Pith. sign in

REVIEW 5 major objections 5 minor 14 references

For a class of slowly rotating linear time-varying systems, shrinking a single flow parameter ε keeps Oja's flow inside an arbitrarily small neighborhood of the time-varying dominant subspace, providing a theoretical basis for low-rank Kalm

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 14:57 UTC pith:UP3HFFKD

load-bearing objection A real time-varying Oja-flow analysis, but the main theorem is false as stated; the proof only controls distance to the subspace, not to a rotating basis. the 5 major comments →

arxiv 2607.29034 v1 pith:UP3HFFKD submitted 2026-07-31 eess.SY cs.SYmath.OC

On the Oja-Flow-Based Low-Rank Approximation of Kalman-Bucy Filters for Linear Time-Varying Systems

Kentaro Ohki This is my paper
classification eess.SY cs.SYmath.OC
keywords Oja flowprincipal component flowlow-rank Kalman-Bucy filterlinear time-varying systemsdominant subspace trackingspectral gapRiccati equation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tackles the computational bottleneck of Kalman-Bucy filtering, where the full matrix Riccati equation is too costly for high-dimensional systems. It analyzes Oja's principal component flow as a way to track the dominant r-dimensional subspace of the time-varying coefficient matrix, then uses that tracked subspace to build a low-rank filter. The central theoretical result is that, under structured assumptions—a uniform spectral gap and a rigidly rotating eigenbasis—the flow stays within a permanently bounded, arbitrarily small neighborhood of the time-varying dominant subspace if the flow parameter ε is chosen small enough. This turns approximate tracking, rather than exact tracking, into a justified design principle for low-rank filtering of certain linear time-varying systems.

Core claim

The paper's central claim is Theorem 3: if A(t) = Ψ(t)Λ(t)Ψ(t)^{-1} with a diagonal Λ(t) whose r-th and (r+1)-th eigenvalues are separated by a uniform gap c1, and if the eigenbasis rotates smoothly while the subspace Ψ(t)^{-1}U_r(A(t)) stays fixed, then there exists ε > 0 such that the Oja flow starting near a dominant subspace at time t0 remains near the time-varying dominant subspace for all t ≥ t0. The tracking error is controlled by ε times the rotation speed divided by the spectral gap, so by taking ε small the error can be made arbitrarily small. The paper applies this to a low-rank Kalman-Bucy filter and demonstrates in a numerical example that smaller ε keeps the filter's error cova

What carries the argument

The central mechanism is the coordinate change K(t) = Ψ(t)^{-1}U(t), which passes to the rotating eigenframe. Under the theorem's assumptions 4 and 5, the target dominant subspace becomes a fixed subspace in this frame, so the time variation of A(t) is converted into a bounded perturbing term involving Ψ(t)^{-1}dΨ(t)/dt, added to the familiar constant-matrix Oja flow. The spectral gap c1 makes the transverse component contract at exponential rate roughly c1/ε, while the rotation term pushes the trajectory out at rate proportional to ε times the rotation speed; choosing ε small lets contraction dominate. This perturbation-coordinate argument is what carries the proof and ties the tracking err

Load-bearing premise

The load-bearing premise is that the time variation is only a rigid rotation: after undoing the rotating eigenbasis, the target dominant subspace does not move; if the subspace drifts within that frame, the coordinate trick collapses and the proof gives no error bound.

What would settle it

Take A(t)=R(t)Λ(t)R(t)^T with R(t) rotating at speed ω and Λ(t) chosen so the spectral gap narrows substantially at some time. Numerically integrate the Oja flow for ε below the paper's threshold, and measure the maximum distance to the dominant subspace. If the error grows like ω/gap rather than εω/gap, or if the bound is violated whenever Ψ(t)^†Ψ(t) is not constant, the theorem's restrictive assumptions are essential and the practical envelope is smaller than the statement suggests.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Choosing ε small enough makes the tracking error permanently small whenever the spectral gap is uniformly positive and the eigenbasis rotates slowly, with an explicit bound in terms of ε, rotation speed, and gap.
  • The low-rank Kalman-Bucy filter built on this flow inherits a near-tracking guarantee, so its error covariance can stay close to the full Kalman-Bucy filter for slowly rotating linear time-varying systems.
  • The proof provides a quantitative tuning rule for ε: it should be small relative to the gap divided by the rotation speed, but not so small that numerical stiffness dominates the discretization.
  • Systems satisfying the assumptions—diagonalizable with a uniform spectral gap and a rigidly rotating eigenbasis—form an explicit, tractable class where low-rank filtering is theoretically grounded.
  • Because the result is approximate rather than exact tracking, it applies to time-varying systems where perfect tracking is impossible, extending the constant-matrix Oja-flow analysis to a limited but nontrivial time-varying setting.

