Tensor-Train Weak SINDy: Identifying High-Dimensional Nonlinear Dynamics
Will Houser, Vanja Dukic, David M. Bortz
cs.LG · cs.CE · stat.ML
Quick verdict
Proposes TT-WSINDy, merging tensor-train MANDy with weak-form WSINDy; theory is sound and code public, but gains shown only to D=12 with modest 2-7x speedups.
Summary
The paper introduces TT-WSINDy, a method for identifying high-dimensional ordinary differential equations that couples the tensor-train (TT) feature construction of MANDy with the weak-form regression and MSTLS sparsification of WSINDy. The test function is folded in as an extra TT core, so weak-form transformation, pseudoinverse regression, and thresholding all operate without materializing the J^D-sized candidate library. The authors prove the construction solves the same weak regression problem as flat WSINDy (Proposition 4.6) and show TT-STLS runs in O((JD)²M³) time and O(JDM²) memory — polynomial in library parameters rather than exponential. A rank-reduced construction (Algorithm 5.2) with randomized SVD further cuts storage an order of magnitude (≈1.0·10⁹ to ≈1.1·10⁸ entries, Table 5.1). On four ODE systems (FPUT, Lorenz 96, Kuramoto, Chua), the weak form outperforms strong-form TT regression across the [0,1] noise-ratio continuum over 40 realizations per level. Scaling Lorenz 96 to D=12 with M=20000, TT-WSINDy runs 2–7× faster than flat WSINDy for D=6–12 (flat wins at D=5) with ~10⁻⁴ relative coefficient error, though the coarse pass at D=12 retains ~2000 spurious terms. Limitations: demonstrations stop at D=12 ODEs, speedups are modest, no statistical testing, and no guarantee the coarse screening avoids false negatives.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- TT-WSINDy solves the same weak-form sparse regression problem as flat WSINDy while never forming the J^D-sized feature matrix (Proposition 4.6).
- TT-STLS runs in O((JD)²M³) time and O(JDM²) memory, polynomial rather than exponential in the library parameters (Proposition 5.4).
- The weak-form TT regression is more robust to noise than strong-form TT regression across noise ratios in [0,1] on four ODE systems (Section 6.1).
- The rank-reduced construction (Algorithm 5.2) with randomized SVD reduces feature-tensor storage by roughly an order of magnitude and handles 2·10⁵-snapshot datasets (Table 5.1, Section 6.1).
- On Lorenz 96 up to D=12, TT-WSINDy is 2–7× faster than flat WSINDy for D=6–12 while recovering identical supports and ~10⁻⁴ relative coefficient error (Section 6.2).
Strengths
Rigorous equivalence to flat WSINDy
Proposition 4.6 (via Lemmas 4.2–4.4) proves the TT weak construction solves the identical weak regression problem — in the original-rank regime this is a reformulation, not an approximation.
Polynomial complexity with honest scoping
Proposition 5.4 gives O((JD)²M³) time / O(JDM²) memory for TT-STLS; Remarks 5.6 and 5.9 explicitly state where flat WSINDy wins (M ≳ JD) and where the advantage vanishes (dense true support).
Novel coarse-support machinery
Low-rank construction enables dense sampling
Transparent empirical reporting
Weaknesses
Limited scale of the headline claim
Modest empirical gains
No false-negative guarantee for screening
Weak strong-form baseline
Thin statistical reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TT-WSINDy (arXiv 2609.09434) targets ODE system identification when the candidate library grows as J^D — all products of J univariate basis functions across D state dimensions (Eq. 3.1). The method folds the WSINDy test function into a tensor train as an extra core (Eqs. 3.4–3.5), so weak-form transformation, pseudoinverse regression, and thresholding run without materializing the J^D-column feature matrix. Three central claims: (1) the TT construction solves exactly the same weak regression problem as flat WSINDy (Proposition 4.6); (2) TT-STLS has polynomial complexity — O((JD)²M³) time, O(JDM²) memory (Proposition 5.4) — subverting the curse of dimensionality; (3) weak-form TT regression is noise-robust and TT-WSINDy beats flat WSINDy in walltime at sufficient dimension. Claims 1–2 are well supported; claim 3 is demonstrated only modestly.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TT-WSINDy in simple terms?
It is a method that learns the equations of a dynamical system from noisy time-series data. It combines two earlier ideas: WSINDy, which avoids differentiating noisy data by using an integral (weak) formulation, and MANDy, which uses tensor-train compression so the method can consider all products of basis functions across dimensions without running out of memory.
Does TT-WSINDy require derivative data?
No. The weak formulation integrates the equations against a smooth test function and applies integration by parts, so state derivatives are never estimated — a key advantage for noisy data (Section 2.2). For second-order systems like FPUT, the data are convolved with the test function's second derivative instead.
How much faster is TT-WSINDy than standard WSINDy?
On the Lorenz 96 benchmark scaled from D=5 to D=12 dimensions with 20,000 snapshots, TT-WSINDy was 2–7 times faster for D=6–12; at D=5 the standard flat solver was faster. The advantage grows with the size of the candidate library (J^D terms per equation) and the paper notes flat methods still win when snapshots outnumber the library (M ≳ JD).
Is the code available?
Yes. The authors state the code is publicly available at github.com/whouser2001/TT-WSINDy, and all ODE systems, parameters, basis functions, and test-function settings are tabulated in the paper (Tables 6.1–6.2, Eq. 6.3).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 5.1, 6.1–6.3, and B.1 are accessible only through captions and prose; exact curve values (per-D walltimes, error-quartile levels) could not be extracted, so quantitative judgments about them rest on the text's stated summaries.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.