Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Convergence Analysis of Natural Power Method and Its Applications to Control

T0 review · 2 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The discrete-time natural power method provably converges to the r-dimensional subspace of the eigenvalues with largest absolute value, enabling control applications.

desk verdict The convergence proof is broken and the result is classical; honest applications don't save it. read the letter →

arxiv 2512.21469 v2 pith:WKP3HYTC submitted 2025-12-25 math.OC cs.SYeess.SY

Daiki Tsuzuki , Kentaro Ohki This is my paper · ORCID
classification math.OCcs.SYeess.SY MSC 65F1593B1193C55
keywords naturalpowermethoddominantsubspacetrackingprincipalcomponentfloweigenvaluemagnitudeorderingmodelorderreductiondiscrete-timeLTIsystemslow-rankcontrollersynthesislineartime-varying
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 proves that the discrete-time natural power iteration — $U[k+1] = A U[k] (U[k]^T A^T A U[k])^{-1/2}$ — converges to the r-dimensional invariant subspace belonging to the r eigenvalues of A with largest absolute value, for almost all initial conditions. This differs from the standard continuous-time principal-component flow, which targets eigenvalues with the largest real part, so the two methods extract genuinely different subspaces for non-symmetric matrices. The convergence theorem, a generalization of a known result for symmetric positive semidefinite matrices, also shows that the projection onto the limit subspace preserves the dominant r eigenvalues. On this basis the paper builds a model order reduction method for discrete-time LTI systems that preserves reachability, observability, and the transfer function, and a low-rank observer/feedback controller for slowly-varying LTV systems that tracks time-varying dominant subspaces.

What carries the argument

The engine of the analysis is the coordinate transformation $\Lambda = \Psi^{-1} A \Psi$, where $\Psi$ is the eigenvector matrix, which sends the iteration into block form with a dominant block $\Lambda_m$ and a minor block $\Lambda_\perp$. The invariant set $V^{(d)}_{m,r}$ ensures that the dominant block of $U[k]$ keeps full rank; the proof shows the minor block decays like $(|\lambda_{m+1}|/|\lambda_m|)^k$, because the normalization factors in the iteration scale both blocks by the same $|\lambda_m|^{-k}$. This block decomposition is what converts the nonlinear matrix iteration into a ratio comparison of eigenvalue magnitudes. A second piece of machinery is the volume argument that $V^{(d)}_{m,r}$ has the same volume as the Stiefel manifold, which is why

What would settle it

Take a defective matrix with a strict spectral gap, e.g., $A = \begin{bmatrix} 2 & 1 & 0 \\ 0 & 2 & 0 \\ 0 & 0 & 1 \end{bmatrix}$ (eigenvalues 2,2,1) with $r=1$ and $m=2$, initialize $U[0]$ as a generic unit vector, and run the iteration (1). If the iterates fail to converge to the dominant eigenspace $\operatorname{span}([1,0,0]^T)$, or if the paper's invariant set $V^{(d)}_{2,1}$ cannot even be defined because the eigenvector matrix is singular, then the 'general square matrix' claim in the abstract is not supported by the proof.

Watch

Extended reading notes

Core claim

The central result is Theorem 3: if A has a gap in the magnitudes of its eigenvalues at index m $(|\lambda_m| > |\lambda_{m+1}|)$ and the initial condition $U[0]$ lies in the invariant set $V^{(d)}_{m,r}$ — the Stiefel matrices whose top block in the eigenvector basis is full-rank — then the iterates $U[k]$ converge to $U^{(d)}_{m,r}$, the set of Stiefel matrices whose bottom block vanishes. In the case $m=r$, the limit subspace is spanned by the r right eigenvectors of A with largest absolute values, and the projected matrix $\bar{U}_r^T A \bar{U}_r$ is similar to the diagonal block of the r dominant eigenvalues, so those eigenvalues are exactly preserved. The paper further shows that the natural power method's fixed

Load-bearing premise

The proof assumes A is diagonalizable — that is, it has a complete basis of eigenvectors — so the convergence theorem does not cover defective matrices, even though the abstract claims general square matrices.

