Algebraic Matrix Completion Framework

• Algebraic matrix completion is a framework that uses inherent algebraic structures, such as determinantal varieties and circuit polynomials, to recover missing matrix entries.
• It employs local completion algorithms based on solving circuits and variance-aware aggregation to achieve robust entrywise reconstruction.
• The approach extends to lifted polynomial models and general linear measurements, facilitating matrix recovery beyond traditional low-rank constraints.

Algebraic matrix completion framework denotes a family of approaches that treats matrix completion as recovery under algebraic structure rather than solely as global optimization over missing entries. In its classical form, the unknown matrix is constrained to the determinantal variety of rank-$\le r$ matrices, so completion is governed by vanishing minors, generic fiber structure of coordinate projections, and combinatorics of the observation mask. In broader formulations, the same viewpoint extends to circuit polynomials, algebraic varieties obtained by polynomial lifting, arbitrary linear measurements, and linearly parameterized factor models (Kiraly et al., 2012, Király et al., 2012, Blythe et al., 2014, Riegler et al., 2015, Ongie et al., 2017, Chen et al., 2020).

1. Determinantal foundations and generic identifiability

The classical low-rank setting starts from an unknown matrix $A \in \mathbb{C}^{m\times n}$ or $A\in \mathbb{K}^{m\times n}$, a set of observed positions $E \subseteq [m]\times[n]$, and a masking operator

$$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$

The model class is the determinantal variety

$$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$

defined by the vanishing of all $(r+1)\times(r+1)$ minors and having dimension

$\dim \mathcal M(r;m\times n)=(m+n-r)r$

when $m,n\ge r$ (Kiraly et al., 2012, Király et al., 2012).

This formulation immediately turns completion into an algebraic-geometric question about the fibers of the restricted projection

$$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$

A central result is that, for a generic rank-$A \in \mathbb{C}^{m\times n}$0 matrix, the dimension of the fiber $A \in \mathbb{C}^{m\times n}$1 depends only on the mask $A \in \mathbb{C}^{m\times n}$2, and if that dimension is zero, then the number of completions also depends only on the mask. Thus generic completability is a property of the observation pattern rather than the observed values (Kiraly et al., 2012).

The same papers stress that injectivity cannot hold uniformly over all low-rank matrices. For $A \in \mathbb{C}^{m\times n}$3, the restricted masking $A \in \mathbb{C}^{m\times n}$4 is injective iff $A \in \mathbb{C}^{m\times n}$5. This is why the basic notion is generic injectivity or generic finiteness rather than worst-case identifiability (Kiraly et al., 2012). Necessary conditions for generic finiteness include the edge-count bound

$A \in \mathbb{C}^{m\times n}$6

minimum degree at least $A \in \mathbb{C}^{m\times n}$7, and $A \in \mathbb{C}^{m\times n}$8-edge-connectivity of the associated bipartite graph (Kiraly et al., 2012).

2. Circuits, matroids, and local algebraic dependence

A distinctive algebraic-combinatorial development replaces global completion by entrywise completability. The key technical object is the Jacobian of the factorization map

$A \in \mathbb{C}^{m\times n}$9

with $A\in \mathbb{K}^{m\times n}$0, $A\in \mathbb{K}^{m\times n}$1. For an observed set $A\in \mathbb{K}^{m\times n}$2, the submatrix $A\in \mathbb{K}^{m\times n}$3 of Jacobian rows indexed by $A\in \mathbb{K}^{m\times n}$4 determines generic finite completability: a missing position $A\in \mathbb{K}^{m\times n}$5 is finitely completable iff its Jacobian row lies in the row span of $A\in \mathbb{K}^{m\times n}$6 (Király et al., 2012).

This induces the rank-$A\in \mathbb{K}^{m\times n}$7 determinantal matroid, with rank function

$A\in \mathbb{K}^{m\times n}$8

Its closure equals the finitely completable closure $A\in \mathbb{K}^{m\times n}$9, and its circuits are minimal dependent sets of entries. In matroid language, a missing entry is finitely completable iff it belongs to a circuit contained in $E \subseteq [m]\times[n]$0 (Király et al., 2012).

