Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Standard LSParameter Estimators Ensure Finite Convergence Time for Linear Regression Equations Under an Interval Excitation Assumption

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proves that the classical least-squares estimator with forgetting factor reconstructs unknown parameters exactly in finite time whenever the regressor is interval-excited and bounded, with the convergence time equal to the…

desk verdict The math is sound but the title overstates: the finite-time convergence holds for a post-processed signal, not the raw LS estimate itself. read the letter →

arxiv 2506.08211 v1 pith:DWHSRRO5 submitted 2025-06-09 eess.SY cs.SYmath.STstat.TH

Romeo Ortega , Jose Guadalupe Romero , Stanislav Aranovskiy , Gang Tao This is my paper
classification eess.SYcs.SYmath.STstat.TH MSC 93B3093C40
keywords Least-squaresestimationfinite-timeconvergencelinearregressionequationintervalexcitationforgettingfactorparameteridentifiability
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 the standard least-squares (LS) estimator, fitted with a forgetting factor, has finite-time convergence for linear regression equations—a property usually obtained only by modifying the estimator with high-gain injection or signal differentiation. Under the assumptions that the regressor is interval-excited (IE) and bounded, the paper constructs a signal from the LS covariance and parameter estimates that equals the unknown parameter exactly for all times after the excitation interval ends. The convergence time is exactly the length of the interval on which the IE condition holds. Since IE is known to be necessary and sufficient for identifiability, this means finite-time identification is achieved under the weakest possible excitation assumption, with no change to the classical algorithm's structure.

What carries the argument

The load-bearing identity is the constancy of $F^{-1}(t)\widetilde\theta(t)$ along the LS flow, where $F(t)$ is the covariance matrix and $\widetilde\theta(t)$ is the parameter error. Rearranged as $\widetilde\theta(t)=F(t)F^{-1}(0)\widetilde\theta(0)$, it lets $\theta$ be recovered as $\theta = [I_q - z(t)f_0F(t)]^{-1}[\widehat\theta(t)-z(t)f_0F(t)\theta_0]$ once the matrix in brackets is invertible. The interval excitation condition $\int_0^{T_c}\varphi(\tau)\varphi^\top(\tau)d\tau \ge \rho I_q$ guarantees that invertibility for every $t\ge T_c$, and the scalar variable $z(t)$, driven by $\chi(t)$, keeps the reconstruction bounded while preserving the standard least-squares structure.

What would settle it

Simulate the standard LS estimator (4a)–(4d) with a regressor that is interval-excited on $[0,T_c]$ but unbounded (for instance, a signal whose energy on the interval satisfies the IE inequality while its amplitude grows), and check whether the signal (5) satisfies $\theta_{\mathrm{FCT}}(t)=\theta$ for all $t\ge T_c$; if not, the boundedness assumption is essential to the claim.

Watch

Extended reading notes

Core claim

The central claim is Proposition 1: for the scalar-output linear regression equation $y(t)=\varphi^\top(t)\theta$ with unknown constant $\theta$, if $\varphi$ is interval-excited and bounded, then the standard LS estimator with forgetting factor (equations (4a)–(4d)) produces a signal $\theta_{\mathrm{FCT}}(t)$, defined in (5), that satisfies $\theta_{\mathrm{FCT}}(t)=\theta$ for all $t\ge T_c$, where $T_c$ is the length of the interval excitation interval. The proof rests on the identity $\frac{d}{dt}\big(F^{-1}(t)\widetilde\theta(t)\big)=0$, which yields $\widetilde\theta(t)=F(t)F^{-1}(0)\widetilde\theta(0)$; rearranging this identity expresses the unknown parameter in terms of measured quantities and the covariance matrix. The IE condition forces the matrix $I_q - z(t)f_0F(t)$ to be full rank for $t\ge T_c$, which converts the usual asymptotic LS convergence into exact finite-time equality. The paper also shows all signals remain bounded, and a simulation comparison indicates that the resulting estimator is far less sensitive to measurement noise than a high-gain FCT alternative.

Load-bearing premise

The argument requires the regressor $\varphi(t)$ to be bounded in addition to being interval-excited; if an IE regressor can grow without bound, the finite-time reconstruction may fail even though the parameters remain identifiable.

Editorial extensions

If this is right

  • Finite-time parameter identification for linear regression equations can be obtained with an off-the-shelf LS estimator; only interval excitation and boundedness of the regressor are required.
  • The convergence time equals the length of the excitation interval, so the speed of identification is governed by the data's excitation, not by aggressive gain tuning.
  • Because IE is necessary and sufficient for identifiability, the finite-time property is realised under the weakest excitation condition that makes the problem solvable at all.
  • When measurement noise is present, the FCT estimate oscillates around the true parameter with a smaller amplitude than the high-gain FCT estimator in the simulated example, suggesting a practical robustness advantage for the standard LS reconstruction.

