Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Application of operator inference to reduced-order modeling of constrained mechanical systems

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Operator inference can learn reduced-order ODE models of constrained mechanical systems from DAE simulation data alone.

desk verdict A useful, reproducible OpInf extension to constrained mechanical DAEs, but the velocity-constraint section silently assumes zero initial position offset. read the letter →

arxiv 2507.05472 v1 pith:EGBNQIOS submitted 2025-07-07 math.DS

Peter Benner , Yevgeniya Filanova , Igor Pontes Duff , Jens Saak This is my paper · ORCID
classification math.DS MSC 65L8070E5593B11
keywords operatorinferencemodelorderreductiondifferential-algebraicequationsconstrainedmechanicalsystemsproperorthogonaldecompositionsemidefiniteprogramming
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 establishes that operator inference can recover the reduced-order ordinary differential equation (ODE) model of a constrained mechanical system directly from simulation data of the differential-algebraic equation (DAE) system, without ever accessing the original mass, damping, stiffness, or constraint matrices. For proper index-2 and index-3 DAEs, such a hidden-manifold ODE realization exists, and the paper argues that POD compression of displacement or velocity snapshots automatically respects the algebraic constraints, so the fitted second-order operators describe the true constrained dynamics. The practical payoff is that proprietary or matrix-free simulation software, which only provides time series, becomes a sufficient source for building stable and interpretable surrogates. Numerical tests on two mass-spring-damper systems, one with position constraints (index-3) and one with velocity constraints (index-2), show accurate output tracking for both the training input and an unseen harmonic input.

What carries the argument

The load-bearing mechanism is the POD-based elimination of algebraic constraints combined with structure-preserving second-order operator inference. Proper orthogonal decomposition here means the singular value decomposition of the displacement snapshot matrix X (for position constraints) or the velocity snapshot matrix Ẋ (for velocity constraints); truncating the left singular vectors gives a basis Vr that satisfies GpVr = 0 or GvVr = 0 because the full snapshot matrices already do. This guarantees the reduced model lives on the constraint manifold. The reduced mass, damping, stiffness, and input matrices are then fitted by minimizing the residual of the second-order ODE on the compressed snapshots, subject to symmetric positive semidefinite constraints; the output matrices are fit by a separate least-squares problem. The identity that carries the argument is therefore the snapshot-to-basis implication 'if constraints hold for the data, they hold for the POD subspace,' which turns a constrained DAE reduction into a standard second-order ODE fit.

What would settle it

Generate snapshots for a small position-constrained DAE using an integrator that does not project onto the constraint manifold (so Gp x ≈ 0 only up to tolerance), build the POD basis Vr, and compute the residual norm ||Gp Vr||_F. If this residual is not at machine precision, the claimed elimination (8) fails, and the operators fitted to the constrained data will not reproduce the hidden ODE, with the output error growing accordingly.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the algebraic constraints of a proper constrained mechanical DAE can be eliminated without the system matrices: if the data come from an integrator that satisfies the constraints exactly, then the displacement snapshots lie in ker(Gp) for position constraints and the velocity snapshots lie in ker(Gv) for velocity constraints, and the POD basis inherits this property. Consequently, projecting the DAE onto that basis yields the reduced second-order ODE (Eq. 11), and solving the semidefinite optimization problem (16) with SPD constraints on the reduced mass, damping, and stiffness matrices recovers this reduced system from data alone. The paper therefore claims that the operator inference problem directly provides the underlying ODE realization of the DAE in the reduced subspace, with structure preservation guaranteeing stability and interpretability. The numerical results support this claim for an n = 600 index-3 position-constrained system reduced to r = 5 and an n = 301 index-2 velocity-constrained system reduced to r = 35.

Load-bearing premise

The load-bearing premise is that the DAE integrator returns solutions that lie exactly on the constraint manifold, so that every snapshot satisfies the algebraic equations and the POD basis inherits the property GpVr = 0 or GvVr = 0; if constraints are only approximately satisfied, the inferred reduced operators can be inconsistent with the mechanics.

Editorial extensions

If this is right

  • The reduced-order model is obtained from time-domain data alone, so the method applies when the full-order equations live inside proprietary or matrix-free simulation software.
  • Because the inferred operators are forced to be symmetric positive semidefinite, the resulting ROM inherits stability and physical interpretability from the data-fitting step.
  • Position-constrained (index-3) and velocity-constrained (index-2) systems are both covered, with the choice of snapshot matrix (displacement vs. velocity) determined by the constraint type.
  • The training-data heuristic — impulse-like excitation with end time chosen from the output response — gives a practical recipe for producing an informative snapshot set.

