Tree Packing in Graph Theory

• Tree packing is a family of graph-theoretic problems that embeds multiple tree structures into a graph while satisfying constraints such as edge-disjointness, capacity, and planarity.
• It serves as a dual framework in min-cut, arboricity, and distributed communication, with methods extending to fractional LP formulations and greedy algorithms.
• Variants include spanning tree packing, rooted and directed packings, and Steiner tree packing, each addressing unique embedding challenges and optimization goals.

Tree packing is a family of graph-theoretic problems centered on placing multiple tree structures into a common graph under edge, capacity, planarity, connectivity, or optimization constraints. In one major sense, it asks for collections of spanning trees, often edge-disjoint or fractional, and relates them to global cut structure. In another, it asks whether prescribed trees can be embedded edge-disjointly into a host such as $K_n$. Further variants include rooted and directed packings, Steiner tree packings, bounded-outdegree rooted packings motivated by peer-to-peer overlays, and low-diameter packings motivated by distributed communication. The term therefore denotes a class of closely related but technically distinct notions rather than a single canonical problem (Vos et al., 2024, Kerivin et al., 2011, Chalise et al., 2024).

1. Core meanings and formal variants

The most classical formulation treats a tree-packing of an undirected graph $G=(V,E)$ as a family of spanning trees, where an edge may appear in multiple trees unless explicit edge-disjointness is required. In the load-based formulation, if $L(e)$ is the number of packed trees containing $e$, the relative load is

$\ell(e)=\frac{L(e)}{|\mathcal T|}.$

This viewpoint is central in min-cut and arboricity theory because the edge loads encode cut information (Vos et al., 2024).

A different but equally classical formulation asks whether a given family of trees $G_1,\dots,G_k$ can be realized as pairwise edge-disjoint subgraphs of a host graph $H$. The Tree Packing Conjecture is the best-known example: trees of orders $2,3,\dots,n$ are to be packed into $K_n$ (Gerbner et al., 2011, Chalise et al., 2024).

Rooted and constrained variants change the object being packed. In the Maximum Bounded Rooted-Tree Packing problem, the input is an undirected connected graph with a designated root $r$, an integer $G=(V,E)$0, and per-vertex capacities $G=(V,E)$1. One seeks $G=(V,E)$2 rooted trees $G=(V,E)$3 maximizing

$G=(V,E)$4

subject to the bounded-outdegree condition

$G=(V,E)$5

where $G=(V,E)$6 is the number of children of $G=(V,E)$7 in $G=(V,E)$8 (Kerivin et al., 2011).

Directed variants replace spanning trees by rooted out-trees through specified terminals. For a digraph $G=(V,E)$9, a set $L(e)$0, and $L(e)$1, an $L(e)$2-tree is an out-tree rooted at $L(e)$3 containing all vertices of $L(e)$4. The parameters $L(e)$5 and $L(e)$6 denote the maximum numbers of internally disjoint and arc-disjoint $L(e)$7-trees, respectively (Sun et al., 2020).

Steiner Tree Packing replaces spanning or rooted spanning trees by edge-disjoint connected subgraphs containing designated terminal sets. In STP an instance is $L(e)$8, and one asks for $L(e)$9 pairwise edge-disjoint trees each containing $e$0. GSTP generalizes this to multiple terminal sets with individual demands (Hastrich et al., 14 May 2025).

Variant	Packed object	Typical constraint or objective
Spanning-tree packing	Spanning trees of $e$1	Maximize packing value or require edge-disjointness
Host-graph tree packing	Prescribed trees $e$2 into host $e$3	Pairwise edge-disjoint embeddings
Bounded rooted packing	Rooted trees through a source $e$4	Capacity bounds $e$5
Directed Steiner tree packing	$e$6-trees in a digraph	Arc-disjointness or internal disjointness
Steiner tree packing	Terminal-containing trees/subgraphs	Edge-disjointness with demand $e$7

2. Min-max theory and fractional tree packings

A fundamental structural principle is the Nash-Williams/Tutte min-max relation between tree packings and vertex partitions. For a partition $e$8 of the vertex set, define

$e$9

and for a tree-packing $\ell(e)=\frac{L(e)}{|\mathcal T|}.$0,

$\ell(e)=\frac{L(e)}{|\mathcal T|}.$1

Then

$\ell(e)=\frac{L(e)}{|\mathcal T|}.$2

This identifies a good tree-packing as a primal certificate for a good partition value, and it underlies the use of tree packings in min-cut, arboricity, and weighted spanning-tree-packing formulations (Vos et al., 2024).

