Mean Subtree Order in Graphs

• Mean subtree order is defined as the average number of vertices in a connected subgraph, computed via generating functions in trees and extended to general graphs.
• The invariant underpins several extremal results, with the path uniquely minimizing it and caterpillar structures conjectured to maximize it.
• Local measures such as mean order and density, along with analysis in k-trees, reveal how edge operations and localized configurations impact global graph structure.

Mean subtree order is the average number of vertices in a subtree, and it sits at the intersection of enumerative graph theory, extremal graph theory, and generating-function methods. In the classical tree setting, a subtree is a connected induced subgraph, so the invariant measures the first moment of the subtree-size distribution. Two later extensions coexist in the literature. One extends the induced-subgraph viewpoint from trees to all connected graphs by averaging over connected induced vertex sets; the other, used for general graphs, counts all tree subgraphs, not necessarily induced. This terminological bifurcation is central: many structural theorems are parallel across the two settings, but edge-addition and extremal phenomena can differ substantially (Vince, 2021, Cambie et al., 2020, Cambie et al., 28 Aug 2025).

1. Definitions and competing conventions

For a tree $T$, the standard definition counts induced connected subgraphs. If $S_k(T)$ denotes the number of subtrees of order $k$, then the mean subtree order is

$$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$

The associated generating polynomial is written as $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$, giving

$\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$

This convention underlies the classical work of Jamison and most of the extremal theory for trees (Cambie et al., 2020).

For connected graphs, one important extension replaces subtrees by connected induced subgraphs. If $c_k(G)$ is the number of connected induced subgraphs of order $k$, then

$$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$

Writing $C_G(z)=\sum_{k=1}^n c_k(G)z^k$, one has $S_k(T)$0. In this setting the quantity is also called the mean connected induced subgraph order or mean CIS order (Vince, 2021, Balodis et al., 2018).

A different general-graph convention counts all tree subgraphs. If $S_k(T)$1 is the number of $S_k(T)$2-vertex subtrees of a connected graph $S_k(T)$3, then

$S_k(T)$4

where $S_k(T)$5. Distinct edge sets on the same vertex set are counted separately, and subtrees are not required to be induced (Cambie et al., 28 Aug 2025, Cambie et al., 2023, Cameron et al., 2019).

Setting	Objects counted	Mean formula
Tree $S_k(T)$6	induced connected subgraphs of $S_k(T)$7	$S_k(T)$8
Connected graph $S_k(T)$9, CIS convention	connected induced subgraphs	$k$0
Connected graph $k$1, subtree-subgraph convention	connected acyclic subgraphs	$k$2

Local variants are equally important. For trees, $k$3 denotes the average order of all subtrees containing a fixed vertex $k$4, and more generally $k$5 denotes the average order of all subtrees containing a fixed subtree $k$6. Normalized versions are called densities: $k$7 for the global tree density, $k$8 for connected induced subgraphs of a graph, and $k$9 for local density at a subtree $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$0 (Cambie et al., 2020, Vince, 2021, Wang, 2023).

2. Classical tree theory and the path minimum

The foundational extremal theorem is due to Jamison: among all trees of order $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$1, the path $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$2 uniquely minimizes the mean subtree order. Explicitly,

$$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$3

with equality if and only if $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$4. Since subtrees of $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$5 are exactly contiguous vertex intervals and there are $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$6 of order $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$7, one obtains

$$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$8

This is the baseline lower bound for the subject (Vince, 2021, Balodis et al., 2018).

A later contraction theorem gives a second route to the same bound. If $$\mathrm{M}(T)=\frac{\sum_{k\ge 1} k\,S_k(T)}{\sum_{k\ge 1} S_k(T)}.$$9 is any edge of a tree $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$0, then contracting $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$1 produces a tree $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$2 satisfying

$F_T(x)=\sum_{k\ge 1} S_k(T)x^k$3

with equality if and only if $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$4 is a path. The inequality is strict whenever $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$5 does not lie on a pendant path. Repeated contraction to a single vertex immediately yields

$F_T(x)=\sum_{k\ge 1} S_k(T)x^k$6

again with equality only for $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$7 (Wang, 2023).

