REVIEW 3 major objections 6 minor 7 references
Interpretable Gradient Descent for Kalman Gain
T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read This paper proves that gradient descent on the innovation loss converges to the optimal Kalman filter gain at an interpretable geometric rate, using a new factorization of the gradient into an observability term and an…
desk verdict A genuinely new gradient decomposition and a surprising counterexample, with a seriously flawed lemma in the rate proof that is likely repairable. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the gradient factorization $\nabla_L J_{\mathrm{innov}}(L) = -2W_o(L)K(L)$: $W_o(L)$ is the observability Gramian of the closed-loop pair $(F(L),CA)$, solving the Lyapunov equation $W_o=F(L)^TW_oF(L)+A^TC^TCA$, and $K(L)$ is the steady-state cross-covariance of the estimation error and the innovation, $K(L)=F(L)P(L)A^TC^T+(I-LC)Q_wC^T-LR_v$. The factorization makes stationarity of the loss equivalent to $W_o(L)K(L)=0$; under the observability assumption $W_o(L)$ is positive definite, so stationary points are exactly zeros of $K(L)$. Coercivity of $J_{\mathrm{innov}}$ on $S$ keeps trajectories in compact level sets, and LaSalle's invariance principle then gives convergence to $L_{KF}$, with the constants $\kappa=\inf_{L\in\mathcal{L}_0}\lambda_{\min}(W_o(L))$ and $c=\sup_{L\ne L_{KF}}(J_{\mathrm{innov}}(L)-J_{\mathrm{innov}}(L_{KF}))/\|K(L)\|_F^2$ determining the geometric rate.
What would settle it
Run a numerical search over stabilizing gains for a system with $(A,CA)$ observable, solving $P(L)=F(L)P(L)F(L)^T+Q_\eta(L)$ and forming $K(L)=F(L)P(L)A^TC^T+(I-LC)Q_wC^T-LR_v$; finding any $L\neq L_{KF}$ in $S$ with $K(L)=0$ would invalidate the LaSalle step of the proof. A cheaper check is to test whether the differential $DK(L_{KF})$ has full rank, since the bound on the constant $c$ uses a quadratic-form comparison that requires the denominator $\|K(L)\|_F^2$ to control the loss's Hessian near the optimum.
Extended reading notes
Core claim
The central claim is that, in the standard (non-look-ahead) Kalman filtering formulation, gradient flow on the innovation loss has no spurious stationary points when $(A,CA)$ is observable, and every trajectory initialized in the stabilizing set $S$ converges to the Kalman gain $L_{KF}$ at a geometric rate. The engine is the identity $\nabla_L J_{\mathrm{innov}}(L) = -2W_o(L)K(L)$, which factors the gradient into an observability Gramian $W_o(L)$ and the steady-state cross-covariance $K(L)$ between the estimation error and the innovation; the orthogonality principle says $K(L_{KF})=0$. The proof uses LaSalle's invariance principle on compact level sets, obtained from a coercivity lemma, and the rate is $J_{\mathrm{innov}}(L(t))-J_{\mathrm{innov}}(L_{KF})\le e^{-4(\kappa^2/c)t}(J_{\mathrm{innov}}(L(0))-J_{\mathrm{innov}}(L_{KF}))$. Example 4.4 shows the stronger observability assumption is necessary: with only $(A,C)$ observable, there is a continuum of critical points.
Load-bearing premise
The argument needs the steady-state cross-covariance between the estimation error and the innovation to vanish at only one stabilizing gain, the Kalman gain; if another gain also made it vanish, the invariance argument would only guarantee convergence to that larger set.
Editorial extensions
If this is right
- From any stabilizing initial gain in $S$, gradient descent on the innovation loss converges to the Kalman gain whenever $(A,CA)$ is observable and $(A,Q_w^{1/2})$ is stabilizable.
- Observability of $(A,C)$ alone is not enough: Example 4.4 exhibits a continuum of spurious stationary points, so the stronger condition is a genuine requirement rather than a proof artifact.
- The convergence rate $4\kappa^2/c$ splits into a worst-case observability factor and a steepness factor, so the geometry of the filtering problem directly controls how fast gradient gain tuning can work.
- The coercivity lemma ensures the innovation loss grows without bound as gains leave the stabilizing region, so the sublevel sets are compact and the trajectory cannot escape to infinity or the boundary.
Reading between the lines
- A direct consequence of the gradient identity is a design principle for adaptive filters: any update that pushes $K(L)$ toward zero while keeping the Gramian $W_o(L)$ positive definite along the trajectory should enjoy a similar geometric guarantee; the paper itself does not analyze the stochastic case, so this is an extension.
