Spatial Factorization of Rank

• Spatial factorization of rank is a framework that decomposes a matrix into structured components reflecting pivot and dependency relations to explicitly define its column, row, and nullspace bases.
• The methodology leverages classical elimination and pivot identification to derive the canonical A = CR decomposition, ensuring minimality and numerical stability.
• This approach extends to structured positive semidefinite matrices and matrix polynomials, offering insights into localized support, factor width, and the uniqueness of spatial representations.

The spatial factorization condition of rank refers to algebraic and geometric structures that relate the rank of a matrix or matrix polynomial to canonical factorizations—typically expressed as a product (or sum) of structured factors whose spatial arrangement or support embodies the dependence relations among the underlying data. This theme appears across Gaussian elimination, block-decompositions, and positive semidefinite matrix rank analogs. Such factorizations provide not only compressions of linear data but also explicit bases for fundamental subspaces (row, column, and nullspace) and, in structured cases, control over uniqueness and the spatial locality of factor supports.

1. Classical Spatial Factorization via Pivot Structures

For $A\in\mathbb{R}^{m\times n}$ of rank $r$, spatial factorization begins with column operations to isolated pivot columns. By performing Gaussian or Gauss–Jordan elimination combined with column permutations $P$, there exists a decomposition

$$UAP = Z = \begin{bmatrix} I_r & F & 0 \end{bmatrix}$$

where $U\in GL(m)$, $I_r$ is an $r\times r$ identity, $F\in\mathbb{R}^{r\times (n-r)}$ encodes dependencies, and $0$ fills out to $m$ rows. The index set $J$ identifies $r$ independent (pivot) columns; its complement $K$ the non-pivot columns. Define $C = A(:,J)$ and $D = A(:,K)$, with $D = CF$. Thus, after permutation,

$AP = [C\,|\,CF] = C[I_r\,F] =: CR$

with $R = [I_r\,F]$. This is the canonical “spatial factorization of rank” for $A$ (Strang, 2023).

2. Formal Spatial Factorization Theorem and Geometric Interpretation

The factorization yields the following equivalence: $A\in\mathbb{R}^{m\times n}$ has rank $r$ if and only if there exist $C\in\mathbb{R}^{m\times r}$ of column rank $r$ and $R\in\mathbb{R}^{r\times n}$ of row rank $r$ such that $A=CR$. Explicitly:

• $C$ forms an ordered basis of $\mathrm{Col}(A)$;
• $R$ encodes the coordinate map from $\mathbb{R}^r$ to the representation of $A$’s columns (as linear combinations of the pivots).
• Minimality is enforced: both $C$ and $R$ use $r$ as the minimal size needed for full spanning.

Geometrically, $C$ is an isomorphism from $\mathbb{R}^r$ onto $\mathrm{Col}(A)$, while $R$ is a surjective map whose kernel is $\mathrm{Null}(A)$ and whose rows span $\mathrm{Row}(A)$. The factorization $A = CR$ reflects a two-step linear process: project onto the $r$-dimensional row space, then embed into $\mathbb{R}^m$ (Strang, 2023).

3. Structure of Subspaces via the Factor Matrix $F$

The matrix $F$ encodes detailed spatial relations:

• Row space: $R = [I_r\,F]$ has $r$ nonzero rows spanning $\mathrm{Row}(A)$; these rows correspond to the nonzero rows of $Z=UAP$ and form an explicit basis.
• Nullspace: The general solution to $Ax=0$ is equivalent to $Rx=0$ (since $C$ has full column rank). Writing $x = (x_J, x_K)$ ($|J|=r, |K|=n-r$), the coordinate system satisfies $x_J = -F x_K$. All $x_K \in \mathbb{R}^{n-r}$ parametrize the nullspace, generating the $n-r$ nullspace basis vectors as the columns of $$N = \begin{bmatrix} -F \ I_{n-r} \end{bmatrix}$$.

This reveals how $F$ acts as a “coordinate map” expressing each dependent (nonpivot) column as a spatial combination of the pivots and simultaneously specifying bases for both $\mathrm{Row}(A)$ and $\mathrm{Null}(A)$ (Strang, 2023).

