---
title: 'CISTs: Completely Independent Spanning Trees'
url: https://www.emergentmind.com/topics/completely-independent-spanning-trees-cists
type: topic
---

# CISTs: Completely Independent Spanning Trees

Completely independent spanning trees (CISTs) are spanning trees $T_1,\dots,T_k$ of a connected graph $G$ such that, for every pair of vertices $u,v\in V(G)$, the unique $u$–$v$ paths in the trees are pairwise edge-disjoint and pairwise internally vertex-disjoint. In the standard root-free formulation, this means
$$
E(P_{T_i}(u,v))\cap E(P_{T_j}(u,v))=\emptyset,\qquad
V(P_{T_i}(u,v))\cap V(P_{T_j}(u,v))=\{u,v\}
$$
for all $i\neq j$. A fundamental equivalent characterization states that $T_1,\dots,T_k$ are CISTs if and only if they are pairwise edge-disjoint and, for every vertex $x\in V(G)$, there is at most one index $i$ with $d_{T_i}(x)>1$ [2512.15486]. This notion is stronger than rooted independent spanning trees, because it requires path independence for all vertex pairs rather than only for root-to-vertex paths [2512.19973].

## 1. Definition, scope, and terminological variants

The standard graph-theoretic definition treats CISTs as a root-free all-pairs object: the independence condition is imposed simultaneously for every pair of vertices of the graph. Hasunuma’s characterization makes the condition local: edge-disjointness plus the rule that any vertex can be internal in at most one tree is equivalent to full complete independence [2512.15486].

This all-pairs requirement is strictly stronger than the rooted independent spanning tree condition. In the rooted setting, one fixes a root $r$ and asks only that the $r$–$u$ paths in distinct trees be internally disjoint for each $u\neq r$. By contrast, CISTs require internal disjointness of the $x_1$–$x_2$ paths for every pair $x_1,x_2\in V(G)$, so the rooted condition does not imply the root-free one [2512.19973].

The literature is not fully uniform in terminology. In random and pseudorandom graph work, “completely independent spanning trees” may refer to rooted vertex-independence, and edge-disjointness is not always built into the definition; the same paper explicitly distinguishes this from stronger edge-disjoint variants [2509.26401]. Related network papers also separate CISTs from weaker notions such as completely edge-independent spanning trees (CEISTs), which require only pairwise edge-disjoint spanning trees [2401.01585]. This suggests that precise statement of the adopted definition is essential in any technical comparison.

## 2. Structural characterizations

A second major characterization is Araki’s partition theorem. A partition $V(G)=V_1\sqcup\cdots\sqcup V_k$ is a $k$-CIST-partition if each induced subgraph $G[V_i]$ is connected and, for every $i<j$, every connected component of the bipartite subgraph $B(V_i,V_j,G)$ satisfies $|E(H)|\ge |V(H)|$, equivalently, no such component is a tree. A connected graph admits $k$ CISTs if and only if it admits a $k$-CIST-partition [2512.15486].

Another characterization uses connected dominating sets. A graph has $k$ CISTs if and only if there exist $k$ disjoint connected dominating sets $V_1,\dots,V_k$ such that every component of the bipartite subgraph $\langle V_i,V_j\rangle_G$ has a cycle for all $i\neq j$ [2209.09565]. This formulation is particularly useful in line graphs and in constructive existence proofs, because it converts path-independence into a combination of domination, connectivity, and cycle conditions across pairwise interfaces.

For split graphs, the structure becomes hypergraph-theoretic. If $G=(D\sqcup I,E)$ is a split graph, its associated hypergraph is $H(G)=(D,\{N_G(x):x\in I\})$. If $G$ has $k$ CISTs, then $H(G)$ is panchromatically $k$-colorable; if $H(G)$ is bipanchromatically $k$-colorable, then $G$ has $k$ CISTs. If $M$ is the maximum number of CISTs in $G$, then
$$
\chi_p^2(H(G)) \le M \le \chi_p^2(H(G))+1,
$$
so the CIST number of a split graph lies in a sharp window controlled by bipanchromatic colorings of its associated hypergraph [2512.15486].