Where Pith is reading between the lines

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

  • The assumptions essentially freeze the dominant subspace in the rotating eigenframe; a natural extension would allow slow drift of Ψ(t)^{-1}U_r(A(t)) and should produce an error bound with an additional drift-over-gap term.
  • For filtering practice, the near-tracking guarantee suggests adapting ε online from a live estimate of the eigenbasis rotation speed, rather than fixing it conservatively; the paper's analysis gives the scaling but not yet an adaptive scheme.
  • The numerical tests use a very slow rotation (speed 10^{-4}); pushing the rotation speed toward the spectral gap would be a sharper test of whether the ε-scaling predicted by the proof holds at the edge of the assumptions.
  • The proof's perturbation technique could likely transfer to other subspace-tracking flows that share the same coordinate-change structure, but the paper itself only claims the Oja-flow case.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper studies the tracking behavior of Oja's principal component flow (2) for bounded linear time-varying matrices. It first gives a two-dimensional rotating-matrix analysis (Proposition 2), then a more general tracking theorem (Theorem 3) under assumptions that make the dominant eigenspace fixed in a rotating eigenframe, and finally applies the result to a low-rank Kalman-Bucy filter for an LTV system. Numerical experiments show that smaller ε yields smaller distance to the dominant subspace and filter covariance closer to the full-order Kalman-Bucy filter. The paper explicitly acknowledges that the assumptions are restrictive and identifies them as a tractable class rather than covering general LTV systems.

Significance. If the main theorem is corrected, the paper would contribute an explicit, quantitative bound for Oja-flow tracking of a time-varying dominant subspace in a nontrivial class of systems, and a useful design rule relating ε, the spectral gap, and the rotation speed. The paper's strengths are its explicit analytical goal, the admission of its assumptions, and numerical evidence. However, the central theorem is stated for tracking a particular time-varying basis while the proof controls only the component orthogonal to the dominant subspace; this distinction is essential and currently breaks the theorem. The filtering section also relies on an unpublished same-author preprint for a key gain formula. These issues are local and fixable, so the manuscript is substantially promising but needs revision.