The same minimization result admits a structural proof via local transformations. The Gluing Lemma shows that when a rooted tree is attached to a path, the mean subtree order is maximized when the attachment point is central; the Strong Gluing Lemma strengthens this to strict monotonicity as the gluing point moves inward. These tools yield a conceptually simple proof that every non-path tree has a standard $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$8-associate with strictly smaller mean subtree order, and iteration leads to the path (Mol et al., 2017).

The path theorem is sometimes treated as merely a tree result. That is misleading. It is the prototype for several later generalizations, and the sharp constant $F_T(x)=\sum_{k\ge 1} S_k(T)x^k$9 persists in broader settings (Vince, 2021, Cambie et al., 28 Aug 2025).

3. Maximal mean subtree order in trees

The opposite extremal direction is structurally subtler. Jamison conjectured that among all trees on $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$0 vertices, a caterpillar attains the maximum mean subtree order. This Caterpillar Conjecture remains open, but the asymptotic order of magnitude is known: $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$1 More precisely, if $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$2 on $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$3, extended $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$4-periodically, then

$\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$5

and

$\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$6

Thus the global maximum is pinned down to within $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$7 (Cambie et al., 2020).

The local extremum is sharper. If $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$8 is maximal over all trees $\mathrm{M}(T)=\frac{F_T'(1)}{F_T(1)}.$9 of order $c_k(G)$0 and vertices $c_k(G)$1, then $c_k(G)$2 must be a broom and $c_k(G)$3 is the end of the handle opposite the leaves. Consequently,

$c_k(G)$4

This local result explains why broom-like and double-broom-like structures recur in extremal constructions (Cambie et al., 2020).

Optimal or near-optimal trees are highly elongated. There exists a constant $c_k(G)$5 such that if a tree $c_k(G)$6 of order $c_k(G)$7 has diameter $c_k(G)$8, then

$c_k(G)$9

As a corollary, if $k$0 is optimal, then $k$1 is $k$2 for every $k$3. The same paper shows that an optimal tree has $k$4 subtrees, and its central part $k$5 induces a connected subtree with at least $k$6 vertices and at most $k$7 leaves (Cambie et al., 2020).

Several structural refinements are known. If $k$8 is optimal in $k$9 or in the caterpillar family $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$0 and $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$1, then every leaf is adjacent to a vertex of degree at least $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$2; equivalently, every limb has order $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$3. In $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$4, the number of leaves in an optimal caterpillar is $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$5, while in $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$6 the best general bound is $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$7 (Mol et al., 2017).

Degree constraints force a different asymptotic regime. For trees whose internal vertices all have degree at least $$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$8, the density satisfies

$$M(G)=\frac{\sum_{k=1}^n k\,c_k(G)}{\sum_{k=1}^n c_k(G)}.$$9

and both bounds are tight. For series-reduced trees, a sequence has density tending to $C_G(z)=\sum_{k=1}^n c_k(G)z^k$0 if and only if the proportion of leaves tends to $C_G(z)=\sum_{k=1}^n c_k(G)z^k$1, and tends to $C_G(z)=\sum_{k=1}^n c_k(G)z^k$2 if and only if the leaf proportion tends to $C_G(z)=\sum_{k=1}^n c_k(G)z^k$3 while the twig proportion tends to $C_G(z)=\sum_{k=1}^n c_k(G)z^k$4 (Haslegrave, 2013, Vince, 2021).

4. Local mean, local density, and anchored subtree structure

Local mean order often reveals more structure than the global average. Jamison’s local monotonicity states that for vertex sets $C_G(z)=\sum_{k=1}^n c_k(G)z^k$5,

$C_G(z)=\sum_{k=1}^n c_k(G)z^k$6

with equality if and only if the smallest subtree containing $C_G(z)=\sum_{k=1}^n c_k(G)z^k$7 equals that containing $C_G(z)=\sum_{k=1}^n c_k(G)z^k$8. In particular,

$C_G(z)=\sum_{k=1}^n c_k(G)z^k$9

for every vertex $S_k(T)$00, with strict inequality when $S_k(T)$01. A further lower bound states that for any subtree $S_k(T)$02,

$S_k(T)$03

These inequalities make local means a natural vehicle for proving global bounds (Cambie et al., 2020, Wang, 2023).

