Laminar Multiway Cuts

Updated 6 July 2026

Laminar multiway cuts are a structured family of non-crossing cuts that separate terminals while controlling edge load in capacitated graphs.
They transform arbitrary LP solutions into laminar fractional surrogates, yielding constant-factor approximations through precise rounding techniques.
This approach enhances algorithmic performance in applications like graph labeling and separator decompositions, reducing dependencies on logarithmic factors.

Searching arXiv for papers on laminar multiway cuts and closely related separator structures. First, I’ll look for the 2008 paper on packing multiway cuts and then for adjacent work on multiway cut structure and laminar separators. Laminar multiway cuts arise in multi-terminal cut settings in which the selected cuts are required, or shown, to form a non-crossing family. In the load-minimization formulation of multiway cut packing, the central role of laminarity is explicit: given an undirected graph $G=(V,E)$ and multiple commodities, one seeks a multiway cut for each commodity while minimizing the maximum edge load, and the key structural fact is that every feasible LP solution admits a near-optimal laminar fractional surrogate. This observation yields the first constant-factor approximation for arbitrary undirected graphs in the capacitated setting and makes laminarity a primary algorithmic device rather than merely a descriptive property (0810.0674).

1. Formal setting: multiway cut packing and capacitated load minimization

In the formulation studied in "Packing multiway cuts in capacitated graphs" (0810.0674), the input is an undirected graph $G=(V,E)$ together with $k$ commodities. Each commodity $a\in[k]$ has a terminal set $S_a$ , and each terminal $i\in S_a$ is located at a vertex $r_i\in V$ . The objective is to choose a collection of cuts

$\{E_1,\dots,E_k\},$

where $E_a$ is a multiway cut for commodity $a$ , meaning that for all $G=(V,E)$ 0, $G=(V,E)$ 1, the cut $G=(V,E)$ 2 disconnects $G=(V,E)$ 3 and $G=(V,E)$ 4.

The cost measure is edge load rather than total cut size. For a collection of cuts $G=(V,E)$ 5, the load of an edge $G=(V,E)$ 6 is

$G=(V,E)$ 7

If edges have capacities $G=(V,E)$ 8, the relevant quantity is the relative load

$G=(V,E)$ 9

and the capacitated objective is

$k$ 0

The paper assumes, without loss of generality, that the optimum relative load is $k$ 1, and then seeks an integral packing with small additive or multiplicative violation.

The LP relaxation assigns to each commodity $k$ 2 edge variables $k$ 3 satisfying metric constraints

$k$ 4

terminal separation constraints

$k$ 5

and capacity constraints

$k$ 6

This metric relaxation is the starting point for laminarization and rounding.

A special case is the common-sink $k$ 7- $k$ 8 cut packing problem, abbreviated CSCP. The paper develops separate laminarization and rounding procedures for CSCP and for the general multiway cut packing problem, abbreviated MCP.

2. Laminarity and fractional laminar cut families

The paper switches from edge cuts to vertex subsets: a cut is represented by a set $k$ 9, with corresponding edge cut $a\in[k]$ 0. Two cuts $a\in[k]$ 1 cross if all of

$a\in[k]$ 2

are non-empty. A family of cuts is laminar if no pair crosses. Algorithmically, this yields the classical nested-or-disjoint structure that supports deterministic rounding (0810.0674).

The paper defines a fractional laminar cut family for a terminal set $a\in[k]$ 3 and weight function $a\in[k]$ 4 as a laminar collection of cuts, each assigned to a unique terminal, such that

$a\in[k]$ 5

and every cut in $a\in[k]$ 6 contains the terminal vertex $a\in[k]$ 7. For MCP, feasibility further requires that terminals in the same commodity are separated in the weaker sense encoded in Definition 2 of the paper, and that edge loads satisfy

$a\in[k]$ 8

Laminarity is operational rather than merely structural. Because no pair of cuts crosses, one can identify innermost cuts, contract saturated regions into meta-nodes, and reassign fractional weight without introducing complex crossing dependencies. The paper’s rounding algorithms depend on precisely this hierarchical organization.

A common misconception is that laminarity should always be achievable at the original optimum load. The paper explicitly rules this out for MCP: not every feasible MCP solution can be made laminar at the same load, and Figure 1 gives an instance with no fractional laminar solution of load $a\in[k]$ 9. Thus laminarity is valuable, but it is not free.

3. Near-optimal laminarization and approximation guarantees

The main structural theorem is that every feasible LP solution can be converted into a near-optimal fractional laminar solution. This is the central reason laminar multiway cuts are algorithmically useful (0810.0674).

