---
title: Arrowhead-Matrix Method Overview
url: https://www.emergentmind.com/topics/arrowhead-matrix-method
type: topic
---

# Arrowhead-Matrix Method Overview

Searching arXiv for recent and foundational papers on arrowhead matrices and related methods.
“Arrowhead-Matrix Method” denotes a family of structure-exploiting techniques built around matrices whose off-diagonal entries are confined to one distinguished row and column, or around closely related generalizations such as diagonal-plus-rank-one and block-tridiagonal-arrowhead forms. In the positive-semidefinite setting, the method can mean an explicit rank-one sparse decomposition on \(2\times 2\) principal submatrices; in numerical linear algebra, it can mean shift-and-invert eigensolvers, \(O(n)\) determinant and inverse formulas, sparse approximation from matrix-vector products, or distributed Schur-complement decompositions that preserve arrowhead structure. A particularly explicit formulation appears for positive semidefinite arrowhead matrices in “The Factor Width Rank of a Matrix,” where every such matrix has factor width at most \(2\) and satisfies \(\mathrm{fran}_k(A)=\mathrm{rank}(A)\) for all \(k\ge 2\) [2405.11556].

## 1. Definition and structural setting

An arrowhead matrix \(A\in M_n\) is one satisfying
\[
a_{i,j}=0 \quad\text{whenever } i\neq 1,\ j\neq 1,\ \text{and } i\neq j.
\]
Equivalently, all off-diagonal entries are confined to the first row and first column. In block form,
\[
A= \begin{bmatrix} \alpha & x^*\\ x & D \end{bmatrix},
\]
where \(\alpha\in\mathbb F\), \(x\in\mathbb F^{n-1}\), and \(D=\mathrm{diag}(d_2,\dots,d_n)\) is diagonal. For positive semidefinite arrowhead matrices, \(\alpha\ge 0\), \(D\succeq 0\), and the Schur-complement inequality imposes the relation
\[
\alpha \ge \sum_{i=2}^n \frac{|a_{1,i}|^2}{a_{i,i}}
\]
when all \(a_{i,i}>0\) [2405.11556].

This zero pattern is naturally identified with a star graph. That viewpoint is important because several results in the literature interpret arrowhead matrices as sparse objects whose nonzeros are concentrated on edgewise \(2\times 2\) principal blocks involving a hub index. A broader but closely related setting includes matrices of the form
\[
A=\begin{bmatrix} D & u\\ v^\star & \alpha \end{bmatrix}
\]
over \(\mathbb F\in\{\mathbb R,\mathbb C,\mathbb H\}\), diagonal-plus-rank-one matrices
\[
A=\Delta+x\rho y^\star,
\]
and block-tridiagonal-arrowhead matrices used in selected inversion and interior-point methods [2212.10966].

The term is also used for generalizations in which a small dense head is combined with a narrow band. In “Arrow Matrix Decomposition,” a matrix has arrow-width \(b\) if all nonzeros are confined to the first \(b\) rows, the first \(b\) columns, and a diagonal band of width \(b\); ordinary arrowhead matrices are the special case \(b=1\) [2402.19364].

## 2. Positive-semidefinite decomposition and factor-width rank

For \(A\in \mathcal M_n^+(\mathbb F)\), factor width at most \(k\) means that
\[
A=\sum_{j=1}^r \mathbf v_j \mathbf v_j^*
\]
for vectors \(\mathbf v_j\in\mathbb F^n\), each having at most \(k\) nonzero entries. Equivalently, \(A\) can be written as a sum of positive semidefinite matrices each supported on a single \(k\times k\) principal submatrix. The factor-width-\(k\) rank is
\[
\mathrm{fran}_k(A) :=\min\left\{r:\; A=\sum_{j=1}^r \mathbf v_j\mathbf v_j^*,\ \mathbf v_j\in\mathbb F^n,\ \|\mathbf v_j\|_0\le k\right\}.
\]
One always has
\[
\mathrm{rank}(A)\le \mathrm{fran}_k(A).
\]

For arrowhead matrices, Theorem 4.4 gives a complete characterization: if \(A\in M_n^+\) is arrowhead, then \(A\) has factor width at most \(2\), and
\[
\mathrm{fran}_k(A)=\mathrm{rank}(A)\qquad\text{for all }k\ge 2.
\]
The proof first removes zero diagonal entries, since for \(A\succeq 0\) any vanishing diagonal entry forces the corresponding row and column to be zero. It then uses the determinant identity
\[
\det(A) = \left(\prod_{i=2}^n a_{i,i}\right)\left(a_{1,1}-\sum_{i=2}^n \frac{|a_{1,i}|^2}{a_{i,i}}\right),
\]
so that \(\mathrm{rank}(A)\) is either \(n\) or \(n-1\) depending on whether the scalar residual at \((1,1)\) is positive or zero [2405.11556].

