Breaking the Exponential Barrier: The First Polynomial-Time Algorithm for the Győri-Lovász Theorem

Abstract: We give the first polynomial-time algorithm, after half a century, for the celebrated Győri-Lovász theorem, which resolved a conjecture of Frank (1975). The theorem, one of the simplest existential theorems to explain, states that every $k$-connected graph can be partitioned into $k$ disjoint connected subgraphs of arbitrary prescribed positive sizes. This is a fundamental structural result with broad applications, such as flexible allocation of connected subnetworks of prescribed sizes in sufficiently connected cloud infrastructures. While Lovász (1977) gave a highly non-constructive proof for a stronger directed version using algebraic topology, Győri's original constructive proof (1976) requires exponential time. Despite more than 50 years of effort, no polynomial-time algorithm was known even for $k>4$. Determining the computational complexity of the Győri-Lovász theorem---whether it admits even a sub-exponential-time algorithm or is computationally hard (in particular, PLS-complete or PPAD)---has remained one of the central open problems in algorithmic graph theory. In this paper, we finally resolve this long-standing problem by a fundamentally new proof of the existential theorem via introducing the novel concept of \emph{flow-essential assignment}, which genuinely marries matching and cut structures and yields the first polynomial-time constructive algorithm for the Győri-Lovász theorem. In fact, we obtain a polynomial-time algorithm for Lovász's stronger directed version, whose proof was non-constructive even for DAGs; for DAGs, we further obtain a near-linear-time algorithm. We also develop polynomial-time algorithms for weighted generalizations where the seminal work of Chen, Kleinberg, Lovász, Rajaraman, Sundaram, and Vetta (JACM'07) on confluent flows established only existential non-constructive results.

• The paper provides the first polynomial-time algorithm for the Győri–Lovász theorem, addressing the prescribed connected partition problem in general graphs.
• The algorithm replaces connectivity preservation with flow-essential assignments, derived from minimum cuts, and combines matching arguments, residual-flow structure, edge deletions, and contractions.
• This method constructs partition effectively for Lovász’s directed theorem, weighted extensions and has near-linear run time for directed acyclic graphs.

The paper resolves the algorithmic status of the Győri–Lovász theorem by giving the first polynomial-time construction for the prescribed connected partition problem in general graphs. Its central contribution is not merely an implementation of the classical proof, but a reformulation of the invariant that makes constructive progress possible. The algorithm replaces direct preservation of connectivity by a capacitated assignment condition derived from minimum cuts, and then combines matching arguments, residual-flow structure, edge deletions, and contractions. The resulting framework also yields constructive algorithms for Lovász’s stronger directed theorem, weighted extensions associated with confluent flows, and a near-linear-time algorithm for directed acyclic graphs (2608.30945).

Problem setting and historical barrier

In its directed form, the problem is specified by a digraph $G=(V,E)$, terminals $T=\{t_1,\ldots,t_k\}$, and nonnegative capacities $c_1,\ldots,c_k$ satisfying

$\sum_{i=1}^k c_i = |V\setminus T|.$

The graph is $k$-connected to $T$ if every non-terminal vertex has $k$ vertex-disjoint directed paths to distinct terminals. The objective is to partition $V$ into sets $V_1,\ldots,V_k$ such that $t_i\in V_i$, each part contains exactly $T=\{t_1,\ldots,t_k\}$0 non-terminals, and every vertex in $T=\{t_1,\ldots,t_k\}$1 has a directed path to $T=\{t_1,\ldots,t_k\}$2 within the induced subgraph $T=\{t_1,\ldots,t_k\}$3.

