pith. sign in

arxiv: 2604.02749 · v1 · submitted 2026-04-03 · 📡 eess.SY · cs.SY

Residual-Aware Distributionally Robust EKF: Absorbing Linearization Mismatch via Wasserstein Ambiguity

Minhyuk Jang , Jungjin Lee , Astghik Hakobyan , Naira Hovakimyan , Insoon Yang This is my paper

Pith reviewed 2026-05-13 19:56 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords extended Kalman filterdistributionally robust estimationWasserstein ambiguity setlinearization errornonlinear state estimationsemidefinite programmingrobust filtering
0
0 comments X

The pith

A residual-aware distributionally robust EKF absorbs linearization residuals into a Wasserstein ambiguity set to provide deterministic bounds on nonlinear estimation errors.

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

This paper presents a modified extended Kalman filter that treats errors from linearizing the system dynamics as part of the uncertainty model. By using a Wasserstein distance-based ambiguity set that changes at each time step, the method combines handling of noise model mismatch and approximation errors in one framework. The result is a filter that computes an effective radius for the ambiguity set and gives guaranteed upper bounds on the mean squared errors for the true nonlinear system. The filter can be solved efficiently using semidefinite programming and maintains the step-by-step recursive nature of the original EKF. Tests in target tracking and robot navigation show better performance when the model does not match reality perfectly.

Core claim

The paper establishes that by absorbing linearization residuals as uncertainty within a stage-wise Wasserstein ambiguity set, the distributionally robust EKF can yield computable effective radii and deterministic upper bounds on the prior and posterior mean-squared errors of the true nonlinear estimation error, while admitting a tractable semidefinite programming reformulation that preserves the recursive structure of the classical EKF.

What carries the argument

Stage-wise Wasserstein ambiguity set that absorbs linearization residuals as part of the uncertainty model in the distributionally robust state estimation.

If this is right

  • Deterministic upper bounds are obtained on the prior and posterior mean-squared errors of the true nonlinear estimation error.
  • The filter admits a tractable semidefinite programming reformulation.
  • The recursive structure of the classical EKF is preserved.
  • Improved estimation accuracy is demonstrated in coordinated-turn target tracking under model mismatch.
  • Better safety is achieved in uncertainty-aware robot navigation compared to standard EKF.

Where Pith is reading between the lines

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

  • The approach could be adapted to other recursive estimators like the unscented Kalman filter by defining similar residual-based ambiguity sets.
  • The computable effective radius might enable online adaptation of the robustness level based on observed residuals.
  • This formulation opens connections to distributionally robust control problems where similar ambiguity sets ensure performance under uncertainty.

Load-bearing premise

Linearization residuals can be captured effectively by the stage-wise Wasserstein ambiguity set without excessive conservatism or loss of the recursive structure.

What would settle it

Observing a nonlinear estimation error that exceeds the provided deterministic upper bound in a controlled simulation with known mismatch would disprove the bound's validity.

Figures

Figures reproduced from arXiv: 2604.02749 by Astghik Hakobyan, Insoon Yang, Jungjin Lee, Minhyuk Jang, Naira Hovakimyan.

Figure 1
Figure 1. Figure 1: Coordinated-turn experiment under nominal covariance misspecification. (a) EKF with [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Closed-loop navigation under uncertainty-aware MPC. (a) Representative trajectories for [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Closed-loop navigation under uncertainty-aware MPC with pedestrian prediction on the [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
read the original abstract

The extended Kalman filter (EKF) is a cornerstone of nonlinear state estimation, yet its performance is fundamentally limited by noise-model mismatch and linearization errors. We develop a residual-aware distributionally robust EKF that addresses both challenges within a unified Wasserstein distributionally robust state estimation framework. The key idea is to treat linearization residuals as uncertainty and absorb them into an effective uncertainty model captured by a stage-wise ambiguity set, enabling noise-model mismatch and approximation errors to be handled within a single formulation. This approach yields a computable effective radius along with deterministic upper bounds on the prior and posterior mean-squared errors of the true nonlinear estimation error. The resulting filter admits a tractable semidefinite programming reformulation while preserving the recursive structure of the classical EKF. Simulations on coordinated-turn target tracking and uncertainty-aware robot navigation demonstrate improved estimation accuracy and safety compared to standard EKF baselines under model mismatch and nonlinear effects.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a residual-aware distributionally robust extended Kalman filter (EKF) that absorbs linearization residuals into a stage-wise Wasserstein ambiguity set, unifying treatment of noise-model mismatch and approximation errors. It claims this yields a computable effective radius, deterministic upper bounds on the prior and posterior mean-squared errors of the true nonlinear estimation error, and a tractable SDP reformulation that preserves the recursive EKF structure. The approach is validated via simulations on coordinated-turn target tracking and uncertainty-aware robot navigation, demonstrating improved accuracy and safety under mismatch and nonlinearity.