For CSCP, given a feasible LP solution $S_a$ 0, algorithm Lam-1 produces a fractional laminar family feasible with capacities

$S_a$ 1

For MCP, given a feasible LP solution $S_a$ 2, algorithm Lam-2 produces a fractional laminar family feasible with capacities

$S_a$ 3

Rounding these laminar fractional solutions yields the final integral guarantees:

Problem	Laminarization bound	Final integral load
CSCP	$S_a$ 4	$S_a$ 5
MCP	$S_a$ 6	$S_a$ 7

Equivalently, the main theorem is

$S_a$ 8

These bounds were the first constant-factor approximations for the problem in arbitrary undirected graphs. The paper situates them against the earlier work of Rabani, Schulman, and Swamy, who introduced multiway cut packing and obtained an $S_a$ 9 approximation in general graphs and an $i\in S_a$ 0 approximation in trees. The laminar approach therefore changes the dependence from logarithmic in graph size to constant in arbitrary undirected graphs.

The guarantees are complemented by lower-bound information. The paper proves that both MCP and CSCP are NP-hard, and that the LP relaxation has integrality gap at least $i\in S_a$ 1. For CSCP, the authors note that the final $i\in S_a$ 2 guarantee is nearly optimal because the problem is NP-hard and edge loads are integral, so a $i\in S_a$ 3-type guarantee is essentially best possible up to additive constants.

4. Rounding mechanisms: innermost terminals, inclusion orders, and uncrossing

In the common-sink case, the rounding algorithm Round-1 repeatedly selects an innermost terminal, defined through vertex depth

$i\in S_a$ 4

where $i\in S_a$ 5 is the set of fractional cuts containing $i\in S_a$ 6. The algorithm assigns the current meta-node $i\in S_a$ 7 as the integral cut for that terminal. As fractional mass is consumed, edges whose fractional load drops to $i\in S_a$ 8 are contracted into meta-nodes, and fractional cuts are reassigned so that the remaining family stays laminar and feasible (0810.0674).

The analysis shows that feasibility is preserved throughout, and that each edge ends with load at most $i\in S_a$ 9 after rounding from a laminar fractional CSCP solution. This estimate underlies the final $r_i\in V$ 0 guarantee once laminarization overhead is included.

For general MCP, the paper introduces a cut-inclusion ordering. If $r_i\in V$ 1 is the outermost cut of terminal $r_i\in V$ 2, then

$r_i\in V$ 3

with ties broken consistently. A preprocessing step enforces an inclusion invariant: if $r_i\in V$ 4, then any cuts for $r_i\in V$ 5 and $r_i\in V$ 6 that both contain their terminals satisfy $r_i\in V$ 7. This ordering is needed because terminals in the same commodity must remain pairwise separated throughout the rounding process.

The laminarization of an arbitrary fractional MCP solution proceeds in several stages: generation of a non-laminar fractional family from the LP metric, rounding weights to a fine grid, conversion to an integral instance with many copies, application of the uncrossing procedure Integer-Lam-2, and projection back to a fractional laminar family. The uncrossing step includes simple pairwise uncrossings, a three-terminal reassignment rule, transformations for same-commodity and different-commodity crossings, a directed auxiliary graph with red and blue edges, and a final acyclic leaf-elimination step on connected components.

The load analysis tracks edges through states $r_i\in V$ 8. The invariant structure ensures that edges in $r_i\in V$ 9 are not loaded further, edges entering $\{E_1,\dots,E_k\},$ 0 can be loaded at most once afterward, and the additive loss from rounding a laminar fractional solution is at most $\{E_1,\dots,E_k\},$ 1. Combined with the laminarization factor $\{E_1,\dots,E_k\},$ 2, this yields the final $\{E_1,\dots,E_k\},$ 3 guarantee.

5. Applications, antecedents, and limits of the laminar viewpoint

A principal motivation for multiway cut packing is graph labeling. Each item carries a string label of length $\{E_1,\dots,E_k\},$ 4, with some coordinates possibly undecided. For a fixed coordinate $\{E_1,\dots,E_k\},$ 5, the set of items already labeled $\{E_1,\dots,E_k\},$ 6 is denoted $\{E_1,\dots,E_k\},$ 7. Completing the labeling for that coordinate becomes a set multiway cut problem: one partitions items into $\{E_1,\dots,E_k\},$ 8 parts so that each set $\{E_1,\dots,E_k\},$ 9 is contained in the $E_a$ 0-part. Under an $E_a$ 1 objective, minimizing the maximum weighted edge disagreement becomes a cut packing problem. MCP is the specific case in which each $E_a$ 2 is a singleton (0810.0674).

