Low-Rank Chow Decomposition

• Low-Rank Chow Decomposition is an explicit representation of homogeneous forms as minimal sums of products of linear forms, revealing algebraic structure and optimal decomposition bounds.
• It employs secant and Chow varieties to determine the smallest number of summands (Chow rank) needed to factor a form into linear components.
• Recent efficient algorithms use eigenvalue decomposition and matrix pencil methods to compute low-rank Chow decompositions for cubic and higher-order cases.

A low-rank Chow decomposition is an explicit representation of a homogeneous form or symmetric tensor as a minimal sum of products of linear forms, where the number of summands—known as the Chow rank—is kept small relative to the ambient dimension. For a homogeneous degree-$d$ form in $n+1$ variables, the Chow decomposition expresses the form as a sum of $s$ completely reducible terms, each being a product of $d$ linear forms. The study of low-rank Chow decompositions, their geometric structure, and efficient algorithms for their computation has significant implications in algebraic geometry, tensor analysis, theoretical computer science, and quantum information.

1. The Chow Variety and Chow Rank

Let $\mathbb{k}$ be an algebraically closed field of characteristic zero and $R = \mathbb{k}[x_0, ..., x_n]$. The Chow variety $\mathrm{Ch}_d(\mathbb{P}^n)$ is the (projective) variety of all degree-$d$ forms which factor completely into linear forms:

$$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$

where $R_1$ denotes the space of linear forms. Its dimension is $n+1$0.

Given a homogeneous form $n+1$1, its Chow rank $n+1$2 is the smallest $n+1$3 such that

$n+1$4

Equivalently, this is the minimal $n+1$5 for which $n+1$6 lies in the $n+1$7th secant variety $n+1$8 of the Chow variety (Torrance, 2015).

2. Secant Varieties and Expected Chow Rank

For any projective variety $n+1$9, the $s$0th secant variety $s$1 is

$s$2

the Zariski closure of all linear spans of $s$3 points of $s$4. For the Chow variety, $s$5 parameterizes all forms of Chow rank at most $s$6.

A standard dimension count yields the expected dimension,

$s$7

The smallest $s$8 where this secant variety fills the space gives a generic upper bound $s$9 for the Chow rank:

$d$0

3. Geometric and Algebraic Results for Low Chow Rank

A principal result is that for all $d$1, the secant variety $d$2 is nondefective for all $d$3 except for the case $d$4 and $d$5, the so-called classical quadratic defect (Torrance, 2015). That is, for $d$6 and $d$7, generic forms of degree $d$8 in $d$9 variables have Chow rank exactly $\mathbb{k}$0. Explicit generic decompositions can be constructed in these cases. For example, generic forms in the following settings satisfy this closed-form bound:

$\mathbb{k}$1	$\mathbb{k}$2	$\mathbb{k}$3	$\mathbb{k}$4	$\mathbb{k}$5	Description (examples)
$\mathbb{k}$6	any	$\mathbb{k}$7	$\mathbb{k}$8	$\mathbb{k}$9	binary forms factor
$R = \mathbb{k}[x_0, ..., x_n]$0	$R = \mathbb{k}[x_0, ..., x_n]$1	$R = \mathbb{k}[x_0, ..., x_n]$2	$R = \mathbb{k}[x_0, ..., x_n]$3	$R = \mathbb{k}[x_0, ..., x_n]$4	plane cubics: $R = \mathbb{k}[x_0, ..., x_n]$5
$R = \mathbb{k}[x_0, ..., x_n]$6	$R = \mathbb{k}[x_0, ..., x_n]$7	$R = \mathbb{k}[x_0, ..., x_n]$8	$R = \mathbb{k}[x_0, ..., x_n]$9	$\mathrm{Ch}_d(\mathbb{P}^n)$0	cubic surfaces
$\mathrm{Ch}_d(\mathbb{P}^n)$1	$\mathrm{Ch}_d(\mathbb{P}^n)$2				See (Torrance, 2015), Cor. 4.3

In these regimes, any generic form of degree $\mathrm{Ch}_d(\mathbb{P}^n)$3 in $\mathrm{Ch}_d(\mathbb{P}^n)$4 variables can be written as a sum of $\mathrm{Ch}_d(\mathbb{P}^n)$5 products of $\mathrm{Ch}_d(\mathbb{P}^n)$6 linear forms.

