pith. sign in

arxiv: 2604.21644 · v1 · submitted 2026-04-23 · 📡 eess.SY · cs.SY· eess.SP

An Adaptive Kalman Filter that Learns the Coloring Dynamics of the Process Noise

Mohammad Almuhaihi , Dennis Bernstein This is my paper

Pith reviewed 2026-05-09 21:07 UTC · model grok-4.3

classification 📡 eess.SY cs.SYeess.SP
keywords adaptive Kalman filtercolored process noiseinnovations sequencestate augmentationonline adaptationwhitening filterstate estimation
0
0 comments X

The pith

An adaptive Kalman filter learns the unknown coloring dynamics of process noise by driving innovations toward whiteness.

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

The paper addresses state estimation when process noise has unknown colored dynamics, which makes standard Kalman filters suboptimal. It shows that if measurement noise is white, then the innovations sequence is white exactly when the process noise is correctly modeled. The proposed Innovations-Whitening Adaptive Kalman Filter augments the state with an unknown coloring filter and tunes its parameters online by minimizing the empirical autocorrelation of the innovations. This approach restores near-optimality without requiring prior knowledge of the coloring dynamics. A sympathetic reader would care because many real-world systems have colored disturbances whose statistics are hard to obtain in advance.

Core claim

By embedding an unknown coloring filter inside a state-augmentation framework, the Innovations-Whitening Adaptive Kalman Filter (IWAKF) adapts its parameters by minimizing the empirical autocorrelation of the innovations sequence. This drives the innovations toward whiteness, which holds if and only if the process noise is white or correctly colored, thereby recovering near-optimal estimates without prior knowledge of the coloring dynamics.

What carries the argument

Innovations-Whitening Adaptive Kalman Filter (IWAKF), which augments the state vector with an unknown coloring filter and adapts its parameters to minimize the sample autocorrelation of the innovations.

If this is right

  • The method applies to any linear time-invariant system whose process noise can be modeled as the output of an unknown linear filter driven by white noise.
  • No separate identification step is required; adaptation occurs simultaneously with state estimation.
  • Near-optimality is restored in the sense that the filter covariance approaches the true error covariance once the innovations become white.
  • The technique requires only the ability to compute the sample autocorrelation of the innovations over a sliding window.

Where Pith is reading between the lines

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

  • The same whitening principle could be extended to time-varying coloring dynamics by allowing the filter parameters to vary slowly.
  • If the measurement noise is only approximately white, the adaptation may still reduce but not eliminate innovation correlation.
  • The approach suggests a general recipe for adaptive filters: embed the unknown dynamics and tune them to enforce a known optimality condition such as whiteness of residuals.

Load-bearing premise

Measurement noise must be white, and the innovations sequence must be white if and only if the process noise is white or correctly modeled by the augmented coloring filter.

What would settle it

Apply the filter to a linear system whose process noise is colored by a known filter but whose parameters are deliberately mis-set; the empirical autocorrelation of the innovations should remain significantly above zero after adaptation.

Figures

Figures reproduced from arXiv: 2604.21644 by Dennis Bernstein, Mohammad Almuhaihi.

Figure 1
Figure 1. Figure 1: Pendulum schematic defining the angle θ and its reference. which implies that Φz(ϕ) is constant across all frequen￾cies, thereby ensuring that the signal z is white. This establishes sufficiency. To prove necessity, let Φz(ϕ) = S. Substituting back into (20) yields S = (I − G)R(I − G) ∗ + HΦv(ϕ)H∗ . Since (C, F) is observable, which is implied by (A, C) detectability, H has full column rank. Thus, its pseu… view at source ↗
Figure 3
Figure 3. Figure 3: Performance ratio comparison. The figure [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Innovations autocorrelation comparison. The [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
read the original abstract

