arxiv: 2604.02735 · v1 · submitted 2026-04-03 · 🧮 math.NA · cs.NA· math.OC

Error Estimates of the Gain Approximation by Hermite-Galerkin Method in Feedback Particle Filter

Ruoyu Wang , Peng Sun , Xue Luo This is my paper

Pith reviewed 2026-05-13 19:03 UTC · model grok-4.3

classification 🧮 math.NA cs.NAmath.OC

keywords feedback particle filterHermite-Galerkin methodgain approximationerror estimateskernel density estimationspectral methodnonlinear filtering

0 comments p. Extension

The pith

A two-step Hermite-Galerkin method provides error bounds for gain approximation in feedback particle filters.

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

Feedback particle filters need to approximate a gain function solving a boundary value problem, which is difficult in practice. The paper introduces a two-step method: kernel density estimation for the density followed by Galerkin spectral approximation of an auxiliary variable using generalized Hermite functions. This auxiliary variable's rapid decay at infinity matches the basis, removing the need for artificial boundaries or truncation. The work establishes error rates of O(N_p^{-s/(2s+1)}) for the kernel step and O(M^{-s+1} log M) for the spectral step, with numerical tests confirming better performance than existing methods.

Core claim

The proposed two-step Hermite-Galerkin spectral method approximates the gain function by first estimating the unknown density in the BVP by a kernel density estimator and then approximating an auxiliary variable via the Galerkin spectral method using generalized Hermite functions. This auxiliary variable inherits the rapid decay property of the density at infinity, which aligns with the exponential decay characteristic of generalized Hermite functions, thereby obviating the need for artificial boundary conditions or domain truncation. Rigorous error estimates are established: the kernel approximation error decays at O(N_p^{-s/(2s+1)}), while the spectral approximation error converges at O(M^

What carries the argument

Galerkin spectral approximation of an auxiliary variable using generalized Hermite functions after kernel density estimation of the density in the gain function BVP

If this is right

The method furnishes complete theoretical guarantees for accuracy through the proven error decay rates.
It outperforms existing gain approximation schemes in accuracy and computational efficiency according to numerical experiments.
Artificial boundary conditions and domain truncation are not needed.
The theoretical results are validated by comprehensive numerical experiments.

Where Pith is reading between the lines

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

The error rates suggest potential for efficient implementation in high-dimensional nonlinear filtering tasks.
Similar spectral techniques could be adapted for other boundary value problems in filtering or PDEs with decaying solutions.
Further analysis might explore the method's performance when the density decay assumption is only approximately satisfied.

Load-bearing premise

The auxiliary variable inherits the rapid decay property of the density at infinity.

What would settle it

An experiment where the observed errors do not decay according to O(N_p^{-s/(2s+1)}) or O(M^{-s+1} log M) as N_p and M increase would falsify the error estimates.

Figures

Figures reproduced from arXiv: 2604.02735 by Peng Sun, Ruoyu Wang, Xue Luo.

**Figure 2.** Figure 2: The log-log plot of the error estimates of the gain function [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 4.** Figure 4: The RMSEs of each MC run of the three FPFs. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 3.** Figure 3: The estimations of the true state (black) obtained by the FPF [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

read the original abstract

The feedback particle filter (FPF) is a promising nonlinear filtering (NLF) method, but its practical implementation is hindered by the intractability of the gain function, which satisfies a boundary value problem (BVP). This paper proposes a novel two-step Hermite-Galerkin spectral method to address this challenge. First, the unknown density in the BVP is approximated by a kernel density estimator, whose error bounds are well-established in the literature. Second, rather than directly approximating the gain function, we approximate an auxiliary variable via the Galerkin spectral method using generalized Hermite functions. This auxiliary variable inherits the rapid decay property of the density at infinity, which aligns perfectly with the exponential decay characteristic of generalized Hermite functions, thereby obviating the need for artificial boundary conditions or domain truncation. Furthermore, we rigorously establish two fundamental error estimates: the kernel approximation error decays at the rate $O(N_p^{-\frac{s}{2s+1}})$, while the spectral approximation error converges at $O(M^{-s+1}\log M)$, providing complete theoretical guarantees for the method's accuracy. Comprehensive numerical experiments validate the theoretical results and demonstrate that the proposed method outperforms existing gain approximation schemes in both accuracy and computational efficiency.

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.

