Arrow Decomposition (AD): A Structured Approach

• Arrow Decomposition (AD) is a family of methods that exploit arrow-shaped coupling patterns to partition large structured objects into local blocks with a small shared interface.
• It is applied across fields such as sparse matrix multiplication, KKT systems, semidefinite and polynomial optimization, and even in arrow-based homological algebra.
• AD improves computational efficiency by reducing communication costs, enabling scalable parallel processing while maintaining numerical stability and practical performance.

Searching arXiv for the cited papers and related terminology to ground the article. Arrow Decomposition (AD) denotes a family of structure-exploiting decompositions organized around arrow, arrowhead, or generalized arrow sparsity patterns. In the arXiv literature considered here, the term is not monosemous: it refers to decomposition of sparse matrices into permuted arrow matrices for communication-efficient sparse matrix–dense matrix multiplication (Gianinazzi et al., 2024), Schur-complement decomposition of arrowhead KKT systems in massively parallel interior-point methods (Kempke et al., 2024), decomposition of arrow-type positive semidefinite and polynomial matrix inequalities into smaller coupled LMIs or PMIs (Kocvara, 2019, Handa et al., 2 Sep 2025), and, in a broader arrow-based interpretation, reformulation of homological data in arrow categories (Fritz, 2022). This suggests a recurrent structural theme: a large coupled object is split into local components plus a comparatively small shared “head,” border, or morphism-level coupling.

1. Terminological scope and canonical forms

In these sources, AD is best understood as a family resemblance rather than a single invariant. The common feature is the exploitation of an arrow-shaped coupling pattern; the algebraic mechanisms, however, differ sharply across HPC, optimization, semidefinite programming, and category theory (Gianinazzi et al., 2024, Kempke et al., 2024, Kocvara, 2019, Handa et al., 2 Sep 2025).

Setting	Canonical object	Decomposition objective
Sparse SpMM	$A=\sum_{i=1}^{l}P_{\pi_i}B_iP_{\pi_i}^{\top}$	Reduce communication in distributed $Y=AX$
Arrowhead KKT systems	$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$	Decouple local solves and reduce border coupling
Arrow-type PSD/PMI constraints	$\begin{bmatrix}A&B\B^T&C\end{bmatrix}$ or $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$	Replace one large matrix inequality by smaller coupled ones
Arrow categories	$\mathrm{Arr}(C)=C^{[1]}$	Encode homology as arrows rather than quotient objects

Two terminological distinctions are essential. First, an arrow matrix in sparse linear algebra is a band-plus-head object indexed by a permutation; it is not the same object as an arrowhead or doubly-bordered block-diagonal KKT matrix in interior-point methods. Second, in semidefinite optimization, AD is not chordal decomposition: chordal decomposition proceeds through clique structure, whereas AD uses a prescribed arrow structure and introduces coupling variables only on overlaps or in the bottom block (Kocvara, 2019, Handa et al., 2 Sep 2025).

2. Sparse-matrix AD for communication-efficient SpMM

For sparse matrix multiplication, AD is defined for a square matrix $A\in\mathbb{R}^{n\times n}$ through the notion of arrow width $b$. A matrix has arrow-width $b$ if all nonzeros are concentrated in the first $b$ rows, the first $Y=AX$0 columns, and a band of width $Y=AX$1 around the diagonal; formally,

$Y=AX$2

An Arrow Matrix Decomposition expresses a sparse matrix as

$Y=AX$3

where each $Y=AX$4 has arrow-width at most $Y=AX$5, $Y=AX$6 is a permutation matrix, and $Y=AX$7 is the order of the decomposition. Algorithmically, this yields

$Y=AX$8

so SpMM becomes a sequence of dense-feature permutations, arrow-matrix products, and inverse permutations with accumulation (Gianinazzi et al., 2024).

The construction is graph-theoretic. The sparse matrix is viewed as the adjacency matrix of a graph $Y=AX$9, and the central tool is a linear arrangement $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$0 with cost

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$1

The decomposition algorithm, LA-Decompose, iteratively prunes the $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$2 highest-degree vertices into the arrow head, computes a linear arrangement on the remaining induced subgraph, extracts the first $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$3 rows and columns together with a symmetric $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$4-wide diagonal band, and subtracts the resulting arrow matrix. Compactness is quantified by the notion of an $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$5-compacting decomposition: $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$6 which implies $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$7. The compactness guarantee is tied to MLA cost through

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$8

The paper also develops Separator-LA(G) and derives MLA bounds for planar graphs, $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$9-minor-free graphs, bounded-treewidth graphs, series–parallel graphs, and trees. For trees, the smallest-first order yields an $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$0-compacting $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$1-arrow decomposition computable in $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$2 work, showing that a tree can have large bandwidth yet still admit a decomposition into $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$3 constant-bandwidth matrices. For power-law graphs, explicit pruning of high-degree vertices into the arrow head is analyzed under a truncated Zipf model, making the effective arrow width independent of the original maximum degree with high probability (Gianinazzi et al., 2024).

