Covariance Localization: Methods & Applications

• Covariance Localization is the imposition of locality, sparsity, or block structure on covariance matrices to suppress spurious long-range correlations and improve estimation.
• It spans methods like distance-based tapering in ensemble filtering, Bayesian truncation in inverse problems, and tensor-aware approaches in high-dimensional settings, each ensuring positive semidefiniteness and accuracy.
• Practical applications include geophysical data assimilation, adaptive machine-learning tuning, and functional data analysis, offering optimality guarantees and scalable performance.

Covariance localization denotes a family of methods and structural phenomena in which locality, sparsity, or block structure is imposed on—or inferred from—a covariance matrix, covariance kernel, or covariance operator. In ensemble data assimilation it usually means suppressing spurious long-range correlations in an ensemble covariance by Schur-product tapering or local analysis; in high-dimensional covariance estimation it means banding or tapering entries according to geometric structure; in some functional and spectral settings it means that localization is an intrinsic consequence of the covariance itself, producing compactly supported eigenfunctions or eigenvectors concentrated on a small subset of components (Gilpin et al., 22 Aug 2025, Sun et al., 11 Jan 2026, Battagliola et al., 3 Jun 2025, Barucca, 2014).

1. Scope of the term

The term is used in several distinct but related senses.

Context	Localized object	Representative formulation
Ensemble data assimilation	Forecast covariance or gain	$$\mathbf{P}^f_{\text{loc}}=\mathbf{T}\circ \mathbf{S}$$
Bayesian inverse problems / MCMC	Prior covariance, prior precision, and observation influence	$[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$
Tensor covariance estimation	Sample covariance entries on a lattice	$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$
Functional data analysis	Eigenfunctions induced by block covariance	$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$
Random-matrix spectral analysis	Covariance eigenvectors	$\mathrm{IPR}^a=\sum_i(v_i^a)^4$

In ensemble filtering, localization is a correction for sampling error caused by small ensembles. In Bayesian inverse problems, it is the replacement of a nearly local problem by an exactly local one through truncation of covariance or precision and restriction of observation influence. In tensor covariance estimation, it is an entrywise regularization of the sample covariance using multi-index geometry rather than scalar index distance. In localized FPCA, localization is inferred from the support pattern of the covariance kernel itself. In random-matrix analysis, by contrast, “localization” means that eigenvectors are concentrated on a small subset of components rather than extended over all variables (Morzfeld et al., 2017, Battagliola et al., 3 Jun 2025, Sun et al., 11 Jan 2026, Barucca, 2014).

This terminological breadth matters because papers with the same phrase often address different mathematical objects. A covariance taper, a localized posterior factorization, a compactly supported eigenfunction, and a high-IPR covariance eigenvector are all called “localized,” but the mechanism and purpose differ.

2. Distance-based localization in ensemble filtering

In stochastic EnKF, the analysis update is

$$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$

When $n_e \ll n_x$, the sample covariance

$$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$

is noisy, rank-deficient, and contaminated by spurious long-range correlations. The standard localization construction is

$$\mathbf{P}^f_{\text{loc}} = \mathbf{T}\circ \mathbf{S},$$

and, in ETKF notation,

$$\mathbf{B}^f_i = \boldsymbol{\rho}_i \circ \mathbf{A}^f_i\mathbf{A}^{f,T}_i.$$

A local-analysis implementation replaces

$[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$0

which is the LETKF-style patch formulation (Gilpin et al., 22 Aug 2025, Popov et al., 2020).

The operational rationale is that, in realistic geophysical applications, the ensemble size $[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$1 is far smaller than the state dimension $[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$2, so the empirical forecast covariance is rank-deficient and contaminated by strong sampling noise. Localization exploits the spatial locality of correlations among model variables and attenuates spurious correlations introduced by finite-sample noise. The most common taper is Gaspari–Cohn; in the 2025 comparison study of high-dimensional covariance estimation and localization for data assimilation, standard distance-based localization remains very hard to beat, and traditional, distance-based localization generally leads to the largest error reduction across the tested cycling EnKF problems (Gilpin et al., 22 Aug 2025).