The fractional LP viewpoint makes this duality explicit. A fractional spanning-tree packing assigns nonnegative weights $\ell(e)=\frac{L(e)}{|\mathcal T|}.$3 to spanning trees and solves

$\ell(e)=\frac{L(e)}{|\mathcal T|}.$4

subject to

$\ell(e)=\frac{L(e)}{|\mathcal T|}.$5

Its optimum equals

$\ell(e)=\frac{L(e)}{|\mathcal T|}.$6

again ranging over vertex partitions $\ell(e)=\frac{L(e)}{|\mathcal T|}.$7. This same LP language extends to minimum $\ell(e)=\frac{L(e)}{|\mathcal T|}.$8-cut: the dual of the Naor–Rabani LP becomes a tree packing into augmented capacities $\ell(e)=\frac{L(e)}{|\mathcal T|}.$9, with objective

$G_1,\dots,G_k$0

In this sense, tree packing is not merely a decomposition question but a dual object in cut optimization (Chekuri et al., 2018).

Classical edge-connectivity guarantees also fit this framework. Every $G_1,\dots,G_k$1-edge-connected graph contains $G_1,\dots,G_k$2 edge-disjoint spanning trees, which is the baseline theorem from which later low-diameter and dynamic refinements proceed (Chuzhoy et al., 2020).

Weighted versions arise naturally when edges have rates rather than unit capacities. In conference key propagation over a QKD network, the maximum conference key rate is formulated as a weighted spanning-tree-packing problem: $G_1,\dots,G_k$3 and, by Nash-Williams–Tutte, equals

$G_1,\dots,G_k$4

Here tree packing becomes an exact information-theoretic capacity formula rather than only a combinatorial abstraction (Trushechkin et al., 4 Jun 2025).

3. Packing prescribed trees into host graphs

In the host-graph formulation, tree packing asks for edge-disjoint copies of specified trees inside a common host graph. The central statement is Gyárfás’s Tree Packing Conjecture: any family $G_1,\dots,G_k$5 with $G_1,\dots,G_k$6 packs into $G_1,\dots,G_k$7. A 2024 paper gives a proof by translating the decomposition problem into a complete labeling problem for augmented functional trees and then applying a polynomial certificate argument (Chalise et al., 2024).

That proof reframes a rooted labeled tree as a functional directed graph $G_1,\dots,G_k$8 with

$G_1,\dots,G_k$9

and reduces packing to the existence of permutations making the induced edge set an orientation of $H$0. The nonvanishing of a canonical representative of the “Tree Packing Polynomial Certificate” is then equivalent to the existence of a complete labeling, hence to the packing itself (Chalise et al., 2024).

Before this full resolution, much of the literature developed partial and asymptotic forms. One strong intermediate result proves that a linear number of the largest trees in a TPC sequence can be packed into $H$1, thereby implying Bollobás’s conjecture on packing any fixed number of the largest trees for sufficiently large $H$2 (Janzer et al., 2024). Earlier results packed $H$3 trees of consecutive orders into $H$4, and into $H$5 when no packed tree is a star; under large-degree or leaf-structure hypotheses, $H$6 such trees pack into $H$7 (Balogh et al., 2012). Another asymptotic confirmation of Bollobás’s conjecture applies when each tree has the required number of leaves or a pending path of the required order, yielding in particular the case $H$8 (Żak, 2015).

Approximate host enlargement yields a different asymptotic regime. For fixed $H$9 and $2,3,\dots,n$0, every family of trees of orders at most $2,3,\dots,n$1, maximum degrees at most $2,3,\dots,n$2, and total edge count at most $2,3,\dots,n$3 packs into $2,3,\dots,n$4. This gives asymptotic forms of both the Tree Packing Conjecture and Ringel’s conjecture for bounded-degree trees (Böttcher et al., 2014).

