Uniquely Restricted Matching in Graphs

• Uniquely restricted matching is a graph matching where the chosen set is the unique perfect matching of its induced subgraph, defined without alternating cycles.
• The topic employs alternating-cycle characterizations and BD-mapping in bipartite graphs along with Gallai–Edmonds decomposition to enable efficient recognition and structural analysis.
• It also examines tight extremal bounds, NP-hardness of related decision problems, and parameterized as well as polynomial-time algorithms for special graph classes.

A uniquely restricted matching in a graph $G=(V,E)$ is a matching $M\subseteq E$ for which there is no other matching $M'\neq M$ with $V(M')=V(M)$; equivalently, $M$ is the unique perfect matching of the induced subgraph $G[V(M)]$. The notion refines ordinary matching while remaining broader than induced matching, and its theory connects alternating-cycle obstructions, Gallai–Edmonds structure, graph-class-specific algorithms, extremal lower bounds, approximation, and hardness of equality problems involving other restricted matching parameters (Penso et al., 2015, Fürst, 2018).

1. Formal definition and equivalent characterizations

Let $G=(V,E)$ be a finite simple undirected graph and let $M\subseteq E$ be a matching. Writing

$$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$

the matching $M$ is uniquely restricted if $M\subseteq E$0 is the unique perfect matching of $M\subseteq E$1. An induced matching is more restrictive: $M\subseteq E$2 is induced if $M\subseteq E$3 is $M\subseteq E$4-regular, that is, every vertex of $M\subseteq E$5 has degree exactly $M\subseteq E$6 inside $M\subseteq E$7 (Fürst, 2018).

A central equivalent formulation is the alternating-cycle characterization: $M\subseteq E$8 is uniquely restricted if and only if $M\subseteq E$9 contains no $M'\neq M$0-alternating cycle. Here an alternating cycle is an even cycle whose edges alternate between $M'\neq M$1 and $M'\neq M$2. This characterization is used throughout the literature because it converts uniqueness of the perfect matching in $M'\neq M$3 into an explicit forbidden configuration (Francis et al., 2016).

In bipartite graphs, the same obstruction admits a directed-graph encoding. For $M'\neq M$4 and a matching $M'\neq M$5, the BD-mapping associates a digraph $M'\neq M$6 whose vertices correspond to the $M'\neq M$7-endpoints of matched edges, with an arc $M'\neq M$8 whenever $M'\neq M$9, $V(M')=V(M)$0, and $V(M')=V(M)$1. The matching $V(M')=V(M)$2 is uniquely restricted if and only if the corresponding BD-mapping digraph is acyclic (Zhu, 2010).

2. Matching parameters and structural decompositions

The standard extremal parameter is

$V(M')=V(M)$3

When induced and acyclic matchings are included, the parameter hierarchy becomes

$V(M')=V(M)$4

where $V(M')=V(M)$5 is the ordinary matching number, $V(M')=V(M)$6 is the maximum acyclic matching size, and $V(M')=V(M)$7 is the maximum induced matching size. Every induced matching is therefore uniquely restricted, but the converse fails in general (Fürst, 2018).

For maximum-matchability questions, the Gallai–Edmonds decomposition provides the main global structure. With $V(M')=V(M)$8 denoting the vertices not covered by at least one maximum matching, $V(M')=V(M)$9 the vertices outside $M$0 adjacent to $M$1, and $M$2, one builds a reduced bipartite graph $M$3 by deleting $M$4, contracting each component of $M$5, and retaining only the edges between $M$6 and the contracted components. In this framework, there exists a maximum matching of $M$7 that is uniquely restricted if and only if three conditions hold: each component of $M$8 has a unique perfect matching; $M$9 has a maximum matching within the admissible edges that is itself uniquely restricted; and each component $G[V(M)]$0 of $G[V(M)]$1 has at least one vertex $G[V(M)]$2 such that $G[V(M)]$3 has a unique perfect matching (Penso et al., 2015).