major comments (5)
  1. [Theorem 3, §3] The statement claims that U(t) remains in Nδ(Ū(t)) for a time-continued basis Ū(t)∈U_r(A(t)). The proof only bounds L'_⊥, the component orthogonal to the dominant subspace, and does not control L_r. This is not a harmless gap: for r>1 every orthonormal basis of the same dominant subspace is a zero of (I-UU^T)AU, so the basis can rotate inside the subspace while the subspace itself is tracked perfectly. Concrete counterexample satisfying Assumptions 1–5: n=3, r=2, A(t)=R(t)diag(1,0.9,0)R(t)^T with R(t) an xy-plane rotation. Take U(0)=R(0)Kbar with Kbar∈O(2). Then the right-hand side of (2) vanishes identically, so U(t)≡U(0), while Ū(t)=R(t)Kbar has distance ∥R(0)Kbar−R(t)Kbar∥ reaching 2. No ε can keep U in a small Nδ(Ū(t)). The theorem should be restated as a bound on distance to the unparametrized dominant subspace (e.g., ∥UU^T−Π_dominant∥), and the proof adjusted accordingly. The numer
  2. [Proposition 2, §3] For K(t)=R(t)^TU(t), direct differentiation gives ε dK/dt = -ε Rdot^T R K + (I-KK^T)ΛK = -εω J K + ..., with constant J=[[0,-1],[1,0]] (sign depending on convention). The displayed term εω[[0,-cos 2ωt],[cos 2ωt,0]]K does not arise from this change of frame; the cos 2ωt factor appears nowhere in the transformation. The subsequent linearization and the claimed solution should be redone. The conclusion of Proposition 2 may still survive with a constant skew perturbation, but the proof as written is incorrect.
  3. [Theorem 3 proof] In the displayed ODE for L'_⊥, the forcing term -C1(t)(Kbar+L'_r(t)) should carry a prefactor ε, since it comes from -ε Ψ^{-1}(dΨ/dt)K. As written, the equation is dimensionally inconsistent with the preceding linearization. The later norm bound still converges as ε→0 because the integral contributes a factor 1/ε, so this error does not by itself invalidate the theorem, but the derivation and the explicit solution formula must be corrected before the proof can be accepted.
  4. [§4] The gain formula G_Uε(t)=(Vε^TUε)^{-1}Vε^TG is central to the numerical filter, but it is introduced by 'From the result of our follow-up paper [7]' with no derivation in this manuscript. Reference [7] is a same-author arXiv preprint. Since the low-rank filter design is one of the two advertised contributions, this is a gap: either prove the formula, or state it explicitly as an assumption with a complete citation, and give conditions under which Vε^TUε is invertible.
  5. [Theorem 3] Assumptions 4 and 5 force the target dominant subspace to be fixed in a rotating eigenframe, so the analysis is a bounded-perturbation result around the constant-matrix Oja flow. Many natural LTV systems, such as slowly varying eigenbases not of the form R(t)Ψ', are not covered. The paper acknowledges this, and the numerical examples are constructed to satisfy the assumptions, so they do not test the general case. This is acceptable for a tractable first class, but the limitation should be stated prominently and the bounds should clearly expose the dependence on the rotation speed c2.
minor comments (5)
  1. [Throughout] There are several typos and notational inconsistencies. For example, 'accerelated' should be 'accelerated' in §3; the paper uses complex eigenbases Ψ(t) while U(t) is real, so the neighborhoods Nδ need to be defined for the appropriate field; and the notation Φ_{C2+Λ⊥/ε}(t) should be defined explicitly as a fundamental matrix.
  2. [§2–§3] The paper repeatedly uses results from the same authors' unpublished preprint [7], including Theorem 12 and the convergence claims for constant matrices. If this is not yet a published/independent source, the necessary statements should be reproduced or proven in this paper, or a published reference should be provided.
  3. [§4] The system matrices include B, but B is never used; the noise gain G is not specified explicitly in the numerical example. This makes the simulation setup harder to reproduce.
  4. [Fig. 1, Fig. 3] The plotted quantities are projection distances (UU^T minus a subspace projector), not basis distances. The captions and the text should be aligned with the corrected subspace-tracking statement, otherwise the figures appear to support a stronger claim than they show.
  5. [Notation] In Theorem 3, Assumption 5 states 'Ψ(t)^{-1}U_r(A(t)) is time invariant,' but U_r(A(t)) is a set of r-dimensional frames, not a single subspace. The precise meaning of this assumption should be stated in terms of the fixed coordinate form of the frames.

Circularity Check

2 steps flagged

The main Oja-flow tracking theorem is not a circular fit, but the low-rank filter step is justified by an unpublished same-author preprint, and Theorem 3's proof proves only subspace tracking for r>1 (a correctness gap, not circularity).

specific steps
  1. self citation load bearing [Section 4, low-rank Kalman-Bucy filter gain definition (after eq. (7) and before eq. (8))]
    "From the result of our follow-up paper [7], the definition of G_Uε(t) is different from the original proposal of the low-rank Kalman-Bucy filter [9]."

    The projected noise gain G_Uε(t)=(V_ε^T U_ε)^{-1}V_ε^T G is the object that makes the reduced Riccati equation a low-rank KBF, and it is not derived in this paper. Its justification is a citation to the same authors' unpublished preprint [7]. The numerical experiment then uses that formula as if it were established. This is a load-bearing self-citation for the filtering application, though not for the main tracking theorem.

  2. self citation load bearing [Section 3, Proposition 1 statement/proof, and Appendix A]
    "The following result is a generalization of Theorem 9 in [7] ... The proof is almost the same as that of Theorem 9 in [7] ... The similar arguments in the proof of Lemma 7 in [7] show that (I_n + G(t,t_0)P_0) is invertible for all t≥t_0."