The undirected Győri–Lovász theorem follows by replacing every undirected edge with two oppositely directed edges. It therefore suffices to solve the directed formulation. The theorem itself has long been known existentially: Győri’s constructive proof recursively manipulates connected parts, but its cascades can have unbounded length and proliferate exponentially. Lovász’s proof establishes a stronger directed statement through algebraic topology, but does not provide an efficient search procedure. Consequently, even though the problem lies in the total-search landscape, no polynomial-time algorithm was known for general $T=\{t_1,\ldots,t_k\}$4; previous polynomial-time results covered only small fixed values of $T=\{t_1,\ldots,t_k\}$5 or restricted graph classes. The paper reports a particularly strong resolution: the first polynomial-time algorithm for arbitrary $T=\{t_1,\ldots,t_k\}$6, including every fixed $T=\{t_1,\ldots,t_k\}$7, rather than only an improvement to the known exponential local-search bounds (2608.30945).

The computational difficulty is not caused by the connectivity hypothesis alone. A natural contraction strategy would preserve $T=\{t_1,\ldots,t_k\}$8-connectivity to the terminal set after every operation, but the paper gives examples of graphs that are $T=\{t_1,\ldots,t_k\}$9-connected to their terminals in which contracting any pre-terminal destroys that property. Thus, the invariant required by a constructive proof cannot simply be the original connectivity assumption.

Flow-essential assignments

The paper’s main conceptual device is the Flow-Essential Assignment Condition. For a non-terminal vertex $c_1,\ldots,c_k$0, let $c_1,\ldots,c_k$1 denote the maximum number of vertex-disjoint paths from $c_1,\ldots,c_k$2 to distinct terminals. A terminal $c_1,\ldots,c_k$3 is flow-essential for $c_1,\ldots,c_k$4 if deleting $c_1,\ldots,c_k$5 decreases this connectivity by exactly one:

$c_1,\ldots,c_k$6

Equivalently, every maximum family of paths from $c_1,\ldots,c_k$7 to distinct terminals must contain a path ending at $c_1,\ldots,c_k$8.

The cut characterization is central. By Menger’s theorem, $c_1,\ldots,c_k$9 equals the minimum size of a vertex separator separating $\sum_{i=1}^k c_i = |V\setminus T|.$0 from $\sum_{i=1}^k c_i = |V\setminus T|.$1. The paper proves that the essential terminals of $\sum_{i=1}^k c_i = |V\setminus T|.$2 are exactly the terminals contained in the separator of the tightest minimum cut, defined as the intersection of all minimum cuts separating $\sum_{i=1}^k c_i = |V\setminus T|.$3 from $\sum_{i=1}^k c_i = |V\setminus T|.$4. Minimum cuts are closed under the paper’s union and intersection operations, which gives this canonical cut and makes essentiality algorithmically accessible.

A flow-essential assignment is a map $\sum_{i=1}^k c_i = |V\setminus T|.$5 satisfying two conditions:

1. $\sum_{i=1}^k c_i = |V\setminus T|.$6 is essential for every non-terminal $\sum_{i=1}^k c_i = |V\setminus T|.$7.
2. Exactly $\sum_{i=1}^k c_i = |V\setminus T|.$8 vertices are assigned to terminal $\sum_{i=1}^k c_i = |V\setminus T|.$9.

The first condition encodes the cut structure; the second is a capacitated bipartite matching constraint. This is the paper’s decisive synthesis: prescribed cardinalities are treated as a matching problem, while connectivity is represented through essential terminals identified by minimum cuts.

The condition is strictly weaker than $k$0-connectivity to $k$1. In a $k$2-connected instance, every terminal is essential for every vertex, so any assignment satisfying the capacities is valid. At the opposite extreme, a disjoint union of $k$3 appropriately sized arborescences can satisfy the condition even though the graph is far from $k$4-connected. This weakening is what allows the invariant to survive reductions that would destroy ordinary connectivity.

The condition is polynomial-time checkable. For each non-terminal vertex, one maximum-flow computation on a vertex-split network identifies the tightest minimum cut and therefore all essential terminals. A single bipartite maximum-flow computation then determines whether the capacity constraints can be simultaneously satisfied. Using the cited deterministic almost-linear-time minimum-cost-flow and maximum-flow machinery, the essential-terminal preprocessing for one vertex takes $k$5 time (2608.30945).