Theoretical support for distance-based localization appears in spatially structured stochastic dynamics. For a class of stochastically coupled nonlinear systems,

$[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$3

and, when the mean-field term vanishes,

$[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$4

This yields a direct truncation model,

$[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$5

with localization bias bounded by

$[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$6

The same analysis identifies the limitation of purely local tapering: mean-field interaction produces a nonlocal floor of order $[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$7, so localization is accurate only when the local decaying component dominates (Chen et al., 2019).

Positive semidefiniteness is a central implementation constraint. The Schur Product Theorem implies that if both $[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$8 and $[C_{\text{loc}}]_{i,j}=[C]_{i,j}1_{|i-j|\le l}$9 are PSD, then $L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$0 is PSD. By contrast, thresholding methods can generate non-PSD covariance estimates; in cycling EnKF experiments, this can destabilize $L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$1 and cause filter blow-up (Gilpin et al., 22 Aug 2025, Webber et al., 2023).

3. Bayesian, adaptive, and learned localization

A Bayesian interpretation is available for hybrid covariance estimation. With an inverse-Wishart prior, the posterior mode is

$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$2

which is exactly the hybrid estimator

$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$3

By contrast, the commonly used Schur estimator

$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$4

is not the MAP estimator for a smooth Bayesian prior. The proposed QC prior instead penalizes entries of the precision matrix,

$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$5

which leads to a Bayesian view of localization rooted in conditional correlations rather than correlations (Webber et al., 2023).

Adaptive localization replaces hand tuning of a static radius by cycle-dependent optimization or learning. In the OED framework, the localization radii $L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$6 are obtained by solving

$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$7

subject to box constraints on each $L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$8. The localized covariance is

$L_h(S_n;k_h)=[\hat\sigma_{ij}h(\delta_{ij}/k_h)]$9

with $\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$0 built from Gaussian or Gaspari–Cohn kernels, and the paper derives closed-form gradients with respect to the radii. In the reported experiments, adaptive localization performs competitively with manually tuned fixed localization, and regularization is not necessarily needed for localization, unlike adaptive inflation (Attia et al., 2018).

A machine-learning approach tunes the localization radius by supervised regression. The paper proposes one algorithm that adapts the localization radius in time and another that tunes the localization radius in both time and space. The winner radius is defined by minimizing

$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$1

and a random-forest regressor is trained to predict that radius or radius vector from forecast/ensemble features. Numerical experiments with Lorenz-96 and a quasi-geostrophic model show that adaptive-in-time localization can slightly outperform the best fixed hand-tuned radius, whereas the time-and-space variant is more mixed and can create physically implausible localization patterns (Moosavi et al., 2018).

Distance-free localization addresses cases in which physical distance is undefined or misleading. In ES-MDA, localization enters as

$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$2

The 2025 study proposes ML-Localization, which estimates $\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$3 from a large synthetic ensemble generated by a surrogate $\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$4, and CM-Localization, which corrects the sample cross-covariance analytically: $\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$5 These methods are designed for nonlocal/scalar parameters and are used inside pseudo-optimal localization rather than geometric tapering (Silva et al., 16 Jun 2025).

4. Localization as structural decomposition

In Bayesian inverse problems, localization is imposed by truncating prior covariance or precision and restricting the influence of observations to local neighborhoods. Covariance localization is

$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$6

precision localization is

$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$7

and observation localization is

$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$8

For linear-Gaussian problems, the paper gives perturbation bounds showing that the localized posterior moments are close to the exact posterior moments when the neglected prior and observation couplings are small. Once the problem is localized, Gibbs sampling becomes natural; for Gaussian targets with block-tridiagonal covariance or precision, the Gibbs convergence factor is dimension-independent (Morzfeld et al., 2017).

A graph-based data-assimilation formulation uses “localization” in a generic sense: breaking a global assimilation/tuning problem into subproblems. The graph is built from the state–observation mapping using

$\mathcal G=\bigoplus_{k=1}^K \mathcal G_k$9

which is essentially the off-diagonal part of $\mathrm{IPR}^a=\sum_i(v_i^a)^4$0. Community detection then partitions the state into clusters, observations are associated with those clusters, and covariance tuning is performed locally through localized DI01 updates. This approach does not require spatial coordinates, a taper matrix, an empirical cut-off radius, or block-diagonal covariance assumptions (Cheng et al., 2020).