Reading between the lines

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

  • The boundedness assumption on $\varphi$ is likely an artifact of the proof technique rather than a fundamental requirement; testing whether IE alone, or a milder growth condition, suffices would clarify the true scope of the result.
  • The same constancy identity may endow other LS variants—normalized, discrete-time, or with alternative forgetting mechanisms—with finite-time convergence under IE, indicating a structural phenomenon rather than a feature of the specific equations (4a)–(4d).
  • In implementations, $T_c$ need not be known in advance: one can monitor the determinant of $I_q - z(t)f_0F(t)$ online and adopt $\theta_{\mathrm{FCT}}(t)$ the moment it leaves a neighbourhood of zero.
  • A formal sensitivity analysis of the reconstruction (5) with respect to noise could quantify when the FCT-LS estimator outperforms high-gain designs, going beyond the paper's single simulation comparison.
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 paper claims that, for a linear regression equation y(t)=φ(t)^Tθ with an interval-excited and bounded regressor φ(t), the standard least-squares estimator with forgetting factor (equations (4a)-(4d)) ensures finite convergence time (FCT) of the estimated parameters. The main formal result, Proposition 1, actually defines a transformed signal θ_FCT(t) in equation (5) and proves, by reference to previous results, that θ_FCT(t)=θ for all t≥T_c. The paper also recalls that interval excitation is necessary and sufficient for identifiability and presents a simulation comparison among the proposed FCT-LS scheme, standard LS, and a high-gain FCT estimator under measurement noise.

Significance. The underlying algebraic observation, that a simple post-processing of the standard LS state can produce an estimate that becomes exact in finite time under interval excitation, is valuable and relevant to parameter estimation. The paper also gives a useful simulation example showing that this FCT-LS scheme is substantially less noise-sensitive than a high-gain FCT estimator. The credit for these contributions is limited, however, because the central advertised claim is overstated: the finite-time property is proved for the transformed signal θ_FCT, not for the raw LS estimate θ̂. If that claim is corrected and the proof is made self-contained, the note would be a worthwhile contribution.

major comments (3)
  1. [Abstract and Proposition 1] The advertised claim is not supported by the proposition. Proposition 1 proves only that θ_FCT(t)=θ for t≥T_c, where θ_FCT is the nonlinear algebraic transformation defined in equation (5), not that the standard LS estimate θ̂(t) converges in finite time. In fact, following the computation in Section 2 and the analogous derivation for (4), one obtains θ̃(t)=z(t)f0F(t)θ̃(0), so θ̂(T_c)≠θ for generic initial errors because z(T_c)f0F(T_c) is generically nonzero. The title and abstract therefore overstate the result. The claim should be reformulated as 'a simple algebraic transformation of the standard LS estimator yields a finite-time convergent estimate' or equivalent.
  2. [Section 3, Proof 1] The proof is not self-contained and does not establish the key step. It cites Tao (2003, Lemma 3.5) and Ortega et al. (2022b, Proposition 1) without deriving the invertibility of I−z(t)f0F(t) for t≥T_c, which is exactly what makes (5) well-defined and equal to θ. Moreover, Proposition 1(ii) states that all signals are bounded, but no argument for that part is provided. Since the result is short, the authors should include the missing derivation (e.g., compute M(t)=z(t)f0F(t), show that M^{-1}(t)=I+(γ_F/f0)∫_0^t z^{-1}(s)φ(s)φ^T(s)ds, and use IE to prove definiteness of M^{-1}(t)−I) or state the referenced lemmas in full with matching assumptions.
  3. [Section 3, Proposition 1 vs. abstract and Lemma 1] The proposition assumes φ(t) bounded, but this assumption is absent from the abstract and from the statement that IE is the weakest sufficient condition (Lemma 1). As written, the main result does not justify the headline claim for all IE regressors. The boundedness condition appears removable because z(t)>0 and IE imply that M^{-1}(t)−I is positive definite for t≥σ0+T_c, but the manuscript does not make this argument. The authors should either state the boundedness assumption in the abstract and main claim or remove it from the proposition if it is unnecessary.
minor comments (5)
  1. [Equation (3) and footnote 2] Equation (3) defines IE with integration from 0 to T_c, while footnote 2 allows the interval [σ0,σ0+T_c]. If σ0>0, the conclusion θ_FCT(t)=θ holds for t≥σ0+T_c, not t≥T_c. Please clarify the convention used in Proposition 1.
  2. [Section 1] There is a typo in 'mathemathically skillful'; it should be 'mathematically skillful'.
  3. [Section 2] The phrase 'equation that appears in (de Larminat, 1984, equation (17))' should be 'an equation that appears'.
  4. [Figures 1 and 3] The captions are inconsistent: Figure 1 refers to 'estimated parameters θ̃_i' while Figure 3 refers to 'estimated parameters θ̂_i'. The tilde denotes the parameter error, not the parameter estimate, so the captions should be aligned.
  5. [Section 4] In the simulation comparison, the 'FCT-LS' curve is the transformed signal θ_FCT computed with a threshold on det(I−z(t)f0F(t)), whereas the 'standard LS' curve is the raw estimate θ̂. This is a comparison of two different objects; the text should state explicitly that the FCT-LS curve is the post-processed estimate, not the output of the LS recursion itself.

