Triangle Matchings in Graph Theory

• Triangle matchings are configurations defined in multiple contexts, ranging from disjoint triangles in Ramsey theory to triangle-free 2-matchings in restricted matching problems.
• They play a crucial role in evaluating graph properties, driving algorithmic frameworks, and establishing bounds in geometric proximity and hypergraph partitioning.
• Recent work leverages local decomposition and shrinking techniques to transform overlapping triangle systems into tractable matching and design models.

Triangle matchings form a cluster of related notions rather than a single standard definition. In graph theory, the term may denote the graph $nK_3$ of $n$ vertex-disjoint triangles, its connected analogue $c(nK_3)$, or perfect matchings in structures whose basic blocks are triangles, such as Steiner triple systems. In matching theory, it also appears through restrictions such as triangle-free $2$-matchings, where a $2$-matching may contain paths and cycles but is forbidden to contain a $3$-cycle. In computational geometry, triangles may define adjacency through empty equilateral witnesses or act as forbidden subgraphs in unions of non-crossing perfect matchings. The literature therefore treats triangle matchings through Ramsey theory, $2$-matching theory, geometric proximity graphs, hypergraph packings, and cell-incidence matchings in triangulations (Gyarfas et al., 2015, Babenko et al., 2010, Babu et al., 2012, Kwan, 2016).

1. Foundational meanings and formal models

The main uses of the term can be organized by the ambient combinatorial object.

Context	Formal object	Representative source
Ramsey theory	$nK_3$ and $c(nK_3)$	(Gyarfas et al., 2015)
Restricted $2$-matching	triangle-free $n$0-matching, or $n$1-free $n$2-matching	(Babenko et al., 2010, Kobayashi, 2019)
Design theory	perfect matching in a Steiner triple system	(Kwan, 2016)
Tripartite triangle systems	Latin squares as triangle decompositions of $n$3	(Kwan et al., 2021)
Computational geometry	matching in $n$4, $n$5-graphs, or geometric perfect matchings with triangle-free union	(Babu et al., 2012, Biedl et al., 2019, Trao et al., 2016)
Cell complexes	complete matching of cells in a triangulation	(Meigniez, 2018)

In Ramsey-theoretic usage, $n$6 means a triangle matching: a graph consisting of $n$7 vertex-disjoint triangles. The connected version $n$8 denotes any connected graph containing $n$9 vertex-disjoint triangles; it need not be exactly $c(nK_3)$0 (Gyarfas et al., 2015). In $c(nK_3)$1-matching theory, a triangle-free $c(nK_3)$2-matching is an edge set $c(nK_3)$3 with $c(nK_3)$4 for every vertex and with no triangle in its support (Babenko et al., 2010). In the weighted $c(nK_3)$5-free formulation, one is given a specified set $c(nK_3)$6 of forbidden triangles and requires $c(nK_3)$7 for every $c(nK_3)$8 (Kobayashi, 2019).

In design theory, a Steiner triple system is a decomposition of $c(nK_3)$9 into triangles, and a perfect matching in that $2$0-uniform hypergraph is a set of vertex-disjoint triangles covering all vertices (Kwan, 2016). In Latin-square language, a Latin square is a triangle decomposition of $2$1; a partial transversal corresponds to a vertex-disjoint triangle matching, whereas the full Latin square itself is an edge-disjoint triangle decomposition (Kwan et al., 2021). In geometric settings, triangles can be either witnesses of adjacency, as in $2$2, or forbidden configurations in the union of several non-crossing perfect matchings (Babu et al., 2012, Trao et al., 2016).

2. Ramsey-theoretic triangle matchings

The classical Ramsey-theoretic triangle matching is $2$3, while the connected analogue $2$4 is defined as any connected graph containing $2$5 vertex-disjoint triangles. The exact two-color Ramsey number of the connected version is

$2$6

whereas the classical Burr–Erdős–Spencer result gives

$2$7

The increase from $2$8 to $2$9 is caused by connectedness: many disjoint monochromatic triangles do not suffice unless they lie in one connected monochromatic subgraph (Gyarfas et al., 2015).

A central auxiliary statement in that theory is

$2$0

This mixed Ramsey number drives the proof of the exact theorem by converting the absence of a connected triangle matching in one color into the presence of a sizable ordinary matching in the other color, which can then be extended to triangles using vertices in other components (Gyarfas et al., 2015). The same paper develops perturbed versions with a third color encoding missing edges and applies them to almost-squares of cycles, proving that for every fixed $2$1 there exists $2$2 such that

$2$3

This places connected triangle matchings in the regularity-and-blow-up pipeline as reduced-graph skeletons for embedding cycle-square-like structures (Gyarfas et al., 2015).

