Pith. sign in

REVIEW 3 major objections 4 minor 25 references

Deterministic Kalman filters for uncertain dynamical systems

T0 review · 3 major / 4 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 claims that, for linear systems whose dynamics and noise covariances are uncertain but drawn from a finite set of candidates, three deterministic Kalman-filter estimators deviate from the ideal filter for the hidden parameter…

desk verdict Solid new expected-energy filter, but the advertised N-independent error bounds are not proven: Lemma 4.9 reverses a max/min inequality and the corrected constant depends on spectral gaps. read the letter →

arxiv 2506.00463 v1 pith:ORJNIN2V submitted 2025-05-31 math.DS math.OCmath.PR

Karl Kunisch , Jesper Schröder This is my paper
classification math.DSmath.OCmath.PR MSC 34F0534H0549N1093B5393C15
keywords KalmanfilterstateestimationuncertainparametersdeterministicfilteringRiccatiequationsMahalanobisdistanceexpectedenergyminimization
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 takes the deterministic, energy-based view of the Kalman filter and asks what can still be guaranteed when the system matrix and the noise covariance matrices are not known exactly but are drawn from a finite set of candidate values. It proposes three estimators: the Kalman filter built from the expected matrices, the plain average of the individual candidate filters, and the minimizer of the expected filter energy (a precision-weighted average of the candidate filters). Its main result is that, measured in the Mahalanobis norm defined by the true filter's precision matrix, all three estimators deviate from the ideal filter of the hidden parameter by at most a constant independent of $N$ times $\mathbb{E}[\|S_\sigma-S_{\bar\sigma}\|_1]$. That is a quantitative guarantee that model uncertainty transmits linearly, rather than exponentially, into filtering error, and it also gives a rigorous reason to prefer the energy minimizer: it minimizes the expected squared Mahalanobis distance within the candidate family. Two numerical examples illustrate the theory and show where the simple average and the weighted average behave differently.

What carries the argument

The central device is the deterministic minimum-energy formulation of the Kalman filter, in which the state estimate is the minimizer of an energy functional whose quadratic value function has Hessian equal to the precision matrix $P(t)=\Pi^{-1}(t)$. For each candidate parameter $k$, this produces a quadratic value function $V_k(t,\xi)$ whose minimizer is the candidate Kalman trajectory $\hat{x}_k(t)$; the expected energy $\mathcal{E}(t,\xi)=N^{-1}\sum_k V_k(t,\xi)$ is then quadratic with Hessian $N^{-1}\sum_k P_k(t)$, so its minimizer is the precision-weighted average of the candidate trajectories. The error analysis works by relating the difference of covariance matrices to the difference of value functions of a dual linear-quadratic control problem, proving a time-uniform Lipschitz estimate for those value functions with respect to the parameters, and then propagating the bound through the filter equations with Gronwall's inequality.

What would settle it

Construct a candidate set whose reference Riccati solution $\Pi_{\bar\sigma}(t)$ reaches a near-zero minimum eigenvalue on $(0,T)$ (for example a nearly unobservable mode combined with very small measurement noise $Q_{\bar\sigma}$), then compute the constants $c_E$, $c_\emptyset$, and $c_E$ from the proof while increasing $N$. If the smallest eigenvalue of the precision sum $P(t)=\sum_k P_k(t)$ decays faster than $O(1/N)$, Proposition 4.10 and Lemma 4.9 cannot hold uniformly, and the claimed bound would fail for finite candidate sets.

Watch

Extended reading notes

Core claim

The paper establishes that the worst-case Mahalanobis distance to the oracle filter $\hat{x}_{\bar\sigma}$ is controlled by the first-order mean deviation of the parameter-dependent matrices. For the expected-matrix filter, the expected filter, and the energy-minimizing filter, Propositions 4.7, 4.8, and 4.10 give bounds of the form $\|\hat{x}(t)-\hat{x}_{\bar\sigma}(t)\|_{P_{\bar\sigma}(t)} \leq c\, \mathbb{E}[\|S_\sigma-S_{\bar\sigma}\|_1]$ with constants $c_E$, $c_\emptyset$, and $c_E$ independent of the number $N$ of candidates. The paper further proves that the energy minimizer is the unique minimizer of the expected squared Mahalanobis distance to the candidate family, characterizes it as the precision-weighted average $(\sum_k P_k(t))^{-1}\sum_k P_k(t)\hat{x}_k(t)$, and confirms the qualitative predictions in two numerical experiments: a harmonic oscillator with uncertain damping and two connected amplidynes with uncertain inductances.

