EM Algorithms for Covariance Tuning

• EM Covariance Tuning is the adaptive estimation of covariance matrices using iterative EM steps combined with regularization to overcome low-sample issues.
• It employs penalized and constrained updates to incorporate structured targets (e.g., isotropic, diagonal, autoregressive), ensuring positive definiteness and numerical stability.
• The approach extends to state-space and online filtering models, enhancing accuracy in estimating process and observation noise in high-dimensional systems.

The Expectation-Maximization (EM) algorithm for covariance tuning refers to a class of methodologies that adaptively or robustly estimate covariance matrices within probabilistic latent-variable models, such as Gaussian mixture models (GMMs), state-space models, or more complex hierarchical structures. Covariance tuning via EM is essential in high-dimensional, low-sample-size scenarios, structured models, and online filtering, where naïve maximum likelihood estimates are singular, non-invertible, or fail to incorporate structural domain knowledge.

1. Penalized and Regularized EM for Covariance Estimation

Standard EM algorithms maximize the observed data log-likelihood, which for mixture and latent variable models leads to closed-form updates for mean and covariance under Gaussian assumptions. However, when the sample size $n$ is not much larger than the dimension $m$, classical EM can produce singular or ill-conditioned covariance estimates.

To address this, Houdouin et al. (2022) introduce a regularized EM (RG-EM) algorithm that modifies the traditional log-likelihood by incorporating a penalization term for each component $k$:

$$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$

Here, $\Pi_{\rm KL}(\Sigma_k, T_k)$ is a Kullback-Leibler type divergence between the current covariance and a structured positive-definite target $T_k$, and $\eta_k$ is a regularization parameter. In the M-step, $\Sigma_k$ is updated as a convex combination (shrinkage) of the empirical covariance and $T_k$, where the relative weight is controlled by $\eta_k$. This ensures positive definiteness and improves numerical conditioning, particularly in the low-sample support regime (Houdouin et al., 2023, Houdouin et al., 2023).

2. Structured Targets, Constraints, and Selection of Tuning Parameters

Tuning the covariance in EM can involve structured target matrices $m$0 reflecting domain knowledge:

• Isotropic ($m$1): Enforces global scaling and rotational invariance.
• Diagonal: Retains variable-specific noise power while ignoring cross-correlation.
• Autoregressive, block-diagonal, or custom: Embeds structural priors, e.g., spatial, temporal, categorical dependencies.

RG-EM and constrained-covariance EM also support direct constraint on the eigenvalues of $m$2 (i.e., minimal and maximal allowed variances) by spectral clipping in the M-step. This approach, as developed in Gaussian Parsimonious Clustering Model (GPCM) settings, eliminates singularities and pathologies by enforcing hard bounds, thereby guaranteeing monotonic ascent of the penalized likelihood and operational robustness (Browne et al., 2013).

Tuning hyperparameters such as $m$3 (regularization strength), or bounds ($m$4, $m$5) in the spectral domain, typically employs cross-validation, Bayesian Information Criterion (BIC), or empirical Bayes. For each candidate, clustering or out-of-sample likelihood is evaluated and the parameter that minimizes validation loss or maximizes BIC is selected (Houdouin et al., 2023, Browne et al., 2013).

Target Type	Example	Induced Structure
Isotropic	$m$6	Equal variance, no correlation
Diagonal	$m$7	No correlation, heteroscedasticity
Block-diagonal		Grouped dependency
Auto-regressive (AR)	Toeplitz(ρ)	Temporal/spatial correlation

3. EM Covariance Tuning in State-Space and Filtering Models

Covariance tuning is fundamental in state-space models (e.g., MARSS, Kalman filters), especially for adapting process noise ($m$8) and observation noise ($m$9). The EM algorithm alternates between smoothing the latent states via the Kalman (or extended/ensemble/particle) smoother (E-step) and maximizing the expected complete-data log-likelihood (M-step) with respect to $k$0, $k$1, and other parameters.

In classical MARSS settings, the unconstrained updates are:

$k$2

where $k$3 and $k$4 are smoothed residual outer products, formed via the Rauch–Tung–Striebel smoother. Constraints on $k$5 and $k$6 can be imposed via linear parameterizations, enabling the enforcement of structure or sparsity (Holmes, 2013).

In nonlinear or high-dimensional DA (data assimilation), EM-based covariance tuning has been extended:

• Online EM for EnKF/particle filters uses a stochastic-approximation update for the sufficient statistic, e.g.,

$k$7

where $k$8 is a decaying step size and $k$9 is a new innovation covariance (Cocucci et al., 2020).

