Identification of a Kalman filter: consistency of local solutions

L\'eo Simpson; Moritz Diehl

arxiv: 2601.04198 · v3 · submitted 2025-12-04 · 📡 eess.SY · cs.SY· math.DS

Identification of a Kalman filter: consistency of local solutions

L\'eo Simpson , Moritz Diehl This is my paper

Pith reviewed 2026-05-17 01:28 UTC · model grok-4.3

classification 📡 eess.SY cs.SYmath.DS

keywords Kalman filtersystem identificationconsistencylocal minimizersprediction error methodasymptotic unimodality

0 comments

The pith

Local minimizers for Kalman gain estimation converge to the true gain as data grows.

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

The paper shows that estimating only the Kalman gain via prediction error methods yields local solutions that become statistically consistent with the true gain. This holds because the non-convex objective converges to a limiting function possessing a single minimizer. Readers care since it indicates that non-convexity need not block reliable Kalman filter tuning in this restricted setting. The result rests on the system being linear time-invariant and the data satisfying standard ergodicity and excitation conditions.

Core claim

We prove that these local solutions are statistically consistent estimates of the true Kalman gain. This follows from asymptotic unimodality: as the dataset grows, the objective function converges to a limit with a unique local (and therefore global) minimizer.

What carries the argument

Asymptotic unimodality of the prediction error objective, which guarantees a unique minimizer in the large-data limit.

Load-bearing premise

The data must be generated exactly by a linear time-invariant system with enough excitation for the objective to converge to a unimodal limit.

What would settle it

A large dataset in which repeated local optimizations from varied starting points produce Kalman gains that remain far from the true value would disprove consistency.

read the original abstract

Prediction error and maximum likelihood methods are powerful tools for identifying linear dynamical systems and, in particular, enable the joint estimation of model parameters and the Kalman filter used for state estimation. A key limitation, however, is that these methods require solving a generally non-convex optimization problem to global optimality. This paper analyzes the statistical behavior of local minimizers in the special case where only the Kalman gain is estimated. We prove that these local solutions are statistically consistent estimates of the true Kalman gain. This follows from asymptotic unimodality: as the dataset grows, the objective function converges to a limit with a unique local (and therefore global) minimizer. We further provide guidelines for designing the optimization problem for Kalman filter tuning and discuss extensions to the joint estimation of additional linear parameters and noise covariances. Finally, the theoretical results are illustrated using three examples of increasing complexity. The main practical takeaway of this paper is that difficulties caused by local minimizers in system identification are, at least, not attributable to the tuning of the Kalman gain.

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

Local Kalman gain solutions are consistent via asymptotic unimodality, though convergence details warrant checking.

read the letter

The one thing to know is that the paper proves local minimizers are statistically consistent for estimating the Kalman gain by itself, based on the prediction error objective becoming asymptotically unimodal. They do a solid job extending the theory for this specific case. By focusing only on the gain, they avoid the full complexity of joint estimation and show that the limit function has a unique minimizer at the true value. The examples help ground the theory, and the guidelines for optimization design are a nice practical addition. It's good to see work that directly addresses why local solutions might still be reliable in this subproblem. Where it could be softer is in the technical conditions for the consistency to transfer from the limit to the finite-sample local solutions. The stress-test raises a valid point here. Even if the limiting objective is unimodal, you need to make sure that the sample objectives converge in a way that their local minimizers don't go to infinity or to other points. Since the space of possible Kalman gains isn't compact, without some coercivity or uniform convergence result, it's not automatic. The paper relies on standard assumptions like ergodicity and excitation, but it would be good to see if they prove the stronger properties needed or if there's an implicit argument I'm missing from the abstract. Overall, this paper is for researchers and practitioners in linear system identification and Kalman filtering who are concerned about optimization difficulties in prediction-error methods. Someone working on theoretical foundations or practical tuning algorithms would get something out of the consistency result and the examples. It has enough substance to merit a serious referee report, particularly to verify the convergence arguments in detail. I would recommend putting it through peer review rather than desk rejecting it.

Referee Report

2 major / 2 minor

Summary. The paper claims that when estimating only the Kalman gain in linear time-invariant systems via prediction-error or maximum-likelihood criteria, local minimizers of the resulting non-convex objective are statistically consistent for the true gain. The argument proceeds by showing that, under standard ergodicity and persistent-excitation conditions, the finite-sample objective converges to a limiting function that is asymptotically unimodal with a unique minimizer at the true Kalman gain; local solutions of the sample problem therefore converge to this global minimizer. The manuscript supplies design guidelines for the optimization, sketches extensions to joint parameter and covariance estimation, and illustrates the results on three numerical examples of increasing complexity.

