Conditional Linear Gaussian Update

• Conditional Linear Gaussian (CLG) update is an analytic mechanism that conditions a multivariate Gaussian prior with linear, additive noise to yield a Gaussian posterior, foundational for Kalman filtering.
• The CLG update admits multiple equivalent interpretations, including probabilistic, optimization-based, and RKHS views, enhancing both theoretical understanding and computational implementation.
• Ensemble methods like the Ensemble Kalman Filter (EnKF) employ the CLG update to adjust empirical means and covariances, bridging theoretical models with high-dimensional data assimilation.

A Conditional Linear Gaussian (CLG) update refers to the analytic mechanism by which a Gaussian prior, combined with linear observations subject to additive Gaussian noise, yields a Gaussian posterior via conditioning. The CLG update constitutes the central step in Kalman filtering, Gaussian process regression, conditional Gaussian process (CGP) inference, and their ensemble analogues, including the Ensemble Kalman Filter (EnKF) and the Ens-CGP framework. This update forms the backbone of inference in conditional linear-Gaussian models, hybrid Bayesian networks (CLG BNs), certain empirical Bayesian classifiers, and high-dimensional data assimilation methods (Ravela et al., 14 Feb 2026, Malik et al., 2024, Mackinlay, 5 Feb 2025, Madsen, 2012).

1. Mathematical Foundations of the CLG Update

The CLG update formalizes the conditioning of a multivariate Gaussian prior on linear, additive-Gaussian observations to yield a Gaussian posterior. Given

• prior $f \sim \mathcal{N}(m, P)$ ($f \in \mathbb{R}^n$),
• observation model $y = H f + \epsilon$, $\epsilon \sim \mathcal{N}(0, R)$, $R \succ 0$, $\epsilon \perp f$,

the joint distribution is

$$\begin{bmatrix} f \ y \end{bmatrix} \sim \mathcal{N}\left( \begin{bmatrix} m \ H m \end{bmatrix}, \begin{bmatrix} P & P H^\top \ H P & H P H^\top + R \end{bmatrix} \right).$$

Conditioning on $y$ yields the posterior

$$f|y \sim \mathcal{N}\left(m_\text{post}, P_\text{post}\right)$$

with

$$m_\text{post} = m + P H^\top (H P H^\top + R)^{-1}(y - H m), \ P_\text{post} = P - P H^\top (H P H^\top + R)^{-1} H P.$$

Alternatively, $f \in \mathbb{R}^n$0, where the Kalman gain $f \in \mathbb{R}^n$1 (Ravela et al., 14 Feb 2026, Mackinlay, 5 Feb 2025). This constitutes both a statistical conditioning and the solution to the strictly convex quadratic program given by Tikhonov-regularized least squares,

$f \in \mathbb{R}^n$2

with unique minimizer $f \in \mathbb{R}^n$3.

2. Interpretations and Representational Equivalences

The CLG update admits multiple, rigorously equivalent characterizations:

• Probabilistic/GP view: Conditioning a joint Gaussian law.
• Kalman Filter (KF) view: The analysis step of classical and high-dimensional Kalman filters; $f \in \mathbb{R}^n$4, $f \in \mathbb{R}^n$5.
• MAP/Quadratic Program (QP) view: Solves a strictly convex quadratic program for the posterior mode; the Hessian $f \in \mathbb{R}^n$6 inverts to $f \in \mathbb{R}^n$7.
• RKHS/Regularization view: Minimization in an RKHS with prior-induced penalty; $f \in \mathbb{R}^n$8 defines the inner product, the update is Tikhonov-regularized regression (Ravela et al., 14 Feb 2026).

These formulations are mathematically identical; the conditional Gaussian law underpins all computational realizations, including variational, optimization-based, and Bayesian inference procedures.

3. CLG Updates in Ensemble and Empirical Settings

Modern high-dimensional data assimilation implements the CLG update empirically through ensemble representations, notably in Ens-CGP and EnKF (Ravela et al., 14 Feb 2026, Mackinlay, 5 Feb 2025, Malik et al., 2024). Given an ensemble $f \in \mathbb{R}^n$9, define:

• Empirical mean: $y = H f + \epsilon$0
• Anomaly matrix: $y = H f + \epsilon$1
• Empirical covariance: $y = H f + \epsilon$2 (rank at most $y = H f + \epsilon$3)

The ensemble CLG (Ens-CGP or EnKF) update applies the exact analytic formulas with $y = H f + \epsilon$4:

$y = H f + \epsilon$5

and updates each ensemble member via deterministic (square-root) or stochastic (perturbed-observation) mappings:

$y = H f + \epsilon$6

or

$y = H f + \epsilon$7

For jointly Gaussian ensembles, the ensemble update asymptotically matches the theoretical CLG posterior (Ravela et al., 14 Feb 2026, Mackinlay, 5 Feb 2025).

4. Role in Complex and