Pith. sign in

REVIEW 2 major objections 5 minor 28 references

Iteratively Saturated Kalman Filtering

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A saturated Kalman filter iterates to stay robust against outlier shocks at steady-state cost.

desk verdict Clean derivation of a cheap robust KF variant; the covariance recursion gap and thin experiments keep it from being more than a good contributed paper. read the letter →

arxiv 2507.00272 v1 pith:IKK27TYB submitted 2025-06-30 eess.SY cs.SY

Alan Yang , Stephen Boyd This is my paper · ORCID
classification eess.SYcs.SY MSC 93E1165K10
keywords iterativelysaturatedKalmanfilterscaledgradientmethodrobustestimationHuberfunctionsteay-stateoutlierrobustnessconvexMAP
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 introduces the iteratively saturated Kalman filter (ISKF), a modification of the standard Kalman filter's update step that adds saturating nonlinearities to damp the influence of both measurement and process-noise outliers. The ISKF is derived as a scaled gradient method for a convex robust maximum a posteriori estimation problem, so each update is a step toward the optimum of a well-defined objective. The authors claim that a steady-state variant matches the standard steady-state Kalman filter's low per-step cost, needing only matrix-vector products, while handling outliers that can break the standard filter.

What carries the argument

The central object is the circular Huber function $\varphi$(a; $\lambda$), which is the partial minimization of (1/2)||a-b||^2 + $\lambda$||b|| over b; it is quadratic inside a ball of radius $\lambda$ and grows linearly outside. Composed with the Kalman-prior and measurement-noise whitening, it forms the convex robust MAP objective. The gradient of $\varphi$ is a saturation mapping, and combining the two saturation terms through the Woodbury identity gives the Kalman gain structure, allowing the scaled gradient update to be written directly in terms of $\sigma$ and rho.

What would settle it

Simulate a linear system with a process-noise outlier massive enough to move the true posterior mean far from the predicted state, then compare the ISKF's reported covariance P_{t|t} to the empirical covariance of the state over many such runs; a systematically too-small P_{t|t} after the outlier would falsify the calibration assumption.

Watch

Extended reading notes

Core claim

The ISKF update is the iteration x^k = $x^{{k-1}}$ + K $\sigma$(y - C $x^{{k-1}}$) + (I - K C) rho(x_pred - $x^{{k-1}}$), where $\sigma$ and rho are saturating functions at thresholds lambda_y and lambda_x. The paper shows this is exactly a scaled gradient step with unit step size for minimizing the convex objective f(x) = $\varphi$($Sigma^{{-1/2}}$(x - x_pred); lambda_x) + $\varphi$($V^{{-1/2}}$(y - C x); lambda_y), where $\varphi$ is the circular Huber function. Because the scalar step size is known and the gradient is Lipschitz continuous with constant one under the Kalman scaling, the iteration is a strict descent method. The authors argue that one or two iterations suffice in practice, yielding outlier robustness at the same order of per-step cost as the Kalman filter, with a steady-state version that requires only O($n^{2}$ + np) flops online.

Load-bearing premise

The filter assumes the standard Kalman covariance recursion still gives the true uncertainty of the state after a saturated, nonlinear update, so that the next step's saturation thresholds and gain stay correctly calibrated.

Editorial extensions

If this is right

  • If the ISKF works as claimed, a real-time estimator can reject both measurement and process-noise outliers without solving an optimization problem online.
  • The steady-state variant can be implemented with precomputed gain matrices, making it suitable for embedded and high-frequency control loops.
  • The derivation implies that robust Kalman filtering can be understood as convex optimization, not just a heuristic covariance inflation, which may guide parameter tuning and analysis.
  • The method extends to time-varying and linearized nonlinear systems, though without the steady-state cost advantage for time-varying cases.

Reading between the lines

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

  • The paper implicitly suggests that the covariance recursion (1.8) remains a reliable uncertainty measure even after a saturated, nonlinear state update; this is a testable assumption that, if violated, would cause the filter to be overconfident after large outliers.
  • A natural extension is to recover the full robust posterior, not just the MAP point, by propagating a heavier-tailed distribution; the current filter only outputs the mode.
  • The one- or two-iteration recommendation is empirical; a deeper analysis of when the gradient method converges in few steps for ill-conditioned systems could sharpen the practical guidance.
  • The connection to Huber regression suggests that the ISKF could be extended to other outlier models, such as group sparsity or structured outliers, by replacing the circular Huber with other convex penalties whose proximal operators are cheap.
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 / 5 minor