Reading between the lines

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

  • A natural diagnostic for the constraint-preservation premise is to measure the residual norm ||Gp Vr||_F (or ||Gv Vr||_F) of the POD basis: near-zero values confirm the elimination step, while larger values signal the need for constraint projection before inference.
  • The method is demonstrated for systems with either position or velocity constraints; combining both would require a basis in ker(Gp) ∩ ker(Gv), and the paper does not address that mixed case.
  • Because the learned model is a standard second-order ODE with SPD operators, it can be handed to downstream control-design or further reduction pipelines that expect ODE structure; this is an implicit benefit.
  • A systematic study of how the choice of training input (impulse vs. random vs. multi-sine) affects the inferred operators would extend the practical guidance beyond the impulse-excitation heuristic used here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript applies operator inference (OpInf) to second-order differential-algebraic equation (DAE) models of constrained mechanical systems of index 2 and 3. The proposed pipeline is: collect displacement, velocity, and acceleration snapshots from a DAE integrator; build a POD basis from displacement snapshots for position constraints or from velocity snapshots for velocity constraints; compress the snapshot data by projection onto this basis; solve a semidefinite program to fit reduced mass, damping, stiffness, and input matrices; solve a least-squares problem for output matrices; and then simulate the resulting reduced ODE. Two numerical examples are presented: an anchored mass-spring-damper system with position constraints and a triple-chain system with velocity constraints. The ROMs are validated on both the training input and a separate harmonic test input, and the code and data are made available.

Significance. The contribution is practically relevant: it extends operator inference to a class of DAE systems without requiring access to the full-order system matrices, removes algebraic constraints through snapshot-based POD, and enforces structure in the reduced operators. The numerical results on the tested examples are encouraging, and the availability of open code and data is a strength. However, the generality of the velocity-constraint elimination is overstated, the stability guarantee is not established by the semidefinite constraints actually enforced, and the paper provides no identifiability or consistency analysis for the claimed 'direct' recovery of the underlying ODE representation. These issues affect the central claim and require revision.

major comments (3)
  1. [Section 2, Eqs. (5), (9)-(10), (13)-(14)] The derivation for velocity-constrained systems assumes that the displacement snapshots lie in the POD subspace obtained from velocity snapshots. The constraint Gv * x_dot(t) = 0 implies Gv * x(t) = Gv * x(0), not Gv * x(t) = 0. Unless x(0) is zero or otherwise lies in range(Vr), the displacement snapshot matrix X contains a component outside range(Vr), and the compression X_r = V_r^T * X discards that component. Consequently, the reduced ODE (11) and the output equation cannot reproduce the position-dependent part of the response. The triple-chain numerical example appears to start from rest, so it does not exercise this failure mode. The claimed general applicability to index-2 velocity-constrained systems is therefore not supported. The authors should either restrict the scope to consistent zero initial positions or enrich the basis with x(0) or an affine offset.
  2. [Section 3, Eqs. (15)-(16), and Abstract] The paper states that stability and interpretability are guaranteed by enforcing symmetric positive definite structure using semidefinite programming. However, the constraints actually imposed in (16) are M_r ⪰ 0, D_r ⪰ 0, K_r ⪰ 0. Positive semidefiniteness is not sufficient for the claim: a singular M_r makes the reduced equation ill-posed as an ODE, and a singular K_r permits rigid-body modes. The strict definiteness required for the stated mechanical interpretation is not enforced. The authors acknowledge the relaxation, but the 'guarantee' language should be removed or replaced by a precise statement about semidefinite structure and the numerical strategy used to keep the matrices non-singular.
  3. [Section 3, Eq. (16), and Section 4] The paper claims that the OpInf optimization 'directly provides the underlying ODE representation' of the DAE system. No identifiability or consistency analysis is given: for a single training trajectory there can be multiple tuples (M_r, D_r, K_r, B_r) that minimize the data-fit objective, and the relation between the fitted operators and the exact reduced operators from (12) is not analyzed. The numerical validation is limited to two examples. A formal discussion, or at least an explicit identifiability condition such as persistent excitation and sufficient snapshot length, is needed to support the general claim.
minor comments (5)
  1. [Section 2, Eq. (4)] In the position-constraint reduction, the output equation reads y(t) = C_p * N * z(t) + C_v * \z_dot(t); the second term should be C_v * N * \z_dot(t), matching the dimensions of the matrices.
  2. [Abstract and Section 3, Eq. (15)] The phrase 'symmetric positive definite structure' conflicts with the semidefinite constraints in Eq. (16); the terminology should be made consistent, e.g., 'positive (semi-)definite'.
  3. [Section 4.2] The heuristic for choosing dt and t_k based on the impulse response is reasonable, but no sensitivity study is reported; a sentence on how sensitive the resulting ROM is to these simulation parameters would be helpful.
  4. [Section 4.2, Figures 4 and 6] The singular value decay and cumulative energy plots are informative, but the chosen truncation point r=35 is justified only by visual saturation; stating the energy threshold used would make the selection reproducible.
  5. [Section 3, first paragraph] The assumption that the DAE integrator produces snapshots that exactly satisfy the constraints is explicit, but the consequences of a violation (e.g., with a constraint-drift integrator) are not discussed; a brief remark on acceptable drift would be useful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; OpInf validation is independent, and the velocity-constraint basis issue is a modeling gap rather than a circular derivation.