## 3. Generalizations and related notions

CISTs are the spanning case of a broader Steiner theory. For a terminal set $S\subseteq V(G)$ with $|S|\ge 2$, an $S$-Steiner tree is a subtree spanning $S$ whose leaves all belong to $S$. A family of $S$-Steiner trees is completely independent if, for every pair of terminals $u,v\in S$, the $u$–$v$ paths in different trees are pairwise edge-disjoint and internally vertex-disjoint. When $S=V(G)$, this definition specializes exactly to CISTs [2512.19973]. The associated packing number is $\kappa_G^*(S)$, and the generalized $k^*$-connectivity is
$$
\kappa_k^*(G)=\min\{\kappa_G^*(S): S\subseteq V(G),\ |S|=k\}.
$$

The Steiner formulation preserves the same structural core. A family of $S$-Steiner trees is completely independent if and only if the trees are pairwise edge-disjoint and their sets of internal vertices are pairwise disjoint [2604.19886]. It also admits a directed analogue: completely independent spanning arborescences (CISA) in a directed $S$-minor are equivalent, in the appropriate sense, to completely independent Steiner trees in the undirected host graph [2604.19886]. This directed correspondence extends the CIST framework beyond undirected spanning trees.

Several relaxations have been studied. The $(i,j)$-disjoint spanning tree framework defines $T_1,\dots,T_k$ to be $(i,j)$-disjoint when at most $i$ vertices are internal in more than one tree and at most $j$ edges are shared by more than one tree. In that language, CISTs are exactly the $(0,0)$-disjoint case; $(0,\ast)$ captures internally vertex-disjoint spanning trees, and $(\ast,0)$ captures edge-disjoint spanning trees [1702.08289]. On the other side, CEISTs impose only pairwise edge-disjointness and are strictly weaker than CISTs [2401.01585].

## 4. Existence theorems, exact values, and graph classes

Research on CISTs combines exact enumeration in highly structured graphs with upper bounds in sparse or topologically restricted families.

| Graph family | Main CIST result | Source |
|---|---|---|
| Complete graph $K_n$ | $\kappa_n^*(K_n)=\lfloor n/2\rfloor$ | [2512.19973] |
| Line graph $L(K_n)$ | $\tau^*(L(K_n))=\lfloor (n+1)/2\rfloor$ for $n\ge 4$ | [2209.09565] |
| Hypercube $Q_n$ | $(\tfrac{1}{12}+o(1))n$ CISTs, each of diameter $(2+o(1))n$ | [2412.11780] |
| Planar graphs | Any planar graph has at most $4$ CISTs | [2604.19886] |
| Treewidth-$w$ graphs | Any graph of treewidth at most $w$ has at most $w+1$ CISTs | [2604.19886] |
| 3-connected 2-outerplanar triangulated discs | Two CISTs always exist | [2606.12827] |

For complete graphs, the Steiner generalization yields the exact formula
$$
\kappa_s^*(K_n)=n-\left\lceil \frac{s}{2}\right\rceil
$$
for $2\le s\le n$ and $n\ge 4$, so the spanning case gives
$$
\kappa_n^*(K_n)=n-\left\lceil \frac{n}{2}\right\rceil=\left\lfloor \frac{n}{2}\right\rfloor.
$$
This agrees with the classical CIST packing number in $K_n$ [2512.19973]. In line graphs, the exact number can increase: for every $n\ge 4$,
$$
\tau^\ast(L(K_n))=\left\lfloor \frac{n+1}{2}\right\rfloor,
$$
and this value is optimal [2209.09565].

Hypercubes provide a second major benchmark. For each fixed $k$, sufficiently high-dimensional hypercubes $Q_n$ contain $k$ CISTs, and more strongly there exist
$$
\left(\frac{1}{12}+o(1)\right)n
$$
pairwise completely independent spanning trees in $Q_n$, each with diameter
$$
(2+o(1))n.
$$
Since every spanning tree of $Q_n$ has diameter at least $2n-1$, this diameter is asymptotically optimal [2412.11780]. A complementary construction shows that $Q_n$ has three CISTs for every $n\ge 7$, with diameter bounds at most $2n+1$, $2n+3$, and $2n+4$ for the three trees [2410.03379].

Topological restrictions yield sharp thresholds. In planar graphs, any CIST packing has size at most $4$; in graphs of bounded treewidth $w$, the CIST number is at most $w+1$ [2604.19886]. For 3-connected 2-outerplanar triangulated discs, two CISTs always exist, while there is a 3-connected 4-outerplanar triangulated disc with no two CISTs [2606.12827]. At the level of degree-sum conditions, if $G$ is a connected graph of order $n\ge 7$ and
$$
\mu_2(G)\ge n,
$$
then $G$ has two CISTs [2502.11522].

