Linear Matroid Parity Problem

• Linear Matroid Parity is a problem that generalizes matching and linear matroid intersection using linearly represented paired columns or 2D subspaces.
• It employs algebraic techniques such as skew-symmetric matrices and Pfaffian expansions to tackle decision, optimization, and counting tasks.
• Deterministic polynomial-time algorithms and quasi-NC approaches demonstrate its tractability in special cases like Pfaffian parities.

Searching arXiv for the cited papers and closely related work on linear matroid parity. The Linear Matroid Parity Problem is the linearly represented form of matroid parity, also called linear matroid matching in the line-based formulation. It asks for a feasible collection of paired columns, or equivalently 2-dimensional subspaces, such that each chosen pair contributes full dimension and the total cardinality is maximized; in weighted forms, the objective is to optimize line weights; in counting forms, the objective is to enumerate feasible or minimum-weight feasible solutions. In the linear setting, the problem is a common generalization of matching and linear matroid intersection, admits algebraic formulations through skew-symmetric matrices and Pfaffians, and has deterministic polynomial-time algorithms for optimization. Exact counting, however, is only known to be tractable on special cancellation-free subclasses such as Pfaffian parities (Matoya et al., 2019).

1. Problem statement and equivalent formulations

A standard matrix formulation uses a matrix $A \in \mathbb{K}^{2r \times 2n}$ whose $2n$ columns are partitioned into $n$ disjoint pairs called lines. Writing $L$ for the set of lines, the pair $(A,L)$ is a linear matroid parity instance. For a subset $J \subseteq L$, the submatrix $A[J]$ consists of all columns in the lines of $J$. A subset $B \subseteq L$ is a parity base if $|B|=r$ and $2n$0 is nonsingular; equivalently, $2n$1 is a $2n$2 invertible matrix made of whole lines. The set of parity bases is denoted $2n$3. This yields three canonical tasks: the decision/search problem of determining whether $2n$4 and finding $2n$5, the optimization problem of minimizing $2n$6 over parity bases, and the counting problem of computing $2n$7 or counting minimum-weight parity bases (Matoya et al., 2019).

An equivalent line-based formulation represents the input as a family of 2-dimensional subspaces

$2n$8

A subset $2n$9 of lines is a matroid matching if

$n$0

A perfect matroid matching has size $n$1. In the pair-partition specialization, each pair $n$2 spans a line $n$3, and selecting a set of pairs is feasible exactly when all selected pair-elements are independent, equivalently when the chosen lines contribute full dimension $n$4. In this sense, linear matroid matching is “also known as the linear matroid parity problem” (Gurjar et al., 2024).

This problem strictly generalizes ordinary matching. For a graph $n$5 with $n$6, one can construct a totally unimodular matrix $n$7 with two columns per edge, viewed as a line, such that perfect matchings of $n$8 are exactly parity bases $n$9. It also generalizes linear matroid intersection: given an $L$0 matrix pair $L$1, Lawler’s reduction produces a $L$2 parity instance $L$3 whose lines correspond to the original columns and such that $L$4 (Matoya et al., 2019).

2. Algebraic and Pfaffian formulations

The algebraic core of linear matroid parity is a skew-symmetric encoding. For $L$5, define the $L$6 skew-symmetric block-diagonal matrix $L$7 whose $L$8 block for a line $L$9 is

$(A,L)$0

and define

$(A,L)$1

A standard characterization is

$(A,L)$2

over the rational function field $(A,L)$3. The skew-symmetric matrix $(A,L)$4 can be written as

$(A,L)$5

which already exhibits the rank-two skew-symmetric coefficient structure characteristic of the problem (Matoya et al., 2019).

Using the Ishikawa–Wakayama Pfaffian Cauchy–Binet formula, the parity bases appear as monomials in a Pfaffian expansion: $(A,L)$6 and

$(A,L)$7

Thus the Pfaffian is a generating polynomial over parity bases, with coefficients $(A,L)$8. In the weighted setting, if $(A,L)$9 and $J \subseteq L$0 is an indeterminate, then

$J \subseteq L$1

The coefficient of $J \subseteq L$2 is therefore

$J \subseteq L$3

This formulation is exact, but in general it does not by itself yield counts, because the coefficients $J \subseteq L$4 can cancel algebraically (Matoya et al., 2019).

A complementary weighted formulation uses the skew-symmetric polynomial matrix

$J \subseteq L$5

where $J \subseteq L$6 is block diagonal with one $J \subseteq L$7 block per line $J \subseteq L$8: $J \subseteq L$9 Its Pfaffian encodes the optimum weight through the identity

$A[J]$0

and $A[J]$1 if there is no parity base. This degree-of-Pfaffian viewpoint underlies the deterministic weighted algorithm described below (Iwata et al., 2019).

