---
title: 'Block/Two-level BPJD: Parallel Eigenvalue Solver'
url: https://www.emergentmind.com/topics/block-two-level-bpjd
type: topic
---

# Block/Two-level BPJD: Parallel Eigenvalue Solver

Block/Two-level BPJD (Block Preconditioned Jacobi-Davidson) refers to an optimal, parallelizable solver for large-scale eigenvalue problems—specifically, the computation of multiple or clustered eigenpairs of symmetric elliptic operators as arise in high-order finite element discretizations. The method leverages block strategies (simultaneous treatment of several eigenpairs), a Jacobi–Davidson correction framework tailored to constrained subspaces, and a two-level domain-decomposition preconditioner to realize mesh-independent convergence and robustness against eigenvalue multiplicities or clusters [2203.06327]. The approach is motivated by and tailored for practical settings with high-dimensionality, tight spectral clustering, and parallel hardware.

## 1. Mathematical Formulation and Block Correction Framework

The central problem is the generalized symmetric eigenproblem in discrete variational form: For $A^h:V^h \rightarrow V^h$ (FE stiffness matrix), and $b(\cdot,\cdot)$ the $L^2$ inner product, compute the $s$ lowest eigenpairs $(\lambda_i^h, u_i^h)$:
\[
A^h u_i^h = \lambda_i^h u_i^h, \quad b(u_i^h, u_j^h) = \delta_{ij}, \quad i=1,\dots,s.
\]
The block Jacobi–Davidson formulation replaces the classical single-vector correction with simultaneous computation of block search vectors. At JD iteration $k$, for approximate Ritz pairs $(\lambda_i^k, u_i^k)$, the block correction equation seeks $t_i^{k+1} \in (U^k)^\perp$:
\[
b\bigl((A^h-\lambda_i^k I)t_i^{k+1}, v\bigr) = b(r_i^k, v), \quad v \in (U^k)^\perp, \quad r_i^k = \lambda_i^k u_i^k - A^h u_i^k,
\]
where $U^k = \mathrm{span}\{u_1^k, ..., u_s^k\}$. Each correction system is preconditioned rather than solved exactly, and updated subspaces $W^{k+1}$ are expanded accordingly.

## 2. Two-Level Overlapping Domain-Decomposition Preconditioning

The key innovation is the construction of an efficient two-level preconditioner, $M_i^{k,-1}$, specifically designed to accelerate convergence for multiple and clustered eigenvalues independently of mesh size. The approach is based on overlapping domain decomposition (a Schwarz-type method), for which:
- The domain $\Omega$ is partitioned into $N$ overlapping subdomains $\{\Omega_l'\}$ (diameter $H$, overlap width $\delta$).
- Local spaces $V^{(l)} = V^h \cap H_0^m(\Omega_l')$ and a coarse space $V^H \subset V^h$ are introduced, each admitting $b$-orthogonal projectors.

The two-level block preconditioner for the $i$th correction equation is:
\[
M_i^{k,-1} := (B_{0,i}^k)^{-1} Q_{s+1}^H Q^H + \sum_{l=1}^N (B_{l,i}^k)^{-1} Q^{(l)},
\]
where $B_{0,i}^k = A^H - \lambda_i^k I$ (coarse grid), $B_{l,i}^k = A^{(l)}-\lambda_i^k I$ (local subdomains). These solves are all low-dimensional, independent, and thus highly parallelizable. The preconditioner alternates a coarse global correction with parallel-local corrections, efficiently targeting both global and local error modes.

## 3. Algorithmic Structure

A typical iteration of the two-level BPJD method includes:
1. **Initial Coarse Solve:** Compute Ritz pairs on the coarse space. Initialize $u_i^0$, $\lambda_i^0$, $U^0$.
2. **Block JD Correction:** For $i=1,\dots,s$, build $r_i^k$ and compute $t_i^{k+1} = (I - Q_{U^k}) M_i^{k,-1} r_i^k$.
3. **Subspace Expansion:** Update $W^{k+1}$ by incorporating all current block corrections.
4. **Ritz Extraction:** Solve the reduced eigenproblem in $W^{k+1}$ to obtain new approximations.
5. **Convergence Check:** Terminate or repeat according to Rayleigh quotient convergence.

All preconditioner applications proceed in parallel, and the overall memory and computation scale favorably with the number of subdomains and block size.

## 4. Theoretical Convergence and Robustness

A central result is the convergence bound for the combined Rayleigh quotient errors across blocks:
\[
\sum_{i=1}^s (\lambda_i^{k+1} - \lambda_i^h) \leq \gamma \sum_{i=1}^s (\lambda_i^k - \lambda_i^h), \qquad \gamma = c(H) [1 - C (\delta/H)^{2m-1}]^2,
\]
with $c(H) \to 1$ as $H \to 0$ and $C$ independent of $h$, $H$, or eigenvalue gaps. The method is robust to clustering or multiplicities: the convergence constant $\gamma$ depends solely on the geometric decomposition parameters $(\delta, H)$, not on spectral separation. Iteration counts are uniform in mesh size $h$, establishing optimality for high-fidelity FE discretizations. Increasing the overlap $\delta$ or partition refinement (smaller $H$) yields improved $\gamma$ and faster convergence.

## 5. Numerical Results and Scalability

Comprehensive numerical investigations for 2D and 3D Laplace and biharmonic eigenproblems demonstrate:
- Uniform iteration counts as the degrees of freedom increase by orders of magnitude, confirming $h$-independence.
- Additional subdomains (smaller $H$) monotonically reduce iteration counts.
- Performance on challenging cases (L-shaped, nonconvex domains) matches that on regular geometries.
- Robustness for tightly clustered and multiple eigenvalues: block sizes $s$ up to 20 exhibit no iteration count increase.

Empirical data show iteration numbers nearly halve as $N$ increases from hundreds to thousands, consistent with the $c(H)$ theory.

## 6. Practical Considerations and Parameter Selection

The method requires only the solution of small-scale eigenproblems (coarse grid) and subdomain Dirichlet eigenproblems per JD correction. No global factorization of $A^h$ is needed. The two-level structure ensures both global and local error components are efficiently damped. Selection of overlap width $\delta$ and subdomain diameter $H$ should balance convergence with communication/computation cost, with larger overlap and finer partition yielding superior performance but higher subproblem redundancy.

## 7. Significance and Applicability

Block/Two-level BPJD provides an optimal and highly parallelizable approach for large-scale symmetric elliptic eigenproblems, particularly in regimes where multiple or tightly clustered eigenvalues arise. Its independence from mesh size, spectral gap robustness, and scalability with subdomain count renders it suitable for exascale parallel architectures and applications in quantum mechanics, materials science, and beyond [2203.06327]. A plausible implication is that future work may further integrate such block preconditioners in multilevel or hierarchical frameworks to address even higher-dimensional or more complex systems.

Source: https://www.emergentmind.com/topics/block-two-level-bpjd