Desk Editor's Note private letter to a colleague

The paper gives a two-step Hermite-Galerkin scheme for gain approximation in feedback particle filters with explicit kernel and spectral error rates, but the claim that the auxiliary variable keeps rapid decay after KDE needs explicit verification to support the bounds.

read the letter

The main takeaway is that this work splits gain approximation into kernel density estimation on the density inside the BVP, then Galerkin projection of an auxiliary variable onto generalized Hermite functions. It derives separate rates—O(N_p^{-s/(2s+1)}) for the kernel step and O(M^{-s+1} log M) for the spectral step—and reports numerical tests that show the method beats existing schemes on accuracy and cost. The choice of generalized Hermite basis is the practical angle, since it matches decay at infinity and skips artificial boundaries or truncation. That combination is not a standard extension of prior spectral work on this problem, and the error split follows from established kernel and approximation theory without obvious circularity or fitted parameters. The numerics appear to back the rates under the stated smoothness assumptions. The soft spot is the inheritance step: the analysis assumes the auxiliary variable after KDE still decays fast enough for the projection bounds to hold without extra remainder terms from slower tails or oscillations. The abstract states this carries over directly, but if the paper only asserts it without a separate estimate on how the kernel estimate affects decay, the spectral rate could need adjustment in practice. No other load-bearing gaps stand out from the description. This is aimed at people working on numerical methods for nonlinear filtering and stochastic control. Readers who already use spectral or kernel tools for BVPs with decaying solutions will get the most from the rates and experiments. It has enough formal grounding and computational checks to deserve peer review rather than a desk reject, though referees will likely press on the decay verification.

Referee Report

2 major / 2 minor

Summary. The paper proposes a two-step Hermite-Galerkin spectral method for approximating the gain function in the feedback particle filter. The unknown density in the governing BVP is first replaced by a kernel density estimator; an auxiliary variable is then approximated via Galerkin projection onto generalized Hermite functions. The manuscript claims to derive rigorous error bounds separating the kernel contribution O(N_p^{-s/(2s+1)}) from the spectral contribution O(M^{-s+1} log M), together with numerical experiments that validate the rates and show improved accuracy and efficiency over existing schemes.

Significance. If the decay-inheritance assumption holds, the work supplies complete, parameter-free theoretical guarantees for a practical gain approximation that exploits the natural decay of the auxiliary variable to avoid artificial boundary conditions. The explicit separation of kernel and spectral errors and the alignment of the basis with the decay property constitute a clear methodological advance for nonlinear filtering implementations.

major comments (2)

[Abstract and spectral-error derivation] The O(M^{-s+1} log M) spectral error bound (stated in the abstract and derived after the definition of the auxiliary variable) presupposes that the KDE-replaced right-hand side of the BVP still decays rapidly enough at infinity for the generalized Hermite projection estimates to hold without additional truncation or remainder terms. No explicit verification, auxiliary decay bound, or perturbation analysis is supplied showing that the KDE error term does not produce slower tails or local oscillations that would invalidate the stated rate.
[Error-estimate section] The total error is presented as the sum of the kernel term O(N_p^{-s/(2s+1)}) and the spectral term O(M^{-s+1} log M). The manuscript does not derive a combined bound that accounts for the interaction between the two approximations (e.g., how the KDE error propagates into the Galerkin matrix and the auxiliary-variable solution), leaving the claimed overall rate for the gain function formally incomplete.

minor comments (2)

The smoothness index s and the precise definition of the auxiliary variable should be restated with equation numbers at the beginning of the error-analysis section for immediate reference.
Numerical tables or figures comparing observed rates against the predicted exponents for several values of s would strengthen the validation section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. The points raised highlight opportunities to strengthen the rigor of the error analysis, and we will revise the paper accordingly to address them.

read point-by-point responses

Referee: [Abstract and spectral-error derivation] The O(M^{-s+1} log M) spectral error bound (stated in the abstract and derived after the definition of the auxiliary variable) presupposes that the KDE-replaced right-hand side of the BVP still decays rapidly enough at infinity for the generalized Hermite projection estimates to hold without additional truncation or remainder terms. No explicit verification, auxiliary decay bound, or perturbation analysis is supplied showing that the KDE error term does not produce slower tails or local oscillations that would invalidate the stated rate.