Load-bearing premise

The proof needs a time-uniform Lipschitz bound on the optimal-control value function when the candidate matrices vary; if some candidate drives the reference Riccati solution near singularity, the constants in the bound can grow without control, so the stated error bound is not genuinely uniform across the whole uncertainty class.

Editorial extensions

If this is right

  • If the bounds hold, then shrinking the uncertainty set linearly shrinks the guaranteed Mahalanobis error of any of the three estimators, no matter how many candidate parameters are used.
  • The energy-minimizing filter has essentially the same computational cost as the plain average filter: it requires solving the same $N$ decoupled Kalman-Riccati systems plus one matrix-weighted combination.
  • Because the energy minimizer minimizes expected squared Mahalanobis distance, it is the natural point estimate for the candidate family under a precision-weighted quadratic loss.
  • The numerical experiments show that the ranking of the estimators depends on the structure of the precision matrices: with diagonally dominant precisions the weighted average favors high-precision candidates, while with strong off-diagonal precision entries it can leave the convex hull of the individual estimates.

Reading between the lines

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

  • For a continuous parameter distribution, the finite average of value functions would become an integral; if the uniform constants survive that limit, a bound of the same linear form should hold for compactly supported parameter densities.
  • A risk-averse variant that minimizes a higher moment of the energy rather than its expectation would lose the closed-form precision-weighted average but could give tighter worst-case guarantees in settings like the oscillator experiment with small true damping.
  • The proven bounds measure distance to the oracle filter, not to the true state; a full state-reconstruction guarantee would need to combine them with the classical filter-error behavior of $\hat{x}_{\bar\sigma}$, which depends on observability and noise levels.
  • The linear dependence on $\mathbb{E}[\|S_\sigma-S_{\bar\sigma}\|_1]$ suggests a practical calibration test: estimate the variation of the uncertain matrices from data and compare it against an acceptable Mahalanobis error tolerance, although the constants in the bound may not be easy to compute in closed form.
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 / 4 minor

Summary. The paper proposes three deterministic Kalman-filter-type estimators for linear systems whose dynamics and noise covariances depend on a finite set of uncertain parameters: the filter built from the expected matrices (\hat x_E), the expectation of the individual parameter-dependent filters (\hat x_\emptyset), and the minimizer of the expected Mortensen energy (\hat x_{\mathcal E}, which is a precision-weighted average of the individual filters). The central theoretical contribution is a set of error bounds, Propositions 4.7, 4.8, and 4.10, claiming that for all t in [0,T], the Mahalanobis distance \|\hat x(t)-\hat x_{\bar\sigma}(t)\|_{P_{\bar\sigma}(t)} is bounded by c\,\mathbb E[\|S_\sigma-S_{\bar\sigma}\|_1] with a constant c independent of the cardinality N of the parameter set. The paper closes with numerical experiments on a harmonic oscillator and on connected amplidynes, with code available on Zenodo.

Significance. If the claimed uniform bounds held, the paper would provide a useful deterministic counterpart to stochastic Kalman filtering under parametric uncertainty, with explicit dependence of the estimation error on the scatter of the uncertain matrices. The paper is largely self-contained, gives explicit formulas for all three estimators (Proposition 3.4 and Corollary 3.6), and includes a reproducible numerical code, which are clear strengths. The proofs are checkable, and the value-function representation of the Kalman filter in Lemma 2.5 is cleanly derived. However, the main theoretical claim rests on constants that are asserted to be independent of N without a proof of uniform spectral bounds; this is a load-bearing gap that affects all three main estimates.