The corresponding characterization for the stronger property that every maximum matching is uniquely restricted is parallel but stricter. It requires that every component of $G[V(M)]$4 have a unique perfect matching, every near-perfect matching of each factor-critical component of $G[V(M)]$5 be uniquely restricted, every maximum matching of $G[V(M)]$6 be uniquely restricted, and every edge $G[V(M)]$7 that occurs in some maximum matching of $G[V(M)]$8 satisfy the condition that $G[V(M)]$9 has exactly one neighbor in the component $G=(V,E)$0. These criteria yield polynomial-time recognition algorithms (Penso et al., 2015).

3. Polynomial-time algorithms on structured graph classes

The bipartite case admits particularly explicit recognition criteria. One direction uses the BD-mapping acyclicity theorem for a fixed matching. A more global result resolves the question of when all maximum matchings of a bipartite graph are uniquely restricted: after extending the BD-mapping with free vertices, the property is equivalent to path-uniqueness conditions such as “for every free vertex $G=(V,E)$1 there is at most one directed path from $G=(V,E)$2 to any sink $G=(V,E)$3,” together with two equivalent dual formulations. Testing the required disjoint-path obstructions is polynomial-time, thereby answering an open question of Levit and Mandrescu in the affirmative (Zhu, 2010).

Interval-type graph classes support optimization rather than mere recognition. For interval graphs, a maximum-cardinality uniquely restricted matching can be computed in polynomial time by reducing the problem to a maximum-cardinality strong independent set in an interval-nest digraph. Given an interval representation $G=(V,E)$4, one forms a digraph $G=(V,E)$5 on the edge set $G=(V,E)$6, with

$G=(V,E)$7

for each edge $G=(V,E)$8, and proves that a set $G=(V,E)$9 is a uniquely restricted matching in $M\subseteq E$0 if and only if it is a strong independent set in $M\subseteq E$1. The resulting dynamic program runs in $M\subseteq E$2 time for interval graphs, while proper interval graphs and bipartite permutation graphs admit linear-time $M\subseteq E$3 algorithms based on successor-style dynamic programming over canonical vertex orderings (Francis et al., 2016).

Subcubic graphs furnish another positive algorithmic regime. For $M\subseteq E$4-connected subcubic graphs of sufficiently large order, equality $M\subseteq E$5 has a complete structural characterization, and as a consequence subcubic graphs with $M\subseteq E$6 can be recognized in polynomial time (Fürst et al., 2018).

4. Equality with induced matchings and hardness phenomena

The problem of deciding whether the induced matching number and the uniquely restricted matching number coincide,

$M\subseteq E$7

was posed by Golumbic, Hirst, and Lewenstein. For general graphs this decision problem is NP-hard, and it remains NP-hard even when the input graph is bipartite. The same work remarks that membership in NP or co-NP is not known, because it is not obvious how to certify in polynomial time that no larger induced or uniquely restricted matching exists (Fürst, 2018).

The hardness reduction proceeds from Exact-SAT under the restriction that every clause has size three, each variable appears positively at most three times, and no literal is negated. From an instance with variables $M\subseteq E$8 and clauses $M\subseteq E$9, one constructs a bipartite graph by attaching a $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$0 variable gadget $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$1 for each variable and a $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$2 clause gadget for each clause, then linking clause leaves to designated leaves of the variable gadgets according to literal occurrence. Two key properties are proved: $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$3, and there is an induced matching of size $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$4 if and only if the Exact-SAT instance is exactly satisfiable. This suggests that a good characterization is unlikely to be possible in full generality (Fürst, 2018).

By contrast, bounded degree yields a structural dichotomy. For $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$5-connected subcubic graphs with $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$6, one has $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$7 if and only if $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$8 belongs to one of two infinite families $$V(M)=\{\,v\in V:\exists\,e\in M,\ v\in e\,\}, \qquad G(M)=G[V(M)],$$9 or $M$0. This characterization underlies a polynomial-time recognition algorithm for arbitrary subcubic graphs, combining exact handling of finitely many small graphs with the large-order structure theorem (Fürst et al., 2018).

5. Extremal bounds, sparse graphs, and approximation

A substantial part of the theory concerns guaranteed lower bounds on $M$1. For graphs with maximum degree at most $M$2 and no isolated vertex,

$M$3

and equality holds if and only if every component of $M$4 is isomorphic to $M$5 for some $M$6. For subcubic graphs,

$M$7

and for graphs of maximum degree $M$8 and girth at least $M$9,

$M\subseteq E$00

These bounds are all stated as tight (Fürst et al., 2018).