Summary. The paper proposes an iteratively saturated Kalman filter (ISKF) that modifies the standard KF update step by iterating a saturated correction: $\hat{x}^k = \hat{x}^{k-1} + K_t \sigma(y_t - C\hat{x}^{k-1}) + (I-K_t C)\rho_t(\hat{x}_{t|t-1} - \hat{x}^{k-1})$, with saturations on the $V^{-1/2}$-normalized innovation and the $P_{t|t-1}^{-1/2}$-normalized deviation from the prediction. It derives this update as a scaled gradient method with step size one for the convex robust MAP problem (1.11), and proves asymptotic convergence of the inner iteration via a Lipschitz argument. A steady-state version with precomputed gain $K$ is proposed, with online cost $O(n^2+np)$ per time step. Two numerical examples (vehicle tracking and a cascade of CSTRs) compare the ISKF with the KF, WoLF, and a Huberized KF, showing robustness gains on outlier-contaminated test trajectories.

Significance. The manuscript is clearly written and the per-step derivation is a strength: the Woodbury identity step in §1.5.2 is algebraically correct, the normalization via $\eta=1$ is natural, and the convergence proof for the inner iteration is sound. The numerical evaluation is honest: parameters are tuned on a separate trajectory, and the comparison against a Huberized KF and WoLF is informative. The steady-state variant's $O(n^2+np)$ online cost is a practically attractive feature. However, the theoretical guarantee stops at the per-step optimization; the recursive use of the standard Kalman covariance update (1.8) after a nonlinear measurement update is acknowledged in §1.5.1 but never analyzed, leaving the closed-loop filter without a consistency or stability statement. This gap is the main barrier to the paper's broader significance.

major comments (2)
  1. [§1.4, Eqs. (1.6) and (1.8); §1.5.1] The filter uses the standard Kalman covariance recursion (1.8) after the nonlinear update (1.6). This recursion is exact only when the update is linear, that is, when the saturations are not active. When $\rho_t$ or $\sigma$ saturate, the posterior over $x_t$ is non-Gaussian and (1.8) is not its conditional covariance. The manuscript explicitly concedes in §1.5.1 that the prior distribution "is no longer necessarily Gaussian," but it does not analyze the closed-loop effect: the next step's $P_{t|t-1}$, the gain $K_t$, and the scaling inside $\rho_t$ are all computed from this potentially miscalibrated covariance. The paper provides no consistency proof, no bound on overconfidence, and no divergence analysis. Because the recursive formulation and the steady-state variant (1.10) both rely on this covariance recursion, the advertised low-cost recursive robustness depends on an unverified summary-statistic approximation.
  2. [§1.5.3 and §1.6] The convergence proof in §1.5.3 shows that the scaled-gradient iteration converges to the minimizer of the per-step problem (1.11) as the number of inner iterations grows, with the step size $\eta$ in $(0,2)$. It does not provide a contraction-rate or error bound for the finite number of iterations used in practice ($\tilde{k}=1$ or $2$), nor does it analyze how inner-iteration errors propagate through the time recursion. The recommendation that one or two iterations suffice rests on the two numerical examples in §1.8. This is not a claim of incorrectness, but it means the practical guidance is empirical rather than theoretically supported, and this distinction should be stated more carefully in the paper.
minor comments (5)
  1. [References] References [23] and [24] are incomplete: the first lacks the book title and the second lacks the publisher and title; please complete them.
  2. [§1.5.1] The circular Huber function is described as "smooth"; it is continuously differentiable with a 1-Lipschitz gradient but is not twice differentiable on the threshold sphere, so "smooth" may be misleading.
  3. [§1.3] The sentence "We assume that the initial state $x_0$, the process noise $w_t$, and the measurement noise $v_t$ are independent and identically distributed (IID)" is imprecise: $x_0$ is a single random vector, and the noises are independent across time; consider rewording to "independent and, for the noises, identically distributed across time."
  4. [§1.8.1] There is a typo in "on a simulated test trajectory of, using the state estimate RMSE"; the phrase should likely read "on a simulated test trajectory of 1000 time steps."
  5. [§1.8.2] The CSTR example is based on a MathWorks web page (reference [28]); to make the example reproducible, consider providing the continuous-time model equations or the numerical parameter values used for linearization.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the ISKF update is an exact scaled-gradient step for the stated robust MAP objective, and the tuned thresholds are standard hyperparameters.