3. Weighted optimization and deterministic polynomial-time algorithms

The weighted linear matroid parity problem remained open for decades in the represented setting, even though the unweighted linear case was already polynomial-time solvable. A deterministic polynomial-time algorithm is now known for the minimum-weight parity base problem. In the formulation with a row-full-rank matrix $A[J]$2, row set $A[J]$3, column set $A[J]$4, and partition of $A[J]$5 into lines $A[J]$6, a parity base is a base $A[J]$7 of $A[J]$8 such that $A[J]$9 is a union of whole lines, and the objective is

$J$0

The main theorem is that Algorithm $J$1 finds a minimum-weight parity base, or proves infeasibility, in

$J$2

arithmetic operations over $J$3. Over a fixed finite field this is strongly polynomial; over $J$4, polynomial-time solvability follows by reduction through a sequence of finite fields (Iwata et al., 2019).

The algorithm is combinatorial and primal-dual in spirit, but it is formulated through a cocircuit matrix and Pfaffian algebra rather than through an explicit LP for matroid parity. For a current base $J$5, the fundamental cocircuit matrix is

$J$6

A standard fact is

$J$7

Thus exchanges of elements are converted into nonsingularity tests on principal submatrices. The algorithm allows the current base to violate parity; lines containing exactly one chosen endpoint are source lines. The goal is to eliminate source lines by augmenting-path operations until the base becomes a parity base (Iwata et al., 2019).

To support matching-style blossom manipulations, the method introduces an enlarged vertex set $J$8, a laminar family of blossoms $J$9, dummy lines formed by bud-tip pairs, a current augmented base $B \subseteq L$0, and an augmented matrix $B \subseteq L$1 obtained by pivoting. Dual variables consist of endpoint potentials $B \subseteq L$2 and blossom weights $B \subseteq L$3, constrained by

$B \subseteq L$4

for every line $B \subseteq L$5, together with inequalities of the form

$B \subseteq L$6

for exchange edges $B \subseteq L$7. Tight edges satisfy equality. Search is performed in the graph of tight edges, and augmenting paths are required to induce a unique tight perfect matching on their support, ensuring that the corresponding pivot is nonsingular (Iwata et al., 2019).

Optimality is certified indirectly. Feasible dual variables $B \subseteq L$8 are embedded into the dual LP of a maximum-weight perfect matching problem on an auxiliary graph built from a transformed skew-symmetric matrix $B \subseteq L$9. Weak duality in that matching LP yields the required upper bound on $|B|=r$0, and hence on $|B|=r$1. This is why the algorithm behaves like a weighted blossom algorithm while remaining fundamentally algebraic (Iwata et al., 2019).

4. Pfaffian parities and exact counting

For exact counting, the decisive obstruction is algebraic cancellation. In a general parity instance,

$|B|=r$2

has mixed coefficients $|B|=r$3, so extracting the number of solutions is not straightforward. The subclass introduced to remove this obstruction is the Pfaffian parity: a matroid parity $|B|=r$4 is Pfaffian if there exists $|B|=r$5 such that

$|B|=r$6

The value $|B|=r$7 is called the constant of $|B|=r$8. In that case every feasible parity base contributes the same nonzero coefficient in the Pfaffian expansion, and the identities simplify to

$|B|=r$9

$2n$00

At $2n$01,

$2n$02

This is the parity analogue of Pfaffian orientations in matching and of Webb’s Pfaffian pairs in matroid intersection (Matoya et al., 2019).

The resulting algorithmic consequences are strong. If $2n$03, one can count the number of parity bases of a $2n$04 Pfaffian parity deterministically in

$2n$05

time, and also construct one parity base in the same time. If the Pfaffian constant $2n$06 is given, the number of parity bases can be computed modulo $2n$07 in deterministic

$2n$08

time, improving to $2n$09 in characteristic $2n$10. For weighted Pfaffian parities with line weights $2n$11, the number of minimum-weight parity bases modulo $2n$12 can be computed deterministically in

$2n$13

For $2n$14, the number of minimum-weight parity bases can be computed deterministically in time polynomial in the binary encoding length of $2n$15 (Matoya et al., 2019).

This tractable subclass includes several structurally important examples. If $2n$16 is a Pfaffian orientation of a graph $2n$17, then the associated parity instance $2n$18 is a Pfaffian parity, embedding classical Pfaffian perfect-matching counting into linear matroid parity. If a 3-uniform hypergraph has a 3-Pfaffian orientation, then Lovász’s reduction to graphic matroid parity yields a parity instance whose parity bases correspond to spanning hypertrees. The paper also proves that, in the $2n$19 case, weighted shortest disjoint $2n$20-$2n$21-$2n$22 paths in an LGV position correspond bijectively to minimum-weight parity bases of a weighted Pfaffian parity (Matoya et al., 2019).