• Particle flow filter EM approximates the expectation by combining filtering ensembles with a fixed-point update for $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$0, eliminating the need for backward smoothing and enabling direct optimization even in high dimensions (Lucini et al., 2019).
• Extended Kalman filtering with EM (left/right invariant EKF): Noise covariances are tuned via batch EM using filtered and smoothed error moments for both process and measurement noise. Explicit closed-form updates for $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$1 and $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$2 are derived under linearized models, leveraging invariant geometry for numerical stability (Pandey et al., 2024, Pandey et al., 2024).

4. Estimation of Parameter Covariance and Standard Errors in EM

The observed Fisher information matrix, necessary for standard errors and uncertainty quantification, is not directly available from the sequence of EM iterates. The Supplemented EM (SEM) and Agile-SEM approaches estimate the parameter covariance matrix $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$3 by reconstructing the observed-data information $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$4 from (i) the complete-data information at the MLE, and (ii) the rate-of-convergence operator (Jacobians of the EM map). Agile-SEM adaptively controls finite-difference step sizes, ensuring stability with IEEE floating-point arithmetic, and requires only three single EM evaluations per parameter, yielding reliable asymptotic covariances even in high dimensions (Pritikin, 2016, Brümmer, 2014).

5. Theoretical Properties and Empirical Results

EM-based covariance tuning possesses several crucial theoretical guarantees:

• Monotonicity: Penalized likelihood or observed-data likelihood is non-decreasing under RG-EM, constrained EM, and general covariance-tuning variants (Houdouin et al., 2023, Browne et al., 2013).
• Positive-definiteness: Covariance updates maintain strict positive-definiteness either by convexity (RG-EM) or by explicit spectral clamping (constrained EM, Kalman filter variants).
• Consistency and Stability: Proper choice of targets and regularization yields consistent covariance estimates, avoids degeneracy, and enhances robustness to sample scarcity.

Empirical results across a variety of domains include:

• In GMM clustering (AR/Toeplitz covariances, UCI benchmarks), RG-EM maintains failure-free operation and high precision clustering as $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$5, outperforming unregularized EM by up to 10% on precision metrics (Houdouin et al., 2023).
• In state-space and DA models (Lorenz-63, Lorenz-96), both batch and online EM approaches accurately recover process and observation noise structures, successfully tracking time-varying covariances with strong stability even for $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$6 and $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$7 of dimension 1600 (Cocucci et al., 2020, Lucini et al., 2019).
• In quaternion-based attitude estimation, adaptive EKFs with EM-tuned covariance matrices achieve filter accuracy and stability essentially indistinguishable from filters initialized with the true noise covariances, even when initial guesses are off by two orders of magnitude (Pandey et al., 2024, Pandey et al., 2024).

6. Extensions, Limitations, and Future Directions

Covariance-tuning EM frameworks readily generalize to:

• Mixtures of heavy-tailed or skewed distributions by adapting the penalty or constraints in the M-step (Houdouin et al., 2023).
• High-dimensional and structured settings, leveraging low-rank, sparse, or block-diagonal targets and penalization, with data-driven or empirical Bayes selection of regularization hyperparameters.
• Unsupervised learning of target structures $$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$8 or (joint) tuning of multiple covariance modules in hierarchical models.

Outstanding challenges include:

• Fully unsupervised and statistically principled selection of tuning parameters ($$\ell_{\text{pen}}(\Theta | X) = L(\Theta | X) - \sum_{k=1}^K \eta_k \; \Pi_{\rm KL}(\Sigma_k, T_k)$$9, spectral bounds).
• Learning structured targets $\Pi_{\rm KL}(\Sigma_k, T_k)$0 directly from data.
• Extending finite-sample error and uncertainty quantification in non-classical regimes (e.g., small $\Pi_{\rm KL}(\Sigma_k, T_k)$1, high $\Pi_{\rm KL}(\Sigma_k, T_k)$2).

These axes represent active research directions across statistical machine learning, signal processing, and applied fields reliant on robust probabilistic modeling.

---

References:

• Regularized EM for GMMs: (Houdouin et al., 2023, Houdouin et al., 2023)
• Constrained covariance EM for mixture models: (Browne et al., 2013)
• EM for MARSS and state-space covariance estimation: (Holmes, 2013)
• EM in ensemble and particle filtering: (Cocucci et al., 2020, Lucini et al., 2019)
• Standard error/parameter covariance via SEM: (Pritikin, 2016, Brümmer, 2014)
• EM-augmented adaptive (left/right) invariant EKF: (Pandey et al., 2024, Pandey et al., 2024)
• Monte Carlo EM for covariance adaptation in distribution estimation: (Brookes et al., 2019)