full rationale

The core identification is not circular: the reduced operators in (16)-(17) are fitted by least squares to compressed snapshots from the impulse training input (Fig. 3a), and the method is then evaluated on a different harmonic input (Fig. 3b), so the validation outputs are not used to set any fitted constants. The POD elimination (8) is an elementary linear-algebra fact (GpX=0 implies GpVr=0) and is stated explicitly rather than resting on the self-citation to [4]; the second-order OpInf with SPD constraints is re-derived in (15)-(17), and no uniqueness theorem or model-selection principle is imported from [7]. Self-citations to [4] and [7] are contextual and not load-bearing. The exact-constraint integrator assumption is stated openly in Section 3 and is a modeling assumption, not a disguised input. One correctness caveat is real but not circular: for velocity constraints, building Vr from an SVD of \dot X (Eqs. (9)-(10)) forces GvVr=0, so any position component outside ker(Gv) is discarded by Xr=Vr^T X; unless initial displacements lie in ker(Gv), the reduced model cannot represent the position state. This is a limitation of the claimed index-2 generality, not an equation that makes the test prediction equal to the training fit, so it does not raise the circularity score.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method rests on standard DAE theory (existence of a hidden-manifold ODE), the full-rank constraint assumption, the assumption that the DAE integrator returns exactly constraint-satisfying solutions, and the usual POD truncation heuristic. The reduced order and snapshot window are chosen by hand from the training data, which are the only tuned quantities; no extra invented physical entities are introduced.

free parameters (2)
  • Reduced order r = r = 5 (position constraints); r = 35 (velocity constraints)
    Chosen by thresholding the cumulative energy of POD singular values; not cross-validated against test data.
  • Snapshot time step dt and end time tk = Not specified numerically; chosen from impulse response analysis
    Selected so that trajectories are captured and vibrations decay; affects the quality of the fit.
assumptions (5)
  • domain assumption There exists an underlying ODE representation on a hidden manifold for proper DAE systems.
    Invoked in the abstract and Section 2 to justify reducing the DAE to a second-order ODE (Eq. 2); standard DAE theory (e.g., Kunkel and Mehrmann).
  • domain assumption Constraint matrices Gp and Gv are full rank.
    Stated in Section 2; required for a well-defined constraint manifold and for the hidden-manifold reduction.
  • domain assumption The mechanical system is linear time-invariant with constant coefficient matrices M, D, K.
    The model (1) assumes LTI structure; the method as presented does not address nonlinear or time-varying systems.
  • domain assumption The DAE integrator computes numerical solutions that exactly satisfy the algebraic constraints.
    Assumed in Section 3; the POD-based elimination GpVr = 0 and GvVr = 0 relies on the snapshots lying on the constraint manifold.
  • ad hoc to paper The leading POD modes span the active constraint manifold and capture the essential dynamics.
    This is the standard POD projection assumption (Section 2), not proven for these systems; the reduced order r is chosen heuristically from singular value decay.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Application of operator inference to reduced-order modeling of constrained mechanical systems." pith.science (2026). https://pith.science/paper/EGBNQIOS

@misc{pith2026250705472,
  author       = {Pith},
  title        = {Pith review of: Application of operator inference to reduced-order modeling of constrained mechanical systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EGBNQIOS}},
  note         = {Machine review of arXiv:2507.05472}
}
read the original abstract

