Randomised SVD of Interface Iteration

• The paper introduces a randomized SVD technique to extract dominant interface modes, enabling efficient coarse space construction with strong probabilistic error guarantees.
• It leverages inexpensive random sampling of interface operators to bypass costly deterministic SVDs, thereby reducing computational complexity.
• The method integrates with hierarchical Schwarz methods, accelerating convergence for challenging PDEs such as the Helmholtz equation.

Randomised SVD of Interface Iteration refers to the use of randomized singular value decomposition (rSVD) techniques within interface iteration mappings, particularly for building coarse spaces in domain decomposition and iterative solvers. In the context of solving PDEs with nontrivial interface structure—exemplified by the Helmholtz equation—this approach enables efficient extraction of dominant interfacial modes and the construction of compact hierarchical bases, ensuring both scalability and rapid convergence of domain decomposition methods. The overarching methodology optimizes coarse spaces through cheap randomized sampling of interface operators, obviating the need for costly deterministic SVDs and offering strong probabilistic approximation guarantees.

1. Interface Iteration Operators and Their Role

Interface iteration operators arise in domain decomposition methods—such as Schwarz algorithms—deployed for large, indefinite, or oscillatory PDE systems (e.g. the Helmholtz problem). For a decomposition into subdomains $\Omega_i$, the local interface operator $T_{i}$ is given algebraically as

$$T_{i} = B_{i}^{o}\, \tilde{A}_{i}^{-1} B_{i}^{T} \in \mathbb{R}^{n_{\rm out} \times n_{\rm in}}$$

where:

• $B_{i}^{T}$ injects interface data (degrees of freedom on $\partial \Omega_i$) into the local right-hand side,
• $\tilde{A}_{i}^{-1}$ solves the local subdomain PDE (e.g., Helmholtz) with appropriate boundary conditions,
• $B_{i}^{o}$ extracts traces (typically impedance-type) on interface boundaries with neighbors.

The operator $T_{i}$ encapsulates the response of subdomain $\Omega_i$ to data on its interface and is central to the propagation of information in Schwarz-type iterations. Its singular vectors encode the principal directions (modes) along which interface error propagates most robustly.

2. Randomised SVD Algorithm for Interface Maps

Randomised SVD enables efficient low-rank approximation of $T_{i}$ without forming or storing the full operator. The key steps are:

1. Test Matrix Generation: Draw a random Gaussian matrix $\Omega \in \mathbb{R}^{n_{\text{in}} \times (r+p)}$ with target rank $r$ and oversampling $p$.
2. Range Sampling: Compute the sample matrix $$Y = T_{i}\, \Omega \in \mathbb{R}^{n_{\text{out}} \times (r+p)}$$ by solving $(r+p)$ local subdomain problems and collecting traces.
3. Orthonormalization: Compute $[Q, R] = \mathrm{qr}(Y)$ to obtain an orthonormal basis $Q$.
4. Projection: Form the small projected operator $$B = Q^T T_i \in \mathbb{R}^{(r+p) \times n_{\text{in}}}$$.
5. SVD on Projected Operator: Compute $B = \widehat{U} \Sigma V^{T}$.
6. Subspace Assembly: Form approximate left singular vectors $U = Q \widehat{U}$.
7. Truncation: Select the leading $r$ singular triplet $(U_r, \Sigma_r, V_r)$ for the rank-$r$ approximation $T_{i} \approx U_r \Sigma_r V_r^T$.

This procedure is a direct application of randomized subspace iteration SVD methodologies as articulated by Halko, Martinsson, Tropp, and others (Tropp et al., 2023, Gu, 2014). For typical interface operators in elliptic and wave problems, a single pass suffices due to spectral decay. Power iterations can be added for improved accuracy if singular values decay slowly.

