Average Kalman Rank Condition

• Average Kalman Rank Condition is a spectral property where error covariances are asymptotically bounded by the number of non-negative Lyapunov exponents.
• The methodology shows that forecast and analysis covariances lose support along stable directions, justifying reduced-order assimilation in the unstable–neutral subspace.
• This result bridges Riccati dynamics with multiplicative ergodic theory, highlighting that long-term uncertainty aligns with inherent dynamical growth rates.

Searching arXiv for the specified paper and closely related terminology to ground the article in the cited work. For a linear, discrete, time-varying, deterministic system with noisy outputs, the central rank-deficiency result associated with the Kalman filter is that the Riccati transformation asymptotically bounds the ranks of both the forecast and analysis error covariance matrices by the number of non-negative Lyapunov exponents of the dynamics (Gurumoorthy et al., 2015). In the formulation treated in "Rank deficiency of Kalman error covariance matrices in linear time-varying system with deterministic evolution" (Gurumoorthy et al., 2015), the error covariances ultimately become supported only on the unstable–neutral subspace spanned by the corresponding backward Lyapunov vectors. This suggests an "average" rank condition in the sense that the effective covariance dimension is controlled not by an instantaneous algebraic test, but by asymptotic growth rates encoded in the Lyapunov spectrum.

1. Dynamical setting and assumptions

The setting is the linear, discrete, time-varying, deterministic system

$$x_{n+1}=A_{n+1}x_n,\qquad y_n=H_nx_n+\eta_n,\qquad \eta_n\sim N(0,R_n),$$

with uniformly bounded sequences $\{A_n\}$, $\{H_n\}$, and $\{R_n\}$, and with the pair assumed uniformly completely observable (Gurumoorthy et al., 2015). The formulation is explicitly a perfect-model setting: the state evolution is deterministic, while uncertainty enters through noisy observations.

Within this setting, the objects of interest are the forecast covariance $P_k^f$ and the analysis covariance $P_k^a$. The result does not assert arbitrary low-rank behavior for all filtering problems. A plausible misconception is to read it as a universal statement about Kalman filtering with model error; the theorem is instead stated for deterministic evolution with noisy outputs, and the observability assumption is part of the hypothesis.

The significance of these assumptions is technical and structural. Uniform complete observability is used to obtain the boundedness properties required in the proof, while the deterministic state propagation allows the asymptotic geometry of the error covariances to be linked directly to Lyapunov directions of the propagator.

2. Rank bound in terms of the Lyapunov spectrum

Let $\mu_1\ge \cdots \ge \mu_d$ denote the Lyapunov exponents of the dynamics, and define

$m:=\#\{j:\mu_j\ge 0\}.$

The main theorem states that

$$\limsup_{k\to\infty}\,\mathrm{rank}\,P_k^f\le m,\qquad \limsup_{k\to\infty}\,\mathrm{rank}\,P_k^a\le m$$

(Gurumoorthy et al., 2015).

This is the core rank condition. The asymptotic rank of each covariance matrix is bounded above by the number of non-negative Lyapunov exponents, rather than by the full state dimension $d$. In particular, if only a small number of Lyapunov exponents are non-negative, then the filter’s long-time uncertainty can occupy only a correspondingly small subspace.

The statement is asymptotic and spectral. It does not say that the covariance matrices are exactly rank $\{A_n\}$0 at every finite time, nor that $\{A_n\}$1 must be attained. It gives an upper bound expressed through the Lyapunov spectrum. This suggests that the effective covariance dimension is determined by long-time average amplification and neutrality, rather than by transient stretching alone.

3. Lyapunov exponents and backward Lyapunov vectors

Fix a reference time $\{A_n\}$2 and define the state-transition operator

$\{A_n\}$3

If $\{A_n\}$4 are its singular values, then Oseledets’ theorem yields the limits

$\{A_n\}$5

which are the Lyapunov exponents (Gurumoorthy et al., 2015).

The backward Lyapunov structure is introduced through

$\{A_n\}$6

whose eigenvectors

$\{A_n\}$7

are the backward Lyapunov vectors at time $\{A_n\}$8. Their eigenvalues satisfy

$\{A_n\}$9

Accordingly, exactly the first $\{H_n\}$0 exponents are $\{H_n\}$1 at the level of $\{H_n\}$2, while the remaining $\{H_n\}$3 satisfy $\{H_n\}$4 (Gurumoorthy et al., 2015).