The distributed multiplication algorithm exploits the tile structure of a single arrow matrix. With $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$4 processors, each rank stores one top-row tile, one left-column tile, one diagonal tile, and one slice of the dense feature matrix. Arrow Matrix Multiply requires one broadcast of the head block $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$5, a reduction for the head output tile $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$6, and local products for the remaining diagonal and column tiles. For a single arrow matrix $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$7 and $\begin{bmatrix}A&B\B^T&C\end{bmatrix}$8, the communication cost is

$\begin{bmatrix}A&B\B^T&C\end{bmatrix}$9

For a full $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$0-compacting decomposition, permutations between successive $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$1 are handled by sorting-network-based sparse scatters, yielding the bound

$\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$2

for computing $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$3 under $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$4 and $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$5. The storage cost is

$\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$6

matching 2D methods in feature storage while improving bandwidth over replicated 1.5D algorithms (Gianinazzi et al., 2024).

Empirically, the decompositions are of small order: typically $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$7–$\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$8, often $\begin{bmatrix}A&B\B^T&\Gamma\end{bmatrix}$9–$\mathrm{Arr}(C)=C^{[1]}$0, even for matrices with tens to hundreds of millions of rows. On SuiteSparse-derived matrices with $\mathrm{Arr}(C)=C^{[1]}$1M–$\mathrm{Arr}(C)=C^{[1]}$2M rows and up to $\mathrm{Arr}(C)=C^{[1]}$3B nonzeros, arrow widths from $\mathrm{Arr}(C)=C^{[1]}$4 to $\mathrm{Arr}(C)=C^{[1]}$5 yield decompositions of order at most $\mathrm{Arr}(C)=C^{[1]}$6; the second matrix has only $\mathrm{Arr}(C)=C^{[1]}$7–$\mathrm{Arr}(C)=C^{[1]}$8 of rows nonzero; and the method produces up to $\mathrm{Arr}(C)=C^{[1]}$9–$A\in\mathbb{R}^{n\times n}$0 fewer nonzero tiles than a comparable 1.5D block decomposition. On $A\in\mathbb{R}^{n\times n}$1 GPUs, communication volume is reduced by $A\in\mathbb{R}^{n\times n}$2–$A\in\mathbb{R}^{n\times n}$3 versus 1.5D; on up to $A\in\mathbb{R}^{n\times n}$4 nodes with matrices exceeding $A\in\mathbb{R}^{n\times n}$5M rows, speedups over the 1.5D baseline are $A\in\mathbb{R}^{n\times n}$6–$A\in\mathbb{R}^{n\times n}$7, and over 1D hypergraph partitioning $A\in\mathbb{R}^{n\times n}$8–$A\in\mathbb{R}^{n\times n}$9. Weak scaling on the MAWI graph series shows only $b$0–$b$1 runtime growth as the graph increases from $b$2M to $b$3M rows (Gianinazzi et al., 2024).

3. AD as Schur-complement decomposition of arrowhead KKT systems

In large-scale linear programming, especially in PIPS-IPM++, AD denotes decomposition of an arrowhead or doubly-bordered block-diagonal linear system of the form

$b$4

Here the $b$5 are local blocks, $b$6 is the border block, and the $b$7 encode coupling. In the LPs considered, the primal variables are partitioned as $b$8, where $b$9 are global or linking variables and $b$0 are local variables; after forming the primal–dual augmented KKT system and eliminating slack directions, the same arrowhead structure is preserved across interior-point iterations (Kempke et al., 2024).

The basic decomposition is a Schur complement reduction. With $b$1 and $b$2, the border Schur complement is

$b$3

Factorization proceeds by factoring each $b$4, assembling the local Schur contributions $b$5, forming $b$6, and factorizing $b$7. The solve phase computes

$b$8

followed by back-substitution

$b$9

This is the flat form of AD for arrowhead systems: decoupled block solves plus a reduced border problem (Kempke et al., 2024).