- The proof's only unstated premise is that $\{L:K(L)=0\}$ is a singleton at $L_{KF}$; closing that gap with an explicit argument from the algebraic Riccati equation would make the theorem fully self-contained.
- Because $\kappa$ and $c$ are defined by optimizations over a level set, the rate is an a posteriori object; an online scheme that estimates these constants from data as the gain evolves would turn the interpretable rate into a practical stopping rule.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper studies gradient descent on the steady-state innovation loss J_innov(L)=Tr(Σ_δ(L)) for the gain L of a linear time-invariant filter. The main results are: (i) a formula ∇_L J_innov(L) = -2 W_o(L) K(L), where W_o is the observability Gramian of ((I-LC)A, CA) and K is the steady-state error-innovation cross-covariance; (ii) a convergence theorem (Theorem 4.3) stating that under observability of (A, CA) and stabilizability of (A, Q_w^{1/2}), the gradient flow from any stabilizing L converges to the Kalman gain with rate e^{-4(κ²/c)t}; (iii) a counterexample showing that observability of (A, C) alone permits spurious stationary points. The proof relies on a coercivity lemma and a technical lemma about quadratic forms. The paper also discusses the interpretability of the rate constants and an open question about level-set-dependent constants.
Significance. If correct, the gradient decomposition is a clean and interpretable identity that explains the presence or absence of spurious stationary points in terms of observability of (A, CA) and violation of the orthogonality principle. The convergence theorem appears to be the first for the standard Kalman filter formulation with an explicit rate, complementing the look-ahead result of Talebi et al. The coercivity analysis is nontrivial, and Example 4.4 is a useful warning about the inadequacy of observability of (A, C) alone. The paper is largely self-contained and does not fit parameters. However, the load-bearing technical gaps in the proof of Theorem 4.3 (uniqueness of the zero set of K and finiteness of c) need to be repaired before the claims are fully established.
major comments (3)
- [4, proof of Theorem 4.3] The statement that the set {L ∈ L0 : K(L) = 0} "is a singleton with L_KF as its only member" is asserted without proof or citation. LaSalle's principle only yields convergence to the set {L ∈ L0 : W_o(L)K(L) = 0}, and since W_o(L) ≻ 0 under Assumption 4.1, the argument needs a proof that the stabilizing gain satisfying the orthogonality condition K(L) = 0 is unique. This follows from standard algebraic Riccati theory, but it is load-bearing and should be stated explicitly.
- [4, Lemma 4.6] Lemma 4.6 is false as stated. For N = diag(1,0) and M = [[0,1],[1,0]], the hypothesis x^T N x = 0 implies x_1 = 0 and hence x^T M x = 0, but for x = (1,t) one has x^T M x = 2t and x^T N x = 1, so no finite β exists. The proof uses the step x^T M x = z^T M z for the orthogonal decomposition x = y + z with y ∈ ker(N); this requires M y = 0, which is not implied by vanishing of the quadratic form on ker(N). In the application to Theorem 4.3 the stronger condition does hold because ∇²J_innov(L_KF)[ΔL] = -2 W_o(L_KF) DK(ΔL), so the gap is repairable, but the lemma and its proof must be corrected.
- [4, proof of c < ∞] The proof that c < ∞ only bounds the ratio of the leading quadratic forms (1/2)⟨ΔL, ∇²J_innov(L_KF)[ΔL]⟩ and ∥DK(ΔL)∥². Even after correcting Lemma 4.6, this does not control directions with DK(ΔL) = 0: in such directions the denominator ∥K(L)∥² is of higher order than ∥ΔL∥², and the first-order expansions around Eq. (18) are insufficient to establish finiteness of the ratio (J_innov(L) - J_innov(L_KF))/∥K(L)∥². The authors need to prove that DK(L_KF) is injective (equivalently, that the Hessian of J_innov at L_KF is positive definite) or provide a higher-order argument. This is load-bearing for the claimed exponential rate.
minor comments (6)
- [Section 1] The word "lookeahead" is a typo; it should be "look-ahead".
- [Section 2, definitions of κ] In the definition of κ, the set S0 appears to be a typo for L0, which is defined immediately above; as written S0 is undefined.
- [Section 5, proof of Lemma 5.1] The displayed expression for L(j)_p uses the index k in the summation, but the sum should run from s = 0 to p-1 with exponent p-1-s; as written it is confusing and appears to be a typo.
- [Section 5] The word "definitie" is a typo; it should be "definite".
- [Section 5, proof of part 2] The implication "observability of (A, CA) implies invertibility of A" is used without proof. It is true, since if Ax = 0 then every block of the observability matrix of (A, CA) annihilates x, but it should be stated explicitly.
- [Section 6] The limitation that κ and c are defined by optimization over a level set is honestly acknowledged; this should perhaps be emphasized in the introduction as a caveat on the claim that the rate is "interpretable".