Constrained mechanical systems occur in many applications, such as modeling of robots and other multibody systems. In this case, the motion is governed by a system of differential-algebraic equations (DAE), often with large and sparse system matrices. The problem dimension strongly influences the effectiveness of simulations for system analysis, optimization, and control, given limited computational resources. Therefore, we aim to obtain a simplified surrogate model with a few degrees of freedom that is able to accurately represent the motion and other important properties of the original high-dimensional DAE model. Classical model reduction methods intrusively exploit the system matrices to construct the projection of the high-fidelity model onto a low-dimensional subspace. In practice, the dynamical equations are frequently an inaccessible part of proprietary software products. In this work, we show an application of the non-intrusive operator inference (OpInf) method to DAE systems of index 2 and 3. Considering the fact that for proper DAEs there exists an ODE realization on the so-called hidden manifold, the OpInf optimization problem directly provides the underlying ODE representation of the given DAE system in the reduced subspace. A significant advantage is that only the DAE solution snapshots in a compressed form are required for identification of the reduced system matrices. Stability and interpretability of the reduced-order model is guaranteed by enforcing the symmetric positive definite structure of the system operators using semidefinite programming. The numerical results demonstrate the implementation of the proposed methodology for different examples of constrained mechanical systems, tested for various loading conditions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

  1. [1]

    M. I. Ahmad and P. Benner. Interpolatory model re- duction techniques for linear second-order descriptor systems. In Proc. European Control Conf. ECC 2014, Strasbourg, pages 1075–1079. IEEE, 2014. doi:10. 1109/ECC.2014.6862210

    work page arXiv 2014

  2. [2]

    A. C. Antoulas. Approximation of Large-Scale Dy- namical Systems , volume 6 of Adv. Des. Control . SIAM Publications, Philadelphia, PA, 2005. doi: 10.1137/1.9780898718713

    doi:10.1137/1.9780898718713 2005

  3. [3]

    Benner and H

    P. Benner and H. Faßbender. Model Order Re- duction: Techniques and Tools. In J. Bail- lieul and T. Samad, editors, Encyclopedia of Sys- tems and Control . Springer, 2019. doi:10.1007/ 978-1-4471-5102-9_142-2

    work page 2019

  4. [4]

    Benner, P

    P. Benner, P. Goyal, J. Heiland, and I. Pontes Duff. Operator inference and physics-informed learning of low-dimensional models for incompressible flows. Electron. Trans. Numer. Anal., 56:28–51, 2022. doi: 10.1553/etna_vol56s28

    work page doi:10.1553/etna_vol56s28 2022

  5. [5]

    C. L. Bottasso, D. Dopico, and L. Trainelli. On the optimal scaling of index three DAEs in multibody dynamics. Multibody Syst. Dyn., 19:3–20, 2008. doi: 10.1007/s11044-007-9051-9

    work page doi:10.1007/s11044-007-9051-9 2008

  6. [6]

    Eich-Soellner and C

    E. Eich-Soellner and C. F¨ uhrer. Numerical Methods in Multibody Dynamics , volume 45. Springer, 1998. doi:10.1007/978-3-663-09828-7

    doi:10.1007/978-3-663-09828-7 1998

  7. [7]

    Filanova, I

    Y. Filanova, I. Pontes Duff, P. Goyal, and P. Ben- ner. An operator inference oriented approach for linear mechanical systems. Mech. Syst. Signal Pro- cess., 200(110620), 2023. doi:10.1016/j.ymssp. 2023.110620

    arXiv 2023

  8. [8]

    Gugercin, T

    S. Gugercin, T. Stykel, and S. Wyatt. Model reduc- tion of descriptor systems by interpolatory projec- tion methods. SIAM J. Sci. Comput. , 35(5):B1010– B1033, 2013. doi:10.1137/130906635

    work page doi:10.1137/130906635 2013

Show all 29 references
  1. [9]

    E. J. Haug. Computer-Aided Kinematics and Dy- namics of Mechanical Systems , volume 1. Allyn & Bacon, Inc., 1998

    1998

  2. [10]

    Heinkenschloss, D

    M. Heinkenschloss, D. C. Sorensen, and K. Sun. Bal- anced truncation model reduction for a class of de- scriptor systems with application to the Oseen equa- tions. SIAM J. Sci. Comput. , 30(2):1038–1063, 2008. doi:10.1137/070681910

    2008 doi

  3. [11]

    Khodabakhshi and K

    P. Khodabakhshi and K. E. Willcox. Non- intrusive data-driven model reduction for dif- ferential–algebraic equations derived from lifting transformations. Comput. Methods Appl. Mech. Eng., 389:114296, 2022. doi:10.1016/j.cma.2021. 114296

    2022 doi

  4. [12]

