Localized Reduced Basis Methods

• Localized Reduced Basis Methods are techniques that construct reduced models using local subspaces, tailored to subdomains or parameter clusters.
• They overcome limitations of global approximation by using online enrichment and adaptive selection to handle multiscale coefficients and moving discontinuities.
• These methods offer improved computational efficiency and accuracy, crucial for simulations of high contrast, complex geometries, and real-time parameter variations.

Searching arXiv for recent and foundational papers on localized reduced basis methods to ground the article in current literature. Localized reduced basis methods are reduced basis or localized model order reduction techniques in which approximation is built from local spaces rather than a single global reduced space. In the literature, locality may refer to overlapping or non-overlapping subdomains in physical space, coarse elements or multiscale patches, or neighborhoods in parameter or solution space selected by clustering or anisotropic distance. The shared objective is to replace a global reduced space that becomes inefficient for multiscale coefficients, high contrast, moving discontinuities, or parameterized geometries by a family of smaller spaces that can be combined, selected, or enriched adaptively (Buhr, 2017, Abdulle et al., 2014, Maday et al., 2012).

1. Forms of locality

A standard reduced basis formulation starts from a parametrized variational problem

$a(u(\mu),v;\mu)=f(v;\mu)\qquad \forall v\in V,$

and replaces the high-dimensional solution by a Galerkin approximation in a low-dimensional space. Localized variants keep this projection structure but alter the architecture of the reduced space. Instead of one global space, they employ local spaces attached either to subdomains, coarse cells, patches, or parameter regions. This suggests that “localized” denotes a family of approximation strategies rather than a single construction principle.

Notion of locality	Typical construction	Representative papers
Spatial or patch-based	Local spaces on subdomains, overlapping patches, or coarse cells; global coupling by Galerkin, DG, or Schwarz ideas	(Buhr, 2017, Abdulle et al., 2014, Ohlberger et al., 2014)
Parameter-space or solution-space	Local spaces selected from nearby snapshots or clusters in parameter/snapshot space	(Maday et al., 2012, Chasapi et al., 2022, Chasapi, 12 Oct 2025)
Two-scale or optimization-integrated	Local reduced spaces embedded in multiscale or trust-region algorithms	(Keil, 2022, Keil et al., 2023)

A common misconception is that localization must mean domain decomposition in physical space. The parameter-local literature instead defines a local approximation space directly in parameter space, sometimes with a fixed online dimension $N$, and chooses for each query parameter only the nearest snapshots according to an anisotropic metric (Maday et al., 2012). By contrast, multiscale and domain-decomposition formulations localize in the computational domain and assemble a global approximation from locally supported contributions (Abdulle et al., 2014).

2. Spatial localization and multiscale constructions

In overlapping-domain formulations, the domain is covered by subdomains $\{\omega_i\}_{i=1}^{N_D}$ with local spaces

$O_i := H^1_0(\omega_i),$

extended by zero outside $\omega_i$. Starting from $\widetilde V_0=\operatorname{span}\{0\}$, residual-based online enrichment computes the reduced solution $\widetilde u_n$, forms the residual $R_n(\cdot)=f(\cdot)-a(\widetilde u_n,\cdot)$, selects the patch with maximal local dual norm $\|R_n\|_{O_i'}$, solves a local correction problem in that patch, and enriches the global reduced space by one localized function. A globally coupled variant instead solves on $\widetilde V_n\oplus O_i$ for every patch and is optimal among all one-function enrichments drawn from local fine spaces (Buhr, 2017).

In localized multiscale settings, locality is built into the trial space itself. The reduced basis localized orthogonal decomposition introduces a quasi-interpolation $N$0, the fine-scale kernel

$N$1

and local correctors $N$2 on coarse patches $N$3. The localized multiscale space is

$N$4

with $N$5. This yields basis functions with local support and sparse global systems while retaining $N$6-accuracy of order $N$7 when $N$8 (Abdulle et al., 2014).

The localized reduced basis multiscale method (LRBMS) adopts a coarse partition $N$9, local reduced spaces $\{\omega_i\}_{i=1}^{N_D}$0 on each coarse element $\{\omega_i\}_{i=1}^{N_D}$1, and a global broken reduced space

$\{\omega_i\}_{i=1}^{N_D}$2

coupled by a generalized SWIPDG formulation. This construction combines numerical multiscale discretization and reduced basis compression, and it is explicitly designed to reduce the computational complexity with respect to both the multiscale parameter $\{\omega_i\}_{i=1}^{N_D}$3 and the online parameter $\{\omega_i\}_{i=1}^{N_D}$4 (Ohlberger et al., 2014).