The subcubic setting has several sharper results. If $M\subseteq E$01 is a connected subcubic graph with $M\subseteq E$02 edges and $M\subseteq E$03 good bridges, and $M\subseteq E$04, then

$M\subseteq E$05

If $M\subseteq E$06 is a connected subcubic graph of order $M\subseteq E$07 and girth at least $M\subseteq E$08, then

$M\subseteq E$09

The proofs proceed by minimal-counterexample arguments in which low-degree vertices or small local configurations are peeled away and the matching is extended back using bridge edges, which cannot lie on alternating cycles (Fürst et al., 2018).

For girth at least $M\subseteq E$10, an even more precise theorem is available: every connected subcubic graph with $M\subseteq E$11 vertices and girth at least $M\subseteq E$12 contains a uniquely restricted matching of size at least $M\subseteq E$13, except for two exceptional cubic graphs $M\subseteq E$14 and $M\subseteq E$15 of orders $M\subseteq E$16 and $M\subseteq E$17, for which $M\subseteq E$18 and $M\subseteq E$19 (Fürst et al., 2018).

Approximation algorithms complement the extremal bounds. For connected bipartite graphs of maximum degree $M\subseteq E$20, there is a polynomial-time algorithm that constructs a uniquely restricted matching $M\subseteq E$21 with

$M\subseteq E$22

that is, a $M\subseteq E$23-approximation, improving over a $M\subseteq E$24-approximation for subcubic bipartite graphs. The same work introduces the uniquely restricted chromatic index $M\subseteq E$25, proves $M\subseteq E$26 for connected graphs with equality if and only if $M\subseteq E$27, and proves $M\subseteq E$28 for connected bipartite $M\subseteq E$29 with $M\subseteq E$30 (Baste et al., 2016).

6. Parameterized complexity and open directions

From a parameterized viewpoint, the decision problem asks whether a graph $M\subseteq E$31 contains a uniquely restricted matching of size at least $M\subseteq E$32. On general graphs, the problem remains W[1]-hard when parameterized by $M\subseteq E$33. Positive results emerge once the graph structure is restricted: on line graphs $M\subseteq E$34, the problem is fixed-parameter tractable in $M\subseteq E$35, using the characterization that $M\subseteq E$36 has a uniquely restricted matching of size $M\subseteq E$37 if and only if the host graph $M\subseteq E$38 contains $M\subseteq E$39 edge-disjoint $M\subseteq E$40's whose union is a forest and such that no two of the paths together induce a $M\subseteq E$41. This yields an algorithm with running time $M\subseteq E$42 (Chaudhary et al., 16 Aug 2025).

Treewidth is another FPT parameter. Given a nice tree decomposition, one may maintain a dynamic-programming table $M\subseteq E$43 in which $M\subseteq E$44 records saturation status of bag vertices and $M\subseteq E$45 records which pairs are already connected by an $M\subseteq E$46-alternating path in the partial solution. The resulting deterministic algorithm solves the problem in time

$M\subseteq E$47

At the same time, the problem does not admit a polynomial kernel with respect to the parameter $M\subseteq E$48 unless $M\subseteq E$49 (Chaudhary et al., 16 Aug 2025).

Several boundary questions remain open. For deciding $M\subseteq E$50, the exact complexity for graphs of maximum degree $M\subseteq E$51 is left open; the known hardness extends to maximum degree $M\subseteq E$52, and by minor modifications to maximum degree $M\subseteq E$53. The same problem is not currently known to lie in NP or co-NP, and a broader open direction is to delineate the boundary between tractable and intractable cases for equality of restricted matching numbers in terms of degree bounds or forbidden induced subgraphs (Fürst, 2018). Related conjectural directions ask whether all bridges, rather than only good bridges, can be used in the bound $M\subseteq E$54, and whether large girth forces $M\subseteq E$55 to approach $M\subseteq E$56 arbitrarily closely for bounded-degree graphs (Fürst et al., 2018, Fürst et al., 2018).