These objects provide the geometric decomposition needed for the rank result. The unstable–neutral directions correspond to non-negative Lyapunov exponents, and the stable directions correspond to negative exponents. The theorem ultimately shows that the error covariances lose support along the stable directions.

4. Riccati recursions and equivalent propagator form

The Kalman-filter Riccati recursions are

$\{H_n\}$5

In data-assimilation notation, one writes

$\{H_n\}$6

so that

$\{H_n\}$7

An equivalent formulation introduces the cumulative gain-modified propagator

$\{H_n\}$8

with

$\{H_n\}$9

(Gurumoorthy et al., 2015).

This equivalent form is central because it separates the roles of dynamical propagation and filtering correction. The operator $\{R_n\}$0 carries the Lyapunov spectrum, while $\{R_n\}$1 captures the accumulated effect of assimilation. The proof relies on combining the asymptotics of $\{R_n\}$2 with boundedness properties of $\{R_n\}$3.

5. Mechanism of the rank-deficiency proof

The proof proceeds by first establishing, through standard Kalman-filter bounds, that $\{R_n\}$4 remain uniformly bounded in operator norm (Gurumoorthy et al., 2015). This step uses observability and prevents the gain-modified propagation from introducing unbounded growth that would obscure the Lyapunov decay structure.

Next, one takes the singular-value decomposition

$\{R_n\}$5

so that

$\{R_n\}$6

The columns of $\{R_n\}$7 converge to the forward Lyapunov vectors $\{R_n\}$8. For any forward Lyapunov direction with index $\{R_n\}$9, the singular values satisfy $P_k^f$0 with $P_k^f$1, hence $P_k^f$2 exponentially (Gurumoorthy et al., 2015).

By Lemma 3.3 in the paper, the projection of $P_k^f$3 onto the subspace spanned by

$P_k^f$4

has operator norm tending to zero. Therefore there are at least $P_k^f$5 directions on which $P_k^f$6 vanishes asymptotically, which yields $P_k^f$7 for large $P_k^f$8. The same argument applies to $P_k^f$9.

A key technical ingredient is the stated $P_k^a$0-eigenvalue collapse lemma: if a symmetric $P_k^a$1 is small $P_k^a$2 on a $P_k^a$3-dimensional subspace, then it has at least $P_k^a$4 eigenvalues in $P_k^a$5 (Gurumoorthy et al., 2015). In context, this converts directional smallness into an eigenvalue-count statement, and hence into the asymptotic rank bound.

6. Confinement to the unstable–neutral subspace

Let

$P_k^a$6

denote the backward Lyapunov vectors associated with the negative exponents, namely the stable subspace. Theorem 3.5 establishes that

$P_k^a$7

(Gurumoorthy et al., 2015).

Equivalently, asymptotically,

$P_k^a$8

lie in the span of the first $P_k^a$9 backward Lyapunov vectors, that is, in the unstable–neutral subspace. This is stronger than a bare rank statement. The result identifies not only how many covariance directions survive asymptotically, but also which directions survive.

The paper states that this provides a rigorous justification for methodologies and algorithms that perform assimilation only in the unstable–neutral subspace (Gurumoorthy et al., 2015). The implication is structural rather than heuristic: under the stated assumptions, covariance mass in the stable subspace is asymptotically extinguished by the combined effect of deterministic dynamics and the Riccati update.

7. Scope, special case, and interpretive significance

The autonomous system is treated as a special case, where the equivalent property is also investigated (Gurumoorthy et al., 2015). No broader generalization is asserted in the summarized result beyond the linear, time-varying, perfect-model framework with noisy outputs and uniform complete observability.

The conceptual importance of the result lies in its synthesis of Riccati dynamics and multiplicative ergodic theory. The controlling quantity is the number of non-negative Lyapunov exponents, and the controlling geometry is furnished by backward Lyapunov vectors. This suggests that the relevant notion of rank in such filtering problems is fundamentally asymptotic and dynamical.

A further implication is methodological. Since both forecast and analysis covariances become supported on the unstable–neutral backward subspace, reduced-order assimilation strategies targeted to that subspace are not merely computational devices but are consistent with the asymptotic structure of the exact Kalman filter in the stated regime (Gurumoorthy et al., 2015). Within that regime, the average rank condition is therefore most naturally understood as a Lyapunov-spectral restriction on the long-time covariance image induced by the Riccati transformation.