---
title: Block Semi-Separable Matrix (Block-SSD)
url: https://www.emergentmind.com/topics/block-semi-separable-matrix-block-ssd
type: topic
---

# Block Semi-Separable Matrix (Block-SSD)

A block semi-separable matrix (block-SSD) is a structured matrix arising as the level-0 (single-level) special case within the broader hierarchically block separable (HBS) framework. Such matrices are typically dense yet data-sparse, characterized by the property that their off-diagonal blocks admit low-rank factorizations. Block-SSD matrices and their associated fast algorithms enable efficient direct and least-squares solvers for problems exhibiting nonoscillatory, asymptotically smooth behavior, including operators derived from radial kernels. Their structure facilitates near-linear complexity in both algorithmic factorization and solution procedures in a dimensionality-dependent manner [1212.3521].

## 1. Mathematical Definition and Structural Properties

Let $A\in\mathbb{C}^{M\times N}$ partitioned into $p\times p$ blocks, where the $i$th row block has size $m_i$ and the $j$th column block has size $n_j$:
$$
A =
\begin{pmatrix}
A_{11} & A_{12} & \cdots & A_{1p} \\
A_{21} & A_{22} & \cdots & A_{2p} \\
\vdots & \vdots & \ddots & \vdots \\
A_{p1} & A_{p2} & \cdots & A_{pp}
\end{pmatrix}.
$$
A matrix is block semi-separable (block-SSD) if every off-diagonal block ($i \neq j$) admits a low-rank factorization:
$$
A_{ij} = U_i S_{ij} V_j^T,
$$
with $U_i\in\mathbb{C}^{m_i \times k_i^r}$, $V_j\in\mathbb{C}^{n_j\times k_j^c}$, and $\operatorname{rank}(A_{ij})\leq \min(k_i^r,k_j^c)\ll\min(m_i,n_j)$. Defining $D=\operatorname{diag}(A_{11},A_{22},\dots,A_{pp})$, $U=\operatorname{diag}(U_1,\dots,U_p)$, $V=\operatorname{diag}(V_1,\dots,V_p)$, and $S$ as the block-matrix with $S_{ii}=0$, the block-SSD admits the canonical single-level semi-separable form:
$$
A = D + U S V^T.
$$
This structure is precisely the level-0 realization of the multilevel HBS decomposition, in which higher levels correspond to further hierarchy and recursive partitioning.

## 2. Connection to Hierarchically Block Separable (HBS) Matrices

Block-SSD matrices function as a foundational element within the HBS paradigm, which embeds blockwise low-rank approximability at multiple levels of partitioning. In HBS, index sets $(I_1 = \{1,\dots,M\},\; J_1=\{1,\dots,N\})$ are recursively split and assembled in a tree $\mathcal{T}$ of depth $\lambda$. At each level $\ell$, off-diagonal blocks, indexed by the corresponding tree nodes, must admit factorizations analogous to the block-SSD form. The telescoping, multilevel decomposition is defined recursively via blockwise interpolative decompositions:
$$
A \approx D^{(\lambda)} + L^{(\lambda)}\Bigl[D^{(\lambda-1)} + L^{(\lambda-1)}[\cdots D^{(1)} + L^{(1)} D^{(0)} R^{(1)} ] R^{(2)} \cdots ] R^{(\lambda)}.
$$
For block-SSD, this hierarchy truncates at level-0, directly yielding $A = D + U S V^T$ without further recursion.

## 3. Recursive Skeletonization and Matrix Compression

The key methodology for exploiting block-SSD/HBS structure is recursive skeletonization. For general HBS matrices, skeletonization proceeds bottom-up from the leaves of the partitioning tree. At the finest level ($\ell=\lambda$), each off-diagonal block is approximated by an interpolative decomposition (ID):
$$
A_{ij}\approx L_i^{(\lambda)} S_{ij}^{(\lambda)} (R_j^{(\lambda)})^T,
$$
with $L_i^{(\lambda)}$ and $R_j^{(\lambda)}$ extracting row/column skeletons. The process ascends levels by forming reduced systems $S^{(\lambda)}$, repartitioning, and recompressing. For block-SSD (level-0), no recursion is needed; the compression reflects the single-level structure, and the explicit factorization $A = D + U S V^T$ suffices.