Ringel-type exact decomposition is also known in quasirandom settings. Any $2,3,\dots,n$5-typical graph $2,3,\dots,n$6 on $2,3,\dots,n$7 vertices of density $2,3,\dots,n$8 can be decomposed into $2,3,\dots,n$9 copies of any tree $K_n$0 with $K_n$1; the special case $K_n$2 yields Ringel’s conjecture for all sufficiently large trees (Keevash et al., 2020).

A planar counterpart asks whether two trees on the same $K_n$3-vertex set can be packed into a planar host on the same vertices. Except for the star obstruction, the answer is affirmative: every two nonstar trees of the same size admit a planar packing, and the constructive proof gives an $K_n$4 algorithm (Geyer et al., 2016).

4. Algorithmic and structural roles in cuts, arboricity, and distributed communication

In algorithm design, tree packing often functions as a structural proxy for cuts. Greedy tree packings are built by repeatedly computing minimum spanning trees with respect to current edge loads. Thorup’s ideal load decomposition $K_n$5 is approximated by actual greedy loads $K_n$6, with guarantee

$K_n$7

A 2024 reanalysis shows that for min-cut purposes one needs only $K_n$8 greedy trees to guarantee either a tree that $K_n$9-respects a min-cut or a trivial cut in an appropriate contracted graph, improving substantially over Thorup’s earlier $r$0 requirement (Vos et al., 2024).

This structural refinement yields faster dynamic algorithms. The same paper gives a deterministic fully dynamic exact min-cut algorithm with worst-case update time

$r$1

for min-cut value bounded by $r$2, and a general fully dynamic exact min-cut algorithm with amortized update time

$r$3

improving on $r$4. It also derives the first fully dynamic deterministic $r$5-approximation of fractional arboricity with amortized update time

$r$6

using the identity

$r$7

Tree packing here serves simultaneously as a min-cut certificate and as a density certificate (Vos et al., 2024).

The LP-based theory of minimum $r$8-cut gives a parallel generalization. The dual tree-packing interpretation, combined with the LP integrality gap

$r$9

implies that for an optimal $G=(V,E)$00-cut there exists a tree in the support of an optimal dual solution crossing the cut at most $G=(V,E)$01 times. The same framework yields the bound

$G=(V,E)$02

on the number of $G=(V,E)$03-approximate $G=(V,E)$04-cuts (Chekuri et al., 2018).

Diameter-sensitive packing adds another algorithmic layer. For an $G=(V,E)$05-vertex $G=(V,E)$06-edge-connected graph of diameter $G=(V,E)$07, there is an efficient randomized algorithm producing $G=(V,E)$08 spanning trees with edge-congestion at most $G=(V,E)$09, each of diameter

$G=(V,E)$10

with high probability. Sampling also yields $G=(V,E)$11 edge-disjoint spanning trees each of diameter

$G=(V,E)$12

and in $G=(V,E)$13-connected graphs there are $G=(V,E)$14 spanning trees of diameter $G=(V,E)$15 with edge-congestion $G=(V,E)$16 (Chuzhoy et al., 2020). These results explain why tree packings are useful in distributed MST, information dissemination, connectivity verification, and secure distributed computation.

5. Rooted, directed, and Steiner generalizations

The Maximum Bounded Rooted-Tree Packing problem models under-provisioned peer-to-peer streaming overlays. The graph represents the overlay, the root is the source peer, and $G=(V,E)$17 rooted delivery trees correspond to stripes in a multiple-description coding scheme. The goal is to maximize total covered appearances $G=(V,E)$18 while respecting total forwarding capacity at each vertex. The decision problem is NP-complete; the MBRT special case $G=(V,E)$19 is NP-complete by reduction from 3-SAT. Nonetheless, optimal solutions are computable in polynomial time on two graph classes: complete graphs in $G=(V,E)$20, via a Hamiltonian-path-based construction followed by greedy expansion, and rooted trees in $G=(V,E)$21, via bottom-up dynamic programming with a non-standard multiple-choice knapsack subproblem (Kerivin et al., 2011).