    The paper delegates the Stiefel-manifold/full-rank preservation and singular-value convergence arguments to Theorem 9 and Lemmas 7-8 of the same authors' preprint [7]. This is not the central tracking theorem and much of it can be proved elementarily, but as written the justification chain passes through an unverified same-author source.

full rationale

The main derivation chain is not circular. Theorem 3 does not fit a parameter and call it a prediction: the ε-dependence and spectral-gap bound are obtained by a perturbation/linearization argument in the rotating frame, and Assumptions 1-5 are stated structural restrictions, not the conclusion. The numerical examples are demonstrations for systems constructed to satisfy those restrictions ('Notice that the matrix A(t) satisfies the conditions of Theorem 3'), so they do not constitute independent validation of the general claim, but they are not fitted outputs either. However, the paper is not fully self-contained: the low-rank filter gain and parts of Proposition 1 rest on same-authors' unpublished preprint [7], which is the reason the score is above 2. I also note a serious proof/statement mismatch that is not circularity: in Theorem 3 the proof bounds only the off-subspace component L_⊥ and concludes tracking of the dominant subspace, while the theorem statement promises a neighborhood of a particular basis Ū(t); for r>1, all orthonormal frames in the dominant invariant subspace are equilibria, so a neutral rotation inside the subspace can keep U exactly on the subspace while moving away from Ū(t). This is a correctness concern, not an equivalence-to-input concern, and I have not counted it as circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No new physical or mathematical entities are postulated and no parameters are fitted to data. The main burden lies in the rigid-eigenframe assumptions 4-5 of Theorem 3, which are custom restrictions that make the proof tractable rather than empirically supported facts.

free parameters (2)
  • ε (Oja flow time constant) = chosen small; e.g. 0.01 in simulations
    Explicit tuning parameter of the flow. Theorem 3 only asserts existence of a sufficiently small ε; no closed-form selection rule is derived. In the numerics ε=1, 0.1, 0.01 are chosen by hand.
  • a (positive shift in Proposition 1) = taken large enough so A_sym(t)+aI_n > cI_n
    A shift added to force convergence to the Stiefel manifold. It is not fitted to data but is assumed to exist by boundedness of A(t).
axioms (5)
  • standard math Riccati comparison lemma (Lemma 4 from [14])
    Used in the proof of Proposition 1 to represent the solution of the P=UU^T Riccati equation. It is quoted from the literature and not reproved.
  • standard math Constant-matrix Oja flow convergence (Theorems 9 and 12 of [7])
    The paper relies on the known exponential convergence of (1) for constant A with a spectral gap, and extends it to the time-varying case by perturbation. [7] is a same-author preprint but the result is treated as established background.
  • domain assumption A(t) is bounded and a shift exists with A_sym(t)+aI_n > cI_n
    Boundedness is stated at the start of Section 3; the shift is used in Proposition 1 to guarantee convergence to the Stiefel manifold. This is a modeling assumption on the system class.
  • domain assumption Λ(t) is diagonal and has a uniform spectral gap c1
    Assumptions 1-2 of Theorem 3; needed for the linearized error dynamics to have a stable contraction with rate c1/ε.
  • ad hoc to paper Ψ(t)^†Ψ(t) is constant and Ψ(t)^{-1}U_r(A(t)) is time-invariant
    Assumptions 4-5 of Theorem 3. These are introduced specifically to make the coordinate transformation K=Ψ^{-1}U reduce the problem to a bounded perturbation of a fixed equilibrium. They define the restrictive class and are not generally true for LTV systems.

pith-pipeline@v1.3.0-daily-deepseek · 9565 in / 24241 out tokens · 242222 ms · 2026-08-03T14:57:21.400985+00:00 · methodology

0 comments
read the original abstract

This paper studies a low-rank Kalman-Bucy filtering framework for linear time-varying systems through the tracking analysis of Oja's principal component flow. Under structured assumptions on the eigenspaces and their time variation, we show that the Oja flow can remain in a neighborhood of the time-varying dominant subspace by tuning a parameter of the flow, rather than tracking it exactly. These restrictive assumptions identify a tractable class of linear time-varying systems and provide a theoretical basis for low-rank filtering, as illustrated by a numerical experiment.

