Minimal Multiway Cut Conjecture

• Minimal Multiway Cut Conjecture is a collection of claims addressing the optimal removal of nonterminal elements to disconnect specified terminal sets in a graph.
• It encompasses diverse frameworks including max-flow min-cut solvability, polynomial-delay enumeration, approximation thresholds, and parameterized cuts above lower bounds.
• Understanding these formulations aids in designing efficient algorithms and reveals key structural insights for network reliability and graph optimization.

Searching arXiv for papers on the Minimal Multiway Cut Conjecture and closely related formulations. arXiv search query: "Minimal Multiway Cut Conjecture multiway cut Mader-Mengerian important separators enumeration approximation" The Minimal Multiway Cut Conjecture is not a single universally fixed statement but a family of conjectural claims centered on separating a prescribed terminal set in a graph. In its common core, the problem asks for a minimum set of nonterminals or edges whose removal destroys all paths between every pair of distinct terminals. In the literature represented by Jost and Naves, by Kanté, Limouzy, Mary, and Nourine, by Ene, Vondrák, and Wu, and by Xiao, the same label is used for at least four distinct research programs: exact polynomial-time solvability for fixed terminal number, polynomial boundedness of the number of inclusion-wise minimal cuts, optimal approximation ratio, and fixed-parameter tractability above the largest minimum isolating cut (Naves et al., 2011, Kurita et al., 2020, Ene et al., 2015, Razgon, 2010).

1. Problem setting and principal formulations

For an undirected graph $G=(V,E)$ and terminals $T=\{t_1,\dots,t_k\}\subseteq V$, a vertex multiway cut is a set $S\subseteq V\setminus T$ such that in $G-S$ no two distinct terminals lie in the same connected component. The edge version asks for a set $M\subseteq E$ whose removal destroys all paths between any two distinct terminals. A cut is minimal if no proper subset is again a multiway cut; in the edge case this is equivalent to requiring that for every $e\in M$, the set $M\setminus\{e\}$ fails to separate some pair of terminals (Kurita et al., 2020).

A second exact formulation uses an independent terminal set $\mathcal{S}\subseteq V(G)$ together with a weight function $w:V(G)\to \mathbb{N}$. An $\mathcal{S}$-path is a simple path whose two distinct ends lie in $T=\{t_1,\dots,t_k\}\subseteq V$0 and whose internal vertices lie in $T=\{t_1,\dots,t_k\}\subseteq V$1. One then compares $T=\{t_1,\dots,t_k\}\subseteq V$2, the maximum $T=\{t_1,\dots,t_k\}\subseteq V$3-packing of $T=\{t_1,\dots,t_k\}\subseteq V$4-paths, with $T=\{t_1,\dots,t_k\}\subseteq V$5, the minimum weight of a vertex cut separating $T=\{t_1,\dots,t_k\}\subseteq V$6 (Naves et al., 2011).

In the approximation literature, Graph-Multiway Cut, Node-weighted Multiway Cut, Hypergraph Multiway Cut, and Submodular Multiway Partition are treated in a common framework. Graph-Multiway Cut removes weighted edges; Node-weighted Multiway Cut removes weighted nonterminals; Hypergraph Multiway Cut removes hyperedges; and Submodular Multiway Partition seeks a partition $T=\{t_1,\dots,t_k\}\subseteq V$7 with $T=\{t_1,\dots,t_k\}\subseteq V$8 minimizing $T=\{t_1,\dots,t_k\}\subseteq V$9 for a submodular function $S\subseteq V\setminus T$0 (Ene et al., 2015).

A further parameterized formulation uses isolating cuts. For a terminal $S\subseteq V\setminus T$1, an isolating cut is a set of nonterminals separating $S\subseteq V\setminus T$2 from $S\subseteq V\setminus T$3. Writing

$S\subseteq V\setminus T$4

the quantity $S\subseteq V\setminus T$5 is a polynomially computable lower bound on the size of any multiway cut (Razgon, 2010).

2. Exact solvability, max-flow–min-cut, and Mader–Mengerian graphs

In the exact-optimization line, the conjecture is formulated as follows: for every fixed $S\subseteq V\setminus T$6 there is a polynomial-time algorithm that, given $S\subseteq V\setminus T$7 and $S\subseteq V\setminus T$8, computes a minimum $S\subseteq V\setminus T$9-vertex-cut. Jost and Naves do not prove this in full generality, but they identify the class on which the problem is exactly “flow-solvable,” namely the Mader–Mengerian graphs (Naves et al., 2011).