Circularity Check

2 steps flagged · score 6.0 of 10

The central claim reduces to an algebraic output transformation plus a self-cited full-rank lemma: FCT is proved for the post-processed θ_FCT in Eq. (5), not for the raw LS estimate, so the advertised result is partly definitional.

  1. self definitional [Section 3, Proposition 1 (Eq. 5)]
    "For t ≥ Tc, define the signal θF CT (t) := [ Iq −z(t)f0F (t)]−1[ˆθ(t)−z(t)f0F (t)θ0]. (i) For all initial conditions this signal verifies θF CT (t) = θ, ∀t ≥ Tc."

    From the invariant (2) in Section 2, θtilde(t)=z(t)f0F(t)θtilde(0), i.e. θhat(t)=[I−zf0F]θ+zf0Fθ0. Equation (5) is exactly the algebraic solution θ=[I−zf0F]^{-1}(θhat−zf0Fθ0). Hence θ_FCT(t)=θ holds at every t for which the matrix is invertible; no convergence mechanism or finite-time property of the LS recursion is used. The 'FCT of the standard LS estimate' is achieved only by redefining the estimated parameter as this nonlinear post-processing of the LS state, so the advertised property is built into the definition.

  2. self citation load bearing [Section 3, Proof 1]
    "The result is contained in (Tao, 2003, Lemma 3.5) for (normalized) LS without forgetting factor and, for LS+Dynamic Regression Extension and Mixing (DREM) of (Ortega et al., 2022b) with forgetting factor, it follows directly from the proof of (Ortega et al., 2022b, Proposition 1), where it is shown that the matrix Iq − z(t)f0F (t) is full rank for t ≥ Tc."

    The only nontrivial step needed to turn the algebraic identity into an FCT statement is full-rankness of I−zf0F under interval excitation. This is not proved in the present note; it is imported from prior work by co-author Tao and by the same Ortega/Romero/Aranovskiy group. The opening of the note also states of Ortega et al. (2022b) that 'the connection with FCT was not established', so the present new step reduces to the cited full-rank lemma plus the definitional rearrangement (5). The central claim therefore rests on a self-citation chain rather than on a self-contained derivation.

full rationale

The paper's derivation is not a data-fitting exercise; there is no fitted-input-called-prediction step and the underlying algebra is correct. The circularity is structural rather than empirical. Proposition 1 defines θ_FCT(t) as the algebraic inverse of the invariant relation (2), so once the matrix I−zf0F is nonsingular, θ_FCT(t)=θ is true by construction; the finite-time property is put into the output definition rather than extracted from the LS dynamics. The only nontrivial input, the full-rank implication from IE, is not proved in the note; it is imported from prior work by co-author Tao and by the same Ortega/Romero/Aranovskiy group. The note itself says of Ortega et al. (2022b) that 'the connection with FCT was not established,' confirming that the new 'connection' is the definitional rearrangement. The boundedness assumption in Proposition 1 is an extra technical condition not needed for the inversion, as the skeptic notes, but this is a correctness/assumption gap rather than a circularity. Lemma 1's self-citation to Wang et al. (2023) is contextual and not load-bearing. Because the paper is transparent that it 'recalls' a known fact and the cited full-rank result is externally checkable, the proper score is partial circularity (6), not complete equivalence (8-10). The title and abstract, however, overstate the result by attributing FCT to the standard LS estimate θhat rather than to the transformed signal θ_FCT.