## 5. Complexity and algorithmic methods

The general decision problem is computationally hard. Determining whether a graph contains $k$ CISTs is NP-complete, even for $k=2$ [2502.11522]. This hardness persists in highly structured classes: deciding whether a split graph has two CISTs is NP-complete, and the related hypergraph problem of bipanchromatic $k$-colorability is also NP-complete [2512.15486].

The Steiner generalization strengthens these hardness results. For any fixed number of terminals $r\ge 3$ and any fixed $k\ge 2$, the decision problem asking whether a graph contains $k$ completely independent Steiner trees is NP-complete. The corresponding maximization problem is NP-hard to approximate within $\Omega(\log n)$ and is APX-hard already for $r=3$ [2604.19886]. These results reinforce the view that CIST packing is substantially harder than ordinary spanning-tree packing.

Despite this, several algorithmic regimes are tractable. For fixed terminal number $r$ and fixed packing size $k$, completely independent Steiner trees can be decided in polynomial time by enumerating constant-size templates and reducing feasibility to a constant-size vertex-disjoint paths problem [2604.19886]. For bounded-treewidth graphs and fixed $k$, existence is expressible in monadic second-order logic, so Courcelle’s theorem yields linear-time algorithms [2604.19886]. In split graphs, integer linear programming formulations are given for $\chi_p(H)$, $\chi_p^2(H)$, and $\alpha_{\chi_p(H)}(H)$, providing exact computation for small-to-medium instances [2512.15486].

On the constructive side, several existence proofs are algorithmic. The complete-graph and complete-bipartite CISST constructions are polynomial-time and therefore yield polynomial constructions of optimal CIST families in the spanning case [2512.19973]. In hypercubes, the explicit deterministic construction producing $2^k$ CISTs can be implemented in total time $O(k\cdot n\cdot 2^n)$ to output all trees [2412.11780].

## 6. Network applications, resilience, and open directions

The principal motivation for CISTs is fault-tolerant multipath communication. Because for every pair $u,v$ the $u$–$v$ paths across the trees are pairwise edge-disjoint and internally vertex-disjoint, up to $k-1$ arbitrary edge failures or up to $k-1$ internal-vertex failures, excluding the endpoints, still leave at least one intact tree-path between $u$ and $v$ [2604.19886]. In the Steiner setting, the same mechanism focuses redundancy on a specified terminal set, which is relevant to multicast, backbone, and service-node architectures [2512.19973].

The Steiner papers also make the design contrast explicit. Completely independent $S$-Steiner trees provide stronger fault tolerance than internally disjoint $S$-Steiner trees, because complete independence forbids shared internal bottlenecks on terminal-to-terminal routes. The emergency communication example with terminals $\{u$ (Command Center), $x$ (Hospital), $v$ (Fire Station)$\}$ shows that an internally disjoint design may still route $u$–$v$ communication through $x$, whereas a CISST design can preserve $u$–$v$ connectivity even if $x$ fails [2512.19973].

Implementation-oriented work in dense Gaussian networks illustrates the performance angle. A recent construction of two CISTs partitions the dense Gaussian network, builds the first tree, obtains the second by rotation, and reports an improvement of at least $33\%$ in the average maximum number of steps required to deliver a message from the root node to all other nodes, relative to existing approaches [2606.23935]. This suggests that CISTs are not only a resilience certificate but also a latency-sensitive routing primitive in symmetric interconnection networks.

Several open directions remain central. For hypercubes, the exact asymptotic constant $c$ such that $Q_n$ has $(c+o(1))n$ CISTs is unknown; the best uniform lower bound is $c\ge 1/12$, while the trivial edge-count upper bound gives $c\le 1/2$ [2412.11780]. The question of whether $Q_6$ has three CISTs was left open by the explicit tri-CIST construction for $Q_n$, $n\ge 7$ [2410.03379]. In locally twisted cubes, the stronger CIST problem remains open even though the optimal number of CEISTs is known [2401.01585]. More broadly, tightening the sufficient connectivity bound
$$
\kappa(G)\ge 10k(|V(G)|-1)
$$
for guaranteeing $k$ CISTs, and obtaining sharper characterizations in planar, bounded-treewidth, and network-specific graph classes, remain active themes [2604.19886].

Source: https://www.emergentmind.com/topics/completely-independent-spanning-trees-cists