IV-matching is strongly NP-hard

IV-matching is a generalization of perfect bipartite matching. The complexity of finding IV-matching in a graph was posted as an open problem at the ICALP 2014 conference. In this note, we resolve the question and prove that, contrary to the expectations of the authors, the given problem is strongly NP-hard (already in the simplest non-trivial case of four layers). Hence it is unlikely that there would be an efficient (polynomial or pseudo-polynomial) algorithm solving the problem.

💡 Research Summary

The paper studies the computational complexity of the IV‑matching problem, a natural generalization of perfect bipartite matching to multi‑layered bipartite graphs. A layered graph consists of ℓ ordered layers V₁,…,V_ℓ; each layer is partitioned into clusters, and edges exist only between consecutive layers. Between two clusters a “macro‑edge” either creates a complete bipartite connection or leaves the pair disconnected. The crucial structural restriction is that macro‑edges between even‑indexed layers (V₂k and V₂k+1) must form a matching, while there is no restriction between odd‑indexed layers (V₂k‑1 and V₂k).

An IV‑matching M ⊆ E is defined by two local degree constraints:

1. Every vertex in an even layer V₂k is incident to exactly one edge, which must connect it either to the previous odd layer V₂k‑1 or to the next odd layer V₂k+1.
2. Every vertex in an odd layer V₂k‑1 must satisfy exactly one of the following exclusive options: – it is incident to two vertices in the preceding even layer V₂k‑2 (forming an “I‑shape”), or – it is incident to a single vertex in the succeeding even layer V₂k (forming a “V‑shape”).

Consequently, edges between V₂k‑1 and V₂k always constitute a matching (the I‑shapes), while edges between V₂k and V₂k+1 consist of independent V‑shapes whose centers lie in V₂k+1.

The decision version asks whether a given layered graph admits an IV‑matching. For ℓ = 2 the problem collapses to ordinary bipartite matching, and for odd ℓ the last layer can be matched uniquely, reducing the instance to ℓ‑1 layers. Hence the first non‑trivial case is ℓ = 4.

The core contribution is a proof that IV‑matching is strongly NP‑complete already for ℓ = 4. The authors give a polynomial‑time reduction from the classic 3‑dimensional matching (3DM) problem, which is known to be NP‑complete (Karp’s 17th problem). An instance of 3DM consists of three equally sized sets X, Y, Z (|X| = |Y| = |Z| = n) and a collection F of hyperedges, each hyperedge being a triple (x, y, z) with x ∈ X, y ∈ Y, z ∈ Z. The question is whether there exists a perfect matching, i.e., a subset of n hyperedges that covers every element exactly once.

Reduction construction:

• Place each vertex of X and Y as singleton clusters in layer V₁, and each vertex of Z as a singleton cluster in layer V₄.
• For every hyperedge e = (x, y, z) ∈ F, create a new cluster {xₑ, yₑ} in layer V₂ (two vertices) and a singleton cluster {zₑ} in layer V₃.
• Add four macro‑edges:
  1. Between the singleton cluster {x} in V₁ and the cluster {xₑ, yₑ} in V₂.
  2. Between the singleton cluster {y} in V₁ and the same cluster {xₑ, yₑ} in V₂.
  3. Between the cluster {xₑ, yₑ} in V₂ and the singleton {zₑ} in V₃.
  4. Between the singleton {zₑ} in V₃ and the singleton {z} in V₄.

All macro‑edges induce complete bipartite connections between the involved clusters. The total number of vertices in the constructed graph is 3n + 3|F|, and the number of macro‑edges is 4|F|, both polynomial in the size of the 3DM instance.

Correctness argument:

• (⇒) Suppose the constructed graph G admits an IV‑matching M. For each hyperedge e, the three new vertices (xₑ, yₑ, zₑ) must be either all matched via I‑shapes (i.e., each paired with its original counterpart in V₁ or V₄) or all matched via a single V‑shape (i.e., the three form a V‑shape centered at V₂ or V₃). Because each vertex in V₁ ∪ V₄ must be incident to exactly one I‑shape, the set of hyperedges whose three new vertices are matched by I‑shapes yields a collection of hyperedges covering all original vertices exactly once. The counting argument shows that exactly n hyperedges are selected, thus forming a perfect 3‑dimensional matching in H.

• (⇐) Conversely, given a perfect matching N ⊆ F in the original hypergraph, for each e ∈ N we realize the three new vertices (xₑ, yₑ, zₑ) as I‑shapes: connect x to xₑ, y to yₑ, and z to zₑ. For each e ∉ N we use a V‑shape: connect xₑ and yₑ to each other (forming the I‑shape between V₁ and V₂) and then connect the pair {xₑ, yₑ} to zₑ (forming the V‑shape between V₂ and V₃). This yields a valid IV‑matching covering all vertices, respecting the degree constraints.

Thus, an IV‑matching exists in G if and only if a perfect 3‑dimensional matching exists in H, establishing NP‑hardness. Since the reduction uses only clusters of size at most two, the problem remains NP‑hard even when the input is encoded in unary; consequently the problem is strongly NP‑complete.

Implications: The result disproves the earlier conjecture that IV‑matching might be solvable in polynomial or pseudo‑polynomial time. It also clarifies the complexity landscape: the problem is trivial for two layers, reducible for odd numbers of layers, and already intractable for four layers. This has direct consequences for algorithms dealing with regular graph covers, as the IV‑matching subroutine appears in the work of Fiala et al. (2014). Future research may focus on restricted graph families, approximation schemes, or parameterized algorithms, but a general efficient algorithm is unlikely unless P = NP.