major comments (3)
  1. [§4.2, Lemma 4.9 and Proposition 4.10] The proof of Lemma 4.9 contains a reversed inequality: with \bar\lambda defined as the maximum of the \lambda_k, the chain (x,P(t)x) \ge \sum_k \lambda_k\|x\|^2 \ge N\bar\lambda\|x\|^2 is false. Replacing the maximum by the minimum fixes the inequality, but then the resulting constant is proportional to 1/\lambda_{\min}, which depends on the worst candidate precision matrix and can diverge as the family grows or as one candidate approaches singularity. Since Lemma 4.9 is used directly in the proof of Proposition 4.10 to obtain a constant c_E independent of N, the N-independence claim in Proposition 4.10 is not established as stated. A uniform positive lower bound on the eigenvalues of P_k(t) over k and t is needed, or the theorem must be restated with constants depending on the family.
  2. [§4.1, Lemma 4.3 and Lemma 4.4] The constants c_* and c_\delta in Lemmas 4.3 and 4.4 depend on spectral properties of Q_*, R_*, and \Gamma_* and on the candidate matrices. In particular, the step \|\bar u_*\|^2_{L^2} \le 2J_*(\bar x_*,\bar u_*;t,x_0) in Eq. (4.8) is only valid if Q_* dominates the identity; for general positive definite Q_* one needs a factor depending on \lambda_{\min}(Q_*)^{-1}, and analogous factors enter the bounds for R_* and \Gamma_*. These factors are not shown to be uniform over the parameter family as N grows. Moreover, the constant in Lemma 4.6 depends on \|y-C\hat x_*\|_{L^2} and \|\Pi_*\|_{C^{n,n}_T}, so the asserted independence of N in Proposition 4.7 for the case *=E is also unsupported without uniform bounds on the expected filter. This issue affects Propositions 4.7, 4.8, and 4.10 simultaneously.
  3. [§4.1, proof of Lemma 4.4] The final displayed inequality in the proof of Lemma 4.4, namely c_1\,\delta J_{\bar\sigma} \le c_2\,(\|A_\delta\|+\|\Gamma_\delta\|+\|R_\delta\|+\|Q_\delta^{-1}\|), is dimensionally inconsistent: the left-hand side is quadratic in the differences while the right-hand side is linear, and the preceding display contains squared norms. As printed, the inequality cannot hold for large parameter differences. It should be corrected to a bound with squared norms on the right-hand side; with that correction the subsequent square-root step in Lemma 4.5 is valid. This is a proof error in a lemma that is central to the main estimates, so it must be fixed.
minor comments (4)
  1. [Abstract and §4] The abstract calls the bound a bound in terms of the 'variance' of the uncertainties, but the theorems bound the mean absolute deviation \mathbb E[\|S_\sigma-S_{\bar\sigma}\|_1]. The terminology should be aligned with the actual statement.
  2. [§1 and §4] The paper should state explicitly and prominently that all error bounds compare the estimators to the oracle Kalman filter \hat x_{\bar\sigma} associated with the hidden parameter, not to the true state trajectory x. This is clear in Section 4 but risks being over-interpreted by readers of the abstract.
  3. [§4.1, Lemma 4.4] The proof of Lemma 4.4 uses both x_\delta(t)=0 and 'x_\delta(0)=0' in the same paragraph; the notation should be made consistent.
  4. [§5] The numerical experiments are illustrative and compare the three estimators on single realizations. A quantitative comparison with the right-hand side of the proven bounds for the same parameter families would help the reader calibrate how sharp the estimates are.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the main error estimates are proven in-paper; the self-citations are not load-bearing. The risks flagged by the skeptic are correctness and uniformity gaps, not circularity.

full rationale

The central claims (Propositions 4.7, 4.8, and 4.10) are genuine theorems: each estimator's error is bounded in terms of the mean absolute parameter deviation E[\|S_\sigma-S_{\bar\sigma}\|_1], and the estimators are defined before, and independently of, the oracle filter to which they are compared. No parameter is fitted from the data used in the error analysis, and the hidden parameter \bar\sigma enters only as the comparator, not as an input to the construction. The value-function and Riccati arguments are developed in the paper, so the bounds do not assume their own conclusions. There are two minor self-citations: Lemma 4.4 says 'We now employ the strategy presented in [8, Lem. 3.2]' (Guth-Kunisch-Rodrigues, with the present first author), and Lemma 2.5 cites [2] (Breiten-Kunisch) for the unweighted value-function identity. Both supply technical ingredients, but the proof of Lemma 4.4 is reproduced in the text via Gronwall/Young estimates, and Lemma 2.5 is a classical LQ identity whose weighted extension is stated rather than being the paper's novel conclusion. These self-citations are therefore not load-bearing in the sense of replacing the derivation with the authors' prior claim. The skeptic's objection to Lemma 4.9 is a rigorousness flaw, not circularity: the displayed inequality '(x,P(t)x) \ge \sum_{k=1}^N \lambda_k\|x\|^2 \ge N\bar\lambda\|x\|^2' with \bar\lambda defined as the maximum is directionally false, and the corrected argument would make the constants depend on the minimal eigenvalue of the precision matrices, weakening the claimed N-independence. Likewise, Lemma 4.3's step '\|\bar u_*\|^2 \le 2J_*' implicitly requires spectral bounds on Q_*, so the constants may inherit dependence on the candidate family. These are correctness concerns about uniformity, not circularity, and do not raise the circularity score above 2.