full rationale

The derivation chain is not circular. In Section 1.5.2 the authors define the objective f in (1.11) using circular Huber functions whose gradients are the saturation functions sigma and rho, then compute the gradient of f explicitly in (1.12) and use the Woodbury identity K = Sigma C^T (C Sigma C^T + V)^-1 = M^-1 C^T V^-1 together with (I - K C) = M^-1 Sigma^-1 to obtain M^-1 grad f(x) = -K sigma(y - C x) - (I - K C) rho(hat x_pred - x). Setting the step size eta = 1 gives exactly the ISKF update (1.6). This is an algebraic identity, not a fitted input renamed as a prediction. The thresholds lambda_x and lambda_y are hyperparameters selected by grid search on a separate tuning trajectory (Sections 1.6 and 1.8) and evaluated on a distinct test trajectory, so there is no fitted-input-called-prediction pattern. The convergence proof in Section 1.5.3 rests on the standard descent lemma and the Lipschitz bound L^T M L = I; the cited [24] is external, and the only author self-citation ([16]) is a survey item in a list of Huberizing works, not load-bearing. The paper explicitly acknowledges in Section 1.5.1 that after process-noise outliers the prior distribution over x_t is no longer necessarily Gaussian and that the optimization is an approximate MAP; reusing the Kalman covariance recursion (1.8) under a nonlinear update is a real modeling approximation, but it is a correctness risk, not a circular reduction. Hence no circular step is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method's central derivation rests on standard LTI system assumptions and on a sparse-outlier noise model; the only fitted quantities are user-chosen thresholds and the iteration count, tuned by grid search in the experiments.

free parameters (4)
  • lambda_x (state saturation threshold) = 0.10 in both examples
    User-tuned robustness parameter controlling how far the estimate may deviate from the prediction before attenuation. Tuned via grid search on a separate trajectory; no closed-form selection rule.
  • lambda_y (measurement saturation threshold) = 1.8 (vehicle), 3.3 (CSTR); 0.89 and 2.34 when eta is tuned
    User-tuned robustness parameter controlling attenuation of large innovations. Tuned via grid search on a separate trajectory.
  • Number of iterations k = 2 (default)
    Chosen empirically; the paper reports good performance at k=2 and diminishing returns beyond k=3.
  • Step size eta (optional) = 1 (fixed); 2.64 and 1.83 in the tuning experiment
    Set to 1 so the filter reduces to the standard KF in the absence of outliers. The paper shows tuning eta can marginally improve outlier performance at the cost of degraded outlier-free performance.
assumptions (5)
  • domain assumption The pair (A,C) is detectable and (A,W^{1/2}) is stabilizable.
    Standard assumptions invoked in Section 1.4 to ensure the steady-state Kalman gain and covariance limits exist.
  • domain assumption Noise covariances W and V are known, and the matrices A and C are known.
    The filter is model-based; mismatch between the assumed and true models is not analyzed in the paper.
  • domain assumption Outliers are sparse additive terms s_t and o_t in the whitened coordinates of the nominal noise, per Eq. (1.2).
    This models sensor glitches and process shocks, but assumes outliers lie in the column space of F and G, which may not hold in general.
  • ad hoc to paper The standard KF covariance recursion (Eqs. 1.4 and 1.8) remains a valid uncertainty model even when the state update is nonlinear (Eq. 1.6).
    The filter uses the Kalman covariance P to form the next prior and the next saturation, but the true posterior is not Gaussian after a saturated update. The paper does not analyze this approximation.
  • domain assumption The circular Huber function phi is an appropriate robust surrogate for the negative log-prior and negative log-likelihood terms (Eq. 1.11).
    A standard robust-statistics choice; the paper itself notes it is an approximate MAP model without a full statistical model of the outliers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Iteratively Saturated Kalman Filtering." pith.science (2026). https://pith.science/paper/IKK27TYB

@misc{pith2026250700272,
  author       = {Pith},
  title        = {Pith review of: Iteratively Saturated Kalman Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IKK27TYB}},
  note         = {Machine review of arXiv:2507.00272}
}
read the original abstract

The Kalman filter (KF) provides optimal recursive state estimates for linear-Gaussian systems and underpins applications in control, signal processing, and others. However, it is vulnerable to outliers in the measurements and process noise. We introduce the iteratively saturated Kalman filter (ISKF), which is derived as a scaled gradient method for solving a convex robust estimation problem. It achieves outlier robustness while preserving the KF's low per-step cost and implementation simplicity, since in practice it typically requires only one or two iterations to achieve good performance. The ISKF also admits a steady-state variant that, like the standard steady-state KF, does not require linear system solves in each time step, making it well-suited for real-time systems.

Figures

Figures reproduced from arXiv: 2507.00272 by the authors.

