Pith. sign in

REVIEW 5 major objections 4 minor 18 references

Tensor-Train Operator Inference

T0 review · 5 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Tensor-train compressed operator inference matches standard accuracy while cutting memory, the paper claims, with algorithms for pseudoinverses in TT format and validation on heat, Burgers, and cylinder flow.

desk verdict Plausible tensorized OpInf with a real bottleneck target, but the compressed pseudoinverse at the core of the paper is unproved as printed and the headline memory claim is unmeasured; worth referee time, not desk reject. read the letter →

arxiv 2509.08071 v1 pith:NXUQOICN submitted 2025-09-09 math.NA cs.NA

classification math.NAcs.NA MSC 15A6965F20
keywords tensor-traindecompositionoperatorinferencemodelreductionpseudoinverseTT-crossinterpolationquantizeddata-drivendynamicsCFDbenchmarks
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

This paper sets out to make operator inference—learning polynomial dynamical-system operators from snapshots—work entirely in the tensor-train (TT) compressed format. The central move is a compressed least-squares solve: compute a regularized pseudoinverse of the data tensor and contract it against the derivative tensor, all in TT cores. Three variants are proposed: full-order TT, quantized TT for unstructured grids, and a reduced-order version where the TT decomposition itself replaces the POD/SVD projection. On the heat equation, viscous Burgers, and laminar cylinder flow, the compressed methods match standard OpInf accuracy while storing far less and scaling to larger data. The load-bearing identity is that the core-wise SVD of the TT representation yields the pseudoinverse of the matricized data tensor.

What carries the argument

The central object is the tensor-train decomposition, in which a high-dimensional tensor is stored as a chain of low-rank cores. The key algorithm, tt_pinv, orthogonalizes the TT cores from both sides, performs a compact SVD on the product of the two interface factors, applies Tikhonov regularization to the singular values, and reassembles a transposed TT representation of the pseudoinverse. A second algorithm contracts this pseudoinverse with the right-hand-side tensor to form the learned operator. Snapshot compression uses TT-SVD or TT-cross interpolation, and the QTT variant uses prime-factorization reshaping to impose tensor structure on unstructured-mesh data.

What would settle it

Take a small random tensor with a known TT representation, compute its pseudoinverse by Algorithm 1, and compute the dense SVD pseudoinverse of the matricized tensor; the entrywise relative difference should be near machine precision. If the two disagree, or if the TT transpose steps produce rank blow-up, the central identity is refuted.

Watch

Extended reading notes

Core claim

The paper claims that the least-squares operator inference problem can be solved without ever expanding the data into dense matrices. Given a snapshot tensor X, the method builds a data tensor D containing linear and quadratic terms, computes D† in TT form via left/right orthogonalization of cores, an SVD of the interface core, regularization of singular values, and TT-mode transposition, then contracts D† with the derivative tensor R to obtain the operator tensor O. The learned O is sliced to extract the linear operator A and quadratic operator F, and predictions are made by time-integrating the tensorized ODE directly in TT format. If this works, full-order and reduced-order polynomial ope

Load-bearing premise

The compressed pseudoinverse built from core-wise orthogonalization and an interface SVD exactly equals the pseudoinverse of the full data tensor, with no uncontrolled rank growth during the TT transposes.

Editorial extensions

If this is right

  • Full-order TT and QTT OpInf can learn full discretization operators for high-dimensional grids where dense SVD or Kronecker-product formation is infeasible.
  • Reduced-order TT-ROM obtains its reduced coordinates directly from a TT core, removing the POD SVD projection step.
  • With TT-cross compression, operator learning can proceed without ever loading the full snapshot dataset into memory.
  • Memory and per-operation cost scale with TT ranks rather than with the full grid size, extending OpInf to much larger snapshot tensors.
  • Regularization can be applied within the compressed pseudoinverse, so ill-conditioned data can be handled without leaving the TT format.