Significance. If the deterministic bounds and radius construction hold, the work would offer a principled way to obtain guaranteed performance for nonlinear filters without sacrificing recursivity or tractability. This could be valuable for safety-critical applications in robotics and tracking where linearization errors compound with model mismatch. The unification of residuals into the ambiguity set and the SDP reformulation are the primary contributions, provided the radius derivation is rigorous.

major comments (2)
  1. [Abstract and ambiguity-set construction] The central claim of deterministic upper bounds on the true nonlinear MSE requires that the stage-wise Wasserstein radius provably contains the actual linearization residual distribution at every step (see skeptic note on remainder bounds). Without an explicit proof that the radius derivation (presumably from Taylor remainder) upper-bounds higher-order terms for all trajectories, the worst-case SDP solution may optimize over an incorrect set and the bounds may not apply to the true error.
  2. [Recursive structure and radius computation] Preservation of the recursive filter structure requires that the per-stage radius be computable from the current predicted covariance alone, without dependence on future states or measurements. The manuscript must clarify whether the residual distribution induced by the unknown true state violates this (see skeptic note on lookahead).
minor comments (2)
  1. [Simulations] The simulation section would benefit from explicit reporting of the computed radii across time steps and trajectories to allow readers to assess conservatism.
  2. [Notation and preliminaries] Notation for the effective radius and the Wasserstein ball should be introduced with a clear equation reference early in the manuscript to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. The concerns regarding the rigor of the radius derivation and preservation of recursivity are important, and we will strengthen the manuscript with additional proofs and clarifications in the revision.

read point-by-point responses

  1. Referee: [Abstract and ambiguity-set construction] The central claim of deterministic upper bounds on the true nonlinear MSE requires that the stage-wise Wasserstein radius provably contains the actual linearization residual distribution at every step (see skeptic note on remainder bounds). Without an explicit proof that the radius derivation (presumably from Taylor remainder) upper-bounds higher-order terms for all trajectories, the worst-case SDP solution may optimize over an incorrect set and the bounds may not apply to the true error.

    Authors: We agree that an explicit proof is required to confirm containment. The radius is derived from the Taylor remainder under the assumption of twice-differentiable dynamics with bounded Hessians (as stated in Assumptions 1-2 of the manuscript). This yields a uniform bound on the remainder term proportional to the squared state deviation, which is controlled by the predicted covariance. In the revision we will insert a dedicated lemma proving that the resulting Wasserstein ball contains the residual distribution for all trajectories, ensuring the SDP solution and MSE bounds apply to the true nonlinear error. revision: yes

  2. Referee: [Recursive structure and radius computation] Preservation of the recursive filter structure requires that the per-stage radius be computable from the current predicted covariance alone, without dependence on future states or measurements. The manuscript must clarify whether the residual distribution induced by the unknown true state violates this (see skeptic note on lookahead).

    Authors: The per-stage radius is computed exclusively from the current predicted covariance, the linearization point, and known model parameters; it has no dependence on future states or measurements. The unknown true state is accounted for by optimizing over the worst-case distribution inside the ambiguity set, but the radius formula itself remains causal. We will revise Section III to include an explicit statement and short proof that the construction is recursive and requires no lookahead, thereby preserving the EKF structure. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper constructs a residual-aware Wasserstein ambiguity set to absorb linearization mismatch, derives a computable effective radius, and obtains deterministic MSE bounds via SDP reformulation while retaining EKF recursion. No quoted step reduces a claimed prediction or bound to a fitted parameter or self-citation by construction; the radius and bounds are presented as following from the ambiguity-set definition and worst-case analysis without tautological re-use of the target nonlinear error. The approach is independent of post-hoc fitting and does not rely on load-bearing self-citations for its central guarantee.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on modeling linearization residuals via a stage-wise Wasserstein ambiguity set and standard properties of distributionally robust optimization; no new entities are postulated.

free parameters (1)
  • ambiguity set radius
    Effective radius is stated as computable but depends on residual characterization, acting as a tunable parameter in the formulation.
axioms (1)
  • standard math Wasserstein distance defines valid ambiguity sets for distributionally robust optimization
    Invoked to construct the uncertainty model capturing both noise mismatch and linearization errors.

pith-pipeline@v0.9.0 · 5477 in / 1188 out tokens · 35393 ms · 2026-05-13T19:56:40.505074+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,”J. Basic Eng., vol. 82, pp. 35–45, 1960

    work page 1960

  2. [2]

    Simon,Optimal State Estimation: Kalman,H ∞, and Nonlinear Approaches

    D. Simon,Optimal State Estimation: Kalman,H ∞, and Nonlinear Approaches. John Wiley & Sons, 2006

    work page 2006

  3. [3]

