needed.

What would settle it

Run esDMD on a trajectory where new directions keep entering after the basis has been compressed to r columns, and at each step test whether every column of Xk and Yk lies in span(Qk); if any non-negligible direction falls outside, the claimed equality of the two column spaces with the single basis does not hold.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is Theorem 1: for an autonomous discrete-time system where streaming snapshot pairs satisfy yi = xi+1, there exists a single orthonormal basis Qk with rank at most r such that span(Qk)=span(Xk)=span(Yk) for every time step. The proof builds Qk inductively by Gram-Schmidt, starting from the QR factorization of the first pair [x1 y1], and appending only the component of the next y snapshot that is not already captured; because the next x snapshot equals the previous y snapshot, the authors conclude no separate x-basis is needed. Consequently the two-basis update of sDMD is redundant, and esDMD computes the same reduced operator \tilde{A}_k = C_k G_{X_k}^+ from one Gram matrix, compressed by a proper orthogonal decomposition step when the basis exceeds r columns. The paper presents this as a constant-factor reduction in computational complexity and memory with accuracy matching the two-basis method.

Load-bearing premise

The premise the whole argument stands on is that a single rank-r orthonormal basis can keep representing the column spaces of both the snapshot matrix and its one-step shift throughout the stream, with the discarded directions contributing no accuracy.

Editorial extensions

If this is right

  • Per-step work drops by a constant factor, so the same streaming hardware can track dynamic modes at a higher sampling rate or with more margin for other processing.
  • Memory footprint for the online basis decreases because one Gram matrix and one basis are stored instead of two.
  • The reformulation inherits the same POD compression step, so existing sDMD implementations can be switched to the single-basis version with minimal algorithmic change.
  • Because the reduced operator is identical in form, all downstream uses of sDMD mode estimates, such as prediction, modal analysis, and Koopman-type filters, carry over unchanged.

Reading between the lines

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

  • The equality span(Qk)=span(Xk)=span(Yk) in Theorem 1 is stronger than what the update actually maintains; what the construction guarantees is that span(Qk) continues to contain span(Yk) and, through the shift relation, span(Xk). If only this weaker containment holds, the constant-factor gain survives but the accuracy guarantee becomes approximate and trajectory-dependent.
  • A natural extension is to quantify accuracy loss when new directions appear after compression; one could instrument esDMD to report the residual norm ||yk+1 - Qk Qk^T yk+1|| and stop compressing when it exceeds a threshold.
  • The same single-basis trick should apply to online Koopman estimation and other two-matrix streaming algorithms, since any method maintaining separate bases for X and Y on shift-structured data inherits the redundancy.
  • The measured constant factor depends on the reference implementation; an optimized sDMD could close some of the gap, so future work should benchmark theoretical operation counts rather than static bytecode instructions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes an efficient streaming dynamic mode decomposition (esDMD) that maintains a single orthonormal basis instead of the two bases used in sDMD, claiming a constant-factor reduction in computational cost and memory with no loss of accuracy. The central theoretical result, Theorem 1 in Section III, asserts that one basis Q_k can simultaneously span X_k and Y_k for every time step. The paper also reports numerical comparisons on a nonlinear oscillatory system and a Kuramoto oscillator network, and it provides a code repository.

Significance. If the single-basis claim were correct, the paper would offer a useful simplification of sDMD with concrete constant-factor savings and would be of genuine interest to the streaming-DMD community. The paper is clearly written, identifies a real redundancy in maintaining two Gram matrices for the same stream, and provides reproducible code and experiments on two canonical systems. However, Theorem 1 is false for generic data, and the numerical validation does not quantify accuracy. The central theoretical and empirical support for the 'no loss of accuracy' claim is therefore not established, so the contribution as stated cannot be accepted.