In many applications of state estimation, the process noise is colored; this case is addressed by applying the standard Kalman filter (KF) to dynamics that are augmented with the coloring dynamics. The present paper considers the case where the coloring dynamics are unknown, which renders the estimates obtained from the standard approach suboptimal. To address this problem, the present paper proposes an adaptive technique based on the principle that, if the measurement noise is white, then the innovations sequence is white if and only if the process noise is white. Leveraging this fact, an Innovations-Whitening Adaptive Kalman Filter (IWAKF) is developed, which learns the process-noise coloring online. By embedding an unknown coloring filter in a state-augmentation framework, IWAKF adapts its parameters by minimizing the empirical autocorrelation of the innovations, thereby driving them toward whiteness and restoring near-optimality without prior knowledge of the coloring dynamics.

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

3 major / 1 minor

Summary. The manuscript proposes the Innovations-Whitening Adaptive Kalman Filter (IWAKF) to address state estimation when process noise is colored but its dynamics are unknown. The approach augments the state with an unknown coloring filter whose parameters are adapted online by minimizing the empirical autocorrelation of the innovations sequence, with the goal of driving innovations toward whiteness and thereby restoring near-optimality of the Kalman filter estimates.

Significance. The method targets a practical gap in adaptive filtering by using the whiteness of innovations as a learning criterion when measurement noise is assumed white. If accompanied by convergence guarantees and validation, it could enable more robust state estimation in applications lacking prior noise-coloring knowledge. The current absence of derivations, proofs, or results, however, prevents a full assessment of its potential contribution.

major comments (3)
  1. [Abstract and Proposed Approach] The central claim that online minimization of empirical innovation autocorrelation restores near-optimality rests on the fixed-model KF property that innovations are white iff the augmented model is correct. No convergence analysis, Lyapunov argument, or error bounds are supplied to establish that the time-varying gain sequence induced by adaptation converges to the true fixed coloring parameters rather than merely whitening observed innovations transiently. (See the IWAKF description in the abstract and the proposed approach.)
  2. [Abstract] The adaptation fits coloring-filter parameters directly to the observed innovations data. This creates a circularity risk: the empirical autocorrelation can be driven to zero by construction without independent verification that the resulting parameters match the true (unknown) coloring dynamics or yield improved state estimates. A formal argument separating the fitting objective from optimality restoration is required. (See the principle stated in the abstract: 'innovations sequence is white if and only if the process noise is white'.)
  3. [Proposed Method] The manuscript provides no derivations of the adaptation law, no stability analysis for the time-varying closed-loop system, no error analysis, and no empirical results. These omissions make it impossible to evaluate whether the claimed restoration of optimality holds under the stated assumptions. (See the full description of IWAKF development.)
minor comments (1)
  1. [Abstract] The abstract uses the term 'near-optimality' without defining the performance metric or providing quantitative bounds on the expected improvement.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. These highlight the need for rigorous theoretical support and validation to fully substantiate the proposed IWAKF method. We address each major comment below and will incorporate the necessary revisions into the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Proposed Approach] The central claim that online minimization of empirical innovation autocorrelation restores near-optimality rests on the fixed-model KF property that innovations are white iff the augmented model is correct. No convergence analysis, Lyapunov argument, or error bounds are supplied to establish that the time-varying gain sequence induced by adaptation converges to the true fixed coloring parameters rather than merely whitening observed innovations transiently. (See the IWAKF description in the abstract and the proposed approach.)

    Authors: We agree that the submitted manuscript provides only a conceptual outline and lacks the requested convergence analysis. In the revision, we will add a dedicated section deriving a Lyapunov stability argument for the adaptive system. Under standard persistence-of-excitation conditions on the regressor formed by past innovations, we will prove asymptotic convergence of the coloring-filter parameters to their true values. This will be complemented by explicit bounds on the state-estimation error that quantify the degradation due to transient parameter mismatch. revision: yes

  2. Referee: [Abstract] The adaptation fits coloring-filter parameters directly to the observed innovations data. This creates a circularity risk: the empirical autocorrelation can be driven to zero by construction without independent verification that the resulting parameters match the true (unknown) coloring dynamics or yield improved state estimates. A formal argument separating the fitting objective from optimality restoration is required. (See the principle stated in the abstract: 'innovations sequence is white if and only if the process noise is white'.)