The constructive decomposition defines, for each \(2\le i\le n\), a vector \(\mathbf v_i\) supported only on coordinates \(1\) and \(i\), with
\[
[\mathbf v_i]_i=\sqrt{a_{i,i}}, \qquad [\mathbf v_i]_1=\frac{a_{1,i}}{\sqrt{a_{i,i}}}.
\]
Then
\[
[\mathbf v_i\mathbf v_i^*]_{1,1}=\frac{|a_{1,i}|^2}{a_{i,i}},\qquad
[\mathbf v_i\mathbf v_i^*]_{1,i}=a_{1,i},\qquad
[\mathbf v_i\mathbf v_i^*]_{i,i}=a_{i,i},
\]
and
\[
A = \sum_{i=2}^n \mathbf v_i\mathbf v_i^* +
\begin{bmatrix}
a_{1,1}-\sum_{i=2}^n \frac{|a_{1,i}|^2}{a_{i,i}} & 0 & \cdots & 0\\
0&0&\cdots&0\\
\vdots&\vdots&\ddots&\vdots\\
0&0&\cdots&0
\end{bmatrix}.
\]
The final correction is rank one if nonzero, so the number of terms is exactly \(n\) or \(n-1\), matching the ordinary rank. This gives a direct factor-width-\(2\) certificate, an explicit minimal sparse rank-one decomposition, and immediate computation of \(\mathrm{fran}_k(A)\) from \(\mathrm{rank}(A)\) [2405.11556].

This behavior is exceptional. For arbitrary positive semidefinite matrices, factor-width rank can be much larger than ordinary rank; for example,
\[
\mathrm{fran}_k(A)\ge \frac{2}{k(k-1)}\,\mathrm{nnzu}(A),
\]
and for dense factor-width-\(2\) matrices with no zero entries and \(n\ge 3\),
\[
\mathrm{fran}_2(A)=\frac{n(n-1)}{2}.
\]
Arrowhead matrices are therefore a favorable class in which factor width \(2\) does not cause combinatorial blowup in sparse rank-one decomposition [2405.11556].

## 3. Numerical linear algebra: eigensolvers, determinants, and inverses

In numerical linear algebra, the Arrowhead-Matrix Method is closely associated with high-accuracy eigensolvers for real symmetric arrowhead matrices
\[
A=\begin{bmatrix} D & z\\ z^T & \alpha \end{bmatrix},
\qquad
D=\operatorname{diag}(d_1,\dots,d_{n-1}),
\]
and with structurally related diagonal-plus-rank-one matrices. The eigenvalues are zeros of the secular function
\[
f(\lambda)=\alpha-\lambda-z^T(D-\lambda I)^{-1}z
=\alpha-\lambda-\sum_{i=1}^{n-1}\frac{\zeta_i^2}{d_i-\lambda},
\]
and the corresponding eigenvectors are
\[
v_i=\frac{x_i}{\|x_i\|_2},\qquad
x_i=\begin{bmatrix}(D-\lambda_i I)^{-1}z\\ -1\end{bmatrix}.
\]
The shift-and-invert algorithm computes each eigenvalue separately by shifting at the nearest pole \(d_i\), forming \(A_i=A-d_iI\), and recovering
\[
\lambda=d_i+\frac{1}{\nu},
\]
where \(\nu\) is an extremal eigenvalue of \(A_i^{-1}\). Higher precision may be needed only for one scalar entry \(b\) of \(A_i^{-1}\), which is the numerically delicate quantity in the inverse representation [1302.7203].

This numerical philosophy extends to diagonal-plus-rank-one matrices
\[
A=D+\rho zz^T.
\]
After shifting to the nearest pole and inverting, the inverse of the shifted matrix becomes a permuted arrowhead matrix,
\[
A_i^{-1}= \begin{bmatrix} D_1^{-1} & w_1 & 0\\ w_1^T & b & w_2^T\\ 0 & w_2 & D_2^{-1} \end{bmatrix},
\]
so the desired eigenpair can be computed by the same arrowhead secular-equation machinery. The resulting algorithm computes each eigenvalue and all components of the corresponding eigenvector with high relative accuracy in \(O(n)\) work per eigenpair and extends naturally to the Hermitian case [1405.7537].