Editorial extensions

If this is right

  • The natural power method extracts the dominant subspace by eigenvalue magnitude, making it the discrete-time analogue of power iteration for absolute values rather than real parts; this is the correct subspace for discrete-time dynamical systems where stability depends on |λ|.
  • For almost every initial condition (in the measure-theoretic sense), the iterates converge to the dominant m-dimensional subspace whenever a strict magnitude gap exists.
  • The reduced-order model formed from the limit subspace preserves the dominant r eigenvalues, and the left/right subspaces allow construction of reduced models that keep reachability, observability, and the transfer function unchanged (Proposition 7).
  • The subspace-dimension reduction trick (Eq. 10) lets one compute a smaller dominant subspace from the projected matrix without revisiting the full n×n matrix, cutting cost when r ≪ n.
  • For slowly-varying LTV systems, the method can track a time-varying dominant subspace; numerical experiments show the controller works when the magnitude gap is large, and degrades as the gap closes.

Reading between the lines

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

  • The convergence theorem as stated in the paper relies on A being diagonalizable, since the proof uses an invertible eigenvector matrix Ψ[A]. The paper's abstract claims general square matrices; extending the proof to defective (non-diagonalizable) matrices, or showing failure, would settle the scope of the result.
  • Because the iteration has no step size or learning rate, it offers a parameter-free way to track time-varying dominant subspaces; the slow-tracking failure at α=0.9 suggests a quantitative bound on the admissible rate of variation could be derived from the magnitude gap.
  • The method's fixed points are not the dominant subspace itself (Remark 4), which is why the paper adds the modified iteration (9). A natural next step is to check whether the original iteration's limit points, rather than just the subspace, can be characterized — they appear to depend on the initial condition.
  • The MOR error analysis is explicitly left for future work; a natural test is whether the transfer-function approximation error inherits the |λ_{r+1}|/|λ_r| convergence rate, as in classical balanced truncation.
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

2 major / 4 minor

Summary. The paper studies the discrete-time natural power method U[k+1] = A U[k] (U[k]^T A^T A U[k])^{-1/2} for general real square matrices. Its central claim, Theorem 3, is that if the initial subspace lies in V^(d)_{m,r} and |λ_m[A]| > |λ_{m+1}[A]|, then the iterates converge to the set U^(d)_{m,r} of r-dimensional subspaces of the dominant m-dimensional invariant subspace, so that for m=r the r eigenvalues of largest absolute value are preserved. The paper then applies this result to model order reduction for discrete-time LTI systems and to low-rank observer/feedback design for a slowly varying LTV example, contrasting the extracted subspace with that of the Oja flow.

Significance. If Theorem 3 were correctly proved, the paper would give a useful convergence result for the natural power method beyond the symmetric positive semidefinite case, with interesting consequences for model reduction and low-rank control. The observation that the method extracts the dominant absolute-value subspace is valuable and complements the Oja-flow results cited in the paper. The MOR transfer-function preservation and the LTV numerical illustration are promising. However, the main proof as written contains a load-bearing algebraic error, and the stated generality is not supported by the proof. The paper does not provide machine-checked proofs or reproducible code; the numerical experiments are illustrative. The underlying convergence claim is very likely true and can be proved by a standard subspace-iteration argument, so the work is potentially salvageable through a substantial revision.