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

The central claim rests on standard LS properties and on the interval excitation condition. The paper does not introduce new entities or fitted parameters; the only extra ingredients are the boundedness of the regressor and the invertibility of the matrix in (5), both asserted with reference to prior work.

assumptions (5)
  • domain assumption The regressor φ(t) satisfies the interval excitation condition (3): there exist Tc > 0 and ρ > 0 such that ∫_0^{Tc} φ(τ)φ^T(τ)dτ ≥ ρ I_q.
    This is the key condition stated in Proposition 1 and is necessary and sufficient for identifiability of the LRE.
  • domain assumption The regressor φ(t) is bounded.
    This is explicitly assumed in Proposition 1 but not justified as necessary; it is used to keep the covariance dynamics and forgetting factor well behaved.
  • domain assumption The matrix I_q - z(t) f0 F(t) is invertible for all t ≥ Tc.
    This invertibility, asserted in the proof by citing (Tao, 2003) and (Ortega et al., 2022b), is required to define θ_FCT in equation (5). The proof is not reproduced in this paper.
  • standard math The relation d/dt(F^{-1}(t)\tilde θ(t)) = 0, i.e., F^{-1}(t)\tilde θ(t) is constant for the LS algorithm, as used in equation (2).
    This is a standard property of LS estimators, derived in Section 2 using equation (1). It underpins the transformation leading to θ_FCT.
  • domain assumption Lemma 1: The LRE is identifiable if and only if the regressor is IE.
    This is cited from (Wang et al., 2023) and justifies why IE is the weakest condition for estimation, but the lemma is not proven in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Standard LSParameter Estimators Ensure Finite Convergence Time for Linear Regression Equations Under an Interval Excitation Assumption." pith.science (2026). https://pith.science/paper/DWHSRRO5

@misc{pith2026250608211,
  author       = {Pith},
  title        = {Pith review of: Standard LSParameter Estimators Ensure Finite Convergence Time for Linear Regression Equations Under an Interval Excitation Assumption},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DWHSRRO5}},
  note         = {Machine review of arXiv:2506.08211}
}
read the original abstract

In this brief note we recall the little-known fact that, for linear regression equations (LRE) with intervally excited (IE) regressors, standard Least Square (LS) parameter estimators ensure finite convergence time (FCT) of the estimated parameters. The convergence time being equal to the time length needed to comply with the IE assumption. As is well-known, IE is necessary and sufficient for the identifiability of the LRE-hence, it is the weakest assumption for the on-or off-line solution of the parameter estimation problem.

Figures

Figures reproduced from arXiv: 2506.08211 by the authors.

Figure 2
Figure 2. White noise signal 0 1 2 3 4 5 -2 -1 0 0 1 2 3 4 5 -3 -2 -1 0 0 1 2 3 4 5 -1 -0.5 0 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Transient behavior of the estimated parameters [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [1]

    Bibitem after note You are using a bibitem after a note in a subbibitems environment; note should the last item in a subbibitems environment \@itemnote @bb@error [] @noitemnote \@tempa \@noitemnote \@noitemnote \@itemnote @bibitem\@bibitem @lbibitem\@lbibitem \@bibitem#1 \@itemnote @bibitem #1 \@tempa @noitemnote \@lbibitem[#1]#2 \@itemnote @lbibitem[#1] ...

  2. [2]

    @stdbsttrue @ctr \@lbibitem [ @ctr] \@bibitem##1 @ctr \@lbibitem[ @ctr] ##1 @bb@error\@mkbberr @filesw @natbibloaded \@auxout \@itemslabel @bibnum a-- @ctr \@auxout \@itemslabel :s @bibnum \@auxout \@itemslabel @bibnum a-- @ctr \@auxout \@itemslabel :s @bibnum @ctr @bibnum @citex\@citex \@tempcntc @citex[#1]#2 @filesw \@auxout #2 \@tempcnta @\@tempcntb @n...

    work page 1991

  3. [3]

    Adetola and M

    V. Adetola and M. Guay. Finite-time parameter estimation in adaptive control of nonlinear systems. , 53 0 (3): 0 807--811, 2008

    work page 2008

  4. [4]

    Andrieu, L

    V. Andrieu, L. Praly, and A. Astolfi. Homogeneous approximation, recursive observer design, and output feedback. , 47 0 (4): 0 1814--1850, 2008

    work page 2008

  5. [5]

    Aranovskiy, R

    S. Aranovskiy, R. Ortega, J. G. Romero, and D. Sokolov. A globally exponentially stable speed observer for a class of mechanical systems: experimental and simulation comparison with high-gain and sliding mode designs. , 92 0 (7): 0 1620--1633, 2019

    work page 2019

  6. [6]

    M. Basin. Finite- and fixed-time convergent algorithms: Design and convergence time estimation. , 48: 0 209--221, 2021

    work page 2021

  7. [7]