3. Probabilistic Error Guarantees and Parameter Choices

The randomized SVD of $T_{i}$ delivers approximation accuracy with high probability, controlled by the oversampling $p$ and the number of power iterations $q$. Standard error bounds are inherited from matrix sketching theory:

$$\|T_i - U_r \Sigma_r V_r^T\|_2 \leq (1 + 9\sqrt{r+p}/p) \sigma_{r+1}$$

with failure probability $O(e^{-p})$ for $q=0$ (Gander et al., 14 Dec 2025). If $q$ subspace iterations are used, the error reduces multiplicatively in $q$ as

$$\|T_i - \widehat{U}_r \Sigma_r \widehat{V}_r^T\|_2 \leq \sigma_{r+1} \left(\frac{r}{p-1}\right)^{1/(2q+1)} + \cdots$$

The coarse rank $r$ must be chosen to capture all physically important interface modes, typically the number of propagating modes per subdomain (scaling with $kH$ in Helmholtz-type problems), plus a margin for evanescent modes. The oversampling $p=5$ is standard and usually sufficient.

4. Computational Complexity and Implementation

The dominating cost in rSVD of interface operators is from local subdomain solves. For each $T_i$, $2(r+p)$ solves are required—$(r+p)$ to sample the range and $(r+p)$ for the projected product $Q^T T_i$. QR and SVD on the small matrices are negligible in comparison once $n_{\text{in}}, n_{\text{out}} \ll N$, where $N$ is the total global degree of freedom count. The approach enables coarse space construction with cost roughly proportional to the sum of the local interface solvers, independent of the global matrix size (Gander et al., 14 Dec 2025). For comparison, deterministic SVD methods would require the explicit assembly and factorization of the full interface operator—a prohibitive cost at large scale.

5. Integration with Hierarchical and Two-Level Schwarz Methods

The randomized coarse basis functions—formed by extending $V_r$ through a single local PDE solve—are assembled into a global coarse matrix $C = [C_1, ..., C_K]$, where each $C_i$ is constructed as $C_i = P_i \tilde{A}_i^{-1} B_i^{T} V_r$. The global coarse operator $A_c = C^T A C$ is then employed in two-level (and recursively, multilevel) Restricted Additive Schwarz (RAS) preconditioners:

• Subdomain correction: Apply local PDE solves and overlap recovery.
• Coarse correction: Compute $$u^{(n)} = \tilde{u}^{(n)} + C A_c^{-1} C^T(f - A \tilde{u}^{(n)})$$.

This eliminates the dominant interfacial error components, leading to iteration counts essentially independent of the fine mesh size $h$ at fixed $kh = O(1)$, and scaling only linearly in $k$ via $r$ (Gander et al., 14 Dec 2025).

6. Relation to General Randomized SVD and Block Krylov Methods

The interface rSVD scheme is an application of general randomized subspace iteration algorithms, whose principles and error theory are robustly established (Gu, 2014, Tropp et al., 2023). The essential innovation is leveraging the block-oriented, pass-efficient nature of randomized SVD—sample, orthogonalize, compress—to treat interface operators, for which deterministic SVD is computationally intractable. For operators with slow spectral decay or poor singular value separation, randomized block Krylov iteration or dynamic-shifted power schemes may be deployed—yielding improved convergence for the same or fewer local matvecs (Feng et al., 2024).

7. Practical Guidelines and Hierarchical Extensions

Empirical studies indicate that oversampling $p=5$ suffices, and power iterations are rarely needed for interface problems in wave physics, due to sharp singular value decay after the propagative modes. In hierarchical settings, one recursively applies randomised SVD at each coarser subdomain level, assembling nested coarse bases $C, C_1, C_2, \ldots$; parallel or sequential application of two-level corrections each eliminate interfacial errors at their respective scales, maintaining mesh-independent convergence even for highly oscillatory or indefinite PDEs (Gander et al., 14 Dec 2025). The overall paradigm ensures practical, scalable coarse space construction with rigorous spectral approximation guarantees grounded in the theory of randomized linear algebra (Gu, 2014, Tropp et al., 2023).

---

References:

• "Hierarchical Coarse Basis by Randomised SVD: the Helmholtz Problem" (Gander et al., 14 Dec 2025)
• "Randomized algorithms for low-rank matrix approximation: Design, analysis, and applications" (Tropp et al., 2023)
• "Subspace Iteration Randomization and Singular Value Problems" (Gu, 2014)
• "Algorithm xxx: Faster Randomized SVD with Dynamic Shifts" (Feng et al., 2024)
• "Faster SVD-Truncated Least-Squares Regression" (Boutsidis et al., 2014)