major comments (2)
  1. [Theorem 3 proof (after Lemma 2)] The proof rewrites the expression for K_perp^(k), correctly obtained in Lemma 2 as Λ_perp^k K_perp Z[0]^{-1/2} Z[1]^{-1/2} ... Z[k-1]^{-1/2}, into Λ_perp^k K_perp (Z[k-1] ... Z[0])^{-1/2}. The matrices Z[i] = U[i]^T A^T A U[i] are symmetric positive definite but do not commute for general non-normal A. Hence the product Z[k-1] ... Z[0] is not symmetric positive definite, and its inverse square root is not defined under the paper's convention. Even with a non-symmetric square root, it is not the ordered product of the individual inverse square roots. The subsequent step of factoring |λ_m|^{-2k} inside the square root and claiming that |λ_m|^{-2k} Z[k-1] ... Z[0] 'remains positive or diverges' is therefore unjustified. This invalidates the contraction argument for K_perp^(k), which is the core of the theorem.
  2. [Lemma 2 / Notation; Theorem 3 statement] The proof assumes that A is diagonalizable, or at least that Ψ[A] is invertible and Λ = Ψ^{-1} A Ψ is block diagonal. This is not stated in Theorem 3 or in the abstract's claim for 'general square matrices'. For defective A, the eigenvector matrix Ψ[A] is singular, the representation U[0] = Ψ [K_{m,r}; K_perp] is unavailable, and the lower bound λ_n[Ψ^†Ψ] used in Lemma 2 is zero. The theorem must either be restricted to diagonalizable A or be supplied with a Jordan/Schur-based proof. As written, the advertised generality is unsupported.
minor comments (4)
  1. [Theorem 3 and Remark 4] The theorem states that U[k] 'converges to U^(d)_{m,r}', but the proof only shows convergence to the set in the sense that K_perp^(k) tends to zero. Since U[k] may rotate within the target subspace (as Remark 4 explains), the convergence statement should be phrased as convergence of the projection U[k]U[k]^T or as distance to the set.
  2. [Lemma 2 / Eq. (5)] The proof uses complex eigenvectors and the Hermitian conjugate while the algorithm is defined on the real Stiefel manifold with transpose. The identities such as Eq. (5) rely on U^T = U^† for real U and on K^†Ψ^†ΨK = I_r. This is correct but should be stated explicitly so that the complex parameterization of real Stiefel matrices is clear.
  3. [End of Theorem 3 proof] The claim that the volume of V^(d)_{m,r} equals that of St(r,n) is imported from a companion paper without proof. Since this is used for the 'almost all initial matrices' statement, a self-contained argument, or at least a precise full-measure statement, should be included.
  4. [Section 4.2 and Abstract] The LTV controller section is explicitly an illustrative numerical study and the paper admits the method fails for α=0.9 and that 'further analysis is required'. The abstract's wording 'showing its utility for tracking time-varying dominant subspaces' overstates what is proved. Please soften the claim or add a rigorous tracking theorem.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central convergence theorem is proved directly from the iteration, with only minor non-load-bearing self-citations in the applications.

full rationale

Theorem 3 is obtained by direct manipulation of the recursion (1): Lemma 2 carries the coordinate representation U[k]=Ψ[K^{(k)}_{m,r}; K^{(k)}_⊥] and expresses K^{(k)}_⊥ as Λ^k_⊥K_⊥ times a product of inverse square roots; Theorem 3 then invokes |λ_m|>|λ_{m+1}| to force K_⊥→0. The conclusion is not an input to the iteration, and no parameter is fitted to the claimed convergence. The main self-citations are not load-bearing for this theorem. The 'almost all' qualification imports a volume statement from (Tsuzuki and Ohki, 2025, §III.C), but that is a parameter-free measure assertion, not a definition of the target subspace and not a fitted value. In the applications, Eq. (11), the reachability/observability preservation claims, and Proposition 7 are referred to the companion paper ('As shown in ... Props. 21–23', 'The proof is identical ... and is therefore omitted'); these are self-citations, but they support peripheral MOR/controller claims rather than the central convergence theorem, and they do not assume the theorem's conclusion. The equilibrium property (Eq. (8)) is likewise cited from Tsuzuki and Ohki 2024a without reduction. The proof of Theorem 3 has a genuine gap—(Z[k−1]...Z[0])^{−1/2} is not defined by the paper's PSD-square-root convention because the Z's need not commute, and the boundedness of the rescaled product is asserted—but this is a correctness risk, not circularity. Similarly, the unstated diagonalizability assumption is a gap in scope, not a self-referential inference. The LTV section is explicitly preliminary ('Further analysis ... is required'), so the simulation is not presented as a derived prediction. Score 2 reflects the minor, non-central self-citations.