A separate structured calculus gives unified \(O(n)\) formulas for matvecs, determinants, and inverses of arrowhead and diagonal-plus-rank-one matrices over \(\mathbb R\), \(\mathbb C\), and \(\mathbb H\). For
\[
A=\operatorname{Arrow}(D,u,v,\alpha)=\begin{bmatrix} D & u\\ v^\star & \alpha \end{bmatrix},
\]
if all shaft entries \(d_i\neq 0\), then
\[
\det(A)=\Big(\prod_i d_i\Big)\big(\alpha-v^\star D^{-1}u\big),
\]
and
\[
A^{-1}=\operatorname{DPR1}(\Delta,x,y,\rho),
\]
with
\[
\Delta=P\begin{bmatrix} D^{-1} & 0\\ 0 & 0\end{bmatrix}P^T,\qquad
x=P\begin{bmatrix} D^{-1}u\\ -1\end{bmatrix},\qquad
y=P\begin{bmatrix} D^{-\star}v\\ -1\end{bmatrix},\qquad
\rho=(\alpha-v^\star D^{-1}u)^{-1}.
\]
All formulas require \(O(n)\) arithmetic operations and remain valid in the quaternionic setting provided multiplication order is preserved [2212.10966].

The same structured viewpoint also underlies a polynomial root-finding method for real polynomials with real distinct roots. By Fiedler’s theorem, such a polynomial can be realized as the characteristic polynomial of a real symmetric arrowhead matrix
\[
A=\begin{bmatrix} D & z\\ z^T & \alpha \end{bmatrix},
\]
with
\[
\alpha = -p-\sum_{j=1}^{n-1} d_j,\qquad
\zeta_j^2 = -\frac{u(d_j)}{v'(d_j)}.
\]
The roots are then computed as eigenvalues of \(A\) by a modified arrowhead eigensolver in \(O(n^2)\) total work, with double the working precision required only in the non-iterative construction phase and occasionally for one scalar in a shifted inverse [1509.06224].

## 4. Sparse approximation and communication-efficient multiplication

For implicit matrices accessible only through matrix-vector products, the arrowhead pattern appears as a distinguished sparse target family. The sparse approximation problem is to recover or approximate
\[
B^\star = S\circ A
\]
for a prescribed sparsity mask \(S\), using right and left matvec queries with \(A\) and \(A^T\). The best approximation supported on \(S\) is characterized by
\[
\min_{B:\,B\circ S=B}\|A-B\|_F = \|A-S\circ A\|_F.
\]
The complexity is governed by the degeneracy \(\mathrm{degen}(S)\), defined as the smallest \(k\) such that iteratively deleting all rows and columns with at most \(k\) ones empties \(S\). The main result gives a polynomial-time algorithm using
\[
O\!\left( \frac{\mathrm{degen}(S)}{\epsilon\delta}\log(m+n) \right)
\]
non-adaptive matvec queries and a matching lower bound \(\Omega(\mathrm{degen}(S))\) [2606.12179].

For the standard \(n\times n\) arrowhead mask,
\[
S_{ij}=\mathbf{1}[i=j \ \text{or}\ i=1\ \text{or}\ j=1],
\]
one has
\[
\mathrm{degen}(S)=2.
\]
Hence exact recovery of an arrowhead-supported matrix needs only
\[
2\,\mathrm{degen}(S)=4
\]
matvecs, while near-optimal approximation needs
\[
O\!\left(\frac{1}{\epsilon\delta}\log n\right)
\]
queries. This sharply improves generic bounds based on maximum row sparsity, total sparsity, or conflict-graph coloring, all of which are larger for the arrowhead pattern [2606.12179].

A different large-scale meaning of the method appears in sparse matrix–dense matrix multiplication. There, a sparse matrix is decomposed as
\[
A=\sum_{i=1}^{l} P_{\pi_i} B_i P_{\pi_i}^{\top},
\]
where each \(B_i\) has arrow-width at most \(b\). Multiplication becomes
\[
AX = \sum_{i=1}^{l} P_{\pi_i} \left( B_i (P_{\pi_i}^{\top} X)\right).
\]
When the decomposition is \(x\)-compacting and \(x\ge \Omega(\log^2 p)\) with \(p=\Theta(n/b)\), one iteration of distributed sparse matrix–dense matrix multiplication has communication cost
\[
O(\alpha \log^2 p + \beta \frac{nk}{p}),
\]
improving bandwidth by a factor \(\Theta(\sqrt p)\) over fully replicated 1.5D methods under the stated assumptions [2402.19364].