For an independent terminal set $G-S$0, the relevant primal–dual pair is the vertex-cut LP

$G-S$1

and the path-packing dual

$G-S$2

The key equivalence is that $G-S$3 is Mader–Mengerian if and only if this system is totally dual integral, equivalently if and only if the polyhedron $G-S$4 is integral (Naves et al., 2011).

The structural theorem states that a graph $G-S$5 is Mader–Mengerian for all independent terminal sets $G-S$6 and all $G-S$7 if and only if $G-S$8 does not contain, as a vertex-minor, any graph in $G-S$9. Here $M\subseteq E$0 is built from an odd cycle with vertices colored by $M\subseteq E$1 colors, together with one terminal per color class, with each cycle vertex adjacent only to the terminal of its own color. The case $M\subseteq E$2 yields an infinite family of minimal obstructions; the simplest is the “net,” a $M\subseteq E$3-cycle with three terminal leaves. Corollary 10 reduces the verification of the property to all terminal triples $M\subseteq E$4 of size $M\subseteq E$5, so the obstruction theory is controlled entirely by $M\subseteq E$6 (Naves et al., 2011).

This characterization has algorithmic consequences. For a fixed pair $M\subseteq E$7 one builds an auxiliary graph $M\subseteq E$8 by deleting terminals and any nonterminal adjacent to at least two terminals, contracting each remaining nonterminal not adjacent to $M\subseteq E$9, and deleting edges whose ends are both neighbors of the same terminal. Then $e\in M$0 is TDI if and only if $e\in M$1 is bipartite. Running this test over all $e\in M$2 terminal triples yields an $e\in M$3 recognition algorithm for the Mader–Mengerian class (Naves et al., 2011).

When $e\in M$4 is Mader–Mengerian, the same auxiliary-graph construction reduces both the maximum packing problem and the minimum cut problem to a single vertex-capacitated max-flow computation in a bipartite network. The resulting max-flow value equals $e\in M$5, and the flow yields both a maximum packing of $e\in M$6-paths and a minimum $e\in M$7-cut. This isolates a largest natural class on which the vertex multiway-cut problem is solvable by a single max-flow (Naves et al., 2011).

3. The enumeration conjecture for minimal multiway cuts

In the enumeration literature, the Minimal Multiway-Cut Conjecture has a different meaning. For every fixed $e\in M$8, it asks whether the number of inclusion-wise minimal $e\in M$9-multiway cuts in an $M\setminus\{e\}$0-vertex graph is bounded by a polynomial in $M\setminus\{e\}$1; equivalently, whether all minimal multiway cuts can be enumerated in total time $M\setminus\{e\}$2 for each fixed $M\setminus\{e\}$3. This is known for $M\setminus\{e\}$4, where there are $M\setminus\{e\}$5 minimal $M\setminus\{e\}$6–$M\setminus\{e\}$7 cuts, but it remains open for $M\setminus\{e\}$8 (Kurita et al., 2020).

The paper on efficient enumeration establishes strong output-sensitive guarantees without resolving the counting question. For minimal node multiway cuts, Algorithm 1 with Neighborhood$M\setminus\{e\}$9 in Algorithm 4 enumerates all minimal node multiway cuts in $\mathcal{S}\subseteq V(G)$0 delay and exponential space. For minimal edge multiway cuts, Algorithm 3 runs in $\mathcal{S}\subseteq V(G)$1 delay and $\mathcal{S}\subseteq V(G)$2 space. The node version uses proximity search plus memoization and a breadth-first traversal of a strongly connected solution graph; the edge version uses reverse-search in the sense of Avis–Fukuda (Kurita et al., 2020).

The structural characterizations are exact. A node cut $\mathcal{S}\subseteq V(G)$3 is minimal if and only if $\mathcal{S}\subseteq V(G)$4 has exactly $\mathcal{S}\subseteq V(G)$5 components $\mathcal{S}\subseteq V(G)$6 with $\mathcal{S}\subseteq V(G)$7, and for each $\mathcal{S}\subseteq V(G)$8 there exist $\mathcal{S}\subseteq V(G)$9 with $w:V(G)\to \mathbb{N}$0 and $w:V(G)\to \mathbb{N}$1. An edge set $w:V(G)\to \mathbb{N}$2 is a minimal edge $w:V(G)\to \mathbb{N}$3-way cut if and only if $w:V(G)\to \mathbb{N}$4 has exactly $w:V(G)\to \mathbb{N}$5 components $w:V(G)\to \mathbb{N}$6 with $w:V(G)\to \mathbb{N}$7 (Kurita et al., 2020).

