GenEO Eigenproblems & Adaptive Coarse Spaces

• GenEO eigenproblems are local generalized eigenproblems in overlapping subdomains that form the spectral basis for adaptive coarse spaces.
• They adaptively select low-energy eigenmodes to capture problematic global error components, ensuring condition number bounds independent of subdomain count and coefficient contrast.
• Variants like R-GenEO and Hk-GenEO optimize computational cost and robustness for SPD, indefinite, Helmholtz, and saddle-point systems.

A GenEO (Generalized Eigenproblems in the Overlap) eigenproblem is a local generalized eigenproblem defined within overlapping subdomain decompositions, serving as the spectral foundation for constructing operator-adapted, robust coarse spaces used in two-level Schwarz-type preconditioners for elliptic and indefinite PDEs. GenEO methods guarantee condition number bounds independent of subdomain count and coefficient contrast by adaptively generating coarse spaces from localized eigenmodes that encapsulate problematic global and near-null error components.

1. Principles of the GenEO Eigenproblem and Coarse Space Construction

The classical GenEO methodology belongs to the family of adaptive or spectral coarse spaces, specifically designed to address the limitations of one-level overlapping Schwarz methods, whose convergence deteriorates with increasing subdomain count $N$ and coefficient heterogeneity. In the two-level Schwarz framework, the key is enriching the one-level local-solve preconditioner with a coarse correction space $V_H$ constructed from selected low-energy eigenmodes of localized generalized eigenproblems.

Let $A$ be the global SPD operator (finite element matrix) arising from discretizing an elliptic PDE on a domain $\Omega$. The domain is covered by overlapping subdomains $\{\Omega_j\}_{j=1}^N$, and the corresponding local finite element spaces are $V_h(\Omega_j)$. The core GenEO eigenproblem on each $\Omega_j$ seeks $(\lambda, t)\in\mathbb{R}\times V_h(\Omega_j)$ satisfying

$$a_{\Omega_j}(u, t) = \lambda\, a_{\Omega_j^\circ}( \xi_j(u), \xi_j(t) ),\quad \forall u \in V_h(\Omega_j),$$

where $a_{\Omega_j}$ and $a_{\Omega_j^\circ}$ denote subdomain and overlap bilinear forms, and $\xi_j$ is a partition-of-unity extension operator tailored to enforce the required boundary and overlap conditions (Spillane, 2021, Seelinger et al., 2019, Bastian et al., 30 Oct 2025).

The set of eigenfunctions with the lowest eigenvalues—those corresponding to slowly-decaying or poorly-represented global error components—is selected, typically using a fixed threshold or by truncating at $m_j$ eigenfunctions per subdomain. The global coarse space is assembled as

$$V_H = \mathrm{span}\{ R_j^T\,\xi_j( t_j^k )\ :\ j=1,\dots,N;\ k=1,\dots,m_j \},$$

where $R_j^T$ is the extension of subdomain functions to the global space (Bastian et al., 30 Oct 2025, Spillane, 2021).

2. GenEO Variants and Adaptations

GenEO eigenproblems have been generalized and adapted to a broad spectrum of problems, including:

• SPD linear elasticity, where the construction and selection of partition-of-unity weights and overlap zones control the representation of near-rigid-body modes (Spillane, 2021).
• Indefinite and non-self-adjoint problems (convection-diffusion-reaction, Helmholtz), where coarse spaces may be based on eigenproblems using the full indefinite operator or combinations thereof—$\Delta$-GenEO employs a positive-definite subproblem, while $H_k$-GenEO uses the indefinite operator directly for wavenumber-robust performance (Bootland et al., 2021, Dolean et al., 2024, Bootland et al., 2021).
• Saddle-point systems, where the GenEO construction is extended to both the primal block and the Schur complement, resulting in dual GenEO eigenproblems in the constraint space (Nataf et al., 2019).
• R-GenEO, which restricts local eigenproblems to thin strips near the subdomain boundary, reducing setup cost while essentially preserving spectral convergence properties (Bastian et al., 30 Oct 2025).

The following table contrasts key features of the classical and R-GenEO approaches in the SPD, elliptic context:

Approach	Local Eigenproblem Domain	Cost per Subdomain	Condition Number Bound
GenEO	Entire subdomain + overlap	$O(h^{-d})$ unknowns	$$\kappa \lesssim (1+k_0)[2+k_0(2k_0+1)(1+1/\lambda)]$$ (Bastian et al., 30 Oct 2025)
R-GenEO	Strip near subdomain boundary	$O(h^{-(d-1)}\delta^{-1})$	$$\kappa \lesssim (1+k_0)[2+k_0(2k_0+1)(2+3/\lambda)]$$ (Bastian et al., 30 Oct 2025)

A plausible implication is that R-GenEO reduces the computational setup cost by a factor of $3$–$4$ in 2D with only a slight possible increase in coarse-space dimension.

3. Algorithmic Realization

Algorithmic steps for GenEO-based coarse space construction and Schwarz preconditioning, as exemplified by large-scale parallel implementations (Seelinger et al., 2019), are:

1. Subdomain-wise: Assemble local stiffness matrices $A_j$ and overlap matrices $A_j^\circ$.
2. Construct discrete partition-of-unity weights and operators $X_j$.
3. Solve the local symmetric generalized eigenproblem $A_j \phi = \lambda B_j \phi$, $B_j = X_j A_j^\circ X_j$, using sparse, shift-invert solvers (e.g., ARPACK).
4. Select $m_j$ low-energy eigenvectors.
5. Lift and glue via partition-of-unity and extension to global coarse basis vectors.
6. Assemble global coarse matrix $A_H$ via local computations and sparse communication.
7. Two-level Schwarz preconditioner $M^{-1}$: apply local Schwarz solves plus a global coarse correction.