Ordered Ramsey theory produces a different triangle-versus-matching problem. For an ordered matching $2$4, the off-diagonal number $2$5 asks how large a complete ordered graph must be before every red-blue coloring contains either a red ordered copy of $2$6 or a blue ordered triangle. For almost all $2$7-vertex ordered matchings with interval chromatic number $2$8,

$2$9

while for every $3$0 there exist $3$1-vertex ordered matchings with $3$2 satisfying

$3$3

These results show that in the ordered setting, sparse matchings can still display substantial Ramsey complexity against triangles (Balko et al., 2023).

3. Triangle-free $3$4-matchings and excluded-triangle constraints

A $3$5-matching allows degree at most $3$6 at each vertex, so its components are paths, cycles, and isolated vertices. Triangle-free $3$7-matching asks for such a subgraph with no $3$8-cycle. This restriction is subtle enough that the classical Cornuéjols–Pulleyblank approach was long central, but the Edmonds–Gallai viewpoint yields a simpler exact framework. In particular, the maximum size $3$9 of a triangle-free $2$0-matching satisfies the min-max theorem

$2$1

where $2$2 is the number of connected components of $2$3 that are triangle clusters. The same paper gives an $2$4-time algorithm for a maximum triangle-free $2$5-matching and proves that every $2$6-regular graph with $2$7 contains a perfect triangle-free $2$8-matching; the specialized algorithms are $2$9 for $nK_3$0, $nK_3$1 for $nK_3$2, and $nK_3$3 for $nK_3$4 (Babenko et al., 2010).

A recent simplification of the local-search structure is a decomposition theorem for triangle-free $nK_3$5-matchings. If $nK_3$6 and $nK_3$7 are triangle-free $nK_3$8-matchings in a simple graph, then $nK_3$9 admits a partition into alternating trails $c(nK_3)$0 such that $c(nK_3)$1 is again a triangle-free $c(nK_3)$2-matching for $c(nK_3)$3. This directly yields a simpler validity proof of the PTAS based on local search: if $c(nK_3)$4 is locally optimal with respect to trails of length at most $c(nK_3)$5, then $c(nK_3)$6 (Kobayashi et al., 2024).

The excluded-$c(nK_3)$7-factor framework gives a unifying formulation. In a bipartite graph, a $c(nK_3)$8-feasible $c(nK_3)$9-matching is a $2$0-matching satisfying

$2$1

so each $2$2 forbids a local $2$3-factor. Triangle-free $2$4-matching appears after reduction to a bipartite $2$5-matching problem in which each forbidden triangle becomes a forbidden $2$6-factor on a 6-vertex set, and the paper develops a common shrinking-and-expansion theory that simultaneously generalizes blossom shrinking, triangle shrinking, square shrinking, and directed-cycle shrinking (Takazawa, 2017).

The weighted problem is harder. For an arbitrary set $2$7 of forbidden triangles, the weighted $2$8-free $2$9-matching problem remains open in full generality when $n$00 is the set of all triangles, but if $n$01 is edge-disjoint then there is a polynomial-time algorithm via an extended formulation with local variables $n$02, strengthened odd-cut inequalities, and a polynomial-time separation oracle (Kobayashi, 2019). In subcubic graphs, there is a different weighted approach based on half-edges: problematic triangles are replaced by gadgets, the problem is reduced to one maximum-weight $n$03-matching computation, and the running time is $n$04, with $n$05 in the unweighted case (Paluch et al., 2020).

4. Geometric formulations and triangle-defined adjacency

In fixed-orientation equilateral-triangle graphs, triangles define adjacency rather than obstruction. For a point set $n$06 in general position, $n$07 joins two points if the smallest downward equilateral triangle containing them is empty. This graph is equivalent to a half-$n$08 graph and to a TD-Delaunay graph. Its main matching bound is

$n$09

and the construction in the paper shows that no universal lower bound can exceed $n$10. For the two-orientation graph $n$11, the block cut point graph is a path, and consequently any $n$12 graph on $n$13 vertices has at most $n$14 edges (Babu et al., 2012).

The full $n$15-graph can also be studied through empty equilateral triangles as witnesses of edges. If $n$16 is the minimum, over all $n$17-point sets $n$18, of the maximum matching size in the corresponding $n$19-graph $n$20, then

$n$21

The same work introduces the blocking number $n$22 and proves

$n$23

The conjectural endpoint is a (near-)perfect matching in every $n$24-graph, equivalently $n$25 (Biedl et al., 2019).

A different geometric regime concerns unions of non-crossing perfect matchings. For $n$26 points in convex position, there exist $n$27 edge-disjoint non-crossing perfect matchings whose union is the maximal triangle-free convex geometric graph $n$28, and any family of $n$29 such matchings has that union. In $n$30-position, the paper gives a sufficient condition for the existence of $n$31 edge-disjoint non-crossing perfect matchings whose union is maximal triangle-free, while for arbitrary point sets in general position it proves the universal lower bound

$n$32

on the number of edge-disjoint non-crossing perfect matchings whose union is triangle-free (Trao et al., 2016).

