LocAteR Regularization

• LocAteR regularization is a framework for location-adaptive penalization that leverages intrinsic geometric and statistical properties to enhance variable selection and support recovery.
• It integrates methods such as Lévy-process-based shrinkage, adaptive grouped regularization for spatial models, and operator-weighted Tikhonov techniques to handle high-dimensional and inverse problems.
• The approach achieves oracle properties, reduced localization errors, and computational efficiency through methods like EM, coordinate descent, and SVD-based reduction.

LocAteR regularization encompasses a diverse class of methodologies for location-adaptive penalization, variable selection, and source identification in high-dimensional statistics, spatially indexed regression, inverse problems, and off-the-grid sparse estimation. Various instantiations of LocAteR share the defining feature of local adaptivity—exploiting geometric or structural information intrinsic to the domain, operator, or data distribution to guide penalization, shrinkage, or support recovery. Prominent examples include Lévy-process-based regression penalties, Fisher–Rao-geometric sparse identification, local adaptive grouped regularization in spatially-varying models, and operator-based interior localization for PDE inverse problems (Polson et al., 2010, Poon et al., 2018, Brooks et al., 2014, Elvetun et al., 2020).

1. Lévy-Process-Based Local–Global Shrinkage (Regression Frameworks)

The principal architectural motif of LocAteR in regression is the synthesis of penalty functions via Lévy subordinators, which enable the construction of a broad spectrum of shrinkage priors as normal-scale mixtures. The general form for the marginal prior on each regression coefficient $\beta_j$ is

$$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$

where $G$ is the mixing measure induced by a subordinator with Lévy density $\mu(dx)$. This admits a representation as increments of a subordinated Brownian motion $Z(s) = W(T(s))$, enabling direct translation from properties of the subordinator to the penalty and shrinkage behavior.

Totally monotone penalty functions $\psi$ are characterized as Laplace exponents of subordinators, producing the general penalty form

$$\rho(\beta_j) = \nu \, \psi\left( \frac{\beta_j^2}{2} \right)$$

whose explicit functional form is dictated by the choice of subordinator (Gamma, inverse-Gaussian, stable, inverted-beta, etc.). The resulting penalties include classical ridge, Lasso, bridge, and horseshoe families as special cases. This formalism provides posterior means (via Tweedie's formula), posterior modes (via EM or coordinate descent), and a structured algorithm for the $p>n$ regime using SVD-based reduction (Polson et al., 2010).

2. Intrinsic Geometry in Off-the-Grid Sparse Recovery

In the context of support localization, LocAteR regularization is manifested by leveraging the Fisher–Rao metric induced by the measurement operator's reproducing kernel $K(x, x')$. This replaces ad hoc Euclidean separation conditions by intrinsic geometry: the separation between true support locations is quantified via the geodesic distance

$$d_G(x, x') = \inf_{\gamma(0) = x}^{\gamma(1) = x'} \int_0^1 \sqrt{ \dot{\gamma}(t)^\top G_{\gamma(t)} \dot{\gamma}(t) } \, dt,$$

where $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$0 is the (information-theoretic) Riemannian metric derived from $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$1. In practical terms, uniqueness and stability of support recovery in the BLASSO (off-the-grid LASSO) are achieved under separation measured by $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$2 and local curvature (negative Hessian of $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$3), yielding reparameterization-invariant, geometry-adaptive guarantees. Canonical cases include translation-invariant convolution, non-stationary kernels (Laplace transforms), and domains such as spheres or tori (Poon et al., 2018).

3. Local Adaptive Grouped Regularization for Spatially-Varying Models

LocAteR also denotes local adaptive grouped regularization (LAGR) schemes for spatially indexed (or more generally, effect-modified) regression. Here, the focus is on varying-coefficient models

$$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$4

where $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$5 varies with a spatial/location parameter $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$6. Around each site $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$7, $$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$8 is locally approximated linearly, and a group-Lasso penalty is placed on the joint block of raw and interaction coefficients

$$p(\beta_j) \propto \int_0^\infty N(\beta_j; 0, \omega_j) \, dG(\omega_j),$$9

The adaptive penalty weights $G$0 are formed from an initial fit, inducing location-sensitive sparsity and oracle properties (asymptotic normality and selection consistency) under standard regularity. Empirical results show improved estimation and support recovery in both synthetic and applied (e.g., Boston housing) datasets, with smooth adaptation of sparsity patterns across space (Brooks et al., 2014).