    Authors: The cited principle is the classical Kalman-filter optimality condition: when measurement noise is white, the innovations are white if and only if the internal model (including process-noise coloring) is correct. The adaptation minimizes a consistent estimator of the innovation autocorrelation; we will augment the revision with a formal identifiability proof showing that the cost function attains its unique global minimum precisely at the true coloring parameters, provided the system is detectable and the measurement noise remains white. This establishes that driving the empirical autocorrelation to zero is equivalent to recovering the correct model and thereby restoring KF optimality. revision: yes

  3. Referee: [Proposed Method] The manuscript provides no derivations of the adaptation law, no stability analysis for the time-varying closed-loop system, no error analysis, and no empirical results. These omissions make it impossible to evaluate whether the claimed restoration of optimality holds under the stated assumptions. (See the full description of IWAKF development.)

    Authors: We acknowledge that the current version contains only a high-level description without the supporting derivations or experiments. The revised manuscript will supply: (i) the explicit stochastic-gradient adaptation law obtained by differentiating the sample autocorrelation cost with respect to the coloring-filter coefficients; (ii) a stability analysis of the resulting time-varying closed-loop system via averaging theory, establishing uniform ultimate boundedness of the parameter and state errors; (iii) an error-propagation analysis that bounds the Kalman-filter estimation error in terms of the instantaneous parameter error; and (iv) Monte-Carlo simulation results on linear and mildly nonlinear systems comparing IWAKF against the non-adaptive KF and existing covariance-matching adaptive filters. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper explicitly constructs IWAKF to minimize empirical autocorrelation of innovations based on the leveraged standard KF principle that white innovations imply correct process-noise modeling (given white measurements). This is a direct algorithmic implementation rather than a reduction of any claimed prediction or first-principles result to its own inputs by construction. No self-citations, self-definitional steps, fitted parameters renamed as independent predictions, or uniqueness theorems are quoted or required in the abstract or description. The central claim follows from the stated adaptation rule and external theory; any questions about time-varying effects during adaptation concern correctness or convergence, not circularity in the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on one domain assumption about innovations whiteness and introduces free parameters for the unknown coloring filter that are learned from data.

free parameters (1)
  • parameters of the unknown coloring filter
    These parameters are adapted online by minimizing the empirical autocorrelation of the innovations sequence.
axioms (1)
  • domain assumption If the measurement noise is white, then the innovations sequence is white if and only if the process noise is white.
    This principle underpins the adaptation strategy for driving the filter toward optimality.

pith-pipeline@v0.9.0 · 5453 in / 1230 out tokens · 64258 ms · 2026-05-09T21:07:02.795476+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

125 extracted references · 125 canonical work pages

  1. [1]

    2014 , volume=

    Polyakov, Andrey and Efimov, Denis and Perruquetti, Wilfrid , booktitle=. 2014 , volume=

    work page 2014

  2. [2]

    1998 , author =

    Automatica , volume =. 1998 , author =

    work page 1998

  3. [3]

    J. Comp. and Appl. Math. , volume =. 2011 , author =

    work page 2011

  4. [4]

    Ibrir, Salim and Diop, Sette , pages =. Int. J. Appl. Math. Computer Sci. , year =

    work page

  5. [5]

    Int. J. Contr. , author =

    work page

  6. [6]

    Mohseni and D

    N. Mohseni and D. S. Bernstein , title=. Proc. Amer. Contr. Conf. , year=

    work page

  7. [7]

    Bruce and A

    A. Bruce and A. Goel and D. S. Bernstein , title=. Automatica , volume =

    work page

  8. [8]

    P. J. Davis and P. Rabinowitz , publisher =. 1984 , edition =

    work page 1984

  9. [9]