A common misconception is that these results make exact counting easy for linear matroid parity in general. They do not. The tractable class is explicitly the class of Pfaffian parities, namely those instances for which feasible determinants are constant and algebraic cancellation disappears (Matoya et al., 2019).

5. Fractional linear matroid parity, non-commutative rank, and quasi-parallelism

The fractional relaxation replaces integral line selection by a half-integral polytope. For lines $2n$23, a vector $2n$24 is a fractional matroid matching if for every subspace $2n$25,

$2n$26

The feasible region $2n$27 satisfies $2n$28, and every vertex of $2n$29 is half-integral: $2n$30 Its integer vertices are exactly the integral matroid matching vertices. The dual problem is the minimum nested 2-cover problem, and the min-max theorem is

$2n$31

(Oki et al., 2022).

The algebraic matrix used here is Lovász’s skew-symmetric symbolic matrix

$2n$32

Ordinary symbolic rank encodes the integral optimum. The key result is that non-commutative rank encodes the fractional optimum: $2n$33 Equivalently, with the second blow-up,

$2n$34

so blow-up size $2n$35 already suffices for this problem family. This matches the half-integrality of the fractional polytope: the rank $2n$36 of a $2n$37 block corresponds to the values $2n$38 (Oki et al., 2022).

These identities yield algebraic algorithms for the fractional problem. A simple search-to-decision reduction finds the lexicographically minimum maximum fractional matroid matching in

$2n$39

field operations with probability at least $2n$40 if $2n$41. A divide-and-conquer sparse-representation algorithm finds a maximum fractional matroid matching in

$2n$42

and a randomized algebraic algorithm finds the dominant 2-cover in

$2n$43

time with probability at least $2n$44. Combined with Gijswijt’s framework, these routines give a weighted fractional algorithm running in

$2n$45

(Oki et al., 2022).

A later development establishes that fractional linear matroid matching is in quasi-NC. In that formulation, the same line-based problem is treated as the fractional relaxation of linear matroid parity, and the optimum is characterized by

$2n$46

The algorithm is deterministic and runs in $2n$47 time using $2n$48 parallel processors. It also yields a deterministic black-box hitting set for non-commutative Edmonds’ problem with rank-two skew-symmetric coefficients. However, the exact integral linear matroid parity problem is not solved in quasi-NC; the conclusion stated is that “The parallel complexity of linear matroid matching is still open” (Gurjar et al., 2024).

6. Generalizations, methodological extensions, and current boundaries

Linear matroid parity also appears as a special case of a broader algebraic framework over linear and projected linear delta-matroids. In this setting, a key representation shift is the contraction representation $2n$49, where $2n$50 is skew-symmetric over $2n$51 and a set $2n$52 is feasible exactly when $2n$53 is nonsingular. This representation is equivalent to the standard twist representation up to $2n$54-time transformations, but is more convenient for algorithmic tasks. For a linear matroid $2n$55 represented by a matrix $2n$56, its basis family becomes a linear delta-matroid, and ordinary matroid parity becomes a special case of DM Parity (Koana et al., 2024).

Within this framework, the union and delta-sum of linear or projected linear delta-matroids remain linear or projected linear, and several decision problems reduce to a single rank computation. In particular, the decision versions of DM Covering, DM Delta-Covering, DM Intersection, DM Partition, and DM Parity reduce to computing the rank of an $2n$57 skew-symmetric matrix. Search versions can then be solved in

$2n$58

field operations with high probability by lifting Harvey’s inverse-update technique, the same algebraic style previously used for fast linear matroid parity algorithms. This improves over the $2n$59-time augmenting path algorithm of Geelen, Iwata, and Murota, and over the $2n$60 bound obtainable by self-reducibility (Koana et al., 2024).

These extensions do not supersede the classical linear matroid parity algorithms; rather, they subsume them into a larger skew-symmetric framework. At the same time, they clarify several current boundaries. Weighted perfect parity in the delta-matroid setting remains harder: the paper states that the more interesting weighted perfect parity version is equivalent to weighted DM intersection and only gets

$2n$61

search via self-reduction after computing the optimum weight (Koana et al., 2024). Likewise, the determinant-degree method for weighted linear matroid intersection gives a linear-algebraic interpretation of Frank’s weight-splitting algorithm and is conceptually close to parity, but it is not itself a parity algorithm and does not provide a direct parity-specific complexity theorem (Furue et al., 2019).

The contemporary boundary is therefore sharply delineated. Exact integral optimization for linearly represented matroid parity is polynomial-time solvable, including the weighted case (Iwata et al., 2019). Exact counting is polynomial-time only for special cancellation-free subclasses such as Pfaffian parities (Matoya et al., 2019). Fractional parity is algebraically characterized by non-commutative rank and lies in quasi-NC (Oki et al., 2022, Gurjar et al., 2024). Exact integral quasi-NC for linear matroid parity remains open (Gurjar et al., 2024).