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

The analysis relies on standard eigenvector assumptions but no fitted parameters or invented entities. The main gaps are the unstated diagonalizability assumption and the invalid product-of-square-roots step.

assumptions (5)
  • domain assumption A is diagonalizable with an invertible eigenvector matrix Ψ[A].
    Used throughout Section 3 to write Λ = Ψ^{-1}AΨ and U[k] = Ψ[K;...]; not stated in the paper, and excludes defective matrices from the 'general square matrices' claim.
  • domain assumption Strict eigenvalue gap |λ_m[A]| > |λ_{m+1}[A]|.
    Stated in Theorem 3; necessary for exponential decay of the lower-block coefficient K_⊥. Without this separation, convergence to a dominant subspace is not guaranteed.
  • domain assumption Initial U[0] belongs to V^{(d)}_{m,r}, i.e., it has full-rank component in the dominant m-dimensional eigensubspace.
    Assumed in Lemma 2 and Theorem 3; avoids initial conditions orthogonal to the dominant subspace, for which convergence fails.
  • domain assumption A is non-singular in Lemma 2.
    Used to guarantee U[k]^T A^T A U[k] is non-singular; the theorem as stated does not require it, but the proof relies on this condition.
  • domain assumption The volume of V^{(d)}_{m,r} equals that of St(r,n).
    Taken from (Tsuzuki and Ohki, 2025, §III.C), a self-cited result; used to claim almost-all initial matrices converge.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Convergence Analysis of Natural Power Method and Its Applications to Control." pith.science (2026). https://pith.science/paper/WKP3HYTC

@misc{pith2026251221469,
  author       = {Pith},
  title        = {Pith review of: Convergence Analysis of Natural Power Method and Its Applications to Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKP3HYTC}},
  note         = {Machine review of arXiv:2512.21469}
}
abstract

This paper analyzes the discrete-time natural power method, demonstrating its convergence to the dominant $r$-dimensional subspace corresponding to the $r$ eigenvalues with the largest absolute values. This contrasts with the Oja flow, which targets eigenvalues with the largest real parts. We leverage this property to develop methods for model order reduction and low-rank controller synthesis for discrete-time LTI systems, proving preservation of key system properties. We also extend the low-rank control framework to slowly-varying LTV systems, showing its utility for tracking time-varying dominant subspaces.

Figures

Figures reproduced from arXiv: 2512.21469 by the authors.

Figure 1
Figure 1. Plot of ∥U[k]U[k] ⊤ − Π12∥ind for each time k. The black solid, red dashed, blue chain, and green dotted lines represent α = 0, 0.2, 0.5, 0.9, respectively. If r = 1, U[k] converges to U (d) 2,1 similarly to the r = 2 case (see [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Plot of ∥U[k]U[k] ⊤ − Π12U[k]U[k] ⊤Π12∥ind for each time k, where U[k] ∈ R 3 . The black solid, red dashed, blue chain, and green dotted lines represent α = 0, 0.2, 0.5, 0.9, respectively. Remark 4. The solution U[k] of (1) generally does not converge to an element of U (d) m,r even if m = r under the conditions of Theorem 3. We briefly show this fact below. If m = r, any element U¯ r of U (d) r can be represented b… view at source ↗
Figure 3
Figure 3. Plot of AU[k] = U[k] ⊤AαU[k] ∈ R at each time k. The black solid, red dashed, blue chain, and green dotted lines represent α = 0, 0.2, 0.5, 0.9, respectively. where AU¯r := U¯ ⊤ r AU¯ r. If U[k] = U¯ r ∈ U(d) r , then U[k + 1] =U¯ rAU¯r (A ⊤ U¯r AU¯r ) −1/2 . Recall that AU¯r ∈ R r×r is invertible. After ℓ steps, we have U[k + ℓ] = U¯ r{AU¯r (A ⊤ U¯r AU¯r ) −1/2 } ℓ . Therefore, U¯ r is a stationary point if and onl… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Plot of ∥x[k]∥ for each time k. The black solid, the red dashed, the blue chain, and green dotted lines represent α = 0, 0.2, 0.5, 0.9, respectively. 0 20 40 60 80 100 Steps 10 -15 10 -10 10 -5 10 0 k x[k] ! ^x[k]k , =0 , =0.2 , =0.5 , =0.9 [PITH_FULL_IMAGE:figures/fu…
Figure 5
Figure 5. Figure 5: Plot of ∥x[k]−xˆ[k]∥ for each time k. The black solid, the red dashed, the blue chain, and green dotted lines represent α = 0, 0.2, 0.5, 0.9, respectively. 5. CONCLUSION We investigated the convergence of the discrete-time nat￾ural power method, demonstrating its abili…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Quantum model reduction based on Oja's flow