A distinct distributed-cooperative-localization literature addresses the impossibility of exact global cross-covariance tracking in ultra large-scale multi-agent systems. The 2026 framework is not covariance localization in the Schur-product sense. Instead, each agent maintains and communicates only local covariance bounds over bounded-size neighborhoods, and overlapping covariance intersection is used to conservatively account for the unknown larger covariance matrix. The resulting filter is fully distributed, recursively consistent, and scalable with per-agent resources independent of total network size (Pedroso et al., 3 Jun 2026).

These formulations share a common principle: a global covariance object is replaced by local blocks, local neighborhoods, or graph-defined subspaces. The mathematical mechanism, however, is not the same as distance-based tapering.

5. High-dimensional covariance estimation on tensors and lattices

For tensor-valued data, covariance localization is generalized from “near the diagonal in vector order” to “near in tensor geometry.” The data live on a $\mathrm{IPR}^a=\sum_i(v_i^a)^4$1-order lattice

$\mathrm{IPR}^a=\sum_i(v_i^a)^4$2

and the tensor-coordinate separation is

$\mathrm{IPR}^a=\sum_i(v_i^a)^4$3

The proposed localization estimator is

$\mathrm{IPR}^a=\sum_i(v_i^a)^4$4

where $\mathrm{IPR}^a=\sum_i(v_i^a)^4$5 is a $\mathrm{IPR}^a=\sum_i(v_i^a)^4$6-variate localization function and $\mathrm{IPR}^a=\sum_i(v_i^a)^4$7 is a vector of scaling parameters. The associated multi-bandable covariance class is controlled by a decay function $\mathrm{IPR}^a=\sum_i(v_i^a)^4$8 over the preserved $\mathrm{IPR}^a=\sum_i(v_i^a)^4$9-zone $$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$0 (Sun et al., 11 Jan 2026).

Two special cases are emphasized. Multi-banding is

$$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$1

and multiplicative tapering uses

$$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$2

The paper also considers Gaspari–Cohn localization in tensor-coordinate space through

$$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$3

Localization is therefore mode-aware and nonseparable by construction; standard univariate banding can fail badly because vectorized index distance $$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$4 does not reflect tensor proximity (Sun et al., 11 Jan 2026).

The theory is explicitly minimax. Under sub-Gaussian assumptions and $$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$5,

$$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$6

and with oracle $$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$7,

$$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$8

The localization estimator is minimax rate-optimal under spectral norm, and analogous Frobenius-norm optimality is also established (Sun et al., 11 Jan 2026).

The ocean-eddy application makes the geometric interpretation concrete. For daily salinity changes on a $$\mathbf{x}_i^a = \mathbf{x}_i^f + \mathbf{K}\left(\mathbf{y} - (\mathbf{H}\mathbf{x}_i^f + \eta_i)\right), \qquad \mathbf{K} = \mathbf{P}^f \mathbf{H}^\text{T}\big(\mathbf{H}\mathbf{P}^f\mathbf{H}^\text{T}+\mathbf{R}\big)^{-1}.$$9 grid, selected localization scales were $n_e \ll n_x$0 longitude, $n_e \ll n_x$1 latitude, and $n_e \ll n_x$2 m depth for banding localization, and $n_e \ll n_x$3 longitude, $n_e \ll n_x$4 latitude, and $n_e \ll n_x$5 m depth for tapering localization. These horizontal scales are around $n_e \ll n_x$6 km, close to the first baroclinic Rossby radius, and tensor-aware localization gave substantially better reconstruction than sample covariance, univariate banding, or univariate tapering (Sun et al., 11 Jan 2026).

6. Covariance-driven localization in functional and spectral analysis

In localized FPCA, localization is not imposed by an explicit sparsity penalty on the eigenfunctions. Instead, it is inferred from the covariance structure of the underlying stochastic process. For a mean-zero process $n_e \ll n_x$7 on $n_e \ll n_x$8 with covariance kernel $n_e \ll n_x$9, the crucial structural condition is

$$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$0

Then the process decomposes into uncorrelated sub-processes on disjoint intervals, the covariance operator becomes

$$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$1

and the eigenfunctions are supported on single blocks. Orthogonality is automatic because supports are disjoint, and the eigenvalues retain their usual variance-explained interpretation (Battagliola et al., 3 Jun 2025).