major comments (3)
  1. [Section III, Theorem 1 and Eq. (8)] Theorem 1 is false as stated. For k=1, Q1=QR([x1 y1]) spans the space generated by x1 and y1=x2, whereas span(X1)=span({x1}) and span(Y1)=span({x2}); unless x2 is parallel to x1, no single vector can equal both column spaces. The induction step repeats the error: after appending the new direction p_{k+1} of y_{k+1}=x_{k+2}, the proof claims span(Q_{k+1})=span(X_{k+1})=span(Y_{k+1}), but y_{k+1} is not in X_{k+1}, so the two spans differ in general. The theorem therefore holds only under an unstated assumption that the trajectory lies in a fixed shift-invariant subspace. Since Theorem 1 is the paper's justification for replacing two bases with one and for the 'no loss of accuracy' claim, the central theoretical support of esDMD is invalid.
  2. [Section III-A and Algorithm 1] Algorithm 1 has a stale-coordinate bug in the compression branch. When cols(Q'_i)>r, the basis is rotated to Qi=Q'_i W', but the stored tilde y_- is expressed in the pre-rotation basis; the subsequent update tilde xi <- tilde y_- (which is taken because new_direction is false in that branch) therefore uses coordinates inconsistent with Qi, and the updates to GXi and Ci mix rotated and unrotated quantities. Additionally, the branch cols(Q'_i)<=r uses G''_Xi, G''_Yi, and C''_i before they are assigned, so the pseudocode is undefined in that branch. Even under a valid low-rank assumption, these issues make the algorithm as written non-reproducible.
  3. [Section IV, Figs. 1 and 2] The numerical validation reports only final-time eigenvalue and frequency overlays together with per-iteration execution times. No quantitative accuracy metric (e.g., eigenvalue error, mode amplitude error, or prediction error), no time-resolved accuracy curve, and no repeated-trial statistics for the spectra are provided. The claim that esDMD 'does not compromise accuracy' is therefore not supported even empirically; the two test systems are low-dimensional and may satisfy the shift-invariant subspace condition that Theorem 1 would require.
minor comments (4)
  1. [Section III, Theorem 1 proof] The statement of Theorem 1 uses r both as an unspecified rank in Qk in R^{n x r} and as a bound in the induction hypothesis cols(Qk)<r; the relationship between these two uses is never defined.
  2. [Algorithm 1] The line 'y_parallel_i <- Qi-1 (Q_i^T yi)' appears to be a typo for Q_{i-1} Q_{i-1}^T y_i; as written, the dimensions are inconsistent.
  3. [Algorithm 2] Algorithm 2 calls UPDATE BASIS((xi, yi), i), but Algorithm 1 declares the procedure with the single argument (xi, yi); the calling convention should be made consistent.
  4. [Section IV-C] The static bytecode-instruction counts (approximately 350 versus 515) are implementation-specific and are not a substitute for a precise per-iteration operation count, especially since the dominant costs depend on r and n.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: esDMD is a standard subspace projection reformulation with external benchmarks; the flawed Theorem 1 is a correctness issue, not a circular reduction.

full rationale

The derivation chain is not circular. The low-dimensional operator identity in Eqs. (9)-(11) is a standard Moore-Penrose projection identity: if Xk=Qk tXk and Yk=Qk tYk with orthonormal Qk, then Qk^T Yk Xk^+ Qk = tYk tXk^T (tXk tXk^T)^+. The algorithm constructs Qk to contain both current and shifted snapshots (initial QR([x1,y1]) and appending components of y_{k+1} not already spanned), so the identity is a property of the construction, not a fitted parameter disguised as a prediction. Efficiency claims are benchmarked externally against sDMD [9] and batch DMD, using execution-time measurements and bytecode instruction counts, so they are not forced by the paper's own assumptions. References to the authors' own prior work ([4], [24]) appear only in introductory and future-work contexts and are not load-bearing. The only substantive concern is that Theorem 1 overstates the property: span(Qk)=span(Xk)=span(Yk) does not follow when y_{k+1} introduces a direction outside Xk, since Qk then spans Xk plus that new direction; this is a correctness/rigor gap in the proof, not a circular step. Consequently, the circularity score is 0.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

No new physical entities are introduced. The main burden is the low-rank simultaneous span assumption and the unspecified hyperparameters.