The contraction-and-deletion algorithm

The unweighted algorithm maintains a graph, terminal capacities, and a witnessing flow-essential assignment. It repeatedly applies one of three operations.

First, a terminal with zero residual capacity is removed. Since no vertex is assigned to it, deleting it preserves the essentiality of all terminals assigned to other vertices.

Second, a pre-terminal of out-degree one is contracted into its unique adjacent terminal. The contraction is reversible at the level of path families: every path using the pre-terminal must continue through its sole outgoing edge, so paths before and after contraction correspond without changing their terminal endpoints. The terminal capacity is decreased by one.

The third operation is the nontrivial one. When all capacities are positive and every pre-terminal has at least two outgoing edges, the algorithm first finds a matching from terminals to distinct pre-terminals. The existence of this matching is proved by Hall’s theorem. If a minimal Hall-deficient terminal set $k$6 existed, the paper shows that no terminal in $k$7 could be essential for a vertex outside its pre-terminal neighborhood. Since every terminal in $k$8 has positive capacity, the witness would need to assign at least one distinct vertex to each of these terminals, producing more neighboring pre-terminals than Hall deficiency permits.

For each matched pair $k$9, the algorithm selects a secondary outgoing edge $T$0 of $T$1, distinct from the matching edge $T$2. It seeks one secondary edge whose deletion preserves the current witness. The obstacle is that every such edge may initially be critical for some assignment. The paper overcomes this by changing the witness rather than insisting on preserving the current assignment.

If $T$3 is critical for assigning a vertex $T$4 to its current terminal, then a cut-based rerouting argument shows that $T$5 is also essential for $T$6. Hence $T$7 can be reassigned to $T$8. These possible reassignments form a directed graph on the terminals, with an arc from the current terminal of $T$9 to $k$0. Every terminal has in-degree one, so the graph contains a directed cycle. Shifting all assignments along such a cycle preserves every capacity exactly and preserves essentiality.

The cycle shift is certified by a potential function. For an assignment $k$1, the criticality cost $k$2 is the number of secondary edges critical for assigning $k$3 to $k$4, and

$k$5

The paper proves that each cycle shift strictly decreases $k$6. The proof depends on two cut lemmas. The matching edge $k$7 ensures that the corresponding secondary edge $k$8 is never critical for assigning a vertex to $k$9. Conversely, any other secondary edge critical for assignment to $V$0 is also critical for the vertex’s previous terminal. Thus the new assignment has a strict criticality-cost advantage.

Since $V$1, at most $V$2 cycle shifts occur before some secondary edge becomes non-critical. That edge can then be deleted while preserving the existence of a valid witness. The algorithm therefore makes progress even in instances where no connectivity-preserving contraction exists. Every recursive call removes at least one vertex or edge, so there are at most $V$3 calls.

The strengthened theorem established by this argument is more general than the original Győri–Lovász statement:

Any directed instance satisfying the Flow-Essential Assignment Condition admits a polynomial-time partition into terminal-rooted connected parts of exactly the prescribed cardinalities.

The classical theorem follows because $V$4-connectivity to $V$5 makes every terminal essential for every vertex.

Weighted generalization

The weighted extension assigns each non-terminal vertex $V$6 a positive integer weight $V$7 and each terminal $V$8 a capacity $V$9. The objective is to produce connected terminal-rooted parts satisfying

$V_1,\ldots,V_k$0

where $V_1,\ldots,V_k$1.

The additive violation is unavoidable. With $V_1,\ldots,V_k$2 terminals of capacity one and a single vertex of weight $V_1,\ldots,V_k$3, the graph can be $V_1,\ldots,V_k$4-connected to the terminals, but the vertex must belong to one part, whose weight exceeds its capacity by $V_1,\ldots,V_k$5. Thus the algorithm achieves the optimal form of the known existential weighted guarantee rather than introducing an artifact of its construction.