    New approaches to robust minimum variance filter design,

    U. Shaked, L. Xie, and Y. C. Soh, “New approaches to robust minimum variance filter design,” IEEE Trans. Signal Process., vol. 49, no. 11, pp. 2620–2629, 2001

    work page 2001

  4. [4]

    Optimization of stochastic linear systems with additive measurement and process noise using exponential performance criteria,

    J. Speyer, J. Deyst, and D. Jacobson, “Optimization of stochastic linear systems with additive measurement and process noise using exponential performance criteria,”IEEE Trans. Autom. Control, vol. 19, no. 4, pp. 358–366, 1974

    work page 1974

  5. [5]

    Risk-sensitive linear/quadratic/Gaussian control,

    P. Whittle, “Risk-sensitive linear/quadratic/Gaussian control,”Adv. Appl. Probab., vol. 13, no. 4, pp. 764–777, 1981

    work page 1981

  6. [6]

    Robustness and risk-sensitive filtering,

    R. K. Boel, M. R. James, and I. R. Petersen, “Robustness and risk-sensitive filtering,”IEEE Trans. Autom. Control, vol. 47, no. 3, pp. 451–461, 2002

    work page 2002

  7. [7]

    Bounding linearization errors with sets of densities in approximate Kalman filtering,

    B. Noack, V. Klumpp, N. Petkov, and U. D. Hanebeck, “Bounding linearization errors with sets of densities in approximate Kalman filtering,” inProc. Int. Conf. Inf. Fusion, 2010, pp. 1–8

    work page 2010

  8. [8]

    High-order extended Kalman filter for state estimation of nonlinear systems,

    L. Ding and C. Wen, “High-order extended Kalman filter for state estimation of nonlinear systems,”Symmetry, vol. 16, no. 5, p. 617, 2024

    work page 2024

  9. [9]

    A simple and efficient tube-based robust output feedback model predictive control scheme,

    J. Lorenzetti and M. Pavone, “A simple and efficient tube-based robust output feedback model predictive control scheme,” inProc. IEEE Eur. Control Conf., 2020, pp. 1775–1782

    work page 2020

  10. [10]

    Robust output feedback MPC for uncertain linear systems with reduced conservatism,

    M. K¨ ogel and R. Findeisen, “Robust output feedback MPC for uncertain linear systems with reduced conservatism,”IFAC-PapersOnLine, vol. 50, no. 1, pp. 10 685–10 690, 2017

    work page 2017

  11. [11]

    Robust state space filtering under incremental model pertur- bations subject to a relative entropy tolerance,

    B. C. Levy and R. Nikoukhah, “Robust state space filtering under incremental model pertur- bations subject to a relative entropy tolerance,”IEEE Trans. Autom. Control, vol. 58, no. 3, pp. 682–695, 2012

    work page 2012

  12. [12]

    Robust Kalman filtering under model perturbations,

    M. Zorzi, “Robust Kalman filtering under model perturbations,”IEEE Trans. Autom. Control, vol. 62, no. 6, pp. 2902–2907, 2016

    work page 2016

  13. [13]

    On the robustness of the Bayes and Wiener estimators under model uncertainty,

    ——, “On the robustness of the Bayes and Wiener estimators under model uncertainty,” Automatica, vol. 83, pp. 133–140, 2017

    work page 2017

  14. [14]

    Wasserstein distributionally robust Kalman filtering,

    S. Shafieezadeh Abadeh, V. A. Nguyen, D. Kuhn, and P. M. Mohajerin Esfahani, “Wasserstein distributionally robust Kalman filtering,”Adv. Neural Inf. Process. Syst., vol. 31, 2018

    work page 2018

  15. [15]

    Bridging Bayesian and minimax mean square error estimation via Wasserstein distributionally robust optimization,

    V. A. Nguyen, S. Shafieezadeh-Abadeh, D. Kuhn, and P. Mohajerin Esfahani, “Bridging Bayesian and minimax mean square error estimation via Wasserstein distributionally robust optimization,”Math. Oper. Res., vol. 48, no. 1, pp. 1–37, 2023

    work page 2023

  16. [16]

    Distributionally robust Kalman filtering with volatility uncertainty,

    B. Han, “Distributionally robust Kalman filtering with volatility uncertainty,”IEEE Trans. Autom. Control, vol. 70, no. 6, pp. 4000–4007, 2025

    work page 2025

  17. [17]

    Distributionally robust Kalman filtering over finite and infinite horizon,

    T. Kargin, J. Hajar, V. Malik, and B. Hassibi, “Distributionally robust Kalman filtering over finite and infinite horizon,”arXiv preprint arXiv:2407.18837, 2024

    work page arXiv 2024

  18. [18]

    On the steady-state distributionally robust Kalman filter,

    M. Jang, A. Hakobyan, and I. Yang, “On the steady-state distributionally robust Kalman filter,” inProc. IEEE Conf. Decis. Control, 2025, pp. 2385–2392

    work page 2025

  19. [19]