The distinctive contribution of the massively parallel interior-point work is a hierarchical Schur complement approach (HSCA). The key structural observation is that many linking constraints are local 2-links, involving $b$0 but no other blocks, while relatively few are genuinely global. The local and global parts are split, the dense global layer is isolated and factorized as a small dense Schur complement, and the remaining inner system retains a banded local-linking structure. The block indices are then grouped recursively,

$b$1

to form new arrowhead systems

$b$2

whose Schur complements have banded form. The same recipe can then be applied recursively to the $b$3, producing a multi-level AD hierarchy (Kempke et al., 2024).

This hierarchy is aligned with MPI parallelization. Leaf systems $b$4 are assigned to groups of MPI processes; local factorizations and backsolves are embarrassingly parallel across groups; intermediate Schur complements are formed by reductions on level-specific communicators; and factorization of intermediate border systems is handled by one process or a small group, often with OpenMP. The dense layer corresponding to $b$5 is treated separately, while the inner banded Schur systems are typically sparse enough for MA57. The paper reports that $b$6–$b$7 hierarchy levels are useful in practice (Kempke et al., 2024).

The complexity discussion is centered on nonzero growth in flat versus hierarchical Schur complements. For the flat local-linking Schur system,

$b$8

where $b$9 is the number of local 2-link constraints, $Y=AX$00 the number of global linking constraints, and $Y=AX$01 the dimension of linking variables. Hierarchically, the corresponding inner Schur systems involve much smaller per-group link sizes. The implementation preserves the augmented-system formulation, uses robust indefinite solvers such as MA57 and PARDISO, and separates dense columns into a dense layer, which the paper notes is beneficial for sparsity and numerical stability (Kempke et al., 2024).

The empirical regime is large-scale energy and unit-commitment optimization. The method is demonstrated on instances with more than $Y=AX$02 nonzeros in the constraint matrix and more than $Y=AX$03 linking constraints. For the instance “MISO_DISP_488,” with about $Y=AX$04M nonzeros, $Y=AX$05M columns, $Y=AX$06M rows, $Y=AX$07 blocks, and about $Y=AX$08k linking constraints, the original flat AD in PIPS-IPM++ cannot solve the problem because of Schur complement memory blowup, whereas HSCA does, showing near-linear speedup up to $Y=AX$09 MPI ranks with two OpenMP threads per process. On mid- to large-scale SIMPLE and REMix instances, HSCA often outperforms commercial solvers by up to an order of magnitude in time and scales efficiently to several thousand cores (Kempke et al., 2024).

4. AD for arrow-type PSD matrices and topology optimization

In semidefinite optimization, AD refers to decomposition of an arrow-type positive semidefinite matrix

$Y=AX$10

where $Y=AX$11 and $Y=AX$12, with each $Y=AX$13 and $Y=AX$14 supported on a local index set $Y=AX$15. The overlaps are $Y=AX$16, and the arrow indices are appended via

$Y=AX$17

The standard chordal decomposition for sparse PSD constraints introduces dense PSD overlap variables $Y=AX$18. AD shows that for arrow-type matrices satisfying suitable assumptions, the linking data can instead be represented through matrices $Y=AX$19, supported only on overlaps, together with small $Y=AX$20; equivalently,

$Y=AX$21

with each $Y=AX$22 containing only local $Y=AX$23, the local bottom block $Y=AX$24, and overlap couplings of the form

$Y=AX$25

Crucially, this theorem requires Assumptions 1–3 on the index sets but not the global chordality assumption needed by the chordal variant (Kocvara, 2019).

The scalar-head case $Y=AX$26 is the central application. Then the $Y=AX$27 are vectors supported on the interface indices $Y=AX$28, and the decomposition replaces full PSD overlap matrices by vector variables $Y=AX$29 and scalars $Y=AX$30. The paper emphasizes that for $Y=AX$31 the necessary linking information is rank one in the arrow direction; hence, in the topology-optimization setting, matrix variables disappear in favor of interface vectors (Kocvara, 2019).

The model problem is minimum-compliance topology optimization. After finite-element discretization,

$Y=AX$32

and the compliance bound is captured by the PSD condition

$Y=AX$33

Partitioning the domain into subdomains $Y=AX$34 with local stiffness contributions $Y=AX$35 and local load pieces $Y=AX$36 yields an arrow-type matrix with $Y=AX$37, $Y=AX$38, and $Y=AX$39. The decomposed formulation enforces local PSD constraints

$Y=AX$40

each involving local stiffness blocks, local load contributions, interface vectors $Y=AX$41, and local scalars $Y=AX$42, with

$Y=AX$43

This formulation has a domain-decomposition interpretation: the interface vectors are discrete Steklov–Poincaré variables, i.e., fictitious interface loads or tractions that allow local subdomain problems to glue together consistently (Kocvara, 2019).