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

14 extracted references

  1. [1]

    R. E. Kalman. A new approach to linear filtering and prediction problems.Journal of Basic Engineering, 82(1):35–45, 1960

    1960

  2. [2]

    R. E. Kalman and R. S. Bucy. New results in linear filtering and prediction theory.Journal of Basic En- gineering, 83(3):95–108, 1961

    1961

  3. [3]

    A. C. Antoulas.Approximation of Large-Scale Dy- namical Systems. SIAM, 2005

    2005

  4. [4]

    A. C. Antoulas, C. A. Beattie, and S. G ¨u˘gercin.Inter- polatory Methods for Model Reduction. SIAM, 2020

    2020

  5. [5]

    Balancing-related model reduction methods: System-and Data-Driven Methods and Algorithms

    Tobias Breiten and Tatjana Stykel. Balancing-related model reduction methods: System-and Data-Driven Methods and Algorithms. InModel Order Reduction, vol. 1, chap. 2, pages 15–56. De Gruyter Berlin, 2021

    2021

  6. [6]

    De- tectability Conditions and State Estimation for Linear Time-Varying and Nonlinear Systems.SIAM J

    Markus Tranninger, Richard Seeber, Martin Stein- berger, Martin Horn, and Christian P ¨otzsche. De- tectability Conditions and State Estimation for Linear Time-Varying and Nonlinear Systems.SIAM J. Con- trol and Optimization, 60(4):2514–2537, 2022

    2022

  7. [7]

    Global Con- vergence of Oja’s Component Flow for General Square Matrices and Its Applications.Preprint arXiv:2510.00801v2, 2025

    Daiki Tsuzuki and Kentaro Ohki. Global Con- vergence of Oja’s Component Flow for General Square Matrices and Its Applications.Preprint arXiv:2510.00801v2, 2025

    arXiv 2025

  8. [8]

    Asymptotic analysis of generalized orthog- onal flows.Linear Algebra and its Applications, 717(15):1–25, 2025

    Yueh-Cheng Kuo, Huey-Er Lin, and Shih-Feng Shieh. Asymptotic analysis of generalized orthog- onal flows.Linear Algebra and its Applications, 717(15):1–25, 2025

    2025

  9. [9]

    On a New Low- Rank Kalman-Bucy Filter and Its Convergence Prop- erty

    Shuto Yamada and Kentaro Ohki. On a New Low- Rank Kalman-Bucy Filter and Its Convergence Prop- erty. InThe 52nd ISCIE International Symposium on Stochastic Systems Theory and Its Applications (SSS ’20), volume 2021, pages 16–20, Osaka, 2020

    2021

  10. [10]

    Comparison of Estimation Error between Two Different Low-Rank Kalman-Bucy Filters

    Shuto Yamada and Kentaro Ohki. Comparison of Estimation Error between Two Different Low-Rank Kalman-Bucy Filters. In2021 60th Annual Confer- ence of the Society of Instrument and Control Engi- neers of Japan (SICE), September 2021

    2021

  11. [11]

    Low-rank ap- proximated Kalman-Bucy filters using Oja’s princi- pal component flow for linear time-invariant systems

    Daiki Tsuzuki and Kentaro Ohki. Low-rank ap- proximated Kalman-Bucy filters using Oja’s princi- pal component flow for linear time-invariant systems. IEEE Control Systems Letters, 8:1583–1588, 2024

    2024

  12. [12]

    The ge- ometry of low-rank Kalman filters

    Sivere Bonnabel and Rodolphe Sepulchre. The ge- ometry of low-rank Kalman filters. InMatrix Infor- mation Geometry, pages 53–68. Springer, 2012

    2012

  13. [13]

    On the finite escape phenomena for matrix Riccati equations.IEEE Trans

    Tetsushi Sasagawa. On the finite escape phenomena for matrix Riccati equations.IEEE Trans. Autom. Contr., 27(4):977–979, 1982

    1982

  14. [14]

    Sinan Kilicaslan and Stephen P. Banks. Exis- tence of solutions of Riccati differential equations. J. Dynamical Systems, Measurement, and Control, 134(3):031001, 2012. APPENDIX A. PROOF OF PROPOSITION 1 We first introduce the following lemma that generalizes the results in [13]. Lemma 4 ( [14, Theorems 13 and 14]) Consider the follow- ing Riccati differen...

    2012