A central technical tool is the index. For a rooted tree $S_k(T)$04,

$S_k(T)$05

and more generally $S_k(T)$06 measures the effect of adding or deleting a leaf relative to a fixed subtree $S_k(T)$07. The Index Lemma gives exact increment formulas: $S_k(T)$08 for a neighbor $S_k(T)$09 of $S_k(T)$10, and

$S_k(T)$11

for a leaf $S_k(T)$12 of $S_k(T)$13. The bound $S_k(T)$14 is especially significant; equality $S_k(T)$15 occurs exactly when the corresponding component is a leaf or a path, and this characterizes limb vertices (Wang, 2023).

These local tools yield precise extremal descriptions of $S_k(T)$16-maximal subtrees, meaning subtrees of order $S_k(T)$17 with largest local mean. Such a subtree has at most one leaf whose degree in $S_k(T)$18 exceeds $S_k(T)$19, and at least one leaf whose degree is at most $S_k(T)$20. If a $S_k(T)$21-maximal subtree has a leaf of degree greater than $S_k(T)$22, then all its other leaves must be leaves of $S_k(T)$23. Degree-$S_k(T)$24 leaves of a $S_k(T)$25-maximal subtree must lie either all in core-paths or all in limbs (Wang, 2023).

Local density normalizes local mean across different anchor sizes: $S_k(T)$26 It is invariant under contraction of a subtree $S_k(T)$27. The fundamental density theorem is

$S_k(T)$28

with equality if and only if $S_k(T)$29 contains the core $S_k(T)$30; equivalently, every component of $S_k(T)$31 is a path. At the other extreme, local density can be arbitrarily close to $S_k(T)$32. Thus $S_k(T)$33 and $S_k(T)$34 play the role of universal lower and asymptotic upper local-density thresholds (Wang, 2023).

5. Extensions from trees to connected graphs

The induced-subgraph extension begins with block graphs. For a connected block graph $S_k(T)$35 of order $S_k(T)$36, the mean connected induced subgraph order satisfies

$S_k(T)$37

with equality if and only if $S_k(T)$38. The proof develops a local-global mean inequality at every vertex and block, together with vertex-gluing, edge-gluing, and stretching lemmas that monotonically reduce the mean until only a path remains (Balodis et al., 2018).

This block-graph theorem was later extended to all connected graphs. If $S_k(T)$39 is connected of order $S_k(T)$40, then

$S_k(T)$41

with equality if and only if $S_k(T)$42. In particular,

$S_k(T)$43

for every connected graph. The proof proceeds by induction on $S_k(T)$44 and combines several ingredients: the conditioned lower bound

$S_k(T)$45

for a connected subset $S_k(T)$46 of size $S_k(T)$47; the inequality

$S_k(T)$48

upper bounds $S_k(T)$49 in $S_k(T)$50-connected regions; a cut-vertex inequality controlling the product $S_k(T)$51; and the identity

$S_k(T)$52

Together these show that the path is the unique minimizer not only among trees or block graphs, but among all connected graphs under the connected-induced-set convention (Vince, 2021).

Canonical examples illustrate the scale of the invariant. Under the connected-induced-set convention, $S_k(T)$53; for the cycle,

$S_k(T)$54

for the star $S_k(T)$55,

$S_k(T)$56

and for the complete graph,

$S_k(T)$57

High connectivity and large blocks therefore push the average well above the path bound (Vince, 2021).

Under the non-induced subtree-subgraph convention, the same path minimum has now also been confirmed for connected graphs: $S_k(T)$58 with equality only for $S_k(T)$59. In this convention the complete graph is conjectured to maximize the mean. Cayley’s formula yields

$S_k(T)$60

hence

$S_k(T)$61

A uniformly random subtree of $S_k(T)$62 therefore has size $S_k(T)$63, with expected deficit tending to $S_k(T)$64 (Cambie et al., 28 Aug 2025).

Several graph classes admit stronger lower bounds than the universal $S_k(T)$65. For connected cographs of order $S_k(T)$66,

$S_k(T)$67

with equality on the right only for $S_k(T)$68. This suggests that forbidding long induced paths or forcing dense module structure raises the mean substantially above the path threshold (Vince, 2021).