The computational effect is a drastic reduction in semidefinite-cone size. In the original formulation, one large PSD block of dimension $Y=AX$44 dominates the interior-point Schur complement. In the chordal formulation, this is replaced by smaller PSD blocks but many dense PSD overlap variables. In the arrow formulation, the same local PSD blocks are retained while the overlap matrices are replaced by vectors. For mesh sequences from $Y=AX$45 to $Y=AX$46, the reported CPU-time fits are approximately

$Y=AX$47

Representative speedups are substantial: for $Y=AX$48, CPU time drops from $Y=AX$49 s without decomposition to $Y=AX$50 s with arrow decomposition using $Y=AX$51 subdomains; for $Y=AX$52, it drops from $Y=AX$53 s to $Y=AX$54 s with $Y=AX$55 subdomains. The paper also reports better numerical conditioning than in heavily decomposed chordal formulations (Kocvara, 2019).

5. AD for PMIs and moment–SOS hierarchies

The 2025 extension revisits the linear-arrow theorem and weakens the assumptions to blockwise positive semidefiniteness: $Y=AX$56 Using the generalized Schur complement with the Moore–Penrose pseudoinverse, the paper proves that for

$Y=AX$57

the condition $Y=AX$58 is equivalent to the existence of overlap matrices $Y=AX$59 and bottom-block pieces $Y=AX$60 such that

$Y=AX$61

with

$Y=AX$62

Because rank-deficient $Y=AX$63 yield decomposed LMIs without interior points, the paper introduces a projection step: with $Y=AX$64 spanning $Y=AX$65, each local LMI is reduced to a smaller projected LMI together with the span condition $Y=AX$66 (Handa et al., 2 Sep 2025).

The same paper extends AD from LMIs to polynomial matrix inequalities. For a POP

$Y=AX$67

with

$Y=AX$68

the polynomial assumption is

$Y=AX$69

Two couplings with moment–SOS hierarchies are studied. The prior approach applies AD first and then builds the mSOS relaxation in the enlarged variable space $Y=AX$70. The paper proves that this is theoretically exact but practically unattractive, because the monomial basis now scales as

$Y=AX$71

which is much larger than the standard $Y=AX$72 when $Y=AX$73 is appreciable (Handa et al., 2 Sep 2025).

The main construction is posterior AD, which preserves the original polynomial variable space and applies AD directly to the localizing matrix. If

$Y=AX$74

then $Y=AX$75 has the same arrow structure in lifted dimension, and the selection matrices become

$Y=AX$76

The AD version of Putinar’s theorem states that a moment sequence $Y=AX$77 has a representing measure supported on $Y=AX$78 if and only if $Y=AX$79 for all $Y=AX$80 and there exist $Y=AX$81 such that the decomposed localizing LMIs hold for all $Y=AX$82. Finite truncations define the posterior hierarchy $Y=AX$83, for which

$Y=AX$84

Thus, at each level the posterior hierarchy is at least as tight as the standard hierarchy and converges to the global optimum, while avoiding monomial-basis blow-up (Handa et al., 2 Sep 2025).

The projection idea also lifts to the moment level. When the ranges of the $Y=AX$85 are constant, one defines $Y=AX$86 and replaces the lifted LMIs by projected ones involving $Y=AX$87, together with lifted span conditions on $Y=AX$88. This removes rank-deficient directions and reduces the number of additional variables (Handa et al., 2 Sep 2025).

The applications are structural optimization problems with natural arrow structure in mixed finite-element formulations. For beam and frame examples, the auxiliary $Y=AX$89-variables acquire a direct mechanical interpretation as interface forces and moments, while the projection constraints become static-equilibrium equations against rigid body modes. In the double-hinged beam under self-weight, combining AD with a non-mixed-term basis yields reported speedups of up to about $Y=AX$90 at relaxation order $Y=AX$91 relative to standard mSOS. In a $Y=AX$92-element modular frame with nine design variables, the hierarchy converges at $Y=AX$93, and the combination of AD with NMT reduces runtime from about $Y=AX$94 s for full mSOS to about $Y=AX$95 s (Handa et al., 2 Sep 2025).