3. Parameter-space localization and moving geometries

Parameter-local reduced basis methods define, for each query $\{\omega_i\}_{i=1}^{N_D}$5, a local parameter ball

$\{\omega_i\}_{i=1}^{N_D}$6

choose exactly $\{\omega_i\}_{i=1}^{N_D}$7 snapshots in $\{\omega_i\}_{i=1}^{N_D}$8, and form a local space $\{\omega_i\}_{i=1}^{N_D}$9. The distance $O_i := H^1_0(\omega_i),$0 is not fixed a priori: it is built from an empirically approximated Hessian $O_i := H^1_0(\omega_i),$1, an anisotropic metric tensor $O_i := H^1_0(\omega_i),$2, and a trapezoidal approximation of the geodesic distance. The method gives explicit control over online cost because the online reduced dimension is the user-chosen $O_i := H^1_0(\omega_i),$3, independent of the total number $O_i := H^1_0(\omega_i),$4 of stored snapshots (Maday et al., 2012).

For parameterized geometries on unfitted meshes, the main difficulty is that the active set of degrees of freedom changes with $O_i := H^1_0(\omega_i),$5. One approach embeds every domain $O_i := H^1_0(\omega_i),$6 in a parameter-independent background domain $O_i := H^1_0(\omega_i),$7, defines a background space $O_i := H^1_0(\omega_i),$8, and extends each snapshot by zero on inactive basis functions, producing vectors $O_i := H^1_0(\omega_i),$9. Locality is then introduced by $\omega_i$0-means clustering in parameter space,

$\omega_i$1

followed by local POD on each cluster and nearest-centroid selection online. In this setting, localization is explicitly parameter-based rather than spatial (Chasapi et al., 2022).

For crack problems with moving discontinuities, the snapshots are first mapped to a common reference configuration,

$\omega_i$2

then partitioned by fuzzy c-means into clusters $\omega_i$3, reduced by local POD into bases $\omega_i$4, and selected online by a classifier $\omega_i$5. This construction is tailored to Heaviside jumps and crack-tip singularities represented by XIGA enrichments, and it is explicitly presented as localization in solution or parameter space rather than domain decomposition (Chasapi, 12 Oct 2025).

4. Hyper-reduction, operator compression, and non-intrusive variants

Localized reduced spaces alone do not remove the online cost if the operators remain nonaffine. For unfitted geometries, one route is matrix and vector DEIM on the extended background mesh: $\omega_i$6 with POD on operator snapshots, greedy selection of “magic points,” and coefficient reconstruction from sampled entries. To avoid intrusive online operator assembly, the DEIM coefficients are interpolated by radial basis functions, and localization is applied cluster-wise to both DEIM and reduced bases (Chasapi et al., 2022).

A unified unfitted framework on parameterized domains combines a deformation-based map from a reference configuration to $\omega_i$7, aggregated unfitted finite element spaces, classical and tensor-train reduced bases, and localized dictionaries of reduced subspaces and hyper-reduction approximations obtained via matrix DEIM. In that formulation, the deformation map provides a common reference configuration, while localization produces separate dictionaries for reduced subspaces and for MDEIM approximations (Mueller et al., 21 Aug 2025).

Localized operator approximation also appears in nonlinear finite volume models. For a spatially resolved three-dimensional lithium-ion battery model, a localized reduced basis scheme is built from non-conforming local approximation spaces and a localized empirical operator interpolation for the model’s nonlinearities, indicating the feasibility of localization beyond linear finite element settings (Ohlberger et al., 2016).

A different path is non-intrusive reduction. In the crack-modeling framework, the local reduced coordinates

$\omega_i$8

are approximated by cluster-wise artificial neural networks $\omega_i$9, so that online evaluation requires classification, network inference, reduced reconstruction, and inverse mapping, without intrusive reduced operator assembly (Chasapi, 12 Oct 2025). This suggests that localized reduced basis methods can be paired either with intrusive hyper-reduction or with data-driven coefficient surrogates.

5. Error control, online enrichment, and stability

One of the strongest theoretical results in the area is exponential convergence of online enrichment on overlapping domains. For the residual-based algorithm,

$\widetilde V_0=\operatorname{span}\{0\}$0

and hence

$\widetilde V_0=\operatorname{span}\{0\}$1

The globally coupled enrichment strategy is moreover optimal among all single-function enrichments drawn from local fine spaces (Buhr, 2017).

For LRBMS, error control is based on conservative flux reconstruction rather than the classical global residual estimator. The resulting localized estimator decomposes into a nonconformity term, a residual term, and a diffusive-flux term,

$\widetilde V_0=\operatorname{span}\{0\}$2

all computable locally on each coarse element. This estimator is rigorous with respect to the weak solution and drives adaptive online enrichment through a localized Solve–Estimate–Mark–Refine loop (Ohlberger et al., 2015).