These results separate delay complexity from counting complexity. The paper states explicitly that the number of outputs can be exponentially many in $w:V(G)\to \mathbb{N}$8, that no nontrivial upper bound on the total number of minimal node or edge multiway cuts is proved beyond the trivial $w:V(G)\to \mathbb{N}$9, and that the conjecture’s strongest form—the existence of a polynomial bound on the total count for fixed $\mathcal{S}$0—remains open. Thus polynomial-delay enumeration does not imply a polynomial counting bound (Kurita et al., 2020).

4. Approximation-ratio interpretation and its resolution for broad generalizations

A third use of the name concerns approximability. In that formulation, the conjecture states that for every fixed integer $\mathcal{S}$1, the best achievable approximation ratio for $\mathcal{S}$2-terminal Multiway Cut, and for Node-weighted Multiway Cut, Hypergraph Multiway Cut, and Submodular Multiway Partition, is exactly $\mathcal{S}$3. Equivalently, the natural LP relaxation has integrality gap $\mathcal{S}$4, and no polynomial-time algorithm can beat $\mathcal{S}$5 assuming the Unique Games Conjecture; from the submodular viewpoint, the Lovász-extension relaxation has the same optimal symmetry gap (Ene et al., 2015).

The Min-CSP formulation uses variables $\mathcal{S}$6 and $\mathcal{S}$7 with

$\mathcal{S}$8

and consistency constraints

$\mathcal{S}$9

Its objective is

$T=\{t_1,\dots,t_k\}\subseteq V$00

For Submodular Multiway Partition, the Lovász relaxation minimizes

$T=\{t_1,\dots,t_k\}\subseteq V$01

subject to $T=\{t_1,\dots,t_k\}\subseteq V$02 and $T=\{t_1,\dots,t_k\}\subseteq V$03, where

$T=\{t_1,\dots,t_k\}\subseteq V$04

The relaxation is convex and solvable in polynomial time via submodular-function oracles (Ene et al., 2015).

The positive result is a randomized rounding with approximation ratio $T=\{t_1,\dots,t_k\}\subseteq V$05 for Submodular Multiway Partition. The matching hardness is twofold. Assuming the Unique Games Conjecture, for every fixed $T=\{t_1,\dots,t_k\}\subseteq V$06 and every $T=\{t_1,\dots,t_k\}\subseteq V$07, it is NP-hard to approximate Hypergraph Multiway Cut or Node-weighted Multiway Cut better than $T=\{t_1,\dots,t_k\}\subseteq V$08. Independently, for any fixed $T=\{t_1,\dots,t_k\}\subseteq V$09 and any $T=\{t_1,\dots,t_k\}\subseteq V$10, no randomized value-oracle algorithm can achieve approximation factor $T=\{t_1,\dots,t_k\}\subseteq V$11 for Submodular Multiway Partition in subexponential query time (Ene et al., 2015).

The paper also proves that integrality-gap instances for the Basic LP and symmetry-gap instances for the multilinear relaxation are equivalent in strength: every Basic-LP integrality gap $T=\{t_1,\dots,t_k\}\subseteq V$12 can be transformed into a symmetry-gap instance of the same value, and conversely. Within the broad classes of Multiway Cut, Node-Weighted and Hypergraph Multiway Cut, and Submodular Multiway Partition, this establishes the conjectured threshold $T=\{t_1,\dots,t_k\}\subseteq V$13 (Ene et al., 2015).

5. Parameterization above the largest isolating cut

A fourth conjectural direction concerns the parameterized complexity of Multiway Cut above the lower bound $T=\{t_1,\dots,t_k\}\subseteq V$14. The paper on computing multiway cut within the given excess over the largest minimum isolating cut gives an $T=\{t_1,\dots,t_k\}\subseteq V$15 algorithm that either computes a multiway cut of size at most $T=\{t_1,\dots,t_k\}\subseteq V$16 or reports that no such cut exists (Razgon, 2010).

The central combinatorial result is an enumeration bound for important separators. For disjoint $T=\{t_1,\dots,t_k\}\subseteq V$17, let

$T=\{t_1,\dots,t_k\}\subseteq V$18

Then for every integer $T=\{t_1,\dots,t_k\}\subseteq V$19, the number of important $T=\{t_1,\dots,t_k\}\subseteq V$20–$T=\{t_1,\dots,t_k\}\subseteq V$21 separators of size at most $T=\{t_1,\dots,t_k\}\subseteq V$22 is at most

$T=\{t_1,\dots,t_k\}\subseteq V$23

The proof proceeds through normalization, witnesses, and compound witnesses. In the normalized instance the only minimum separator is $T=\{t_1,\dots,t_k\}\subseteq V$24; each important separator of excess $T=\{t_1,\dots,t_k\}\subseteq V$25 is encoded by an attribute of rank at most $T=\{t_1,\dots,t_k\}\subseteq V$26; and each subset of total size at most $T=\{t_1,\dots,t_k\}\subseteq V$27 can serve as an attribute for at most one compound witness (Razgon, 2010).