Reading between the lines

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

  • If the TT pseudoinverse identity holds generally, the same compressed least-squares machinery could serve other tensor regression and inverse problems, not only operator inference.
  • The success of TT-cross for snapshot compression suggests a possible streaming or online variant of OpInf where snapshots are sampled adaptively rather than stored wholesale; the paper demonstrates only the offline setting.
  • The practical ceiling is likely TT-rank growth during time integration, so testing rank-controlled integrators beyond the adaptive Runge–Kutta scheme used here could extend prediction horizons.
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

5 major / 4 minor

Summary. The paper proposes a tensor-train (TT) framework for nonintrusive operator inference (OpInf), with three variants: full-order TT-OpInf, full-order quantized TT (QTT)-OpInf, and reduced-order TT-ROM. The central idea is to represent the snapshot data as a TT tensor and solve the regularized least-squares problem for the polynomial operators directly in compressed form, avoiding the formation of dense snapshot and operator matrices. The key claimed novelty is an algorithm for computing a regularized pseudoinverse of a matricized TT tensor and for contracting it with the right-hand side to construct the learned operator in TT format. Numerical experiments on the 2D heat equation, 2D viscous Burgers' equation, and compressible laminar flow over a cylinder compare the TT methods against standard ROM-OpInf and a full-tensor formulation, reporting timings, ranks, and (for the first two problems) relative errors.

Significance. If the TT pseudoinverse and contraction algorithms are correct and the scalability claim holds, this would be a valuable contribution: it would extend operator inference to settings where the state dimension is too large for dense matrix operations, and the use of TT-cross interpolation to avoid forming full snapshot tensors is an attractive idea. The paper is clearly written in its problem statement and experiments are presented for three nontrivial benchmarks. However, the central algorithmic steps are not proven or independently verified, and the numerical evidence for the headline scalability and memory advantages is incomplete. The significance is therefore conditional on closing these gaps.

major comments (5)
  1. [II.D.3, Algorithms 1 and 2] The central novelty is the claim that tt_pinv computes the regularized pseudoinverse of the matricized TT tensor entirely in compressed form. This is not proven, and the algorithm as stated is underspecified. Steps 8–9 say to 'transpose by exchanging input and output dimensions in its mode ordering'; mode permutation is not a local TT operation in general and can cause rank growth, yet no rounding is applied. The input tolerance tt_tol is never used. The combination step ('Combine the transposed tensors...') is not specified. Algorithm 2's contraction is also not verified; the paper does not show that the assembled cores equal D†R. Since these algorithms are the claimed enabler of scalable compressed OpInf, a rigorous derivation or an independent numerical check (e.g., comparing against dense pseudoinverse on small tensors) is required.
  2. [II.C, Eqs. (14)–(15)] The tensor formulation is self-inconsistent in its index labels. Eq. (14) defines D_{i,j,k,q,i',j',k',q',n}, with the first four modes labeled i,j,k,q. In Eq. (15) the contraction uses D_{i',j',k',q',i'',j'',k'',q'',n} and O_{i,j,k,q,i',j',k',q',i'',j'',k'',q''}. These labelings are shifted by one prime level and do not match the stated dimensions of O (Eq. after (15)). As written, the tensor indices in Eq. (15) cannot be contracted to reproduce the least-squares objective. This is a load-bearing inconsistency in the formulation and must be corrected.
  3. [II.E] The claim that 'with G_i for i=1,...,4 being orthogonalized, the reduced snapshot matrix \hat{X} in the standard OpInf method is equivalent to G5' is not correct as stated: G5 is of size r4×K, whereas \hat{X} in Section II.A is K×n. Even accounting for transposition, TT-SVD truncation of a 5D tensor is not equivalent to POD truncation of the flattened N×K snapshot matrix. If TT-ROM is intended to use the last TT core as a reduced coordinate matrix, this is a choice of subspace and should be stated as such, not as equivalence to POD. The comparison with standard ROM in Tables 1–3 may still be meaningful, but the claimed equivalence needs to be clarified or removed.
  4. [III.C, Table 3] For the cylinder test case, no quantitative relative error is reported. The text only states that the predicted pressure fields are 'nearly identical' or show 'minor differences'. Given that the paper's accuracy claims rest on these results, a quantitative error measure (e.g., relative L2 error of the predicted field over the full domain or at the monitoring points) is required. Also, the regularization parameters for the cylinder case (10^5 for ROM/TT-ROM and 5×10^9 for QTT) are extremely large, and no justification (e.g., L-curve or cross-validation) is given; with such strong regularization the learned operators may be dominated by the penalty term, which would undermine the claim of learning the true dynamics.
  5. [III, Tables 1–3] The paper repeatedly claims 'memory efficiency' and 'scalability', but no memory measurements are reported, only timings. The grid sizes used (20×20 and 3,280 cells) are far too small to demonstrate asymptotic scalability of TT methods; for such small problems TT overhead can dominate and the advantages only appear at larger scales. A memory comparison (e.g., snapshot storage and operator storage for each method) and/or a scaling experiment with increasing grid size is necessary to support the central claim.
minor comments (4)
  1. [Throughout] There are numerous typos and spacing issues: 'the the' in Section II, 'Labarotory', 'tt_left_orthogonolize' (spelling), inconsistent use of 'TT' vs 'tt', and inconsistent comma placement. The manuscript would benefit from a careful proofreading pass.
  2. [II.D.2] The finite-difference derivative on the last TT core is said to be described in [16,17]; however, those references are about tensor-train mimetic finite differences and Boltzmann transport, not about time-derivative computation in OpInf. Please provide a description or a more specific reference.
  3. [II.A.1, Eq. (1)] The quadratic term F(x⊗x) uses a Kronecker product; later the tensor formulation uses elementwise products. This is fine under the full-order tensor interpretation, but the connection between the Kronecker-structured F in Eq. (1) and the tensor operator F in Eq. (16) should be made explicit.
  4. [III.A, Table 1] The text says that the TT method operates on a 5-dimensional tensor X_{i,j,k,q,n}, but for a 2D heat equation N_z=1 and N_q=1; this is fine, but the notation is slightly confusing. Clarify that singleton modes are retained for uniformity.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: TT-OpInf fits operators to training snapshots and validates by extrapolating past the training horizon; self-citations are implementation references, not load-bearing.