    quant-ph 2026-07 conditional novelty 6.0 of 10

    Oja's flow applied to Lindblad generators gives reduced quantum models of the slow subspace, and a tensor-product-constrained version preserves complete positivity and finds approximate decoherence-free subspaces.

Reference graph

Works this paper leans on

16 extracted references · cited by 1 Pith paper

  1. [1]

    and Abed-Meraim, K

    Attallah, S. and Abed-Meraim, K. (2002). Fast algorithms for subspace tracking . IEEE Signal Processing Letters, 8(7), 203--206

    2002

  2. [2]

    Badeau, R., David, B., and Richard, G. (2005). Fast approximated power iteration subspace tracking . IEEE Transactions on Signal Processing, 53(8), 2931--2941

    2005

  3. [3]

    Chen, T., Hua, Y., and Yan, W.Y. (1998). Global convergence of Oja's subspace algorithm for principal component extraction . IEEE Transactions on Neural Networks, 9(1), 58--67

    1998

  4. [4]

    Hua, Y. (2004). Asymptotical orthonormalization of subspace matrices without square root . IEEE Signal Processing Magazine, 21(4), 56--61

    2004

  5. [5]

    Hua, Y., Nikpour, M., and Stoica, P. (2001). Optimal reduced-rank estimation and filtering . IEEE Transactions on Signal Processing, 49(3), 457--469

    2001

  6. [6]

    Hua, Y., Xiang, Y., Chen, T., Abed-Meraim, K., and Miao, Y. (1999). A new look at the power method for fast subspace tracking . Digital Signal Processing, 9(4), 297--314

    1999

  7. [7]

    Kuo, Y.C., Lin, H.E., and Shieh, S.F. (2025). Asymptotic analysis of generalized orthogonal flows . Linear Algebra and its Applications, 717(15), 1--25

    2025

  8. [8]

    Oja, E. (1982). Simplified neuron model as a principal component analyzer . Journal of Mathematical Biology, 15, 267--273

    1982

Show all 16 references
  1. [9]

    Oja, E. (1989). Neural networks, principal components, and subspaces . International Journal of Neural Systems, 1(01), 61--68

    1989

  2. [10]

    and Karhunen, J

    Oja, E. and Karhunen, J. (1985). On stochastic approximation of the eigenvectors and eigenvalues of the expectation of a random matrix. Journal of Mathematical Analysis and Applications, 106(1), 69--84

    1985

  3. [11]

    Tranninger, M., Seeber, R., Steinberger, M., Horn, M., and P \"o tzsche, C. (2022). Detectability Conditions and State Estimation for Linear Time-Varying and Nonlinear Systems . SIAM Journal on Control and Optimization, 60(4), 2514--2537

    2022

  4. [12]

    and Ohki, K

    Tsuzuki, D. and Ohki, K. (2024 a ). Low-rank approximated Kalman-Bucy filters using Oja's principal component flow for linear time-invariant systems . IEEE Control Systems Letters, 8, 1583--1588

    2024

  5. [13]

    and Ohki, K

    Tsuzuki, D. and Ohki, K. (2024 b ). Low-rank approximated Kalman filter using Oja's principal component flow for discrete-time linear systems . In 2024 SICE Festival with Annual Conference (SICE FES) , 242--247

    2024

  6. [14]

    and Ohki, K

    Tsuzuki, D. and Ohki, K. (2025). Global Convergence of Oja's Component Flow for General Square Matrices and Its Applications . Preprint arXiv:2510.00801v2

    2025

  7. [15]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence afte...

  8. [16]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.