    Kunisch and S

    K. Kunisch and S. Volkwein. Proper orthogonal de- composition for optimality systems. ESAIM: Math. Model. Numer. Anal. , 42(1):1–23, 2008. doi:10. 1051/m2an:2007054

    2008

  5. [13]

    Kunkel and V

    P. Kunkel and V. Mehrmann. Differential-Algebraic Equations. European Mathematical Society Z¨ urich,

  6. [14]

    L¨ ofberg

    J. L¨ ofberg. YALMIP: A toolbox for modeling and op- timization in matlab. In Proceedings of the CACSD Conference, Taipei, Taiwan, 2004. doi:10.1109/ CACSD.2004.1393890

    2004 arXiv

  7. [15]

    Mehrmann and T

    V. Mehrmann and T. Stykel. Balanced trun- cation model reduction for large-scale systems in descriptor form. In P. Benner, V. Mehrmann, and D. C. Sorensen, editors, Dimension Reduction of Large-Scale Systems , volume 45 of Lect. Notes Comput. Sci. Eng. , pages 83–115. Springer...

    2005

  8. [16]

    Mehrmann and T

    V. Mehrmann and T. Stykel. Descriptor systems: A general mathematical framework for modelling, sim- ulation and control. at - Automatisierungstechnik , 54(8):405–415, 2006. doi:10.1524/auto.2006.54. 8.405

    2006 doi

  9. [17]

    N. M. Newmark. A method of computation for struc- tural dynamics. ASCE J. of the Engrg. Mech. Di- vision, 85(EM 3, July):67–94, 1959. doi:10.1061/ JMCEA3.0000098

    1959

  10. [18]

    Peherstorfer and K

    B. Peherstorfer and K. Willcox. Data-driven opera- tor inference for nonintrusive projection-based model reduction. Comput. Methods Appl. Mech. Eng. , 306:196–215, 2016. doi:10.1016/j.cma.2016.03. 025. Preprint (Max Planck Institute for Dynamics of Complex Technical Systems, M...

    2016 doi

  11. [19]

    Rentrop, O

    P. Rentrop, O. Scherf, and B. Simeon. Mechanical multibody systems with deformable components. In H.-J. Bungartz, F. Durst, and C. Zenger, editors, High Performance Scientific and Engineering Com- puting, pages 143–155. Springer Berlin Heidelberg,

  12. [20]

    Saak and M

    J. Saak and M. Voigt. Model reduction of con- strained mechanical systems in M-M.E.S.S. IF AC- PapersOnLine 9th Vienna International Conference on Mathematical Modelling MATHMOD 2018, Vi- enna, Austria, 21–23 February 2018 , 51(2):661–666,

    2018

  13. [21]

    Salimbahrami and B

    B. Salimbahrami and B. Lohmann. Order reduction of large scale second-order systems using Krylov sub- space methods. Linear Algebra Appl., 415(2–3):385– 405, 2006. doi:10.1016/j.laa.2004.12.013

    2006 doi

  14. [22]

    A. A. Shabana. Dynamics of Multibody Systems . Cambridge University Press, 2020. doi:10.1017/ CBO9781107337213

    2020

  15. [23]

    B. Simeon. Numerical analysis of flexible multibody systems. Multibody Syst. Dyn., 6:305–325, 2001. doi: 10.1023/A:1012096631385

    2001 doi

  16. [24]

    J. J. Thomson. Vibrations and Stability: Advanced Theory, Analysis and Tools . Springer Cham, third edition, 2021. doi:10.1007/978-3-030-68045-9

    2021 doi

  17. [25]

    Truhar and K

    N. Truhar and K. Veseli´ c. An efficient method for estimating the optimal dampers’ viscosity for linear vibrating systems using Lyapunov equation. SIAM J. Matrix Anal. Appl. , 31(1):18–39, 2009. doi:10. 1137/070683052

    2009

  18. [26]

    M. M. Uddin. Computational Methods for Model Reduction of Large-Scale Sparse Struc- tured Descriptor Systems . Dissertation, Otto- von-Guericke-Universit¨ at, Magdeburg, Germany,

  19. [1999]

    doi:10.1007/978-3-642-60155-2_12

    doi

  20. [2015]

    Preprint (Max Planck Institute for Dynamics of Complex Technical Systems, Magdeburg)

    URL: http://nbn-resolving.de/urn:nbn: de:gbv:ma9:1-6535. Preprint (Max Planck Institute for Dynamics of Complex Technical Systems, Magdeburg). 2025-08-21

    2025

  21. [2018]

    doi:10.1016/j.ifacol.2018.03.112

    2018 doi

Pith tools

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