Circularity Check
No significant circularity: the gradient formula and convergence proof are derived from explicit first-principles equations, and the cited prior work is not load-bearing.
full rationale
The paper's main derivation is self-contained. Theorem 3.1 is proved directly from the Lyapunov equation for P(L), the definition of the innovation covariance, and the cross-covariance K(L); no fitted parameter is introduced and no result is imported from the authors' own prior work. The convergence proof of Theorem 4.3 is a standard LaSalle argument using V = Jinnov, with the observability assumption used to make W_o(L) positive definite; the assertion that {L : K(L) = 0} is the singleton {L_KF} is implicit and not proved, but this is a rigor gap rather than a circular step. The rate estimate follows from the constants kappa and c, which are defined directly as an infimum and supremum over the level set; the proof that c is finite invokes Lemma 4.6, which is false as stated, although in the application the additional identity grad^2 J(L_KF)[Delta] = -2 W_o(L_KF) DK(Delta) supplies the needed kernel invariance and makes the argument repairable. These are correctness concerns, not cases where a prediction reduces by construction to an input. The references to prior work [3,5] are by other authors and are contextual rather than load-bearing, and the paper explicitly discloses the limitation that kappa and c are defined through optimization over a level set. No self-definitional, fitted-input, or renamed-known-result pattern is present.
Assumptions & free parameters
assumptions (4)
- domain assumption The process noise covariance Q_w and measurement noise covariance R_v are positive definite.
- domain assumption The pair (A, Q_w^{1/2}) is stabilizable.
- domain assumption The pair (A, CA) is observable (Assumption 4.1).
- standard math There exists a unique stabilizing gain L with K(L)=0, namely L_KF.
Cite this review
Pith. "Pith review of Interpretable Gradient Descent for Kalman Gain." pith.science (2026). https://pith.science/paper/HGD53ZXO
@misc{pith2026250714354,
author = {Pith},
title = {Pith review of: Interpretable Gradient Descent for Kalman Gain},
year = {2026},
howpublished = {\url{https://pith.science/paper/HGD53ZXO}},
note = {Machine review of arXiv:2507.14354}
}
read the original abstract
We derive a decomposition for the gradient of the innovation loss with respect to the filter gain in a linear time-invariant system, decomposing as a product of an observability Gramian and a term quantifying the ``non-orthogonality" between the estimation error and the innovation. We leverage this decomposition to give a convergence proof of gradient descent to the optimal Kalman gain, specifically identifying how recovery of the Kalman gain depends on a non-standard observability condition, and obtaining an interpretable geometric convergence rate.
Reference graph
Works this paper leans on
-
[1]
Optimizing static linear feedback: Gradient method
Ilyas Fatkhullin and Boris Polyak. Optimizing static linear feedback: Gradient method. SIAM Journal on Control and Optimization, 59(5):3887–3911, 2021
work page 2021
-
[2]
Global convergence of policy gradient methods for the linear quadratic regulator
Maryam Fazel, Rong Ge, Sham Kakade, and Mehran Mesbahi. Global convergence of policy gradient methods for the linear quadratic regulator. In International conference on machine learning, pages 1467–1476. PMLR, 2018
work page 2018
-
[3]
Policy Optimization of Finite-Horizon Kalman Filter with Unknown Noise Covariance
Haoran Li and Yuan-Hua Ni. Policy optimization of finite-horizon kalman filter with unknown noise covariance. arXiv preprint arXiv:2310.15467, 2023
work page Pith review arXiv 2023
-
[4]
On the linear convergence of random search for discrete-time lqr
Hesameddin Mohammadi, Mahdi Soltanolkotabi, and Mihailo R Jovanovi´ c. On the linear convergence of random search for discrete-time lqr. IEEE Control Systems Letters, 5(3):989–994, 2020
work page 2020
-
[5]
Data-driven optimal filter- ing for linear systems with unknown noise covariances
Shahriar Talebi, Amirhossein Taghvaei, and Mehran Mesbahi. Data-driven optimal filter- ing for linear systems with unknown noise covariances. Advances in Neural Information Processing Systems, 36:69546–69585, 2023
work page 2023
-
[6]
Pol- icy optimization in control: Geometry and algorithmic implications
Shahriar Talebi, Yang Zheng, Spencer Kraisler, Na Li, and Mehran Mesbahi. Pol- icy optimization in control: Geometry and algorithmic implications. arXiv preprint arXiv:2406.04243, 2024
arXiv 2024
-
[7]
Introduction to Applied Nonlinear Dynamical Systems and Chaos
Stephen Wiggins. Introduction to Applied Nonlinear Dynamical Systems and Chaos. Springer New York, 2006. 15
work page 2006
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.