    R. W. Hamming , publisher =. 1973 , edition =

    work page 1973

  10. [10]

    2008 , publisher=

    Aided Navigation: GPS with High Rate Sensors , author=. 2008 , publisher=

    work page 2008

  11. [11]

    M. S. Grewel and A. P. Andrews and C. G. Bartone , year=

    work page

  12. [12]

    A. G. Ramm and A. Smirnova , journal=. 2003 , volume=

    work page 2003

  13. [13]

    Numerical differentiation from a viewpoint of regularization theory , volume =

    Lu, Shuai and Pereverzyev, Sergei , year =. Numerical differentiation from a viewpoint of regularization theory , volume =

    work page

  14. [14]

    Numerical Differentiation of Experimental Data: Local Versus Global Methods , volume =

    Ahnert, Karsten and Abel, Markus , year =. Numerical Differentiation of Experimental Data: Local Versus Global Methods , volume =

    work page

  15. [15]

    VAN Nieuwstadt and M

    M. VAN Nieuwstadt and M. Rathinam and R. M. Murray , year =. Differential Flatness and Absolute Equivalence of Nonlinear Control Systems , volume =

    work page

  16. [16]

    2016 , volume=

    Bania, Piotr and Baranowski, Jerzy , booktitle=. 2016 , volume=

    work page 2016

  17. [17]

    , journal=

    Schafer, Ronald W. , journal=. 2011 , volume=

    work page 2011

  18. [18]

    , volume =

    Fire Safety J. , volume =. 2005 , author =

    work page 2005

  19. [19]

    S. Proc. Amer. Contr. Conf. , year=

    work page

  20. [20]

    Walker, Jeffrey A , year =. J. Exp. Biol. , volume =

    work page

  21. [21]

    and Grizzle, J.W

    Diop, S. and Grizzle, J.W. and Chaplais, F. , booktitle=

    work page

  22. [22]

    M. Medit. Conf. Contr. Autom. , title=. 2007 , pages=

    work page 2007

  23. [23]

    Chartrand, Rick , journal=

    work page

  24. [24]

    Mojallizadeh, Mohammad Rasool and Brogliato, Bernard and Acary, Vincent , journal =

    work page

  25. [25]

    Dogan and Girard, Anouck and Bernstein, Dennis S

    Verma, Shashank and Sanjeevini, Sneha and Sumer, E. Dogan and Girard, Anouck and Bernstein, Dennis S. , booktitle=. 2022 , volume=

    work page 2022

  26. [26]

    Dogan and Girard, Anouck and Bernstein, Dennis S

    Verma, Shashank and Sanjeevini, Sneha and Sumer, E. Dogan and Girard, Anouck and Bernstein, Dennis S. , booktitle=. On the Accuracy of Numerical Differentiation Using High-Gain Observers and Adaptive Input Estimation , year=

    work page

  27. [27]

    and Patton, R

    Hou, M. and Patton, R. J. , journal=

    work page

  28. [28]

    Corless and J

    M. Corless and J. Tu. State and Input Estimation for a Class of Uncertain Systems. Automatica. 1998

    work page 1998

  29. [29]

    and Barbot, J

    Floquet, T. and Barbot, J. P. State and Unknown Input Estimation for Linear Discrete-Time Systems. Automatica. 2006

    work page 2006

  30. [30]

    Palanthandalam-Madapusi and D

    H. Palanthandalam-Madapusi and D. S. Bernstein. A Subspace Algorithm for Simultaneous Identification and Input Reconstruction. Int. J. Adapt. Contr. Sig. Proc. 2009

    work page 2009

  31. [31]

    and Palanthandalam-Madapusi, H

    Kirtikar, S. and Palanthandalam-Madapusi, H. and Zattoni, E. and Bernstein, D. S. , journal=

    work page

  32. [32]