This step structure enables full parallelization; the only required global communication is the assembly of the coarse matrix.

4. Theoretical Properties and Condition Number Estimates

GenEO-type spectral coarse spaces enjoy explicit quantitative condition number bounds. For classical GenEO (and SPD $A$), under standard assumptions on overlap and mesh geometry, the preconditioned operator's condition number satisfies

$\kappa(M_\mathrm{GenEO}^{-1}A) \leq C\,(1+1/\tau),$

where $C$ depends only on the overlap number, not on mesh size or coefficient contrast, and $\tau$ is the eigenvalue cutoff (Spillane, 2021). For R-GenEO, the condition number bound retains the same independence, with slightly larger constants due to localization of the eigenproblem (Bastian et al., 30 Oct 2025).

For indefinite and non-self-adjoint problems, theory demonstrates that using suitable operator-adapted eigenproblems (e.g., $H_k$-GenEO for Helmholtz, or dual and primal eigenproblems in saddle point systems (Dolean et al., 2024, Nataf et al., 2019)), robust GMRES convergence is achieved, and the required coarse-space dimension scales linearly in wave number for Helmholtz, outperforming classical approaches.

5. Extensions: Indefinite, Non-Hermitian, and Saddle Point Problems

The GenEO paradigm has successfully been generalized beyond SPD settings:

• For heterogeneous indefinite elliptic and Helmholtz problems, $\Delta$-GenEO uses a positive-definite subproblem for the eigenproblem, providing mesh- and contrast-robustness, while $H_k$-GenEO and $\mathcal{H}$-GenEO, using the full indefinite operator, deliver convergence with iteration counts independent or only mildly dependent on wavenumber at the cost of modestly larger coarse spaces (Bootland et al., 2021, Dolean et al., 2024).
• In saddle-point systems, including Stokes and mixed elasticity, both primal and dual spectral coarse spaces are built from local eigenproblems in the displacement and constraint variables, with spectral equivalence established a priori (Nataf et al., 2019).
• Extended GenEO-type eigenproblems with non-Hermitian sesquilinear forms are used in optimized restricted additive Schwarz for non-Hermitian Helmholtz, providing wavenumber-robust preconditioners (Parolin et al., 30 Nov 2025).

6. Computational Performance and Practical Considerations

GenEO methods are characterized by strong scalability and parallel efficiency. The local nature of the eigenproblems ensures embarrassingly parallel setup phase; communication is limited to local neighborhoods and sparse gather operations. Numerical studies report:

• Setup time for R-GenEO is $3$–$4\times$ less than classical GenEO, with identical or nearly identical iteration counts and conditioning, as demonstrated on high-contrast PDEs with up to $10^6$ coefficient jumps (Bastian et al., 30 Oct 2025).
• Weak- and strong-scaling to $10^4$–$10^5$ cores and hundreds of millions of degrees of freedom, retaining iteration counts independent of subdomain number, mesh size, and coefficient contrast, have been reported (Seelinger et al., 2019).
• For Helmholtz and indefinite problems, the coarse-space dimension of $H_k$-GenEO grows linearly with wavenumber, with robust iteration counts; classical coarse spaces' performance deteriorates at high frequency due to missing oscillatory null modes (Dolean et al., 2024).

Trade-offs include coarse-space size versus convergence, with operator-adapted eigenproblems (e.g., $H_k$-GenEO, $\mathcal{H}$-GenEO) enabling smaller coarse spaces for strongly indefinite regimes at the expense of more complex eigenproblem assembly.

7. Numerical Experiments and Comparative Results

Key numerical findings across the literature include:

• For SPD, high-contrast diffusion, and elasticity, GenEO preconditioning reduces the condition number from thousands (one-level) to $\mathcal{O}(10)$–$\mathcal{O}(50)$, with iteration counts of $10$–$40$, independent of subdomain count and heterogeneity (Spillane, 2021, Bastian et al., 30 Oct 2025).
• For Helmholtz, classical coarse spaces require $O(k^2)$ modes per subdomain; $H_k$-GenEO achieves $O(k)$ scaling in coarse-space size and $k$-independent GMRES iteration counts (Dolean et al., 2024).
• For indefinite elliptic problems, $\mathcal{H}$-GenEO outperforms $\Delta$-GenEO as indefiniteness increases, maintaining low iteration counts at higher values of $\kappa$ (Bootland et al., 2021).
• R-GenEO delivers nearly identical iteration counts to GenEO with reduced setup cost:
  • For $N=16$ subdomains, GenEO: $27$ iterations, $\kappa \approx 13.7$, $t_\text{setup} \approx 3.7$ s; R-GenEO: $27$ iterations, $\kappa \approx 12.7$, $t_\text{setup} \approx 1.21$ s (Bastian et al., 30 Oct 2025).

These results confirm the adaptability and robustness of GenEO-type eigenproblems in constructing scalable preconditioners for a wide class of large-scale, heterogeneous, and challenging PDE systems.