4. Operator-Weighted LocAteR for PDE Source Identification

The LocAteR framework extends to inverse problems—specifically, source identification in elliptic PDEs—via operator-weighted Tikhonov regularization. Given a forward operator $G$1 with large nullspace, standard quadratic regularization is insufficient: reconstructions drift toward domain boundaries. LocAteR regularization is realized by scaling each control basis function $G$2 with the norm of its projection orthogonal to $G$3,

$G$4

yielding a modified quadratic penalty $G$5 in the Tikhonov functional

$G$6

The resulting minimizer focuses on interior sources, suppressing boundary artifacts and exhibiting superior localization accuracy in both noiseless and noisy data and in various domain geometries. Theoretical analysis confirms that LocAteR's principal component magnitude peaks at the correct source, and numerical experiments show $G$7 localization error reductions by factors of 2–5 over baseline regularization (Elvetun et al., 2020).

5. Theoretical Properties and Comparative Analysis

LocAteR regularization methods universally exploit structural adaptivity to underlying geometry, sparsity, or operator properties.

• Tail Bounds and adaptivity: Lévy-process-based LocAteR achieves heavy-tailed shrinkage (Student-t, stable, horseshoe), facilitating minimal bias for large signals and strong shrinkage for noise. Infinite-activity subordinators adaptively concentrate mass near zero while maintaining heavy tails, outperforming classical Lasso and ridge penalties in both theory and empirical prediction tasks (Polson et al., 2010).
• Geometric Invariance: Fisher–Rao LocAteR enables support recovery in non-Euclidean parameter spaces, under minimal separation in the operator-induced intrinsic metric, yielding uniform error bounds and dimension-independent rates (modulo polynomial factors) even in complex measurement models (Poon et al., 2018).
• Oracle Properties: Adaptive grouped LocAteR achieves both local selection consistency (probability of excluding truly zero groups tending to 1) and asymptotic normality of active group estimates, matching the efficiency of oracle estimators under mild regularity (Brooks et al., 2014).
• Localization Error Reduction: Operator-weighted LocAteR for PDE inverse problems provides systematic interior localization, robustness to geometric degeneracies (e.g., near domain corners), and enhanced stability under noise (Elvetun et al., 2020).

6. Algorithmic and Computational Considerations

All discussed LocAteR variants admit efficient implementations tailored to their domain structure:

• Regression (Lévy process): Posterior mode finding reduces to weighted ridge or weighted Lasso updates within EM or coordinate descent, leveraging scale mixture structure; high-dimensional Gaussian problems are efficiently addressed via SVD-based dimension reduction (Polson et al., 2010).
• Geometric BLASSO: Solution uniqueness and support stability are governed by geometric certificates—linear systems within local kernel Gram blocks—and efficient empirical approximations via random-feature sampling (Poon et al., 2018).
• Spatial LAGR: Groupwise coordinate descent admits blockwise soft-thresholding updates in penalized problems, with bandwidth selection via locally weighted AIC, exploiting kernel structures for computational tractability (Brooks et al., 2014).
• PDE Inverse: Diagonal penalty operators allow direct sparse factorizations or preconditioned conjugate gradients without the need for advanced acceleration tactics; finite-element packages provide assembly and export to standard solvers (Elvetun et al., 2020).

7. Extensions and Open Problems

Ongoing efforts in LocAteR regularization explore several advanced directions:

• Extension to non-Euclidean parameter spaces (e.g., geometric FLASSO), metric learning for unknown feature distributions, and plug-in information metrics (Poon et al., 2018).
• Higher-order structured sparsity, including support on manifolds, ridges, or stratified sets through generalized Fisher metrics.
• Tighter sample-complexity bounds and removal of random-sign assumptions in geometric sparse recovery (Poon et al., 2018).
• Full-Bayes/variational implementations and integration with uncertainty quantification for priors on infinite-dimensional objects (Polson et al., 2010).
• Algorithmic innovations for large-scale spatial models, manifold-based optimization, and grid-free localization.

A plausible implication is that LocAteR's unifying perspective—local adaptivity grounded in operator or domain geometry—enables robust, interpretable, and theoretically sound solutions to a broad class of inverse, regression, and selection problems, especially where intrinsic locality or geometry is present or must be respected (Polson et al., 2010, Poon et al., 2018, Brooks et al., 2014, Elvetun et al., 2020).