    Bhat and D

    S. Bhat and D. S. Bernstein. Geometric homogeneity with applications to finite-time stability. Mathematics of Control, Signals and Systems, 17: 0 101--127, 2005

    work page 2005

  8. [8]

    Y. Chen, F. Wang, and B. Wang. Fixed-time convergence in continuous-time optimization: A fractional approach. , 7: 0 631--636, 2023

    work page 2023

Show all 28 references
  1. [9]

    Cruz-Zavala, E

    E. Cruz-Zavala, E. Nuno, and J. A. Moreno. Trajectory-tracking in finite-time for robot manipulators with bounded torques via output-feedback. , 96 0 (4): 0 907--921, 2023

    2023

  2. [10]

    de Larminat

    Ph. de Larminat. On the stabilizability condition in indirect adaptive control. , 20 0 (6): 0 793--795, 1984

    1984

  3. [11]

    Efimov and A

    D. Efimov and A. Polyakov. Finite-time stability tools for control and estimation. Foundations and Trends in Systems and Control, 9 0 (2-3): 0 171--364, 2021

    2021

  4. [12]

    Huang, H

    C. Huang, H. Shim, S. Yu, and B. D. O. Anderson. Mode consensus algorithms with finite convergence time. , arXiv:2403.00221 , 2024

    2024 arXiv

  5. [13]

    M. Kawski. Families of dilations and asymptotic stability. Analysis of Controlled Dynamical Systems , pages 285--294, 1991

    1991

  6. [14]

    Krause and P

    J. Krause and P. Khargonekar. Parameter information content of measurable signals in direct adaptive control. , 32 0 (9): 0 802--810, 1987

    1987

  7. [15]

    Kreisselmeier and G

    G. Kreisselmeier and G. Rietze-Augst. Richness and excitation on an interval---with application to continuous-time adaptive control. , 35 0 (2): 0 165--171, 1990

    1990

  8. [16]

    Lozano and X

    R. Lozano and X. Zhao. Adaptive pole placement without excitation probing signals. , 39 0 (1): 0 47--58, 1994

    1994

  9. [17]

    Marino and P

    R. Marino and P. Tomei. On exponentially convergent parameter estimation with lack of persistency of excitation. , 159: 0 105080, 2022

    2022

  10. [18]

    H. Min, S. Shi, J. Gu, and N. Dua. Further results on fixed-time control for nonlinear systems with asymmetric output constraints. , 171: 0 111933, 2025

    2025

  11. [19]

    R. Ortega. An on-line least-squares parameter estimator with finite convergence time. Proc. IEEE , 76 0 (7): 0 847 -- 848, 1988

    1988

  12. [20]

    Ortega, A

    R. Ortega, A. Bobtsov, and N. Nikolayev. Parameter identification with finite-convergence time alertness preservation. , 6: 0 205--210, 2022 a

    2022

  13. [21]

    Ortega, J.G

    R. Ortega, J.G. Romero, and S. Aranovskiy. A new least squares parameter estimator for nonlinear regression equations with relaxed excitation conditions and forgetting factor. Systems and Control Letters , 169: 0 105377, 2022 b

    2022

  14. [22]

    Y. Pan, T. Shi, and R. Ortega. Comparative analysis of parameter convergence for several least-squares estimation schemes. , 69 0 (5): 0 3341--3348, 2023

    2023

  15. [23]

    Rao and H

    C. Rao and H. Toutenburg. Linear Models: Least Squares and Alternatives. Springer Series in Statistics (3rd ed.), 2008

    2008

  16. [24]

    H. Rios, D. Efimov, J. A. Moreno, W. Perruquetti, and J. G. Rueda-Escobedo. Time-varying parameter identification algorithms: Finite and fixed-time convergence. , 62 0 (7): 0 3671--3678, 2017

    2017

  17. [25]

    J.-J. E. Slotine and W. Li. Applied Nonlinear Control. Prentice-Hall, New Jersey, USA, 1991

    1991

  18. [26]

    G. Tao. Adaptive Control Design and Analysis. John Wiley & Sons, New Jersey, 2003

    2003

  19. [27]

    J. Wang, D. Efimov, and A. Bobtsov. On robust parameter estimation in finite-time without persistence of excitation. , 65 0 (4): 0 1731--1738, 2020

    2020

  20. [28]

    L. Wang, R. Ortega, A. Bobtsov, J. G. Romero, and B. Yi. Identifiability implies robust, globally exponentially convergent on-line parameter estimation. , 97 0 (9): 0 1967--1983, 2023

    1967

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.