Forced Optimal Covariance Adaptive Learning (FOCAL)

• FOCAL is a methodology that adaptively tunes covariance matrices to recover inverse Hessian information in evolutionary optimization and to optimize inflation and localization in ensemble filtering.
• It employs forced step-size adaptation and enhanced covariance learning rates to maintain exploration and accurately extract curvature data even in ill-conditioned, high-dimensional landscapes.
• In data assimilation, FOCAL uses analytic gradients within an A-optimal experimental design framework to continuously minimize state uncertainty by updating inflation factors and localization radii.

Forced Optimal Covariance Adaptive Learning (FOCAL) is a family of methodologies designed to optimally adapt covariance matrices within stochastic search or estimation frameworks, with the explicit goal of either high-fidelity Hessian matrix recovery in black-box optimization or adaptive tuning of covariance inflation and localization in ensemble-based data assimilation. The term FOCAL was independently introduced in the contexts of evolutionary optimization (Shir et al., 2011) and ensemble Kalman filtering (Attia et al., 2018), each leveraging forced adaptation mechanisms to overcome the limitations of conventional strategies in high-dimensional, ill-conditioned, or spatially inhomogeneous scenarios.

1. Problem Formulation and Motivation

Two primary FOCAL frameworks exist:

a) Evolution Strategies (ES) and Inverse Hessian Learning

The central objective is to recover the Hessian matrix $\mathbf{H}$ at the global basin of attraction of a continuous, noisy, black-box objective function, without explicit access to derivatives. Classical Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is known to adapt its $n \times n$ covariance matrix $\mathbf{C}$ such that, under idealized conditions, $\mathbf{C} \propto \mathbf{H}^{-1}$. However, in high-dimensional or highly ill-conditioned landscapes ($n \gtrsim 30$, condition number $\xi \gtrsim 10^4$), the global step-size $\sigma$ of CMA-ES typically collapses, resulting in premature sampling freeze and failure to accurately recover $\mathbf{C}$ as the true inverse Hessian (Shir et al., 2011).

b) Ensemble Filters in Data Assimilation

The goal is to adaptively select spatially and temporally varying covariance inflation factors $\lambda(x, t)$ and localization radii $\rho(x, t)$ at each analysis cycle to minimize posterior state uncertainty in ensemble-based filters (e.g., EnKF). Traditional methods rely on empirical, fixed choices, poorly adapting to non-stationary or heterogeneous observational networks and frequently requiring substantial manual tuning (Attia et al., 2018).

2. Theoretical Foundation

Covariance–Hessian Duality in ES

Near an optimum, the local objective $f$ can be Taylor expanded as $$J(\mathbf{x}) = f_{\max} - f(\mathbf{x}) \approx \frac{1}{2}\mathbf{x}^T\mathbf{H}\mathbf{x}$$. For a rank-based, non-elitist $(\mu, \lambda)$ selection, the distribution of selected points around the optimum observes a memoryless exponential distribution $\mathcal{P}(J) \propto \exp(-\gamma J)$. The induced sample covariance is

$$\mathbf{C} = \int \mathbf{x}\mathbf{x}^T \mathcal{P}[J(\mathbf{x})] d\mathbf{x} \propto \mathbf{H}^{-1}$$

guaranteeing, under sufficient exploration, that the covariance structure encodes inverse Hessian information (Shir et al., 2011).

A-Optimal Experimental Design for EnKF

In data assimilation, the analysis-error covariance $A(\lambda, \rho)$ of the updated state after assimilation is given by

$$A(\lambda, \rho) = \bigl( \widetilde{B}(\lambda, \rho)^{-1} + H^T R^{-1} H \bigr)^{-1}$$

where $\widetilde{B}(\lambda, \rho)$ denotes the inflated and localized prior ensemble covariance. The FOCAL approach frames the tuning of $\lambda$ and $\rho$ as an A-optimal experimental design problem that seeks to minimize $$\Psi(\lambda, \rho) = \operatorname{tr}\,A(\lambda, \rho)$$ using analytic gradients (Attia et al., 2018).

3. Core Methodologies

3.1 FOCAL in Evolutionary Optimization

The FOCAL algorithm modifies the standard CMA-ES as follows:

• Enhanced Covariance Learning Rate: Increase $c_{\text{cov}}$ from $O(1/n^2)$ (CMA-ES default) to $O(10^{-2})$, accelerating convergence of the covariance matrix.
• Forced Step-Size Adaptation: Replace cumulative step-size adaptation (CSA) with a forced step-size:

$$\sigma \gets \frac{\sigma_0}{(\lambda_{\min})^{\alpha}}, \quad 0<\alpha \leq \frac12$$

where $\lambda_{\min}$ is the smallest eigenvalue of $\mathbf{C}$. This update guarantees a finite root-mean-square step $$\delta_p \approx \sigma \sqrt{\operatorname{tr}(\mathbf{C})}$$, sustaining exploration even as the objective approaches its maximum and $\lambda_{\min}$ shrinks.