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

The central theorems are derived from first principles in the paper and do not introduce fitted constants. The covariance matrices and parameter sets in the experiments are user-selected inputs. The main assumptions are the finite-uniform-independence structure of the uncertainty and the L2-regularity of disturbances; the main background result is the value-function framework (Lemma 2.5), partly cited from the authors' own [2], and the value-function Lipschitz estimate whose strategy is taken from the authors' prior [8].

free parameters (2)
  • Covariance/weighting matrices Gamma, R, Q in the numerical examples = Oscillator: Gamma=0.1*I2, R=0.05, Q=0.05; Amplidyne: Gamma=diag(0.125,0.25,2.5,5), R=0.01, Q=1600
    Chosen by hand in Section 5.1 to match component scales. They are inputs to the experiments, not fitted to data, and do not enter the theoretical error bounds.
  • Finite parameter sets Sigma_A (and the true hidden parameter barsigma) = Oscillator: 101 damping values in [0.1,3]; Amplidyne: 125 combinations of three inductances from {10,12.5,15,17.5,20}…
    User-chosen uncertainty class. The theorems are independent of N but the constants c_E, c_emptyset, c_E depend on the family through the bounds in Lemmas 4.3-4.5.
assumptions (5)
  • domain assumption The uncertain matrices A_sigma, Gamma_sigma, R_sigma, Q_sigma are mutually independent random variables on a finite discrete probability space with uniform distribution (Section 3.1).
    The expected-energy estimator and the error bounds in Corollary 4.11 are defined with respect to this finite uniform distribution; continuous or correlated uncertainties are excluded.
  • domain assumption Disturbances v, eta, mu have L2 regularity and the measurement y is in L2 (Remark 2.2).
    The Mortensen-Willems deterministic filter and all value-function identities in Sections 2 and 4 require this regularity; white noise realizations do not satisfy it.
  • standard math The Riccati equations (2.11) and (3.4) have unique symmetric positive definite solutions on [0,T] (Dieci-Eirola [5], Prop. 1.1).
    Used to ensure P_k(t) is positive definite (Prop. 3.4) and to justify the invertibility of P(t) in Lemma 4.9.
  • domain assumption The representation of the value function in Lemma 2.5 holds for the general weighting case, extended from the authors' [2] by a continuity/density argument.
    This is a key structural input; the proof in the paper is a citation to a prior paper by Breiten and Kunisch with a sketch for the general case.
  • domain assumption The value functions nu_* and nu_barsigma for the dual optimal control problems depend Lipschitz-continuously on the data (A,Gamma,R,Q) with a constant c_J independent of t (Lemma 4.4), following the strategy of the authors' prior work [8, Lem. 3.2].
    This is the load-bearing technical estimate; its proof in the paper reproduces the argument from the authors' own earlier publication, so the present paper's error bounds inherit the assumptions of that technique, including uniform a priori bounds on the optimal trajectories (Lemma 4.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deterministic Kalman filters for uncertain dynamical systems." pith.science (2026). https://pith.science/paper/ORJNIN2V

@misc{pith2026250600463,
  author       = {Pith},
  title        = {Pith review of: Deterministic Kalman filters for uncertain dynamical systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ORJNIN2V}},
  note         = {Machine review of arXiv:2506.00463}
}
read the original abstract

The Kalman(-Bucy) filter is the natural choice for the state reconstruction of disturbed, linear dynamical systems based on flawed and incomplete measurements. Taking a deterministic viewpoint this work investigates possible extensions of the concept to systems with uncertain dynamics and noise covariances. In a theoretical analysis error bounds in terms of the variance of the uncertainties are derived. The article concludes with a numerical implementation of two example systems allowing for a comparison of the estimators.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    A. C. Aitken , On least squares and linear combination of observations, Proceedings of the Royal Society of Edinburgh, 55 (1936), p. 42–48

    work page 1936

  2. [2]

    Breiten and K

    T. Breiten and K. Kunisch , Neural network based nonlinear observers, System & Control Letters, 148 (2021)

    work page 2021

  3. [3]

    P. G. Ciarlet , The Finite Element Method for Elliptic Problems, Studies in Mathematics and Its Applications, North-Holland, Amsterdam New York Oxford, first ed., 1978

    work page 1978

  4. [4]

    M. H. DeGroot , Optimal statistical decisions, McGraw-Hill, New York a. o., first ed., 1970

    work page 1970

  5. [5]

    Dieci and T

    L. Dieci and T. Eirola , Positive definiteness in the numerical solution of Riccati differential equations, Numerische Mathematik, 67 (1994), p. 303–313

    work page 1994

  6. [6]