    Exactly or approximately Wasserstein distribu- tionally robust estimation according to Wasserstein radii being small or large,

    X. Ding, E. Song, D. Niu, Z. Cao, and Q. Shi, “Exactly or approximately Wasserstein distribu- tionally robust estimation according to Wasserstein radii being small or large,”arXiv preprint arXiv:2510.01763, 2025

    work page arXiv 2025

  20. [20]

    Distributionally robust state estimation for linear systems subject to uncertainty and outlier,

    S. Wang and Z.-S. Ye, “Distributionally robust state estimation for linear systems subject to uncertainty and outlier,”IEEE Trans. Signal Process., vol. 70, pp. 452–467, 2021

    work page 2021

  21. [21]

    Distributionally robust Kalman filtering for INS/GPS tightly coupled integration with model uncertainty and measurement outlier,

    K. Si, P. Li, Z.-P. Yuan, K. Qiao, B. Wang, and X. He, “Distributionally robust Kalman filtering for INS/GPS tightly coupled integration with model uncertainty and measurement outlier,”IEEE Trans. Instrum. Meas., vol. 72, pp. 1–13, 2023

    work page 2023

  22. [22]

    Sinkhorn distributionally robust state estimation via system level synthesis,

    Y. Feng, X. Li, S. X. Ding, H. Ye, and C. Shang, “Sinkhorn distributionally robust state estimation via system level synthesis,”arXiv preprint arXiv:2602.08018, 2026

    work page arXiv 2026

  23. [23]

    Distributionally robust state estimation for nonlinear systems,

    S. Wang, “Distributionally robust state estimation for nonlinear systems,”IEEE Trans. Signal Process., vol. 70, pp. 4408–4423, 2022

    work page 2022

  24. [24]

    Villaniet al.,Optimal transport: old and new

    C. Villaniet al.,Optimal transport: old and new. Springer, 2009, vol. 338

    work page 2009

  25. [25]

    Wasserstein distri- butionally robust optimization: Theory and applications in machine learning,

    D. Kuhn, P. M. Esfahani, V. A. Nguyen, and S. Shafieezadeh-Abadeh, “Wasserstein distri- butionally robust optimization: Theory and applications in machine learning,” inOper. Res. Manag. Sci. Age Analyt.INFORMS, 2019, pp. 130–166

    work page 2019

  26. [26]

    Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations,

    P. Mohajerin Esfahani and D. Kuhn, “Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations,”Math. Pro- gram., vol. 171, no. 1, pp. 115–166, 2018

    work page 2018

  27. [27]

    Distributionally robust Kalman filter,

    M. Jang, A. Hakobyan, and I. Yang, “Distributionally robust Kalman filter,”arXiv preprint arXiv:2512.06286, 2025

    work page arXiv 2025

  28. [28]

    Stochastic stability of the discrete-time extended Kalman filter,

    K. Reif, S. G¨ unther, E. Yaz, and R. Unbehauen, “Stochastic stability of the discrete-time extended Kalman filter,”IEEE Trans. Autom. Control, vol. 44, no. 4, pp. 714–728, 1999

    work page 1999

  29. [29]

    Best choice of coordinate system for tracking coordinated turns,

    F. Gustafsson and A. J. Isaksson, “Best choice of coordinate system for tracking coordinated turns,” inProc. IEEE Conf. Decis. Control, vol. 3, 1996, pp. 3145–3150

    work page 1996

  30. [30]

    EKF/UKF maneuvering target tracking using coordinated turn models with polar/Cartesian velocity,

    M. Roth, G. Hendeby, and F. Gustafsson, “EKF/UKF maneuvering target tracking using coordinated turn models with polar/Cartesian velocity,” inProc. Int. Conf. Inf. Fusion, 2014, pp. 1–8

    work page 2014

  31. [31]

    Crowds by example,

    A. Lerner, Y. Chrysanthou, and D. Lischinski, “Crowds by example,” inComputer graphics forum, vol. 26, no. 3. Wiley Online Library, 2007, pp. 655–664

    work page 2007

  32. [32]

    You’ll never walk alone: Modeling social behavior for multi-target tracking,

    S. Pellegrini, A. Ess, K. Schindler, and L. Van Gool, “You’ll never walk alone: Modeling social behavior for multi-target tracking,” in2009 IEEE 12th international conference on computer vision. IEEE, 2009, pp. 261–268

    work page 2009

  33. [33]

    Trajectron++: Dynamically- feasible trajectory forecasting with heterogeneous data,

    T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone, “Trajectron++: Dynamically- feasible trajectory forecasting with heterogeneous data,” inEuropean conference on computer vision. Springer, 2020, pp. 683–700

    work page 2020