Authors: We appreciate the referee's observation. The original analysis assumes that the kernel density estimator, under the stated smoothness s > 1/2 and standard kernels, produces a perturbation whose weighted Sobolev norm remains controlled, thereby preserving the applicability of the generalized Hermite projection estimates. To make this fully rigorous, we will add a supporting lemma that quantifies the tail behavior of the KDE and shows that the right-hand side perturbation stays within the function space required for the O(M^{-s+1} log M) bound. This addition will eliminate any ambiguity regarding the validity of the spectral rate. revision: yes
Referee: [Error-estimate section] The total error is presented as the sum of the kernel term O(N_p^{-s/(2s+1)}) and the spectral term O(M^{-s+1} log M). The manuscript does not derive a combined bound that accounts for the interaction between the two approximations (e.g., how the KDE error propagates into the Galerkin matrix and the auxiliary-variable solution), leaving the claimed overall rate for the gain function formally incomplete.

Authors: We agree that an explicit combined error bound is needed for completeness. Because the BVP solution operator is continuous in the relevant norm and the Galerkin projection is stable, the propagation of the KDE error through the discrete system can be bounded by a constant multiple of the kernel error. In the revision we will insert a theorem that applies the triangle inequality to the auxiliary-variable error and then transfers the bound to the gain function via the continuous dependence of the gain on the auxiliary variable, thereby establishing the overall rate O(N_p^{-s/(2s+1)} + M^{-s+1} log M) with all interaction terms accounted for. revision: yes

Circularity Check

0 steps flagged

Error bounds drawn from standard KDE and spectral theory; auxiliary decay treated as structural consequence, not fitted input

full rationale

The kernel error rate O(N_p^{-s/(2s+1)}) is explicitly attributed to 'well-established' literature bounds rather than derived or fitted inside the paper. The spectral rate O(M^{-s+1} log M) follows from conventional Galerkin projection error estimates once the auxiliary variable is assumed to inherit rapid decay; this inheritance is asserted from the BVP definition after KDE replacement, without reducing the claimed rates to any parameter fitted on the target data. No self-citation chain, ansatz smuggling, or renaming of known results is required for the central claims. The derivation therefore remains externally anchored and does not collapse to its own inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on (1) standard error bounds for kernel density estimators that are taken from the literature and (2) the rapid-decay property of the auxiliary variable that is asserted to match the decay of generalized Hermite functions. No free parameters are fitted inside the derivation itself; s is a smoothness index chosen by the user.

free parameters (1)

s (smoothness index)
Controls the rate in both error bounds; chosen according to assumed regularity of the density.

axioms (2)

standard math Kernel density estimator error bounds hold at the stated rate under standard assumptions on the kernel and bandwidth.
Invoked for the first step; cited as well-established in the literature.
domain assumption The auxiliary variable inherits the rapid decay at infinity of the underlying density.
Central to the claim that generalized Hermite functions can be used without truncation or artificial boundaries.

pith-pipeline@v0.9.0 · 5524 in / 1518 out tokens · 45563 ms · 2026-05-13T19:03:48.973865+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

auxiliary variable inherits the rapid decay property of the density at infinity... generalized Hermite functions... error estimates: kernel O(N_p^{-s/(2s+1)}), spectral O(M^{-s+1} log M)
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

generalized Hermite functions... Sobolev space H^r(R)... projection error Lemma II.1

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

Feedback particle filter,

T. Yang, P. Mehta, and S. Meyn, “Feedback particle filter,”IEEE Transactions on Automatic Control, vol. 58, no. 10, pp. 2465–2480, 2013

work page 2013
[2]

How to avoid the curse of dimensionality: scalability of particle filters with and without importance weights,

S. Surace, A. Kutschireiter, and J.-P. Pfister, “How to avoid the curse of dimensionality: scalability of particle filters with and without importance weights,”SIAM Review, vol. 61, no. 1, pp. 79–91, 2019

work page 2019
[3]

Gain function approximation in the feedback particle filter,

A. Taghvaei and P. Mehta, “Gain function approximation in the feedback particle filter,” inProceedings of 2016 IEEE Conference on Decision and Control, pp. 5446–5452, 2016

work page 2016
[4]

Multivariable feedback particle filter,

T. Yang, R. Laugesen, P. Mehta, and S. Meyn, “Multivariable feedback particle filter,”Automatica, vol. 71, pp. 10–23, 2016

work page 2016
[5]

Multivariable feedback particle filter,

T. Yang, R. Laugesen, P. Mehta, and S. Meyn, “Multivariable feedback particle filter,” inProceedings of 2012 IEEE Conference on Decision and Control, pp. 4063–4070, 2012

work page 2012
[6]