A parallel formulation appears in the algebraic-combinatorial framework based directly on compatibility with rank $E \subseteq [m]\times[n]$1. A subset $E \subseteq [m]\times[n]$2 is a circuit of rank $E \subseteq [m]\times[n]$3 if every proper subset is unconstrained, while the full set satisfies a minimal algebraic dependence. Every such circuit carries a unique irreducible circuit polynomial $E \subseteq [m]\times[n]$4, up to scalar multiple, such that values on $E \subseteq [m]\times[n]$5 are compatible with rank $E \subseteq [m]\times[n]$6 iff

$E \subseteq [m]\times[n]$7

For minor supports, $E \subseteq [m]\times[n]$8 is the determinant polynomial itself (Blythe et al., 2014).

The dual viewpoint uses stresses. A rank-$E \subseteq [m]\times[n]$9 stress is a vector in the left kernel of the Jacobian, equivalently a matrix $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$0 satisfying

$$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$1

For generic data, maximal stress rank depends only on $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$2 and $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$3, and if it reaches $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$4, then finite and unique completion coincide on the closure (Király et al., 2012). In rank $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$5, circuits reduce to simple cycles in the bipartite observation graph, and the associated circuit polynomial is a binomial relation between products of entries on alternating cycle edges (Király et al., 2012).

3. Local completion algorithms and entrywise uncertainty

The circuit viewpoint yields explicitly local algorithms. In the graph-closure algorithm, one searches for a subgraph isomorphic to $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$6, equivalently an $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$7 submatrix with exactly one missing entry. The vanishing determinant of that minor is then linear in the missing entry, so the entry can be solved and the missing edge added to the observation graph; iterating this realizes $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$8-closure (Kiraly et al., 2012).

The more general framework based on solving circuits treats a missing entry $$\Omega : \mathbb{C}^{m\times n} \to \mathbb{C}^{\alpha}, \qquad (a_{ij}) \mapsto (a_{i_1j_1},\dots,a_{i_\alpha j_\alpha}).$$9 via a solving circuit

$$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$0

with all other positions observed. If $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$1 has degree $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$2 in $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$3, then $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$4 is a unique solving circuit and yields a rational reconstruction formula

$$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$5

This turns completion into a local algebraic inference problem rather than a global convex program (Blythe et al., 2014).

Noise leads naturally to variance-aware aggregation. The variance-minimizing local completion scheme in (Blythe et al., 2014) finds several solving circuits for the same target entry, computes candidate estimates, estimates their variances or covariances, and returns a linear combination

$$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$6

with minimal variance. For determinant-based circuits, the appendix derives first-order error surrogates from the perturbation of the solving equation $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$7, and for an almost-complete minor with missing entry set to $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$8 and $$\mathcal M(r;m\times n)=\{A:\operatorname{rank}(A)\le r\},$$9, the corresponding determinants $(r+1)\times(r+1)$0 lead to explicit weighting formulas (Blythe et al., 2014).

Two concrete algorithms instantiate this program. For positive rank-$(r+1)\times(r+1)$1 data under multiplicative noise, fACCRO uses $(r+1)\times(r+1)$2 minors in log-space, where

$(r+1)\times(r+1)$3

and aggregates many such local estimates. For general rank $(r+1)\times(r+1)$4, vm-Closure searches for almost-complete $(r+1)\times(r+1)$5 minors through the target entry, solves the determinant equation, and combines the resulting candidates with inverse-variance weights (Blythe et al., 2014). The same paper further introduces SMCB and algebraically initialized meta-OptSpace, which use local algebraic completion as an initialization for spectral refinement (Blythe et al., 2014).

4. Lifted variety models and completion beyond low rank

A major extension replaces linear low-rank structure in ambient coordinates by low-rank structure after polynomial lifting. In the algebraic variety model, the data matrix

$(r+1)\times(r+1)$6

has columns lying on an affine algebraic variety

$(r+1)\times(r+1)$7

For degree bound $(r+1)\times(r+1)$8, one lifts each column by the monomial feature map

$(r+1)\times(r+1)$9

and completion is performed by minimizing the rank of the lifted matrix $\dim \mathcal M(r;m\times n)=(m+n-r)r$0 subject to consistency with the observed entries (Ongie et al., 2017).