• Covariance Regularization and Hessian Extraction: On convergence, regularize (e.g., Tikhonov regularization with $\epsilon \approx 10^{-7}$) and invert $\mathbf{C}$ to recover $\mathbf{H}$.

Schematically:

Initialize mean, σ, C = I, set c_cov, σ₀, α Repeat: - Sample λ points ~ N(mean, σ² C) - Select μ best, update mean and covariance C - Eigendecompose C, extract λ_min - Set σ ← σ₀ / (λ_min)^α (forced update) Until convergence Regularize and invert C → estimate H

(Shir et al., 2011)

3.2 FOCAL for Adaptive Ensemble Filtering

At each EnKF analysis cycle, FOCAL performs:

• Control variables: Set per-node inflation $\lambda = (\lambda_1, \ldots, \lambda_N)$ and localization radii $\rho = (\rho_1, \ldots, \rho_N)$.
• Objective: Minimize $$\Psi(\lambda, \rho) = \operatorname{tr}\,A(\lambda, \rho)$$ plus regularizers ($\ell_1$) and box constraints.
• Gradient-based Optimization: Compute analytic derivatives of $\Psi$ with respect to each control variable, then employ gradient-based constrained solvers (e.g., SLSQP) to update $\lambda$ or $\rho$.
• Field Update: Replace background ensemble, apply updated inflation/localization, and proceed with standard EnKF assimilation.

This process ensures adaptation to spatial and temporal variability in uncertainty, ensemble spread, and observational density (Attia et al., 2018).

4. Empirical Performance and Benchmarks

Evolution Strategies

• Noisy, Separable Ellipse ($n=80$, $\xi=10^4$): Standard CMA-ES fails to recover the analytical Hessian spectrum; FOCAL achieves high-fidelity recovery, with $h_i \propto 1/\lambda_i$.
• Atomic Rubidium Control (Rank-deficient, $n=80$): FOCAL uncovers a Hessian with effective rank 6, aligning top eigenvectors with known physical resonances.
• Second Harmonic Generation (Full Rank, $n=80$): FOCAL accurately captures the full eigenspectrum, matching analytic forms, including off-diagonal structure (Shir et al., 2011).

Data Assimilation

• Two-Layer Lorenz-96 (EnKF, $N=25$, partial observations):
  • Fixed inflation ($\lambda = 1.5$), localization $L=0.5$: RMSE $\approx 0.74$.
  • FOCAL-adaptive inflation: RMSE reduced to $\approx 0.65$, with $\lambda(x,t)$ adapting to ensemble spread.
  • FOCAL-adaptive localization: RMSE $\approx 0.70$, with $\rho(x,t)$ adapting to uncertainty structure.
  • Robustness: Superior to fixed-parameter tuning across $N=5\ldots25$ and observation noise $2.5$–$40\%$ (Attia et al., 2018).

5. Algorithmic and Practical Considerations

Key Hyperparameters and Complexity

Parameter	Description	Typical Value / Role
$c_{\text{cov}}$	Covariance learning rate (ES-FOCAL)	$0.01$–$0.1$
$\sigma_0$	Forced step size (ES-FOCAL)	$5$–$10$% of domain span
$\alpha$	Spectral pressure (ES-FOCAL)	$<\frac12$; controls step scaling
$\lambda, \rho$	Inflation, localization (DA-FOCAL)	Adapted within prescribed bounds

FOCAL methods typically require per-update eigendecompositions ($O(n^2)$) for ES implementations and a few dozen iterations of smooth, box-constrained minimization in the data assimilation context. Regularization (e.g., Tikhonov) is essential for inverting empirically estimated covariances.

Operational Guidelines

• For ES, ensure the base optimizer reaches the global optimum before switching to FOCAL updates.
• For EnKF, FOCAL is applied at each assimilation cycle; careful choice of regularization and bounds prevents overfitting to noise or spurious features.

6. Limitations and Open Research Questions

• FOCAL for ES estimates the Hessian only in the local basin reached by the optimizer; it does not survey multiple optima or global non-quadraticity.
• Parameter tuning (e.g., $c_{\text{cov}}$, $\sigma_0$, $\alpha$) remains essential, with recommended values provided for typical settings.
• In extremely high-dimensional scenarios, computational costs can be significant, motivating ongoing investigation into reduced-rank and diagonal-covariance variants.
• Open directions include extension to first-order evolution strategies, establishing conditions under which standard CMA-ES suffices for Hessian learning, and joint inflation-localization adaptation in EnKF settings (Shir et al., 2011, Attia et al., 2018).

7. Impact and Future Developments

FOCAL methodologies in both evolutionary Hessian learning and adaptive ensemble filtering enable systematic, analytic, and robust exploitation of covariance adaptation, substantially improving upon ad hoc or parameter-tuned approaches. They facilitate recovery of landscape curvature or optimal assimilation parameters even in regimes where traditional methods fail. Promising avenues for future research include joint optimization of inflation and localization, use of alternative optimality criteria (e.g., D-optimality), incorporation of Bayesian/smoothness regularization, and development of computationally efficient schemes for large-scale geophysical or quantum optimal control applications (Shir et al., 2011, Attia et al., 2018).