## 5. Large-scale structured solvers and high-performance computing

In direct sparse factorizations, arrowhead structure is exploited by sTiles, a tiled Cholesky framework for symmetric positive definite sparse matrices whose nonzeros lie primarily in the last block row, the last block column, and along the block diagonal. The method combines heuristic reorderings, symbolic factorization, and a left-looking tiled Cholesky using POTRF, TRSM, SYRK, and GEMM kernels. The left-looking formulation enables tree reductions for update accumulations and is specifically motivated by the thin task graph of arrowhead matrices, where parallelism is limited by the structure. The framework reported speedups of up to \(8.41\mathrm{X}\), \(9.34\mathrm{X}\), \(5.07\mathrm{X}\), and \(11.08\mathrm{X}\) compared to CHOLMOD, SymPACK, MUMPS, and PARDISO, respectively, and a \(5\mathrm{X}\) speedup over a 32-core AMD EPYC CPU on an NVIDIA A100 GPU [2501.02483].

For block-tridiagonal arrowhead matrices, selected inversion is developed in Serinv. A block-tridiagonal arrowhead matrix consists of a block-tridiagonal core plus an arrow tip block \(A_{n,n}\) and rectangular couplings \(A_{n,i}\), \(A_{i,n}=A_{n,i}^\dagger\). The selected inversion computes only the inverse blocks on the original sparsity pattern:
\[
\{X_{i,i}\},\quad \{X_{i+1,i}\},\quad \{X_{n,i}\},\quad X_{n,n}.
\]
After block Cholesky factorization \(A=LL^\dagger\), the backward recurrences are specialized to the arrowhead couplings; for example,
\[
U_{i+1,i} = -X_{i+1,i+1}L_{i+1,i} - X_{n,i+1}^{\dagger}L_{n,i},
\qquad
X_{i+1,i} = \operatorname{TRSM}(L_{i,i}, U_{i+1,i}),
\]
and
\[
U_{n,i} = -X_{n,i+1}L_{i+1,i} - X_{n,n}L_{n,i},
\qquad
X_{n,i} = \operatorname{TRSM}(L_{i,i}, U_{n,i}).
\]
The distributed algorithm uses a partition-plus-reduced-system strategy; its sequential complexity is \(O(nb^3)\), the reduced system has \(n_r=2P-1\) main blocks, and the library reports \(32.3\%\) strong-scaling efficiency, \(47.2\%\) weak-scaling efficiency, and up to two orders of magnitude speedup over PARDISO and MUMPS on 16 GPUs [2503.17528].

A related but optimization-centered HPC use appears in PIPS-IPM++. For doubly-bordered block-diagonal linear programs, the interior-point Newton system can be written as
\[
\begin{bmatrix} K_1 & & & L_1\\
& \ddots & & \vdots\\
& & K_N & L_N\\
L_1^T & \dots & L_N^T & K_0
\end{bmatrix}
\begin{bmatrix} z_1\\ \vdots\\ z_N\\ z_0 \end{bmatrix}
=
\begin{bmatrix} b_1\\ \vdots\\ b_N\\ b_0 \end{bmatrix},
\]
with Schur complement
\[
S = K_0 - \sum_{i=1}^N L_i^T K_i^{-1} L_i.
\]
The new contribution is a hierarchical Schur complement decomposition that exploits local border constraints, recursively partitions the inner systems, and isolates a dense layer associated with global links. This enables distributed solution of very large LPs, including instances with more than \(10^9\) nonzeros in the constraint matrix [2412.07731]. A complementary distributed presolve framework for arrowhead linear programs keeps the same block-arrowhead organization, classifies reductions as local or global, and preserves the structure required by the Schur-complement interior-point solve; on the reported benchmark set it outperformed PaPILO by a factor of \(18\) and Gurobi’s presolve by a factor of \(6\) in shifted geometric mean runtime on a single machine, and by a factor of \(13\) against Gurobi in a distributed environment [2603.03498].

## 6. Broader theoretical extensions and domain-specific uses

Beyond decomposition and HPC, arrowhead structure supports exact analysis in several specialized settings. In random-matrix physics, the disordered single-excitation Tavis–Cummings Hamiltonian becomes an \((N+1)\times(N+1)\) arrowhead matrix with diagonal disorder and uniform cavity coupling. Its eigenvalues satisfy the secular equation
\[
\varepsilon_a=\frac1N\sum_{j=1}^N \frac{g^2}{\varepsilon_a-\omega_j},
\]
with interlacing
\[
\varepsilon_1\le \omega_1\le \varepsilon_2\le \omega_2\le \cdots \le \varepsilon_N\le \omega_N\le \varepsilon_{N+1},
\]
and eigenvectors
\[
\psi_a=N_a
\begin{pmatrix}
\dfrac{g/\sqrt N}{\varepsilon_a-\omega_1}\\
\vdots\\
\dfrac{g/\sqrt N}{\varepsilon_a-\omega_N}\\
1
\end{pmatrix}.
\]
This exact arrowhead solution yields asymptotically exact formulas for polaritons, dark states, multifractal participation ratios, and cavity-protected transport [2105.08444].