Triangle meshes give an applied dual-graph interpretation. If adjacent triangles of a triangulation are merged into quadrilaterals, the dual graph is modeled as a bridgeless cubic graph and a set of pairwise disjoint triangle-pairs is a matching in that dual. The paper studies the ratio

$n$33

between the maximum-weight perfect matching and the maximum-weight unrestricted matching. For bridgeless cubic graphs,

$n$34

and this bound is tight for broad classes including planar bridgeless cubic graphs, hamiltonian cubic graphs, generalized Petersen graphs, and nonhamiltonian bipartite cubic graphs (Brazil et al., 2012).

5. Hypergraphs, designs, and higher-dimensional analogues

A Steiner triple system of order $n$35 is exactly a partition of the edges of $n$36 into triangles. A perfect matching in that $n$37-uniform hypergraph is therefore a set of vertex-disjoint triangles covering all vertices. For $n$38, almost all order-$n$39 Steiner triple systems have not merely one such triangle matching but

$n$40

perfect matchings, while every Steiner triple system has at most

$n$41

perfect matchings. Thus almost all Steiner triple systems asymptotically attain the extremal scale for the number of spanning triangle matchings (Kwan, 2016).

Latin squares give the tripartite analogue. A Latin square is a triangle decomposition of $n$42, and a partial Latin square is an edge-disjoint triangle packing in that complete tripartite graph. The paper relates uniformly random Latin squares to the triangle removal process on $n$43, proves quasirandom triangle-count estimates of the form

$n$44

and uses this to compare random Latin-square initial segments with random greedy triangle packings. A transversal is the corresponding vertex-disjoint triangle matching, although the paper’s direct emphasis is edge-disjoint triangle decomposition rather than transversals (Kwan et al., 2021).

In regular hypergraphs, triangle-and-edge partitions arise as a hypergraph analogue of perfect matching theory. If $n$45 is $n$46-uniform and $n$47-regular, then either $n$48 has a perfect matching when $n$49 is even, or $n$50 is factor-critical when $n$51 is odd; in the latter case the vertex set can be partitioned into one triangle $n$52 and a perfect matching of $n$53. The paper extends this to $n$54-uniform, $n$55-regular hypergraphs: $n$56 can always be partitioned into hyperedges of the hereditary closure of size $n$57 and at most one hyperedge of size $n$58 (Sebő, 2019).

A higher-dimensional analogue appears in cell complexes. In a triangulated surface, a triangle is matched not to another triangle but to an incident edge in the Hasse-diagram sense. A matching on a pair $n$59 is a partition of the cells of $n$60 into incident codimension-1 pairs. The relative Euler characteristic condition $n$61 is necessary, and for compact connected $n$62-manifolds with $n$63, every polyhedral cellulation is matchable absolutely and relatively to $n$64 (Meigniez, 2018).

6. Triangle-restricted matching regimes and recurring structural themes

Several adjacent theories study ordinary matchings under triangle restrictions rather than triangle packings themselves. In triangle-free graphs, the number of maximal induced matchings is at most

$n$65

and this bound is tight for disjoint unions of $n$66. As an algorithmic consequence, all maximal induced matchings can be listed in

$n$67

yielding the same running time for maximum induced matching in triangle-free graphs (Basavaraju et al., 2013). This is an induced analogue of the broader phenomenon that forbidding triangles can substantially improve matching structure and enumeration bounds.

Equimatchability imposes a different global restriction. A graph is equimatchable if all maximal matchings have the same size. For connected triangle-free equimatchable graphs, the complete structural classification is: either the graph is bipartite equimatchable, or it is $n$68, or $n$69, or it belongs to one of the explicit non-bipartite families $n$70 and $n$71 built from complete bipartite pieces attached to an induced odd cycle. For non-bipartite inputs, this yields a linear-time recognition algorithm via modular decomposition and a fixed twin-free template $n$72 (Büyükçolak et al., 2018).

Another line of work studies mixed triangle-and-matching constraints in edge-colored complete graphs. For a family $n$73 of color classes in which each $n$74 is either a 2-edge matching or a triangle, the rainbow girth satisfies

$n$75

for some constant $n$76. In the two auxiliary regimes with singleton edges, the paper identifies exact threshold phenomena: when color classes are 2-edge matchings or single edges, logarithmic rainbow girth begins precisely above matching proportion $n$77; when they are triangles or single edges, logarithmic rainbow girth holds for every positive linear proportion of triangles, and the threshold is $n$78 (Guo, 2022). This suggests that structurally constrained triangle classes behave very differently from arbitrary edge sets, often replacing linear extremal scales by logarithmic ones.

Across these strands, a common pattern recurs. Triangle matchings are tractable when triangles supply controlled local structure—connected blocks in Ramsey theory, excluded local cycles in $n$79-matchings, empty geometric witnesses, or rigid $n$80-uniform hyperedges—but become much harder when those local triangle configurations overlap freely. A plausible implication is that the decisive distinction is often not between “triangles present” and “triangles absent,” but between triangle systems that admit local decomposition, shrinking, or parity control and those that do not.