Data-driven gain computation in the feedback particle filter,

K. Berntorp and P. Grover, “Data-driven gain computation in the feedback particle filter,” inProceedings of the American Control Conference, pp. 2711–2716, 2016

work page 2016
[7]

Diffusion map-based algo- rithm for gain function approximation in the feedback particle filter,

A. Taghvaei, P. Mehta, and S. Meyn, “Diffusion map-based algo- rithm for gain function approximation in the feedback particle filter,” SIAM/ASA Journal on Uncertainty Quantification, vol. 8, no. 3, pp. 1090–1117, 2020

work page 2020
[8]

Comparison of gain function approximation methods in the feedback particle filter,

K. Berntorp, “Comparison of gain function approximation methods in the feedback particle filter,” in2018 21th International Conference on Information Fusion, pp. 123–130, 2018

work page 2018
[9]

A decomposition approach for the gain function in the feedback particle filter,

R. Wang, H. Miao, and X. Luo, “A decomposition approach for the gain function in the feedback particle filter,” inProceedings of 2025 IEEE Conference on Decision and Control, pp. 2378–2384, 2025

work page 2025
[10]

A decomposition method in the multivari- ate feedback particle filter via tensor product hermite polynomials,

R. Wang and X. Luo, “A decomposition method in the multivari- ate feedback particle filter via tensor product hermite polynomials,” arXiv:2511.01227v1, 2025

work page arXiv 2025
[11]

Hermite spectral method to 1-d forward kol- mogorov equation and its application to nonlinear filtering problems,

X. Luo and S. S.-T. Yau, “Hermite spectral method to 1-d forward kol- mogorov equation and its application to nonlinear filtering problems,” IEEE Transactions on Automatic Control, vol. 58, no. 10, pp. 2495 – 2507, 2013

work page 2013
[12]

Solving nonlinear filtering problems in real time by Legendre Galerkin spectral method,

W. Dong, X. Luo, and S. S.-T. Yau, “Solving nonlinear filtering problems in real time by Legendre Galerkin spectral method,”IEEE Transactions on Automatic Control, vol. 66, no. 4, pp. 1559 – 1572, 2021

work page 2021
[13]

Splitting-up spectral method for nonlinear filtering problems with correlation noises,

F. Zhang, Y . Zou, S. Chai, R. Zhang, and Y . Cao, “Splitting-up spectral method for nonlinear filtering problems with correlation noises,” Journal of Scientific Computing, vol. 93, no. 1, 2022

work page 2022
[14]

Solving nonlinear filtering problems with correlated noise based on Hermite-Galerkin spectral method,

Z. Sun and S. S.-T. Yau, “Solving nonlinear filtering problems with correlated noise based on Hermite-Galerkin spectral method,”Auto- matica, vol. 156, 2023

work page 2023
[15]

DGLG: A novel deep general- ized Legendre-Galerkin approach to optimal filtering problem,

J. Shi, X. Jiao, and S. S.-T. Yau, “DGLG: A novel deep general- ized Legendre-Galerkin approach to optimal filtering problem,”IEEE Transactions on Automatic Control, vol. 70, no. 4, pp. 2584 – 2590, 2025

work page 2025
[16]

J. Shen, T. Tang, and L. Wang,Spectral Methods: Algorithms, Analysis and Applications. Springer Berlin, Heidelberg, 2011

work page 2011
[17]

Hermite spectral method with hyperbolic cross approximations to high-dimensional parabolic PDEs,

X. Luo and S. S.-T. Yau, “Hermite spectral method with hyperbolic cross approximations to high-dimensional parabolic PDEs,”SIAM Journal on Numerical Analysis, vol. 51, no. 6, pp. 3186 – 3212, 2013

work page 2013
[18]

A. B. Tsybakov,Introduction to Nonparametric Estimation. Springer New York, 2009

work page 2009
[19]

Devroye and L

L. Devroye and L. Gy ¨orfi,Nonparametric Density Estimation: TheL 1 View.John Wiley, New York., 1985. VI. APPENDIX Sketch of the proof of Lemma II.3.We start from ∥pNp,ϵNp −p∥ 2 L1 ≤2∥EpNp,ϵNp −p∥ 2 L1 + 2∥pNp,ϵNp −Ep Np,ϵNp ∥2 L1 .(46) Bias term.For a kernel of orders, the standard bias expansion (using vanishing moments and Taylor expansion) gives ∥EpNp...

work page 1985