6. Arrow categories and arrow-based homological decomposition

In category theory, the term Arrow Decomposition is not explicit in the cited paper, but an arrow-based decomposition of homological data is a plausible interpretation of the construction. For a category $Y=AX$96 with an ideal $Y=AX$97 of null morphisms and kernels and cokernels relative to $Y=AX$98, the arrow category

$Y=AX$99

has objects $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$00 and morphisms given by commutative squares

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$01

A square is null when its diagonal

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$02

lies in $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$03. The paper proves that if $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$04 has kernels and cokernels with respect to $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$05, then $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$06 is a homological category in Grandis’s sense (Fritz, 2022).

The resulting homology objects are themselves arrows. For composable morphisms

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$07

embedded as identity arrows in $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$08, the homology object is the arrow

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$09

More generally, for composable squares, the homology object is an arrow from $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$10 to $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$11. Rather than identifying homology with a quotient object $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$12, the construction records the cycles object, the chains-modulo-boundaries object, and the canonical morphism between them. This is the sense in which the paper decomposes homological information into arrow data (Fritz, 2022).

The same framework is then applied to $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$13-curved homological algebra. In the category $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$14 of seminormed spaces and non-expansive maps, one fixes the ideal $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$15 of morphisms of norm at most $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$16. An $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$17-curved chain complex satisfies

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$18

The paper constructs explicit $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$19-kernels and $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$20-cokernels, derives the exactness criterion

$\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$21

and notes that $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$22 itself is not homological. Passing to $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$23 restores a homological setting, so $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$24-curved homology can again be represented as arrows from cycle-like objects to cokernel-like objects (Fritz, 2022).

7. Cross-domain structure, distinctions, and limitations

Across these literatures, AD consistently separates a large structured object into local pieces plus a small coupling layer, but the mathematical implementation differs. In sparse SpMM, AD is a sum of permuted arrow matrices with compactness controlled by MLA or separator quality (Gianinazzi et al., 2024). In arrowhead LP solvers, AD is a Schur-complement factorization of a single doubly-bordered system, often recursively (Kempke et al., 2024). In semidefinite and polynomial optimization, AD is an equivalent reformulation of matrix inequalities into smaller coupled LMIs or PMIs with overlap variables $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$25 and bottom-block pieces $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$26 (Kocvara, 2019, Handa et al., 2 Sep 2025). In the arrow-category setting, a plausible implication is that homological information is decomposed into kernels, cokernels, and the arrow connecting them, rather than compressed into a quotient object (Fritz, 2022).

Several common misconceptions are therefore inaccurate. AD is not synonymous with bandwidth reduction: in sparse SpMM, the point is precisely to circumvent single-permutation bandwidth limits by representing a matrix as a sum of several narrow arrow matrices (Gianinazzi et al., 2024). AD is not identical to chordal decomposition: the semidefinite-optimization papers stress that chordal methods introduce dense PSD overlap matrices on cliques, whereas AD exploits a specific arrow structure and, in the scalar-head case, replaces overlap matrices by vectors (Kocvara, 2019, Handa et al., 2 Sep 2025). AD is also not a single complexity claim: the communication-optimality statement $\begin{bmatrix}K_1&&&L_1\&K_2&&L_2\&&\ddots&\vdots\L_1^T&L_2^T&\cdots&K_0\end{bmatrix}$27, the near-linear speedup of hierarchical Schur complement methods, and the near-linear empirical scaling of arrow-decomposed SDOs refer to different models and should not be conflated (Gianinazzi et al., 2024, Kempke et al., 2024, Kocvara, 2019).

The limitations are equally domain-specific. Sparse-matrix AD depends on good approximate linear arrangements or separator heuristics, and its theory is centered on square sparse matrices and tall-skinny dense right-hand sides (Gianinazzi et al., 2024). Hierarchical arrowhead decomposition in LPs assumes a doubly-bordered block-diagonal structure with substantial 2-link locality; when the flat Schur complement is already moderate and blocks are large, flat PARDISO-based AD may be preferable (Kempke et al., 2024). PSD and PMI variants require an authentic arrow decomposition with blockwise PSD submatrices, and when blocks are rank-deficient they need projection to restore strict feasibility and eliminate redundant variables (Handa et al., 2 Sep 2025). The arrow-category construction, finally, depends on the existence of kernels and cokernels relative to a chosen null ideal and should be read as an arrow-based reformulation of homology rather than as the same decomposition used in matrix computation (Fritz, 2022).

Taken together, these sources show that “Arrow Decomposition” is a structurally descriptive term rather than a single canonical algorithm. Its precise meaning is determined by the ambient object: sparse matrices, KKT systems, PSD or polynomial matrix inequalities, or morphisms in a category. The unifying idea is narrow and technical: exploit an arrow-shaped coupling pattern so that the global object can be handled through smaller local structures and a controlled interface.