4. Efficient Algorithms for Low-Rank Chow Decomposition

The computational problem of finding, for a given form or symmetric tensor, a minimal Chow decomposition was for a long time approached by symbolic or algebraic methods with high complexity. Recent advances have produced the first algorithms with linear or sub-quadratic time in the ambient tensor size for low-rank cases.

For order-3 symmetric tensors $\mathrm{Ch}_d(\mathbb{P}^n)$7 (i.e., cubic polynomials), a Chow decomposition is an expression

$\mathrm{Ch}_d(\mathbb{P}^n)$8

with each set of linear forms $\mathrm{Ch}_d(\mathbb{P}^n)$9. For concise tensors of Chow rank $d$0, and provided the linear forms are in general position, there exists a deterministic $d$1-time algorithm for computing this decomposition (Blomenhofer et al., 12 Sep 2025):

• Generic “probe” vectors $d$2 are selected.
• Contractions $d$3, $d$4 form a matrix pencil.
• The generalized eigenvalue problem $d$5 yields $d$6 eigenpairs corresponding to the underlying linear forms.
• Inner product tests cluster eigenvectors, and rank-1 factors are reconstructed via intersections in 3D subspaces.

The uniqueness (identifiability) of the decomposition is certified under these genericity conditions (Blomenhofer et al., 12 Sep 2025).

5. Extensions: Higher Orders and Degeneracies

The pencil-based eigenvalue method generalizes to higher odd-order symmetric tensors $d$7 of Chow rank $d$8, with runtime $d$9, which is sub-quadratic in the ambient dimension $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$0. The contraction is performed along one copy of $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$1, forming pencils of size $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$2. For certain non-generic orbits (e.g., decompositions into "W-tensors" of the form $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$3), specialized algorithms partition the eigenstructure and identify factors efficiently in $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$4 time (Blomenhofer et al., 12 Sep 2025).

Decompositions for even orders or for non-concise tensors remain open challenges; the current techniques crucially rely on properties of odd-order pencils and concise support.

6. Applications and Computational Complexity

Low-rank Chow decompositions correspond to highly nontrivial decompositions in algebraic geometry, depth-3 $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$5 circuits in computational complexity, and sums of symmetric product states in quantum information. For generic settings where the Chow rank achieves its expected value $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$6, the optimal homogeneous circuit size for evaluation is $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$7 (Torrance, 2015). The new algorithms for cubic and higher-order cases, achieving runtimes linear or sub-quadratic in the tensor support, represent major improvements over prior algebraic and symbolic approaches.

Potential directions for further research include:

• Extending computational certificates beyond the $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$8 regime,
• Establishing analogous induction and decomposition theory in the superabundant regime,
• Decomposing mixed-degree split varieties,
• Exploring the relevance of Chow decompositions in the study of star configurations and hypersurface intersections (Torrance, 2015, Blomenhofer et al., 12 Sep 2025).

7. Proof Techniques, Limitations, and Open Problems

The study of Chow ranks and decompositions is grounded in secant variety theory. Terracini’s lemma provides the differential geometric underpinning for dimension counts and proofs of nondefectivity: tangent spaces at generic points combine to fill the ambient tangent space unless coalescence (defectivity) occurs. Direct sum decompositions and parameter counts, coupled with computer-aided rank calculations (e.g., Macaulay2 computations up to $$\mathrm{Ch}_d(\mathbb{P}^n) = \mathrm{Split}_d(\mathbb{P}^n) = \left\{ [\ell_1 \cdots \ell_d] \in \mathbb{P}(R_d) : \ell_i \in R_1 \right\},$$9), form the foundation of recent results (Torrance, 2015).

Known limitations include the classical quadratic defect, the complexity of even-order decompositions, and the treatment of non-concise or ill-conditioned tensors. Extending pencil techniques, classifying degeneracies, and improving practical robustness and scalability are ongoing challenges. The synthesis of geometric theory with algorithmic advances continues to shape the field and suggests wider applicability to computational algebraic geometry and tensor theory.