    Fang and R

    H. Fang and R. A. de Callafon and J. Cortes. Simultaneous Input and State Estimation for Nonlinear Systems with Applications to Flow Field Estimation. Automatica. 2013

    work page 2013

  33. [33]

    and Zhu, F

    Yang, J. and Zhu, F. and Sun, X. , journal=

    work page

  34. [34]

    and van Kampen, E.-J

    Lu, P. and van Kampen, E.-J. and de Visser, C. C. and Chu, Q. , journal=

    work page

  35. [35]

    Sanjeevini and D

    S. Sanjeevini and D. S. Bernstein , journal=

    work page

  36. [36]

    Arie Levant , title =. Int. J. Contr. , volume =. 2003 , publisher =

    work page 2003

  37. [37]

    , journal=

    Bogler, P.L. , journal=. 1987 , volume=

    work page 1987

  38. [38]

    Bar-Shalom and X

    Y. Bar-Shalom and X. R. Li and T. Kirubarajan , publisher =

    work page

  39. [39]

    Khaloozadeh and A

    H. Khaloozadeh and A. Karsaz , title =. IET Radar Sonar Nav. , issue =. 2009 , pages =

    work page 2009

  40. [40]

    1999 , volume=

    Hungu Lee and Min-Jea Tahk , journal=. 1999 , volume=

    work page 1999

  41. [41]

    2019 , volume=

    Cui, Henggang and Radosavljevic, Vladan and Chou, Fang-Chieh and Lin, Tsung-Han and Nguyen, Thi and Huang, Tzu-Kuo and Schneider, Jeff and Djuric, Nemanja , booktitle=. 2019 , volume=

    work page 2019

  42. [42]

    and Desbiens, Andre and Gagnon, Eric , booktitle=

    Prevost, Carole G. and Desbiens, Andre and Gagnon, Eric , booktitle=. 2007 , volume=

    work page 2007

  43. [43]

    2017 , volume=

    Altché, Florent and de La Fortelle, Arnaud , booktitle=. 2017 , volume=

    work page 2017

  44. [44]

    2009 , volume=

    Ammoun, Samer and Nashashibi, Fawzi , booktitle=. 2009 , volume=

    work page 2009

  45. [45]

    Algorithms , VOLUME =

    Jia, Zhen and Balasuriya, Arjuna and Challa, Subhash , TITLE =. Algorithms , VOLUME =. 2008 , NUMBER =

    work page 2008

  46. [46]

    Song, Haoran and Ding, Wenchao and Chen, Yuxuan and Shen, Shaojie and Wang, Michael Yu and Chen, Qifeng. Comp. Vision , year=

    work page

  47. [47]

    2020 , volume=

    Rana, Md Masud and Halim, Nazia and Rahamna, Md Mostafizur and Abdelhadi, Ahmed , booktitle=. 2020 , volume=

    work page 2020

  48. [48]

    , journal=

    Kalata, Paul R. , journal=. 1984 , volume=

    work page 1984

  49. [49]

    A H Mohamed and K P Schwarz , pages =. J. Geodesy , year =

    work page

  50. [50]

    Christopher Hide and Terry Moore and Martin Smith , issue =. J. Nav. , keywords =

    work page

  51. [51]

    IEEE Transactions on Automatic Control , pages =

    Lei Guo , doi =. IEEE Transactions on Automatic Control , pages =

    work page

  52. [52]

    Ali Almagbile and Jinling Wang and Weidong Ding , issue =. J. Global Pos. Sys. , month =

    work page

  53. [53]

    Mehra , issue =

    Raman K. Mehra , issue =. IEEE Trans. Autom. Contr. , pages =

    work page

  54. [54]

    Y Yang and H He and G Xu , title =

    work page

  55. [55]

    An optimal Adaptive Kalman Filter , volume =

    Yuanxi Yang and Weiguang Gao , doi =. An optimal Adaptive Kalman Filter , volume =. J. Geodesy , month =

    work page

  56. [56]