6. Local mean order beyond trees: $S_k(T)$69-trees and related structures

The theory extends naturally to $S_k(T)$70-trees. A sub-$S_k(T)$71-tree of a $S_k(T)$72-tree $S_k(T)$73 is a subgraph that is itself a $S_k(T)$74-tree, and the global mean order is

$S_k(T)$75

For a fixed $S_k(T)$76-clique $S_k(T)$77, the local mean $S_k(T)$78 averages over all sub-$S_k(T)$79-trees containing $S_k(T)$80. The shifted local mean $S_k(T)$81 is often used because the clique $S_k(T)$82 is mandatory in every counted object (Cambie et al., 2023).

The principal local extremal theorem is that for $S_k(T)$83, if a $S_k(T)$84-clique $S_k(T)$85 maximizes $S_k(T)$86 in a $S_k(T)$87-tree $S_k(T)$88 of order $S_k(T)$89, then $S_k(T)$90 has degree $S_k(T)$91. A related formulation states that the maximum local mean order occurs at a $S_k(T)$92-clique that is not a major $S_k(T)$93-clique, so a maximizer must have degree $S_k(T)$94 or $S_k(T)$95. The proof passes through the characteristic $S_k(T)$96-tree $S_k(T)$97, for which

$S_k(T)$98

and then uses tree-level Kelmans and partial Kelmans operations to move away from major $S_k(T)$99-cliques while increasing local mean (Cambie et al., 2023, Li et al., 2023).

A second global-local comparison generalizes Jamison’s tree inequality. For any $k$00-tree $k$01 and any $k$02-clique $k$03,

$k$04

This bound is sharp, with $k$05-brooms providing asymptotically extremal examples. In the same framework, if $k$06 has no $k$07-clique of degree $k$08, then

$k$09

and for sufficiently large $k$10 the $k$11-star is the unique minimizer of $k$12 in this class (Cambie et al., 2023).

These $k$13-tree results are not merely analogical. They show that the mean subtree order belongs to a broader family of clique-anchored mean-order invariants controlled by recursive decompositions, characteristic trees, and local-global inequalities. A plausible implication is that many structural phenomena first observed for trees are manifestations of a more general elimination-order geometry, although a full transfer to arbitrary chordal graphs is not stated in the cited work.

7. Edge operations, nonmonotonicity, and open problems

A persistent misconception is that adding edges should always increase mean subtree order. Under the non-induced subtree-subgraph convention this is false. There are connected graphs $k$14 on the same vertex set with $k$15, and the decrease from adding a single edge can be as large as

$k$16

asymptotically. The construction uses a long path with large leaf-stars at its ends; adding one closing edge creates a family in which almost all subtrees contain the new edge and have mean order asymptotic to $k$17, whereas the original tree has mean close to $k$18 (Cameron et al., 2019).

The nonmonotonicity is robust. For every positive integer $k$19, there exist infinitely many pairs of connected graphs $k$20 with $k$21 such that $k$22. The same work confirms a conjectural comparison $k$23 for fixed $k$24 and sufficiently large $k$25, again illustrating that adding edges can lower the mean rather than raise it (Cambie et al., 2023).

Even the weaker statement that some edge addition must help is delicate. It is proved for trees: every tree $k$26 of order at least $k$27 has nonadjacent vertices $k$28 such that

$k$29

But universal edge-addition monotonicity fails, and edge deletion can also move the mean in either direction in general graph families (Cameron et al., 2019, Cambie et al., 28 Aug 2025).

Several open problems remain central. Under the non-induced subtree-subgraph convention, the complete graph $k$30 is conjectured to maximize $k$31 among connected graphs of order $k$32. A reduction shows that it would suffice to prove

$k$33

which would imply both $k$34 and an upper bound on the probability that a random subtree is spanning (Cambie et al., 28 Aug 2025). Under the connected-induced-set convention, two problems highlighted after the path-minimum theorem ask whether minimum degree at least $k$35 forces $k$36, and whether there is an absolute upper bound less than $k$37 on $k$38 for graphs with minimum degree at least $k$39 (Vince, 2021).

Taken together, these results position mean subtree order as a sharply structured, but convention-sensitive, graph invariant. In trees it is controlled by path-minimization, broom-like local maximizers, and near-caterpillar global maximizers. In connected graphs it admits both induced and non-induced extensions, each with a unique path minimizer but with different monotonicity behavior under edge operations. The invariant therefore serves simultaneously as a measure of combinatorial cohesion and as a test case for how local substructure counts propagate through global graph architecture.