Exact assignments are insufficient in the weighted case because one vertex’s weight may be too large to assign integrally to a single terminal without violating capacities. The paper therefore introduces a Flow-Essential Split-Assignment. A vertex can distribute its weight among several essential terminals, subject to terminal capacity upper bounds. This is represented as a capacitated bipartite flow: each vertex supplies $V_1,\ldots,V_k$6 units, each terminal accepts at most $V_1,\ldots,V_k$7 units, and an arc exists only for an essential vertex-terminal pair.

The weighted potential is

$V_1,\ldots,V_k$8

Unlike the unweighted potential, this quantity can be exponentially large when weights are encoded in binary. The paper therefore avoids unit-by-unit local search and computes a minimum-potential split-assignment directly by one minimum-cost-flow computation. If every secondary edge were critical for some positive-flow assignment pair, the same reassignment-cycle argument would produce a strictly cheaper split-assignment, contradicting optimality. Consequently, a removable secondary edge is found in polynomial time despite exponentially large numerical weights.

The weighted case also introduces a new situation: the matching from terminals to pre-terminals may fail to exist. Let $V_1,\ldots,V_k$9 be an inclusion-minimal Hall-deficient set. The algorithm matches all but one terminal in $t_i\in V_i$0 to distinct pre-terminals in its neighborhood, creates singleton or two-vertex completed parts, and removes $t_i\in V_i$1 together with its pre-terminal neighborhood. The paper proves that the restricted split-assignment remains valid after this rounding operation. The proof uses a two-stage transformation of minimum cuts: first replace the deficient terminals by their pre-terminal neighborhood, then remove those temporary terminals one at a time.

This rounding step is the sole source of the $t_i\in V_i$2 additive error. Each rounded terminal receives at most one pre-terminal, whose weight is at most $t_i\in V_i$3, while its positive integral capacity is at least one. The resulting general weighted algorithm runs in

$t_i\in V_i$4

time, as stated in the paper (2608.30945). The dependence on $t_i\in V_i$5 rather than $t_i\in V_i$6 is important: the algorithm remains polynomial in the binary input size even when vertex weights are exponentially large.

The weighted theorem yields a constructive version of the existence result underlying the confluent-flow bounds of Chen, Kleinberg, Lovász, Rajaraman, Sundaram, and Vetta. In that setting, the partition corresponds to routing demand toward sinks under confluent-flow constraints. The paper supplies the algorithmic component that had been absent from the earlier topological proof.

Near-linear time for DAGs

For directed acyclic graphs, the paper obtains a substantially faster algorithm that bypasses the general flow-essential machinery. A DAG is $t_i\in V_i$7-connected to its terminal set if and only if every non-terminal vertex has out-degree at least $t_i\in V_i$8. This converts the global linkage condition into a local degree invariant.

The algorithm fixes a canonical topological order in which every non-terminal precedes all terminals. For each terminal, it maintains a heap of pre-terminals currently adjacent to its growing part. Whenever the terminal has positive residual capacity, the algorithm contracts the earliest available pre-terminal. The earliest choice is safe: if a remaining vertex lost an outgoing neighbor under the contraction, it would have had edges both to the contracted pre-terminal and to the terminal’s current part. Such a vertex would be an earlier candidate in the same heap, contradicting the extraction rule.

Thus every contraction preserves the out-degree lower bound and hence $t_i\in V_i$9-connectivity. The heap implementation performs $T=\{t_1,\ldots,t_k\}$00 insertions and extractions, each in $T=\{t_1,\ldots,t_k\}$01 time. The resulting running time is

$T=\{t_1,\ldots,t_k\}$02

for both the weighted and unweighted settings (2608.30945). In the unweighted case, the weighted bound reduces to exact prescribed cardinalities because the total capacity equals the number of non-terminals and $T=\{t_1,\ldots,t_k\}$03.