full rationale

The derivation chain is: snapshots -> TT compression -> assemble D_TT and R_TT -> solve regularized least squares via TT pseudoinverse -> extract A_TT and F_TT -> advance in time. The learned operators are not defined to be the predicted states; accuracy is assessed on time intervals beyond the training data (heat: predict t=0.9 after training to t=0.25; Burgers: predict t=0.09 after training to t=0.05; cylinder: predict t=15 after training to t=10). The TT pseudoinverse in Algorithm 1 is an algorithmic assertion about how to compute the pseudoinverse in compressed form; even if unproved or incorrect, it is not an identity that reduces the predicted output to the training input. The only tunable quantities are the regularization parameter gamma and the TT truncation tolerance, which are standard hyperparameters rather than results presupposed by the derivation. Self-citations [16,17] are referenced for TT finite-difference implementation details and [18] for the cylinder benchmark setup; none of these supplies a load-bearing theorem or smuggles in the central claim. The paper is therefore self-contained against external benchmarks, and no circular step can be exhibited. The lack of proof or independent verification of Algorithms 1 and 2 is a correctness risk, not a circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. The load-bearing assumptions are the low-rankness of the snapshot data, the quadratic model form, the accuracy of TT cross interpolation, and the correctness of the TT pseudoinverse contraction.