free parameters (2)
  • truncation rank r = 10
    Chosen by hand in experiments; controls basis size and no selection procedure is given.
  • tolerance epsilon = not specified
    Used in the new-direction test in Algorithm 1; its value is not stated, so the behavior of the basis growth is underspecified.
assumptions (2)
  • domain assumption A single rank-r basis can represent span(Xk) and span(Yk) simultaneously without meaningful accuracy loss.
    Used in Theorem 1 and the POD compression step; the attempted proof is false, so this remains an unproved structural assumption about the trajectory.
  • standard math Xk has full row rank, so X_k^+ = X_k^T (X_k X_k^T)^+.
    Invoked in Eq. (9); holds for the experiments with k=1200 snapshots and n=100 states, but is not guaranteed in general streaming applications.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient streaming dynamic mode decomposition." pith.science (2026). https://pith.science/paper/U7XLJ2TA

@misc{pith2026250703770,
  author       = {Pith},
  title        = {Pith review of: Efficient streaming dynamic mode decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7XLJ2TA}},
  note         = {Machine review of arXiv:2507.03770}
}
read the original abstract

We propose a reformulation of the streaming dynamic mode decomposition method that requires maintaining a single orthonormal basis, thereby reducing computational redundancy. The proposed efficient streaming dynamic mode decomposition method results in a constant-factor reduction in computational complexity and memory storage requirements. Numerical experiments on representative canonical dynamical systems show that the enhanced computational efficiency does not compromise the accuracy of the proposed method.

Figures

Figures reproduced from arXiv: 2507.03770 by the authors.

Figure 2
Figure 2. Comparing results obtained with DMD (grey), sDMD [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages

  1. [1]

    J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor, Dynamic mode decomposition: Data-driven modeling of complex systems. SIAM, 2016

    work page 2016

  2. [2]

    Discovering dynamic patterns from infectious disease data using dynamic mode decomposition,

    J. L. Proctor and P. A. Eckhoff, “Discovering dynamic patterns from infectious disease data using dynamic mode decomposition,” Int. Health, vol. 7, no. 2, pp. 139–145, 02 2015

    work page 2015

  3. [3]

    Extract- ing spatial–temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition,

    B. W. Brunton, L. A. Johnson, J. G. Ojemann, and J. N. Kutz, “Extract- ing spatial–temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition,” J. Neurosci. Methods , vol. 258, pp. 1–15, 2016

    work page 2016

  4. [4]

    On analytical construc- tion of observable functions in extended dynamic mode decomposition for nonlinear estimation and prediction,

    M. Netto, Y . Susuki, V . Krishnan, and Y . Zhang, “On analytical construc- tion of observable functions in extended dynamic mode decomposition for nonlinear estimation and prediction,” IEEE Control Syst. Lett., vol. 5, no. 6, pp. 1868–1873, 2021

    work page 2021

  5. [5]

    Dynamic mode decom- position with control,

    J. L. Proctor, S. L. Brunton, and J. N. Kutz, “Dynamic mode decom- position with control,” SIAM J. Appl. Dyn. Syst. , vol. 15, no. 1, pp. 142–161, 2016

    work page 2016

  6. [6]

    Dynamic mode decomposition of numerical and experi- mental data,

    P. J. Schmid, “Dynamic mode decomposition of numerical and experi- mental data,” J. Fluid Mech. , vol. 656, p. 5–28, 2010

    work page 2010

  7. [7]

    The multiverse of dynamic mode decomposition algo- rithms,

    M. J. Colbrook, “The multiverse of dynamic mode decomposition algo- rithms,” in Numerical Analysis Meets Machine Learning , ser. Handbook of Numerical Analysis, S. Mishra and A. Townsend, Eds. Elsevier, 2024, vol. 25, pp. 127–230

    work page 2024

  8. [8]

    Evaluation of stream processing frameworks,

    G. van Dongen and D. Van den Poel, “Evaluation of stream processing frameworks,” IEEE Trans. Parallel Distrib. Syst. , vol. 31, no. 8, pp. 1845–1858, 2020

    work page 2020

Show all 24 references
  1. [9]

    Dynamic mode decomposition for large and streaming datasets,

    M. S. Hemati, M. O. Williams, and C. W. Rowley, “Dynamic mode decomposition for large and streaming datasets,” Phys. Fluids, vol. 26, no. 11, p. 111701, 2014

    2014

  2. [10]

    Streaming dynamic mode decomposition for short-term forecasting in wind farms,

    J. Liew, T. G ¨oc ¸men, W. H. Lio, and G. C. Larsen, “Streaming dynamic mode decomposition for short-term forecasting in wind farms,” Wind Energy, vol. 25, no. 4, pp. 719–734, 2022