## 4. Equality-Constrained Least Squares Embedding

Solving least-squares problems, $\min_{x\in\mathbb{C}^N} \|Ax-b\|_2$, with block-SSD structure invokes a sparse, equality-constrained embedding. The telescoping HBS form introduces auxiliary variables at each level. Collecting these in block vectors, the constrained system is:
$$
\min_{\mathbf{C}\mathbf{x}=0} \|\mathbf{E} \mathbf{x} - b\|_2,
$$
where both $\mathbf{E}$ and $\mathbf{C}$ are sparse matrices derived from the telescoped ID compressions, and $\mathbf{x}$ comprises the multilevel variables. For block-SSD, the embedding reduces to:
$$
\begin{pmatrix}
D & U \\
V^T & -I
\end{pmatrix}
\begin{pmatrix}
x \\
y
\end{pmatrix}
=
\begin{pmatrix}
b \\
0
\end{pmatrix}, \quad y = V^T x,
$$
yielding a system of size $(M+K)\times(N+K)$, $K = \sum k_i \ll N$, thus maintaining sparsity and computational favorability.

## 5. Semi-Direct Least Squares Solver

The fast semi-direct solver operates in two phases: direct precomputation and iterative least-squares refinement.

### Precomputation (Direct Phase)
1. Compress $A$ via (possibly recursive) skeletonization to precision $\epsilon$, yielding sparse $\mathbf{E},\mathbf{C}$.
2. Form the weighted matrix:
   $$
   \mathbf{A}(\tau)=\begin{pmatrix}\mathbf{E} \\ \tau\mathbf{C}\end{pmatrix},\quad \tau=\epsilon_{\text{mach}}^{-1/3}.
   $$
3. Compute sparse QR decomposition of $\mathbf{A}(\tau)$.

### Solve Phase (Deferred-Correction Refinement)
For each right hand side $b$:
1. Solve $\min\|\mathbf{A}(\tau)x - [b;0]\|$ for $x^{(0)}$ using back-substitution.
2. Initialize residuals and Lagrange multipliers.
3. Perform up to two deferred-correction steps:
   - Augment the right-hand side and solve for correction $\Delta x$.
   - Update the solution, residuals, and multipliers.

It is shown that for well-conditioned $A$ no more than two correction steps suffice [1212.3521].

## 6. Computational Complexity and Dimensional Scaling

The asymptotic complexity depends on both the block ranks and the ambient dimension ($d$), with the depth $\lambda\sim (1/d)\log N$. Let $k_\ell$ denote the off-diagonal rank at level $\ell$; empirical scaling for singular kernels is:
$$
k_\ell \sim
\begin{cases}
O(\ell), & d=1 \\
O(2^{(d-1)\ell}), & d>1
\end{cases}
$$
Specific complexity results are:

| Phase          | $d=1$           | $d=2$                | $d=3$             |
|----------------|-----------------|----------------------|-------------------|
| Compression, QR| $O(M+N)$        | $O(M+N^{3/2})$       | $O(M+N^2)$        |
| Solve          | $O(M+N)$        | $O(M+N\log N)$       | $O(M+N^{4/3})$    |

If sources and targets are well-separated, $k_\ell=O(1)$ and all complexities collapse to $O(M+N)$.

## 7. Specialization and Significance

For matrices that are exactly block-SSD (level-0), recursive skeletonization is unnecessary; the semi-separable $A=D+U S V^T$ representation is direct, and the equality-constrained solver operates natively. The sparse system increases only by $K=\sum k_i$ auxiliary variables, preserving scalability:
$$
\text{Complexity}: \quad O(M+N).
$$
This specialization demonstrates that the classical fast SSD direct solvers are recovered within the HBS formalism, providing a unified theoretical and algorithmic framework. The block-SSD class is thus pivotal for efficient numerical solution of dense operator equations with underlying data sparsity, especially in contexts involving radial basis function approximation, updating, and downdating [1212.3521].

---

Block semi-separable matrices represent the atomic case of hierarchically structured data-sparse matrices. Their innate low-rank block structure, efficient skeletal compressions, and compatibility with modern sparse linear algebra techniques enable rapid dense linear-solving procedures with near-linear complexity scaling in problem size and spatial dimension.

Source: https://www.emergentmind.com/topics/block-semi-separable-matrix-block-ssd