This yields the XP algorithm. One computes a smallest important isolating cut for each terminal in $T=\{t_1,\dots,t_k\}\subseteq V$28, chooses a terminal $T=\{t_1,\dots,t_k\}\subseteq V$29 attaining the lower bound $T=\{t_1,\dots,t_k\}\subseteq V$30, enumerates all important isolating cuts $T=\{t_1,\dots,t_k\}\subseteq V$31 of $T=\{t_1,\dots,t_k\}\subseteq V$32 of size at most $T=\{t_1,\dots,t_k\}\subseteq V$33, and recurses on $T=\{t_1,\dots,t_k\}\subseteq V$34. Since there are at most

$T=\{t_1,\dots,t_k\}\subseteq V$35

such cuts, and each can be generated in $T=\{t_1,\dots,t_k\}\subseteq V$36, the total running time is $T=\{t_1,\dots,t_k\}\subseteq V$37 (Razgon, 2010).

The associated open problem is whether Multiway Cut parameterized by the excess $T=\{t_1,\dots,t_k\}\subseteq V$38 above the largest isolating-cut lower bound is fixed-parameter tractable or W[1]-hard. The result places the problem in XP but does not supply an FPT algorithm. The paper notes that achieving an $T=\{t_1,\dots,t_k\}\subseteq V$39 bound would require either a combinatorial improvement reducing the number of important separators of excess $T=\{t_1,\dots,t_k\}\subseteq V$40 to a function of $T=\{t_1,\dots,t_k\}\subseteq V$41 alone, or a different technique that bypasses the enumeration of all size-$T=\{t_1,\dots,t_k\}\subseteq V$42 attributes (Razgon, 2010).

6. Relation among the conjectures and current status

The different statements called the Minimal Multiway Cut Conjecture are linked by subject matter but are not equivalent. One concerns exact polynomial-time solvability for fixed $T=\{t_1,\dots,t_k\}\subseteq V$43 via max-flow–min-cut structure; one concerns the number of inclusion-wise minimal cuts; one concerns the optimal approximation ratio; and one concerns fixed-parameter tractability above a lower bound. A plausible implication is that the phrase should be interpreted from local context rather than treated as a single canonical conjecture.

The exact-solvability program has a complete answer on the Mader–Mengerian class: the relevant graphs are exactly those excluding $T=\{t_1,\dots,t_k\}\subseteq V$44 as a vertex-minor, they can be recognized in $T=\{t_1,\dots,t_k\}\subseteq V$45 time, and on them the cut problem reduces to a single max-flow computation (Naves et al., 2011). The enumeration program has polynomial-delay algorithms for both node and edge versions, but it does not yet establish any polynomial upper bound on the total number of minimal multiway cuts for fixed $T=\{t_1,\dots,t_k\}\subseteq V$46 (Kurita et al., 2020). The approximation program has the sharp threshold $T=\{t_1,\dots,t_k\}\subseteq V$47 for the classes studied, with matching upper and lower bounds under the stated assumptions (Ene et al., 2015). The parameterized “above lower bound” program has an XP algorithm with running time $T=\{t_1,\dots,t_k\}\subseteq V$48, while FPT and W[1]-hardness both remain unresolved (Razgon, 2010).

Two misconceptions are especially common. First, polynomial-delay enumeration does not prove that only polynomially many minimal multiway cuts exist; the enumeration paper explicitly leaves the total-count conjecture open (Kurita et al., 2020). Second, the exact max-flow–min-cut equality does not hold in general graphs; Jost and Naves show that it holds precisely on the vertex-minor-closed class excluding $T=\{t_1,\dots,t_k\}\subseteq V$49 (Naves et al., 2011).

Taken together, these results show that multiway cut admits several distinct “minimality” paradigms: minimal separating sets, minimum separators for fixed terminal sets, integrality of the standard relaxation, and excess over a natural lower bound. The conjectural frontier has therefore become stratified rather than uniform. Some versions are resolved on broad classes or under standard assumptions, while others remain open even for fixed terminal number.