This model strictly generalizes ordinary low-rank completion: $\dim \mathcal M(r;m\times n)=(m+n-r)r$1 recovers the affine-subspace case, while $\dim \mathcal M(r;m\times n)=(m+n-r)r$2 includes unions of affine subspaces, quadratic surfaces, and higher-degree varieties. For a union of $\dim \mathcal M(r;m\times n)=(m+n-r)r$3 affine subspaces of dimension at most $\dim \mathcal M(r;m\times n)=(m+n-r)r$4, the lifted rank obeys

$\dim \mathcal M(r;m\times n)=(m+n-r)r$5

whereas the ambient matrix can be high-rank. A degrees-of-freedom heuristic then yields the per-column sampling law

$\dim \mathcal M(r;m\times n)=(m+n-r)r$6

when enough columns are available, contrasting with the $\dim \mathcal M(r;m\times n)=(m+n-r)r$7-scale implicit in ordinary low-rank methods for unions of $\dim \mathcal M(r;m\times n)=(m+n-r)r$8 subspaces (Ongie et al., 2017).

Optimization is carried out by variety-based matrix completion (VMC),

$\dim \mathcal M(r;m\times n)=(m+n-r)r$9

together with a kernelized IRLS scheme using the polynomial kernel

$m,n\ge r$0

to avoid explicitly constructing the lifted monomial matrix (Ongie et al., 2017).

5. Information-theoretic generalization and support complexity

A different generalization broadens algebraic matrix completion from entrywise observation of low-rank models to arbitrary linear measurements of arbitrary low-description-complexity matrix ensembles. The unknown random matrix

$m,n\ge r$1

may have continuous, discrete, mixed, or singular distribution, and measurements are

$m,n\ge r$2

Classical matrix completion is the special case $m,n\ge r$3, while rank-one sensing uses

$m,n\ge r$4

(Riegler et al., 2015).

The structural quantity is no longer rank alone but the lower Minkowski dimension of a bounded $m,n\ge r$5-support set $m,n\ge r$6 satisfying

$m,n\ge r$7

With covering number

$m,n\ge r$8

the lower Minkowski dimension is

$m,n\ge r$9

The main achievability theorem states that if $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$0, then for Lebesgue almost all measurement matrices $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$1, there exists a measurable decoder with error probability at most $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$2 (Riegler et al., 2015).

For low-rank matrices this recovers the familiar determinantal count. If $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$3, then

$$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$4

hence

$$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$5

measurements suffice, both for general sensing matrices and for rank-one sensing matrices (Riegler et al., 2015). The same paper also constructs a class of rank-$$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$6 matrices, of the form $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$7 with sparse nonzero columns in $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$8 and $$\Omega:\mathcal M(r;m\times n)\to \mathbb C^\alpha.$$9, whose support has Minkowski dimension at most $A \in \mathbb{C}^{m\times n}$00, so recovery is possible from

$A \in \mathbb{C}^{m\times n}$01

measurements. This shows that generic determinantal dimension is a worst-case model-class bound rather than a universal distribution-dependent threshold (Riegler et al., 2015).

6. Exact, parameterized, and projection-based formulations

Algebraic matrix completion also includes exact algorithmic formulations outside the usual real-valued low-rank setting. One strand studies maximum rank matrix completion for linear symbolic matrices

$A \in \mathbb{C}^{m\times n}$02

When $A \in \mathbb{C}^{m\times n}$03 all have rank one, matrix completion for $A \in \mathbb{C}^{m\times n}$04 can be done deterministically in $A \in \mathbb{C}^{m\times n}$05 field operations over any field. The analysis is phrased in terms of matrix spaces, enveloping algebras, idempotents, and rank-maximizing elements of a linear space $A \in \mathbb{C}^{m\times n}$06 (0907.0774).

A second strand studies incomplete matrices over $A \in \mathbb{C}^{m\times n}$07 with missing entries $A \in \mathbb{C}^{m\times n}$08. For $A \in \mathbb{C}^{m\times n}$09-RMC, the task is to complete the matrix to rank at most $A \in \mathbb{C}^{m\times n}$10. The structural parameters are $A \in \mathbb{C}^{m\times n}$11, $A \in \mathbb{C}^{m\times n}$12, and

$A \in \mathbb{C}^{m\times n}$13

the minimum number of rows plus columns covering all missing entries. The parameter $A \in \mathbb{C}^{m\times n}$14 is exactly the size of a minimum vertex cover in the bipartite graph whose edges are missing positions, and can be computed in time

$A \in \mathbb{C}^{m\times n}$15

For bounded-domain $A \in \mathbb{C}^{m\times n}$16, $A \in \mathbb{C}^{m\times n}$17-RMC[comb] is in randomized FPT: the completion constraints are reduced to at most $A \in \mathbb{C}^{m\times n}$18 quadratic equations after branching over dependency signatures and eliminating linear equations (Ganian et al., 2018).