Figure 1.1
Figure 1.1. True vehicle position over time, along with measurements and position estimates produced by the (steady-state) ISKF and KF. 0 200 400 600 800 1000 t 0 5 ξ1 error 0 200 400 600 800 1000 t 0 10 ξ 2 error 0 200 400 600 800 1000 t 0 5 ν1 error 0 200 400 600 800 1000 t 0 5 ν 2 error KF ISKF (˜k = 2) [PITH_FULL_IMAGE:figures/full_fig_p014_1_1.png] view at source ↗
Figure 1.2
Figure 1.2. Vehicle tracking state estimate errors (absolute values) for the KF and the two￾iteration ISKF [PITH_FULL_IMAGE:figures/full_fig_p014_1_2.png] view at source ↗
Figure 1.3
Figure 1.3. Parameter tuning for vehicle tracking example. Left: state estimate RMSE vs. number of iterations k˜. A grid search was performed to choose λx and λy for each value of k˜. Right: contour plot of state estimate RMSE (1.17) for different values of λx and λy. Effect of parameters [PITH_FULL_IMAGE:figures/full_fig_p015_1_3.png] view at source ↗
Figures from the paper (2 more)
Figure 1.4
Figure 1.4. Figure 1.4: True reagent concentration and temperature values over time, along with measure￾ments and state estimates produced by the (steady-state) ISKF and KF. The process noise wt and measurement noise vt are distributed according to wt ∼    N (0, F FT ) with probability 0…
Figure 1.5
Figure 1.5. Figure 1.5: State estimate errors (absolute values) for the KF and the two-iteration ISKF [PITH_FULL_IMAGE:figures/full_fig_p018_1_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 24 canonical work pages

  1. [21]

    Robustifying the Kalman filter against measurement outliers: An innovation saturation mechanism

    “Robustifying the Kalman filter against measurement outliers: An innovation saturation mechanism”, in 2018 IEEE Conference on Decision and Control (CDC), 2018

    work page 2018

  2. [1]

    A new approach to linear filtering and prediction problems

    R. Kalman, “A new approach to linear filtering and prediction problems”,Journal of Basic Engineering82.1 (1960), pp. 35–45, 1960

    work page 1960

  3. [2]

    Efficient decoding with steady-state Kalman filter in neural interface systems

    W. Malik, W. Truccolo, E. Brown, and L. Hochberg, “Efficient decoding with steady-state Kalman filter in neural interface systems”,IEEE Transactions on Neural Systems and Rehabilitation Engineering19.1 (2010), pp. 25–34, 2010

    work page 2010

  4. [3]

    Shocks and frictions in US business cycles: A Bayesian DSGE approach

    F. Smets and R. Wouters, “Shocks and frictions in US business cycles: A Bayesian DSGE approach”,American economic review97.3 (2007), pp. 586–606, 2007

    work page 2007

  5. [4]

    Huber,An augmented steady-state Kalman filter to evaluate the likelihood of linear and time: Invariant state-space models, tech

    J. Huber,An augmented steady-state Kalman filter to evaluate the likelihood of linear and time: Invariant state-space models, tech. rep., Volkswirtschaftliche Diskussionsreihe, 2022

    work page 2022

  6. [5]

    Robust Bayesian estimation for the linear model and robustifying the Kalman filter

    C. Masreliez and R. Martin, “Robust Bayesian estimation for the linear model and robustifying the Kalman filter”,IEEE transactions on Automatic Control 22.3 (1977), pp. 361–371, 1977

    work page 1977

  7. [6]

    Davidon,Variable metric method for minimization, tech

    W. Davidon,Variable metric method for minimization, tech. rep., May 1959,doi: 10.2172/4252678, url: https://www.osti.gov/biblio/4252678

    work page arXiv 1959

  8. [7]

    A Rapidly Convergent Descent Method for Minimization

    R. Fletcher and M. J. D. Powell, “A Rapidly Convergent Descent Method for Minimization”, The Computer Journal6.2 (Aug. 1963), pp. 163–168, Aug. 1963, doi: 10.1093/comjnl/6.2.163, eprint: https://academic.oup.com/comjnl/article- pdf/6/2/163/1041527/6-2-163.pdf, url: https://doi.org/10.1093/comjnl/6.2.163

    doi:10.1093/comjnl/6.2.163 1963

Show all 28 references
  1. [8]

    Robust estimation with unknown noise statistics

    Ž. Ðurović and B. Kovačević, “Robust estimation with unknown noise statistics”, IEEE Transactions on Automatic Control44.6 (1999), pp. 1292–1296, 1999

    1999

  2. [9]

    Robust Kalman filtering based on Mahalanobis distance as outlier judging criterion