    , journal=

    Hsieh, C.-S. , journal=

    work page

  57. [57]

    and Saif, M

    Xiong, Y. and Saif, M. Unknown Disturbance Inputs Estimation Based on a State Functional Observer Design. Automatica. 2003

    work page 2003

  58. [58]

    Palanthandalam-Madapusi, H. J. and Bernstein, D. S. , booktitle=

    work page

  59. [59]

    Joseph and Junkins, John L

    Mook, D. Joseph and Junkins, John L. , title =. J. Guid. Contr. Dyn. , volume =

    work page

  60. [60]

    2008 , volume=

    Itzhak Yaesh and Uri Shaked , journal=. 2008 , volume=

    work page 2008

  61. [61]

    , title =

    Moghe, Rahul and Zanetti, Renato and Akella, Maruthi R. , title =. J. Guid. Contr. Dyn. , volume =

    work page

  62. [62]

    2009 , volume=

    Yong Shi and Chongzhao Han and Yongqi Liang , booktitle=. 2009 , volume=

    work page 2009

  63. [63]

    AIAA Guid

    Venkatesh Madyastha and Vishal Ravindra and Srinath Mallikarjunan and Anup Goyal , title =. AIAA Guid. Nav. Contr. Conf. , chapter =

    work page

  64. [64]

    and Bar-Shalom, Yaakov and Kleinman, David L

    Zhang, Lingyi and Sidoti, David and Bienkowski, Adam and Pattipati, Krishna R. and Bar-Shalom, Yaakov and Kleinman, David L. , journal=. 2020 , volume=

    work page 2020

  65. [65]

    Xie and Dennis S

    Yousaf Rahman and A. Xie and Dennis S. Bernstein , journal=

    work page

  66. [66]

    Sanjeevini, Sneha and Bernstein, Dennis S , booktitle=

    work page

  67. [67]

    2013 , issn =

    Automatica , volume =. 2013 , issn =

    work page 2013

  68. [68]

    1970 , publisher=

    Stochastic Processes and Filtering Theory , author=. 1970 , publisher=

    work page 1970

  69. [69]

    2004 , publisher=

    Optimal Estimation of Dynamic Systems , author=. 2004 , publisher=

    work page 2004

  70. [70]

    2006 , publisher=

    Optimal State Estimation: Kalman, H-infinity, and Nonlinear Approaches , author=. 2006 , publisher=

    work page 2006

  71. [71]

    IEEE Trans

    On the identification of variances and adaptive Kalman filtering , author=. IEEE Trans. Autom. Contr. , volume=. 1970 , publisher=

    work page 1970

  72. [72]

    IEEE Conf

    Data-enhanced Kalman filtering of colored process noise , author=. IEEE Conf. Dec. Contr. , pages=

    work page

  73. [73]

    , author=

    Estimation using sampled data containing sequentially correlated noise. , author=. Journal of Spacecraft and Rockets , volume=

    work page

  74. [74]

    Journal of Geodesy , volume=

    On kalman filter for linear system with colored measurement noise , author=. Journal of Geodesy , volume=. 2014 , publisher=

    work page 2014

  75. [75]

    2013 , publisher=

    Optimal Control: Linear Quadratic Methods , author=. 2013 , publisher=

    work page 2013

  76. [76]

    Kalata , journal=

    Paul R. Kalata , journal=. 1983 , pages=

    work page 1983

  77. [77]

    and Singh, T

    Tenne, D. and Singh, T. , journal=. 2002 , volume=

    work page 2002

  78. [78]

    Dogan Sumer and Dennis S

    Shashank Verma and Sneha Sanjeevini and E. Dogan Sumer and Dennis S. Bernstein , title =. Int. J. Control , publisher =

    work page

  79. [79]

    Lai and Bernstein, Dennis S

    Verma, Shashank and B. Lai and Bernstein, Dennis S. , booktitle=. 2024 , pages=

    work page 2024

  80. [80]

    Bernstein , year=

    Shashank Verma and Brian Lai and Dennis S. Bernstein , year=

    work page

Showing first 80 references.