    Evensen, Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics, J

    G. Evensen, Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics, J. Geophys. Res., 99 (1994), pp. 10143–10162

    work page 1994

  7. [7]

    V. L. Girko , Random matrices, in Handbook of Algebra, M. Hazewinkel, ed., vol. 1 of Handbook of Algebra, North-Holland, 1996, pp. 27–78

    work page 1996

  8. [8]

    P. A. Guth, K. Kunisch, and S. S. Rodrigues , Tracking optimal feedback control under uncertain parameters, Physica D: Nonlinear Phenomena, 467 (2024), p. 134245

    work page 2024

Show all 25 references
  1. [9]

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

    1960

  2. [10]

    R. E. Kalman and R. S. Bucy , New results in linear filtering and prediction theory, Transactions of the ASME–Journal of Basic Engineering, 83 (1961), pp. 95–108

    1961

  3. [11]

    Kempka, C

    H. Kempka, C. Schneider, and J. Vybiral , Path regularity of the Brownian motion and the Brownian sheet, Constructive Approximation, 59 (2024), pp. 485–539

    2024

  4. [12]

    Kwakernaak and R

    H. Kwakernaak and R. Sivan , Linear optimal control systems, Wiley-Interscience, a division of John Wiley & Sons, Inc., New York, Chichester, Brisbane, Toronto, Singapore, first ed., 1972

    1972

  5. [13]

    On the generalised distance in statistics

    P. C. Mahalanobis , Reprint of: Mahalanobis, P. C. (1936) "On the generalised distance in statistics", The Indian Journal of Statistics, 80 (2019), pp. S1–S7

    1936

  6. [14]

    Mehra, On the identification of variances and adaptive Kalman filtering, IEEE Transactions on Automatic Control, 15 (1970), pp

    R. Mehra, On the identification of variances and adaptive Kalman filtering, IEEE Transactions on Automatic Control, 15 (1970), pp. 175–184

    1970

  7. [15]

    R. E. Mortensen , Maximum-likelihood recursive nonlinear filtering, Journal of Optimization Theory and Applications, 2 (1968), pp. 386–394

    1968

  8. [16]

    Øksendal, Stochastic differential equations, Universitext, Springer-Verlag, Berlin, fifth ed., 1998

    B. Øksendal, Stochastic differential equations, Universitext, Springer-Verlag, Berlin, fifth ed., 1998

    1998

  9. [17]

    I. R. Petersen and D. C. McF arlane , Optimal guaranteed cost control and filtering for uncertain linear systems, IEEE Transactions on Automatic Control, 39 (1994), pp. 1971–1977

    1994

  10. [18]

    I. R. Petersen and A. V. Savkin , Robust Kalman Filtering for Signals and Systems with Large Uncertainties, Birkhäuser, Boston, first ed., 1999

    1999

  11. [19]

    Deterministic Kalman filters for uncertain dynamical systems

    J. Schröder, Code for the paper "Deterministic Kalman filters for uncertain dynamical systems",

  12. [20]

    L. Shi, K. H. Johansson, and R. M. Murray , Kalman filtering with uncertain process and measurement noise covariances with application to state estimation in sensor networks, in 2007 IEEE International Conference on Control Applications, 2007, pp. 1031–1036

    2007

  13. [21]

    E. D. Sontag, Mathematical control theory: Deterministic finite dimensional systems, Texts in Applied Mathematics 6, Springer, New York, second ed., 1998

    1998

  14. [22]

    Wiener, Extrapolation, Interpolation, and Smoothing of Stationary Time Series, The MIT Press, Cambridge, first ed., 1949

    N. Wiener, Extrapolation, Interpolation, and Smoothing of Stationary Time Series, The MIT Press, Cambridge, first ed., 1949

    1949

  15. [23]

    J. C. Willems , Deterministic least squares filtering, Journal of Econometrics, 118 (2004), pp. 341–373

    2004

  16. [24]

    Xiong, An introduction to stochastic filtering theory, Oxford University Press, Oxford, first ed., 2008

    J. Xiong, An introduction to stochastic filtering theory, Oxford University Press, Oxford, first ed., 2008. ⋆ Institute of Mathematics and Scientific Computing, University of Graz, A-8010 Graz, Austria 26 KARL KUNISCH ⋆,† AND JESPER SCHRÖDER† † Johann Radon Institute, Austrian...

    2008

  17. [2025]

    doi: 10.5281/zenodo.15519854

    doi

Pith tools

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