    G. Chang, “Robust Kalman filtering based on Mahalanobis distance as outlier judging criterion”,Journal of Geodesy88.4 (2014), pp. 391–401, 2014

    2014

  3. [10]

    Outlier-robust Kalman Filtering through Generalised Bayes

    G. Duran-Martin, M. Altamirano, A. Shestopaloff, L. Sánchez-Betancourt, J. Knoblauch, M. Jones, F.-X. Briol, and K. Murphy, “Outlier-robust Kalman Filtering through Generalised Bayes”,arXiv preprint arXiv:2405.05646(2024), 2024

    2024 arXiv

  4. [11]

    A Kalman filter for robust outlier detection

    “A Kalman filter for robust outlier detection”, in 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2007

    2007

  5. [12]

    An outlier-robust Kalman filter

    “An outlier-robust Kalman filter”, in 2011 IEEE international conference on robotics and automation, 2011

    2011

  6. [13]

    Robust Estimation of a Location Parameter

    P. Huber, “Robust Estimation of a Location Parameter”,The Annals of Mathe- matical Statistics35.1 (1964), pp. 73–101, 1964,doi: 10.1214/aoms/1177703732, url: https://doi.org/10.1214/aoms/1177703732

    1964

  7. [14]

    Robust Kalman filter and its application in time series analysis

    T. Cipra and R. Romera, “Robust Kalman filter and its application in time series analysis”, Kybernetika 27.6 (1991), pp. 481–494, 1991

    1991

  8. [15]

    On robust Kalman filtering

    B. Kovačević, Ž. Ðurović, and S. Glavaški, “On robust Kalman filtering”,Interna- tional Journal of Control56.3 (1992), pp. 547–562, 1992

    1992

  9. [16]

    CVXGEN: A code generator for embedded convex optimization

    J. Mattingley and S. Boyd, “CVXGEN: A code generator for embedded convex optimization”, Optimization and Engineering13 (2012), pp. 1–27, 2012. 22 References

    2012

  10. [17]

    Random sample consensus (RANSAC)

    H. Cantzler, “Random sample consensus (RANSAC)”,Institute for Perception, Action and Behaviour, Division of Informatics, University of Edinburgh3 (1981), 1981

    1981

  11. [18]

    KALMANSAC: Robust filtering by consensus

    “KALMANSAC: Robust filtering by consensus”, in Tenth IEEE International Conference on Computer Vision (ICCV’05) Volume 1, 2005

    2005

  12. [19]

    Recursive outlier-robust filtering and smoothing for nonlinear systems using the multivariate Student-t distribution

    “Recursive outlier-robust filtering and smoothing for nonlinear systems using the multivariate Student-t distribution”, in 2012 IEEE International Workshop on Machine Learning for Signal Processing, 2012

    2012

  13. [20]

    Nonlinear regression Huber–Kalman filtering and fixed-interval smoothing

    C. Karlgaard, “Nonlinear regression Huber–Kalman filtering and fixed-interval smoothing”, Journal of guidance, control, and dynamics38.2 (2015), pp. 322–330, 2015

    2015

  14. [22]

    Generalised Bayesian filtering via sequential Monte Carlo

    A. Boustati, O. D. Akyildiz, T. Damoulas, and A. Johansen, “Generalised Bayesian filtering via sequential Monte Carlo”,Advances in neural information processing systems 33 (2020), pp. 418–429, 2020

    2020

  15. [23]

    Simon, John Wiley & Sons, 2006

    D. Simon, John Wiley & Sons, 2006

    2006

  16. [24]

    Polyak, Optimization Software, 1987

    B. Polyak, Optimization Software, 1987

    1987

  17. [25]

    Goulart and Y

    P. Goulart and Y. Chen,Clarabel: An interior-point solver for conic programs with quadratic objectives, 2024, arXiv: 2405.12762 (math.OC)

    2024 arXiv

  18. [26]

    Bequette, Prentice Hall Englewood Cliffs, NJ, 1998

    B. Bequette, Prentice Hall Englewood Cliffs, NJ, 1998

    1998

  19. [27]

    Seborg, T

    D. Seborg, T. Edgar, D. Mellichamp, and F. Doyle III, John Wiley & Sons, 2016

    2016

  20. [28]

    About the Authors Alan Yang is a Ph.D candidate in Electri- cal Engineering at Stanford University

    MathWorks, Inc.,CSTR Model, https://www.mathworks.com/help/mpc/gs/cstr- model.html, 2025. About the Authors Alan Yang is a Ph.D candidate in Electri- cal Engineering at Stanford University. He received the B.S. degree in Electrical En- gineering from the University of Illinois...

    2025

Pith tools

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