4. Uniqueness, Minimality, and Canonical Choices

The spatial factorization condition of rank possesses uniqueness properties contingent on the selection of pivot indices. When one adopts the canonical reduced-row-echelon form of $A$, $J$ and hence $F$ are unique; otherwise, for alternate bases or left-inverses of $C$, the factorization $A=CR$ is generally non-unique.

Minimality is inherent, as $C$ (and $R$) uses as few columns (resp. rows) as necessary to span $\mathrm{Col}(A)$ (resp. $\mathrm{Row}(A)$). This minimality is critical in compression, subspace identification, and numerical stability. A plausible implication is that redundancy in basis selection increases the degrees of freedom available for matrix descriptions, but only the canonical structure induced by elimination is uniquely tied to the spatial arrangement of independence and dependence (Strang, 2023).

5. Spatial Factorization for Structured Positive Semidefinite Matrices

For $A\in M_n^+(\mathbb{F})$ (Hermitian PSD matrices), the spatial factorization extends through the concept of factor width: $A$ has factor width at most $k$ if it can be written as

$A = \sum_{j=1}^r v_j v_j^*$

with each $v_j$ supported on at most $k$ entries. Equivalently, $A$ is a conic sum of PSD $k\times k$ principal submatrices (“spatial blocks”).

The factor width-$k$ rank $\operatorname{fran}_k(A)$ is the minimal $r$ in such a decomposition. For many matrix classes (banded, arrowhead), $\operatorname{fran}_k(A)$ coincides with usual rank, but for generic patterns it differs—tight connections are established to graph clique covering numbers and sparse covering designs (Johnston et al., 2024).

Property/Class	Factor Width-$k$	Rank Equivalence
Banded, bandwidth $\leq k$	Yes	$\operatorname{fran}_k(A) = \operatorname{rank}(A)$
Arrowhead	$k=2$	$\operatorname{fran}_k(A) = \operatorname{rank}(A)$ for $k\geq 2$
General pattern	Not necessarily	Bound by clique covering, may differ

The spatial factorization condition here demands not just low rank, but that the rank-one or low-rank summands possess localized spatial support—crucial in applications leveraging locality or exploiting sparsity for PSD matrices (Johnston et al., 2024).

6. Rank-One Factorization of Matrix Polynomials and Spatial Uniqueness

For auto-correlation matrix polynomials $\Gamma(z)$ of signature $\operatorname{rank}(\Gamma(z))=1$, the spatial factorization condition involves the existence (and possibly uniqueness) of a polynomial vector $X(z)$ such that

$\Gamma(z) = X(z)\,\widetilde{X(z)}^T$

The uniqueness of this spatial factorization is precisely characterized by the absence of roots off the unit circle in the greatest common divisor $H(z) = \gcd_{i,j} \Gamma_{ij}(z)$:

• If $H(z)$ has no off-unit-circle roots, the factorization is unique up to a global constant phase.
• The number of nontrivially different spatial factorizations is $\prod_{i=1}^P (\mu_i+1)$ given $P$ reciprocal pairs of off-circle roots of multiplicity $\mu_i$.

Spatially, the condition means that all nontrivial spatial factors must reside on the complex unit circle, corresponding to physically meaningful propagation modes in array and multichannel processing. Off-circle roots introduce an ambiguity in partitioning these spatial modes between analysis and synthesis, rendering the decomposition non-unique (Usevich et al., 2023).

7. Broader Implications and Operations

Spatial factorization conditions constrain not only static matrix structure but also matrix operations:

• For PSD matrices, factor width is stable under Hadamard product and integer Hadamard powers, with explicit bounds on the resulting factor width rank (Johnston et al., 2024).
• For block-structured or banded matrices, spatial factorizations provide computational and interpretive advantages in both theory and applications where subspace and support structure are essential.

A plausible implication is that spatial factorization conditions function as a bridge between algebraic rank and physically or computationally meaningful decompositions, dictating both the minimality and interpretability of representations in high-dimensional data and operator theory.