The paper explicitly builds on the work of Rabani, Schulman, and Swamy, who introduced the problem and obtained the earlier $E_a$ 3 and $E_a$ 4 approximation bounds. It also notes that the graph-labeling reduction used there for $E_a$ 5 objectives does not directly preserve the $E_a$ 6 objective. The laminar method is therefore not just an improvement in constants; it is a different structural route tailored to worst-edge congestion.

Several lemmas define the exact scope of the laminar paradigm. From any feasible LP solution, one can construct a fractional laminar family with capacities $E_a$ 7 for CSCP and $E_a$ 8 for MCP. A fractional laminar CSCP family with integral capacities can be rounded to load at most $E_a$ 9, and a fractional laminar MCP family with integral capacities can be rounded to load at most $a$ 0. The Integer-Lam-2 lemma further shows that an integral family satisfying the initial per-commodity separation property can be converted into a laminar one while increasing edge load by at most a factor of $a$ 1: $a$ 2

At the same time, the paper establishes a genuine laminarity gap. Laminarity is not synonymous with optimality at unit load, and the need for bounded capacity blow-up is a theorem rather than a normalization trick. This point is central to an accurate understanding of laminar multiway cuts: the laminar family is a controlled surrogate for arbitrary feasible solutions, not an equivalent reformulation at identical load.

The laminar perspective in multiway cut packing sits within a broader family of separator formalisms in which non-crossing structure yields tractable decomposition. In planar graph theory, "Finding Maximal Sets of Laminar 3-Separators in Planar Graphs in Linear Time" studies a different but closely related structural problem: in a $a$ 3-connected planar graph $a$ 4, one seeks a maximal laminar family of $a$ 5-cutsets and converts it into a tree decomposition of adhesion three. The paper defines laminarity for $a$ 6-cutsets, represents cutsets by canonical $a$ 7-cycles in the barycentric subdivision $a$ 8, builds a conflict graph on non-laminar pairs, and shows that maximal laminar families can be found in linear time, including for all $a$ 9-cuts, non-trivial $G=(V,E)$ 00-cuts, and $G=(V,E)$ 01-non-shiftable $G=(V,E)$ 02-cuts (Eppstein et al., 2018). This is not a multiway cut problem, but it illustrates the same algorithmic theme: non-crossing separators support tree-like decomposition.

In vertex multiway cut, the laminar theme is more indirect. "Computing multiway cut within the given excess over the largest minimum isolating cut" and "Large Isolating Cuts Shrink the Multiway Cut" do not define laminar families as the main object, but both exploit a nested order on important separators and isolating cuts. The first paper proves that the number of important $G=(V,E)$ 03- $G=(V,E)$ 04 separators of size at most $G=(V,E)$ 05 is at most

$G=(V,E)$ 06

and uses this to obtain an $G=(V,E)$ 07 algorithm for finding a multiway cut of size at most $G=(V,E)$ 08, where $G=(V,E)$ 09 is the largest minimum isolating cut size (Razgon, 2010). The second develops the IS-family abstraction, proves that the union of important separators of excess at most $G=(V,E)$ 10 has size at most

$G=(V,E)$ 11

and derives kernelization when $G=(V,E)$ 12, where $G=(V,E)$ 13 is the size of the smallest isolating cut (Razgon, 2011). In both cases, the separator order is laminar-like rather than pairwise laminar in the set-system sense.

More recent work on representative-based generalizations of multiway cut also exhibits separator hierarchies without formal laminar theorems. "Multiway Cuts with a Choice of Representatives" studies variants such as All-to-All, Single-to-All, Single-to-Single, Some-to-Single, Some-to-Some, and Some-to-All. For fixed $G=(V,E)$ 14, it obtains $G=(V,E)$ 15-approximations matching the best known guarantee for Multiway Cut, where $G=(V,E)$ 16. For general $G=(V,E)$ 17, it gives a $G=(V,E)$ 18-approximation for Single-to-All and a tight $G=(V,E)$ 19-approximation for Single-to-Single, with the latter derived via a gammoid characterization on trees and a Gomory–Hu-tree-based $G=(V,E)$ 20-approximation in general graphs (Bérczi et al., 2024). The paper states that these tree and GH-tree arguments are laminar in spirit rather than formal laminar-cut results. This suggests that laminar organization remains a recurring structural principle even when the optimization problem is not explicitly posed as laminar multiway cut packing.