    2022

  3. [11]

    Robust dynamic mode decomposition,

    A. Hossein Abolmasoumi, M. Netto, and L. Mili, “Robust dynamic mode decomposition,” IEEE Access, vol. 10, pp. 65 473–65 484, 2022

    2022

  4. [12]

    De- biasing the dynamic mode decomposition for applied Koopman spectral analysis,

    M. S. Hemati, C. W. Rowley, E. A. Deem, and L. N. Cattafesta, “De- biasing the dynamic mode decomposition for applied Koopman spectral analysis,” Theor. Comput. Fluid Dyn., vol. 31, no. 4, pp. 349–368, 2017

    2017

  5. [13]

    Improving separation control with noise-robust variants of dynamic mode decomposition,

    M. Hemati, E. Deem, M. Williams, C. W. Rowley, and L. N. Cattafesta, “Improving separation control with noise-robust variants of dynamic mode decomposition,” in AIAA Aerospace Sciences Meeting , 2016

    2016

  6. [14]

    Incremental singular value decomposition of uncertain data with missing values,

    M. Brand, “Incremental singular value decomposition of uncertain data with missing values,” in Eur. Conf. on Computer Vision, 2002, pp. 707– 720

    2002

  7. [15]

    Large-scale sparse singular value computations,

    M. W. Berry, “Large-scale sparse singular value computations,” Int. J. Supercomput. Appl., vol. 6, no. 1, pp. 13–49, 1992

    1992

  8. [16]

    On-the-fly algorithm for dynamic mode decomposition using incremental singular value decomposition and total least squares,

    D. Matsumoto and T. Indinger, “On-the-fly algorithm for dynamic mode decomposition using incremental singular value decomposition and total least squares,” preprint arXiv:1703.11004, 2017

    2017 arXiv

  9. [17]

    Online dynamic mode decomposition for time-varying systems,

    H. Zhang, C. W. Rowley, E. A. Deem, and L. N. Cattafesta, “Online dynamic mode decomposition for time-varying systems,” SIAM J. Appl. Dyn. Syst., vol. 18, no. 3, pp. 1586–1609, 2019

    2019

  10. [18]

    Online Koopman mode decomposition for power system synchrophasor data,

    Y . Susuki, T. Shimomura, Y . Ota, and A. Ishigame, “Online Koopman mode decomposition for power system synchrophasor data,” IFAC- PapersOnLine, vol. 55, no. 9, pp. 54–58, 2022

    2022

  11. [19]

    An incremental approach to online dy- namic mode decomposition for time-varying systems with applications to EEG data modeling,

    M. Alfatlawi and V . Srivastava, “An incremental approach to online dy- namic mode decomposition for time-varying systems with applications to EEG data modeling,” J. Comput. Dyn. , vol. 7, no. 2, pp. 209–241, 2020

    2020

  12. [20]

    Online dynamic mode decomposition: An alternative approach for low rank datasets,

    G. Nedzhibov, “Online dynamic mode decomposition: An alternative approach for low rank datasets,” Ann. Acad. Rom. Sci. Ser. Math. Appl , vol. 15, pp. 229–249, 2023

    2023

  13. [21]

    On dynamic mode decomposition: Theory and applications,

    J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N. Kutz, “On dynamic mode decomposition: Theory and applications,” J. Comput. Dyn., vol. 1, no. 2, pp. 391–421, 2014

    2014

  14. [22]

    Python implementation of the Kuramoto model,

    F. Damicelli, “Python implementation of the Kuramoto model,” https: //github.com/fabridamicelli/kuramoto, 2019

    2019

  15. [23]

    Projection approximation subspace tracking,

    B. Yang, “Projection approximation subspace tracking,” IEEE Trans. Signal Process., vol. 43, no. 1, pp. 95–107, 1995

    1995

  16. [24]

    A robust data-driven Koopman Kalman filter for power systems dynamic state estimation,

    M. Netto and L. Mili, “A robust data-driven Koopman Kalman filter for power systems dynamic state estimation,” IEEE Trans. Power Syst., vol. 33, no. 6, pp. 7228–7237, 2018

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.