The computational procedure is covariance-first and eigendecomposition-second. One first identifies the sparse structure of $$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$2 and the supports $$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$3, then computes FPCA within each sub-process, pools the blockwise eigenvalues, and selects the leading localized eigenfunctions. In practice, exact block structure is blurred by noise, so the paper uses BD-SVD on the empirical covariance after smoothing. Simulations show that L-FPCA achieves the best specificity and precision across the reported designs, and in real data the method yields blockwise explained-variance decompositions such as $$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$4 for the pinch interval in pinch-force data (Battagliola et al., 3 Jun 2025).

A different meaning of localization appears in random-matrix and covariance-spectrum analysis. For coupled heterogeneous Ornstein–Uhlenbeck processes with stationary covariance $$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$5, localization refers to eigenvectors concentrated on a small subset of components rather than extended across all variables. It is quantified by the inverse participation ratio

$$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$6

and the component participation ratio

$$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$7

In this model, heterogeneity broadens the spectrum on both lower and upper edges, and eigenvectors near the small and large eigenvalue edges become localized, while bulk eigenvectors remain comparatively extended. The paper also reports an “inverted-bell” relation between localization and temperature: both high-temperature and low-temperature components are the most localized (Barucca, 2014).

This spectral usage is explicitly distinct from tapering or truncation. The paper states that it studies localization in the eigensystem of covariance/correlation matrices, not localization as a computational method. That distinction is essential when the same term appears across functional data analysis, ensemble filtering, and random-matrix theory (Barucca, 2014).

7. Conditions, limitations, and open directions

Covariance localization is not a single universally valid recipe. Distance-based tapering is strongest when real correlations decay with distance and the dominant problem is finite-ensemble noise. When mean-field interaction is strong, covariance includes a nonlocal $$\mathbf{S} = \frac{1}{n_e-1}\sum_{i=1}^{n_e}(\mathbf{x}_i-\overline{\mathbf{x}})(\mathbf{x}_i-\overline{\mathbf{x}})^\text{T}$$8 component, so hard truncation can remove physically meaningful structure rather than only spurious covariance (Chen et al., 2019). In ensemble filtering, more general methods can occasionally do slightly better than standard distance-based localization, but usually only marginally and only with extra tuning and/or stronger prior information; traditional, distance-based localization remains very hard to beat in the tested cycling EnKF problems (Gilpin et al., 22 Aug 2025).

Hybridization and shrinkage do not eliminate the need for localization in undersampled operational settings. In ETKF, shrinkage toward a climatological covariance can enrich the ensemble subspace and improve stability, but the paper’s conclusion is explicit: shrinkage alone is not enough in operational settings, and localization is still needed (Popov et al., 2020). Adaptive and learned localization introduce additional issues of optimizer cost, feature design, training-set construction, and stability of space-varying radius fields (Attia et al., 2018, Moosavi et al., 2018).

When physical distance is unavailable or misleading, alternative structures become relevant. Graph-based localization uses the state–observation Jacobian rather than geometry, and distance-free ES-MDA localization uses improved parameter–data covariance estimates rather than taper radii. These approaches are particularly relevant for nonlocal observation operators, scalar parameters, and heterogeneous sensitivity patterns (Cheng et al., 2020, Silva et al., 16 Jun 2025). In tensor covariance estimation, the central lesson is analogous: the wrong notion of distance creates a large irreducible bias, so geometry-aware localization is essential (Sun et al., 11 Jan 2026).

Covariance-driven localization in FPCA depends on genuine block or uncorrelated-interval structure in the covariance kernel. If the process has smoothly decaying long-range covariance rather than near-block structure, localization may not arise. The main paper emphasizes structural validity and empirical performance rather than a full asymptotic theory for support recovery (Battagliola et al., 3 Jun 2025). In distributed cooperative localization, exact global cross-covariance propagation remains infeasible at scale, so the price of consistency is conservative local covariance bounding rather than exact optimality (Pedroso et al., 3 Jun 2026).

This suggests that covariance localization is best regarded as a family of structure-exploiting reductions rather than a single algorithm. Across data assimilation, inverse problems, tensor covariance estimation, functional data analysis, and random-matrix theory, the common thread is the use of local structure—distance decay, graph neighborhoods, overlap constraints, block covariance, or conditional independence—to make covariance information estimable, interpretable, or computationally manageable.