Directed tree packing extends undirected tree connectivity to digraphs. For fixed integers $G=(V,E)$22 and $G=(V,E)$23, deciding whether $G=(V,E)$24 or $G=(V,E)$25 is NP-complete on general digraphs. On symmetric digraphs, deciding $G=(V,E)$26 is polynomial-time solvable for fixed $G=(V,E)$27, while the complexity of $G=(V,E)$28 splits: polynomial for fixed $G=(V,E)$29, but NP-complete when $G=(V,E)$30 is part of the input. On Eulerian digraphs, arc-disjoint packing admits an exact criterion,

$G=(V,E)$31

whereas the internally disjoint version remains NP-complete (Sun et al., 2020). A common misconception is that arc-disjoint and internally disjoint directed tree packings behave similarly; these results show that their complexity and structure can diverge sharply.

Steiner Tree Packing generalizes both spanning-tree packing and edge-disjoint paths. An STP instance $G=(V,E)$32 asks for $G=(V,E)$33 pairwise edge-disjoint trees each containing $G=(V,E)$34. The GSTP formulation permits multiple terminal sets with individual demands and uses an augmented graph $G=(V,E)$35 obtained by adding a vertex $G=(V,E)$36 adjacent to every vertex in each terminal set $G=(V,E)$37. The 2025 structural-parameterized results show that GSTP is fixed-parameter tractable by the tree-cut width and by the fracture number of the augmented graph, and by the slim tree-cut width of the input graph. As corollaries, STP is fixed-parameter tractable by the tree-cut width of the host graph, and GSTP is FPT by $G=(V,E)$38 with runtime

$G=(V,E)$39

where $G=(V,E)$40 (Hastrich et al., 14 May 2025).

6. Probabilistic, random, and applied network forms

Tree packing also appears in probabilistic decomposition results. For a fixed tree $G=(V,E)$41 on $G=(V,E)$42 vertices, the random graph $G=(V,E)$43 and suitable pseudo-random graphs contain many edge-disjoint $G=(V,E)$44-factors covering almost all edges. In an $G=(V,E)$45-regular graph, if

$G=(V,E)$46

there is a collection of edge-disjoint $G=(V,E)$47-factors covering all but a $G=(V,E)$48-fraction of the edges asymptotically. In $G=(V,E)$49, if

$G=(V,E)$50

then all but an $G=(V,E)$51-fraction of edges can be covered by edge-disjoint $G=(V,E)$52-factors with high probability; under stronger divisibility assumptions this is pushed down to

$G=(V,E)$53

This work places tree-factor packing near the connectivity threshold scale in random graphs (Bal et al., 2013).

In peer-to-peer streaming, rooted tree packing models the distribution of $G=(V,E)$54 stripes in an under-provisioned overlay, where average upload capacity is below the video bitrate. Since a peer’s video quality depends on how many rooted delivery trees contain it, the objective $G=(V,E)$55 measures achievable aggregate service under bounded upload capacities (Kerivin et al., 2011).

In quantum networks, spanning-tree packing becomes an exact protocol for conference key propagation. A single spanning tree of bipartite secret bits yields one conference secret bit by chained one-time-pad propagation, and the asymptotic achievable conference key rate equals the weighted spanning-tree-packing number

$G=(V,E)$56

The same criterion identifies bottleneck partitions and guides optimal placement of new QKD links: links that bridge the minimizing partition increase the rate most directly, whereas links that merely strengthen already dense regions can fail to improve the bottleneck (Trushechkin et al., 4 Jun 2025).

Across these settings, the unifying pattern is that tree packing converts global network function into combinatorial structure. In decomposition problems it organizes edge-disjoint embeddings; in cut problems it exposes bottlenecks through loads and respecting cuts; in rooted and Steiner settings it captures constrained multicommodity connectivity; and in communication systems it serves as an exact throughput or capacity certificate rather than only a heuristic design principle (Vos et al., 2024, Trushechkin et al., 4 Jun 2025).