For parabolic problems, elliptic reconstruction yields offline/online decomposable a posteriori bounds on both model reduction and full approximation error. The localized RB method is thereby extended to scalar parabolic evolution equations with true error certification, including discretization and model reduction contributions in one bound (Ohlberger et al., 2016).

Localized reduced basis methods also have a close relationship with domain decomposition. The Localized Reduced Basis Additive Schwarz method interprets local residual-corrector problems as subdomain solves in additive Schwarz, and the resulting scheme can be viewed as a locally adaptive multi-preconditioning scheme for CG (Gander et al., 2021).

Stability, however, is not automatic. In ArbiLoMod for time-harmonic Maxwell’s equations in $\widetilde V_0=\operatorname{span}\{0\}$3, the reduced inf-sup constant can drop under localized Galerkin projection, producing error spikes and motivating future work on localized test spaces for non-coercive problems (Buhr et al., 2016). By contrast, for unfitted Stokes and Navier–Stokes problems on deformed domains, supremizer enrichment is adapted to the reference configuration, and the paper states that the supremizer operator can be defined on the reference configuration without loss of stability (Mueller et al., 21 Aug 2025).

6. Applications, reported performance, and current directions

The application range in the cited literature includes stationary heat conduction with high contrast and channels, parametric multiscale diffusion, Poisson and linear elasticity on trimmed or embedded geometries, time-harmonic Maxwell equations, pore-scale lithium-ion battery models, crack mechanics with XIGA, and PDE-constrained optimization with localized trust-region surrogates (Ohlberger et al., 2016, Keil et al., 2023, Chasapi et al., 2022).

Setting	Reported outcome	Paper
2D Poisson on trimmed domains	Global $\widetilde V_0=\operatorname{span}\{0\}$4, $\widetilde V_0=\operatorname{span}\{0\}$5, $\widetilde V_0=\operatorname{span}\{0\}$6; local max $\widetilde V_0=\operatorname{span}\{0\}$7, max $\widetilde V_0=\operatorname{span}\{0\}$8, max $\widetilde V_0=\operatorname{span}\{0\}$9; global online $\widetilde u_n$0 ms vs local $\widetilde u_n$1–95 ms	(Chasapi et al., 2022)
Quarter-ring linear elasticity	Local RB speedup $\widetilde u_n$2 vs FOM; global $\widetilde u_n$3	(Chasapi et al., 2022)
Crack ROMs on splines/XIGA	2D edge crack: $\widetilde u_n$4 s vs FOM $\widetilde u_n$5 s, speedup $\widetilde u_n$6; 3D edge crack: $\widetilde u_n$7 s vs FOM $\widetilde u_n$8 s, speedup $\widetilde u_n$9	(Chasapi, 12 Oct 2025)
Large-scale PDE-constrained optimization	Relaxed TR-LRBM used $R_n(\cdot)=f(\cdot)-a(\widetilde u_n,\cdot)$0 DG-MsFEM evaluations, $R_n(\cdot)=f(\cdot)-a(\widetilde u_n,\cdot)$1 LRBM evaluations, and $R_n(\cdot)=f(\cdot)-a(\widetilde u_n,\cdot)$2 local patch evaluations, versus $R_n(\cdot)=f(\cdot)-a(\widetilde u_n,\cdot)$3 DG-MsFEM evaluations for BFGS with DG-MsFEM	(Keil et al., 2023)

The broader literature positions localized reduced basis methods at the intersection of reduced basis methods, domain decomposition, multiscale discretization, and hyper-reduction. Representative neighboring frameworks named explicitly in the data include LRBMS, GMsFEM, CEM-GMsFEM, ArbiLoMod, Localized DEIM, hp-RB methods, multi-domain empirical interpolation, and the Localized Orthogonal Decomposition (Abdulle et al., 2014, Buhr, 2017, Buhr et al., 2016).

Several open issues recur. One is certification: some geometry-localized and non-intrusive approaches report numerical error and singular value decay but do not implement explicit certified a posteriori estimators (Chasapi et al., 2022, Chasapi, 12 Oct 2025). Another is stability for non-coercive or saddle-point systems, where localized test-space design and supremizer strategies remain essential (Buhr et al., 2016, Mueller et al., 21 Aug 2025). A further direction is optimization and multiscale control: adaptive localized reduced basis methods have been embedded in trust-region algorithms, and the thesis literature presents TSRBLOD and TR-LRB methods that avoid FEM evaluations of the involved systems during the optimization loop (Keil, 2022). This suggests that the field has evolved from localized approximation spaces toward localized, certified, and optimization-aware reduced models.