This DAG result is notable because acyclicity does not make the underlying confluent-flow problem trivial; the paper explicitly treats the DAG algorithm as a distinct structural simplification rather than as a direct consequence of the general algorithm.

Relaxed connectivity conditions

The paper further separates the connectivity needed for existence from the stronger invariant needed by the algorithm. It defines local connectivity and compact connectivity, obtaining the hierarchy

$T=\{t_1,\ldots,t_k\}$04

Under local connectivity, the paper adapts Győri’s cascade argument to prove existence of a partition with prescribed sizes. The proof is constructive in principle but remains exponential because the cascade configuration space can be explored only through a lexicographically increasing potential bounded by the number of vertices, without yielding a polynomial bound on the number or cost of configuration updates.

Compact connectivity admits a stronger algorithmic treatment. It can be recognized in polynomial time through flow computations and a final capacitated matching problem. The paper proves that compact connectivity implies the Flow-Essential Assignment Condition, so the main polynomial-time algorithm applies to compact-connected instances even when they are not $T=\{t_1,\ldots,t_k\}$05-connected.

The distinction between compact connectivity and flow-essential assignment is substantive. The paper constructs a compact-connected instance with nine terminals in which every pre-terminal has out-degree exactly two and every edge is critical for compact connectivity. Therefore, no iterative algorithm based on preserving compact connectivity under edge deletion can make progress through the same reduction scheme. Nevertheless, the instance satisfies the flow-essential assignment condition, whose weaker essentiality invariant supports the required edge deletions and contractions. This example demonstrates that the algorithm’s success depends on selecting the right invariant, not merely on weakening connectivity quantitatively.

The paper leaves two questions open within this hierarchy. It conjectures that every locally connected instance admits either a removable zero-capacity terminal or a contractible pre-terminal preserving local connectivity. It also conjectures that checking local connectivity may be computationally difficult when $T=\{t_1,\ldots,t_k\}$06 is part of the input. These questions are not required for the main polynomial-time result, because the flow-essential condition is already efficiently testable and maintainable.

Limitations and open questions

The headline polynomial-time result should be interpreted with the assumptions made explicit. The algorithm is polynomial in the standard binary encoding of graph data and weights, but the paper does not present a sharply optimized low-degree polynomial bound for the general weighted case; its stated bound contains repeated near-linear flow computations and a quadratic dependence on the edge count. The near-linear $T=\{t_1,\ldots,t_k\}$07 result is specific to DAGs and relies on the equivalence between terminal connectivity and minimum out-degree, which has no analogue in general directed graphs.

The weighted algorithm guarantees an additive $T=\{t_1,\ldots,t_k\}$08 violation, and the paper proves that this error is unavoidable in general. It does not establish exact weighted capacities under the same hypotheses. The flow-essential split-assignment condition is stronger than mere existence of a partition, so the strengthened theorem does not characterize all feasible instances. Finally, the local-connectivity conjecture and the complexity of recognizing local connectivity remain unresolved in the paper’s framework.

Conclusion

The paper breaks the long-standing exponential barrier for the Győri–Lovász theorem by replacing direct connectivity preservation with a flow-essential assignment invariant. Tight minimum cuts identify permissible terminal assignments, capacitated matching enforces prescribed sizes, and reassignment cycles expose edges that can be safely deleted until contractions become possible. This yields the first polynomial-time construction for arbitrary $T=\{t_1,\ldots,t_k\}$09, extends constructively to Lovász’s directed theorem and weighted confluent-flow generalizations, and gives an $T=\{t_1,\ldots,t_k\}$10 algorithm for $T=\{t_1,\ldots,t_k\}$11-connected DAGs. The principal methodological conclusion is that the prescribed-size partition problem is governed jointly by cut structure and matching structure; treating either component in isolation is insufficient for polynomial-time construction (2608.30945).