Significance. If the central consistency result holds, the work is significant for system-identification practice: it removes a principal practical objection to local optimization when tuning the Kalman gain alone, thereby simplifying filter design and providing a theoretical basis for the common heuristic of accepting local solutions. The manuscript earns credit for grounding the claim in an explicit asymptotic-unimodality argument rather than simulation alone and for supplying concrete optimization-design recommendations that follow directly from the analysis.

major comments (2)

[§3, Theorem 1] §3, Theorem 1 (or the main consistency statement): the passage from pointwise convergence of the objective to consistency of its local minimizers requires either uniform convergence on a compact set or an explicit coercivity/boundedness argument for the sequence of sample minimizers. The current sketch invokes only ergodicity and excitation; these guarantee pointwise convergence but do not automatically deliver the stronger mode of convergence needed on the non-compact matrix manifold of Kalman gains. An explicit lemma establishing that sample minimizers remain in a compact set (or that the convergence is uniform on large balls) is therefore load-bearing for the claim.
[§4] §4, the extension discussion: the argument for the pure-gain case relies on the objective becoming unimodal in the gain alone. When additional linear parameters or noise covariances are estimated jointly, the limiting objective is no longer guaranteed to be unimodal in the enlarged parameter vector; the manuscript should state whether the same local-to-global transfer still holds or whether additional structural assumptions are required.

minor comments (2)

Notation: the distinction between the finite-sample objective J_N(K) and its limit J(K) should be introduced once and used consistently; several passages mix the two without explicit qualification.
Figure 2 and 3: axis labels and legend entries are too small for print; increasing font size would improve readability without altering content.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and valuable suggestions. We address the major comments below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

Referee: [§3, Theorem 1] the passage from pointwise convergence of the objective to consistency of its local minimizers requires either uniform convergence on a compact set or an explicit coercivity/boundedness argument for the sequence of sample minimizers. The current sketch invokes only ergodicity and excitation; these guarantee pointwise convergence but do not automatically deliver the stronger mode of convergence needed on the non-compact matrix manifold of Kalman gains. An explicit lemma establishing that sample minimizers remain in a compact set (or that the convergence is uniform on large balls) is therefore load-bearing for the claim.

Authors: We agree that an explicit argument for the boundedness of the sample minimizers or uniform convergence is necessary to rigorously establish consistency from pointwise convergence. While the manuscript sketches the consistency using ergodicity and persistent excitation leading to asymptotic unimodality, we acknowledge that this step requires strengthening. In the revised manuscript, we will add a new lemma that shows the sample objective function is coercive outside a compact set containing the true gain, ensuring that local minimizers remain bounded with high probability as the sample size grows. This will close the gap in the proof of Theorem 1. revision: yes
Referee: [§4] the argument for the pure-gain case relies on the objective becoming unimodal in the gain alone. When additional linear parameters or noise covariances are estimated jointly, the limiting objective is no longer guaranteed to be unimodal in the enlarged parameter vector; the manuscript should state whether the same local-to-global transfer still holds or whether additional structural assumptions are required.

Authors: We concur that the unimodality property is specific to the gain-only estimation case. The manuscript's discussion of extensions to joint estimation is preliminary and does not assert that local minimizers remain consistent without further conditions. In the revision, we will explicitly state that for joint estimation, additional assumptions such as separate identifiability of the linear parameters or convexity in the covariance parameters may be needed to preserve the local-to-global consistency. We will also clarify the limitations of the current analysis in this regard. revision: yes

Circularity Check

0 steps flagged

No significant circularity; consistency follows from standard asymptotic analysis

full rationale

The paper's derivation establishes consistency of local minimizers for the Kalman gain by proving that the sample objective converges to a limiting function that is asymptotically unimodal with unique minimizer at the true gain. This relies on ergodicity, excitation, and data generated by the true LTI system, which are external assumptions not defined in terms of the target result. No equation reduces the claimed consistency to a fitted parameter or self-referential definition, and no load-bearing step invokes a self-citation whose validity depends on the present work. The argument applies general results from asymptotic statistics on empirical risk minimization and is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on standard domain assumptions of linear system identification rather than new free parameters or invented entities.

axioms (2)

domain assumption The plant is linear time-invariant and stable.
Required for the Kalman filter to exist and for the prediction-error criterion to be well-defined.
domain assumption Observed data are generated by the true system under persistent excitation.
Needed for the sample objective to converge to a population limit with the claimed unimodal shape.

pith-pipeline@v0.9.0 · 5477 in / 1171 out tokens · 92054 ms · 2026-05-17T01:28:48.076210+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.

as the dataset grows, the objective function converges to a limit with a unique local (and therefore global) minimizer

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.