free parameters (4)
  • TT truncation tolerance ε_TT = 1e-12 (heat), 1e-5 (Burgers), 1e-4 (cylinder)
    Chosen per test case; controls TT rank, accuracy, and memory, and hence directly shapes the reported accuracy/scalability trade-off.
  • Regularization parameter γ = 1e-6 for ROM/TT-ROM (heat and Burgers); 1e5 for ROM/TT-ROM (cylinder); 5e9 for QTT (cylinder); none for full-order TT/QT
    Set by hand for each problem and method; no cross-validation or selection rule is given.
  • Number of POD modes n for ROM = 1 (heat), 5 (Burgers), 33 (cylinder)
    Set equal to the number of modes produced by TT-ROM at the chosen tolerance, not selected independently.
  • RK45 and finite-difference accuracy parameters = not specified
    Order of finite differencing for R_TT and RK45 tolerances are not reported, though they affect the learned operators.
assumptions (5)
  • domain assumption The snapshot dynamics are, or are well approximated by, polynomials of degree at most two.
    Assumed in Eq. (1) and throughout; OpInf's standard premise. If the dynamics are not quadratic, the operator model is misspecified.
  • domain assumption The snapshot tensors have low TT rank at the chosen truncation tolerances.
    Required for TT-SVD/cross compression to capture the data; Section II.B and II.E. No rank analysis or convergence study is provided.
  • domain assumption TT cross interpolation selects enough entries to reproduce X_TT accurately without full data access.
    Invoked in Section II.B.2 and used for the cylinder case; accuracy of this heuristic is not quantified.
  • ad hoc to paper The TT pseudoinverse algorithm and the contraction order in Algorithm 2 yield the exact least-squares solution.
    Stated in Section II.D.3 with the detailed algorithms, but no proof or error analysis is provided; this is the mathematical core of the method.
  • domain assumption Time derivative tensor R_TT can be computed by applying finite-difference stencils directly on the last TT core without losing accuracy.
    Section II.D.2; implementation details are delegated to references [16,17], with no error analysis here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tensor-Train Operator Inference." pith.science (2026). https://pith.science/paper/NXUQOICN

@misc{pith2026250908071,
  author       = {Pith},
  title        = {Pith review of: Tensor-Train Operator Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NXUQOICN}},
  note         = {Machine review of arXiv:2509.08071}
}
read the original abstract

In this study, we present a tensor--train framework for nonintrusive operator inference aimed at learning discrete operators and using them to predict solutions of physical governing equations. Our framework comprises three approaches: full--order tensor--train operator inference, full--order quantized tensor--train operator inference, and reduced--order tensor--train operator inference. In each case, snapshot data is represented in tensor--train format--either through compression or cross interpolation--enabling the efficient handling of extremely large datasets with significantly reduced computational effort compared to standard methods. The effectiveness of each approach is demonstrated through numerical experiments related to Computational Fluid Dynamics and benchmarked against the standard reduced--order operator inference method, highlighting the advantages of the tensor--train representations in both accuracy and scalability.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages

  1. [1]

    A survey of projection-based model reduction methods for parametric dynamical systems,

    Benner, P., Gugercin, S., and Willcox, K., “A survey of projection-based model reduction methods for parametric dynamical systems,”SIAM review, Vol. 57, No. 4, 2015, pp. 483–531

  2. [2]

    92, Springer, 2015

    Quarteroni, A., Manzoni, A., and Negri, F.,Reduced basis methods for partial differential equations: an introduction, Vol. 92, Springer, 2015

  3. [3]

    L., and Kutz, J

    Brunton, S. L., and Kutz, J. N.,Data-driven science and engineering: Machine learning, dynamical systems, and control, Cambridge University Press, 2022

  4. [4]

    Data-drivenoperatorinferencefornonintrusiveprojection-basedmodelreduction,

    Peherstorfer,B.,andWillcox,K.,“Data-drivenoperatorinferencefornonintrusiveprojection-basedmodelreduction,”Computer Methods in Applied Mechanics and Engineering, Vol. 306, 2016, pp. 196–215. https://doi.org/https://doi.org/10.1016/j.cma. 2016.03.025, URL https://www.sciencedirect.com/science/article/pii/S0045782516301104

  5. [5]

    Projection-basedmodelreduction: Formulationsforphysics-based machine learning,

    Swischuk,R.,Mainini,L.,Peherstorfer,B.,andWillcox,K.,“Projection-basedmodelreduction: Formulationsforphysics-based machine learning,”Computers & Fluids, Vol. 179, 2019, pp. 704–717

  6. [6]

    An enhanced algorithm for online proper orthogonal decomposition and its parallelization for unsteady simulations,