A third formulation abstracts completion as repeated projection onto any structured set for which the full-matrix Frobenius projection is known. The generic missing-entry approximation problem

$A \in \mathbb{C}^{m\times n}$19

is solved by the iteration

$A \in \mathbb{C}^{m\times n}$20

where $A \in \mathbb{C}^{m\times n}$21 is the full-matrix projection oracle. This covers low-rank truncation, spectral-norm balls, nuclear-norm balls, Ky-Fan norms, and orthogonality constraints. In the convex case, the projected-gradient step size is $A \in \mathbb{C}^{m\times n}$22, and exact completion with spectral or nuclear norm bounds is obtained by binary search on the constraint radius $A \in \mathbb{C}^{m\times n}$23 (Shabat et al., 2013).

7. Structured nonconvex, statistical, and algorithmic extensions

Later work broadens the algebraic framework in several directions while retaining low-dimensional structure as the governing principle. In nonconvex structured completion with linearly parameterized factors, the unknown matrix is represented as

$A \in \mathbb{C}^{m\times n}$24

with $A \in \mathbb{C}^{m\times n}$25 and $A \in \mathbb{C}^{m\times n}$26 linear in a parameter $A \in \mathbb{C}^{m\times n}$27. The central condition is Correlated Parametric Factorization (CPF), which requires that for any $A \in \mathbb{C}^{m\times n}$28 there exists $A \in \mathbb{C}^{m\times n}$29 with

$A \in \mathbb{C}^{m\times n}$30

Under CPF, every local minimum of the regularized nonconvex completion objective is statistically accurate, and in the noiseless case there are no spurious local minima. The condition is verified for subspace-constrained completion and skew-symmetric completion (Chen et al., 2020).

A complementary statistical generalization replaces deterministic exact constraints by exponential-family likelihoods and general structural regularizers. In this framework the unknown object is a natural-parameter matrix $A \in \mathbb{C}^{m\times n}$31, observed only on a subset $A \in \mathbb{C}^{m\times n}$32, and the estimator is

$A \in \mathbb{C}^{m\times n}$33

with $A \in \mathbb{C}^{m\times n}$34 decomposable. Low-rank completion reappears as the nuclear-norm special case, for which the sample size scale is

$A \in \mathbb{C}^{m\times n}$35

in the stated corollary (Gunasekar et al., 2015). The mixed-type extension partitions the columns into blocks $A \in \mathbb{C}^{m\times n}$36 with different exponential-family observation models and estimates a single low-rank parameter matrix by

$A \in \mathbb{C}^{m\times n}$37

solved by an ADMM scheme based on semidefinite reformulation (Sun et al., 2020).

Structured latent-factor completion pushes this further to model classes

$A \in \mathbb{C}^{m\times n}$38

covering Gaussian mixture models, mixed membership models, bi-clustering, stochastic block models, and sparse dictionary learning. Under completion with Bernoulli sampling rate $A \in \mathbb{C}^{m\times n}$39, the minimax Frobenius-risk scale is

$A \in \mathbb{C}^{m\times n}$40

where

$A \in \mathbb{C}^{m\times n}$41

This makes explicit that complexity is not governed by rank alone but also by sparse and discrete latent structure (Klopp et al., 2017).

Recent algorithmic work returns to the low-rank setting from a different angle: partial completion is obtained on $A \in \mathbb{C}^{m\times n}$42 of rows and columns from about $A \in \mathbb{C}^{m\times n}$43 samples and $A \in \mathbb{C}^{m\times n}$44 time, and under a regularity assumption on row and column spans, full completion is achieved with sample complexity $A \in \mathbb{C}^{m\times n}$45 and runtime $A \in \mathbb{C}^{m\times n}$46. In the incoherent case, the same framework yields $A \in \mathbb{C}^{m\times n}$47 observations and $A \in \mathbb{C}^{m\times n}$48 time, together with noisy recovery error approximately $A \in \mathbb{C}^{m\times n}$49 (Kelner et al., 2023). This suggests that algebraic span constraints, spectral residual estimation, and regression-based reconstruction can be combined into a unified completion framework whose runtime approaches the cost of verifying a proposed low-rank factorization.