In model reduction, arrowhead realizations provide direct access to the sign parameters governing exactness of the balanced truncation error bound. For an arrowhead realization
\[
A=
\begin{bmatrix}
d_1 & \alpha_2 & \cdots & \alpha_n\\
\beta_2 & d_2 & & \\
\vdots & & \ddots & \\
\beta_n & & & d_n
\end{bmatrix},
\qquad
b=\gamma e_1,\qquad c=e_1^\top,
\]
the sign parameters satisfy, up to permutation,
\[
s_{\pi_1}=\operatorname{sign}(\gamma),\qquad
s_{\pi_i}=\operatorname{sign}(\gamma\alpha_i\beta_i),\quad i=2,\dots,n.
\]
If all signs in the truncated block are identical, then the balanced truncation and singular perturbation approximation error bounds are attained with equality:
\[
\|G-G_r\|_{\mathcal H_\infty}=2(\sigma_{k+1}+\cdots+\sigma_q).
\]
Thus the arrowhead structure converts an abstract balanced-canonical invariant into an explicit entrywise sign test [2011.07170].

In matrix completion, a width-one arrowhead specification pattern supports new sufficient conditions for set-completely positive completion. The main theorem concerns partial matrices
\[
M_{G_{n+1,1}^S}= \begin{bmatrix}
1 & x^\top & y_1 & \dots & y_S\\
x & X & z_1 & \dots & z_S\\
y_1 & z_1^\top & Y_1 & * & *\\
\vdots & \vdots & * & \ddots & *\\
y_S & z_S^\top & * & * & Y_S
\end{bmatrix},
\]
and shows that under boundedness, extreme-point, and linear consistency conditions, every such partial matrix in
\[
PCP_{G_{n+1,1}^S}(\mathbb R_+\times \mathcal K\times \mathbb R_+^S)
\]
is completable to
\[
CPP(\mathbb R_+\times \mathcal K\times \mathbb R_+^S).
\]
The method derives completion from exactness of a sparse conic relaxation rather than from graph-theoretic completion theory alone [2509.19348].

A different algebraic use appears in exact rank-one matrix completion via sum-of-squares relaxations. There the coefficient matrix \(Q\) in the objective is chosen to be arrowhead,
\[
Q=
\begin{bmatrix}
q_{11} & q_{12} & \cdots & q_{1N}\\
q_{12} & 1 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots\\
q_{1N} & 0 & \cdots & 1
\end{bmatrix},
\]
so that the associated Gram matrix has rank \(N-1\). The one-dimensional nullspace then isolates the true monomial vector and yields exact completion, including cases where connectivity is hidden in linear combinations of constraints [2311.14882].

Finally, arrowhead matrices also admit geometric analysis through numerical range. For
\[
A=\begin{bmatrix}
a_1&0&\ldots&0&b_1\\
0&a_2&\ldots&0&b_2\\
\vdots&\vdots&\ddots&\vdots&\vdots\\
0&0&\ldots&a_{n-1}&b_{n-1}\\
c_1&c_2&\ldots&c_{n-1}&a_n
\end{bmatrix},
\]
the Gau–Wu number can be computed for broad classes by rotating so that the real part becomes diagonal and the imaginary part remains Hermitian arrowhead. If
\[
\arg b_j+\arg c_j=2\theta+\pi
\]
for all \(j\) with \(b_j,c_j\neq 0\), then \(k(A)\) equals the sum of multiplicities of the largest and smallest eigenvalues of
\[
\operatorname{Im}(e^{-i\theta}A).
\]
The same structural machinery yields a characterization of dichotomous arrowhead matrices and a criterion for their unitary irreducibility [2105.11860].

The cumulative picture is that arrowhead structure is valuable precisely because it collapses dense global behavior into a small border, a rank-one perturbation, or a star-graph interaction. This suggests a unifying interpretation: whenever the nonzero pattern can be localized to one hub row and column, or to a small arrowhead generalization, exact formulas, sparse decompositions, or recursive eliminations often become available in closed form or with sharply reduced complexity.

Source: https://www.emergentmind.com/topics/arrowhead-matrix-method