    Li, X., Hulshoff, S., and Hickel, S., “An enhanced algorithm for online proper orthogonal decomposition and its parallelization for unsteady simulations,”Computers & Mathematics with Applications, Vol. 126, 2022, pp. 43–59

  7. [7]

    Dynamic programming,

    Bellman, R., “Dynamic programming,”science, Vol. 153, No. 3731, 1966, pp. 34–37

  8. [8]

    O(dlogN)-quanticsapproximationofN-dtensorsinhigh-dimensionalnumericalmodeling,

    Khoromskij, B.N., “O(dlogN)-quanticsapproximationofN-dtensorsinhigh-dimensionalnumericalmodeling,”Constructive Approximation, Vol. 34, No. 2, 2011, pp. 257–280

Show all 18 references
  1. [9]

    TT-cross approximation for multidimensional arrays,

    Oseledets, I., and Tyrtyshnikov, E., “TT-cross approximation for multidimensional arrays,”Linear Algebra and its Applications, Vol. 432, No. 1, 2010, pp. 70–88

  2. [10]

    Fast adaptive interpolation of multi-dimensional arrays in tensor train format,

    Savostyanov, D., and Oseledets, I., “Fast adaptive interpolation of multi-dimensional arrays in tensor train format,”The 2011 International Workshop on Multidimensional (nD) Systems, IEEE, 2011, pp. 1–8

  3. [11]

    Learning Nonlinear Reduced Models from Data with Operator Inference,

    Kramer, B., Peherstorfer, B., and Willcox, K. E., “Learning Nonlinear Reduced Models from Data with Operator Inference,” Annual Review of Fluid Mechanics, Vol. 56, No. Volume 56, 2024, 2024, pp. 521–548. https://doi.org/https://doi.org/10.1146/ annurev-fluid-121021-025220, URL...

  4. [12]

    Tensor-train decomposition,

    Oseledets, I. V., “Tensor-train decomposition,”SIAM Journal on Scientific Computing, Vol. 33, No. 5, 2011, pp. 2295–2317

  5. [13]

    CUR matrix decompositions for improved data analysis,

    Mahoney, M. W., and Drineas, P., “CUR matrix decompositions for improved data analysis,”Proceedings of the National Academy of Sciences, Vol. 106, No. 3, 2009, pp. 697–702

  6. [14]

    How to find a good submatrix,

    Goreinov, S. A., Oseledets, I. V., Savostyanov, D. V., Tyrtyshnikov, E. E., and Zamarashkin, N. L., “How to find a good submatrix,”Matrix Methods: Theory, Algorithms And Applications: Dedicated to the Memory of Gene Golub, World Scientific, 2010, pp. 247–256

  7. [15]

    Alternating minimal energy methods for linear systems in higher dimensions,

    Dolgov, S. V., and Savostyanov, D. V., “Alternating minimal energy methods for linear systems in higher dimensions,”SIAM Journal on Scientific Computing, Vol. 36, No. 5, 2014, pp. A2248–A2271

  8. [16]

    Tensor networks for solving the time-independent Boltzmann neutron transport equation,

    Truong, D. P., Ortega, M. I., Boureima, I., Manzini, G., Rasmussen, K. Ø., and Alexandrov, B. S., “Tensor networks for solving the time-independent Boltzmann neutron transport equation,”Journal of Computational Physics, Vol. 507, 2024, p. 112943

  9. [17]

    The tensor-train mimetic finite difference method for three- dimensional Maxwell’s wave propagation equations,

    Manzini, G., Truong, P., Vuchkov, R., and Alexandrov, B., “The tensor-train mimetic finite difference method for three- dimensional Maxwell’s wave propagation equations,”Mathematics and Computers in Simulation, Vol. 210, 2023, pp. 615–639

  10. [18]

    A new direct discontinuous Galerkin method with interface correction for two-dimensional compressible Navier-Stokes equations,

    Danis, M. E., and Yan, J., “A new direct discontinuous Galerkin method with interface correction for two-dimensional compressible Navier-Stokes equations,”Journal of Computational Physics, Vol. 452, 2022, p. 110904. https://doi.org/https: //doi.org/10.1016/j.jcp.2021.110904, U...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.