---
title: Horton–Strahler Number (HS)
url: https://www.emergentmind.com/topics/horton-strahler-number-hs
type: topic
---

# Horton–Strahler Number (HS)

The Horton–Strahler number (HS), also called the Strahler number or the register function, is an integer-valued invariant of a rooted tree that measures branching complexity rather than mere depth. In its standard recursive form, a node inherits the maximum order of its children unless that maximum is attained by at least two children, in which case the order increases by one; in complementary formulations, HS is characterized by pruning dynamics or by the largest complete or perfect binary tree embeddable in the given tree [1511.01558] [2406.03025] [2003.08627]. Across the literature, HS serves as a canonical coordinate for hierarchical branching in hydrology, random-tree asymptotics, analytic combinatorics, compressed-tree algorithms, formal-language theory, parity games, planar reticular networks, and empirical studies of sentence structure [2010.08613] [1504.05535] [1110.1413].

## 1. Core definitions and equivalent characterizations

For rooted directed binary trees, the most common convention assigns order \(1\) to leaves and then defines the order of an internal node \(v\) with children \(ch_1(v),ch_2(v)\) by
\[
S(v)=S(ch_1(v))+1 \quad \text{if } S(ch_1(v))=S(ch_2(v)),
\]
and otherwise
\[
S(v)=\max\big(S(ch_1(v)),S(ch_2(v))\big).
\]
The HS number of the tree is the value at the root, and this bottom-up computation is unique [2307.02697]. In full binary-tree settings the same rule is often written
\[
k(p)=\max(i,j)+\delta_{ij},
\]
where \(i\) and \(j\) are the orders of the two children and \(\delta_{ij}\) is the Kronecker delta [1511.01558].

A second convention, standard in the register-function literature, assigns value \(0\) to a leaf. For a binary tree \(t\) with subtrees \(t_1,t_2\),
\[
\Reg(\square)=0,\qquad
\Reg(t)=
\begin{cases}
\max\{\Reg(t_1),\Reg(t_2)\}, & \Reg(t_1)\neq \Reg(t_2),\\[4pt]
\Reg(t_1)+1, & \Reg(t_1)=\Reg(t_2).
\end{cases}
\]
This is the same invariant up to a shift in indexing convention [1602.06200]. A similar \(0\)-based formulation is used for ranked trees in compressed-representation settings, where internal nodes are evaluated by taking the maximum child value and incrementing iff that maximum is not unique [1504.05535].

Several papers emphasize equivalent structural interpretations. For a rooted tree \(T\), the Horton–Strahler number is the height of the tallest complete binary tree that can be homeomorphically embedded in \(T\), with the embedding required to preserve ancestral structure, lexicographic order, and most recent common ancestors [2406.03025]. In ordered rooted trees, the Strahler number is also described as the largest height of a perfect binary tree that is a minor of the tree [2003.08627]. In pruning-based treatments of full binary trees, the order of a vertex \(v\) and its parental edge is the minimal number of prunings needed to eliminate the subtree rooted at \(v\):
\[
k(v)=\min\{k\ge 1:\ R^k(T_v)=\varnothing\}.
\]
This makes pruning an intrinsic dynamical interpretation of HS [1511.01558].

The notion extends naturally from vertices to branches. In self-similar-tree and random-binary-tree analyses, a branch of order \(r\) is a maximal connected path, or maximal connected sequence, of nodes or vertices and parental edges all having order \(r\) [1907.13346] [1511.01558]. This branch-centric perspective underlies Horton laws, Tokunaga statistics, and branch-count asymptotics.

## 2. Pruning dynamics, Tokunaga structure, and Horton laws

In finite rooted full binary trees, pruning \(R\) is defined by cutting all leaves and their parental edges and then removing the resulting chains of degree-two vertices by series reduction. A basic structural fact is that pruning decreases the Horton–Strahler order of every vertex and branch by unity. At the branch-count level,
\[
N_k[T]=N_{k-1}[R(T)],\qquad k\ge 2,
\]
and for side branches,
\[
N_{ij}[T]=N_{i-1,j-1}[R(T)],\qquad 2\le i<j.
\]
This order shift is the mechanism that makes pruning the natural dynamics behind self-similarity [1511.01558].

For a probability measure \(\mu_K\) on the set \(\mathcal T_K\) of full binary trees of order \(K\), the expected number of order-\(k\) branches is
\[
N_k[K]=\mathbb E_K(N_k),
\]
and the expected number of order-\(i\) branches that merge order-\(j\) branches is
\[
N_{ij}[K]=\mathbb E_K(N_{ij}).
\]
The Tokunaga coefficients are then
\[
T_{ij}[K]=\frac{N_{ij}[K]}{N_j[K]},
\]
measuring the expected number of order-\(i\) side branches attached to a typical order-\(j\) branch [1511.01558].

A tree ensemble is called coordinated when \(T_{ij}[K]\) does not depend on \(K\), and mean self-similar when the coefficients depend only on order difference:
\[
T_{ij}=T_{j-i},
\qquad\text{equivalently}\qquad
T_{i,i+k}[K]=T_k.
\]
In this form, self-similarity is pruning invariance of the expected side-branching statistics. The strong Horton law is the asymptotic geometric decay
\[
\frac{N_k[K]}{N_1[K]}\to R^{\,1-k}\qquad\text{as }K\to\infty
\]
for every fixed \(k\ge 1\), equivalently \(N_{k+1}[K]/N_k[K]\to R\) [1511.01558].

A central theorem identifies the exact condition under which the strong Horton law holds for self-similar trees:
\[
\limsup_{k\to\infty}(T_k)^{1/k}<\infty
\]
is necessary and sufficient. When this condition holds, the Horton exponent is determined by the unique real root \(w_0\in(0,1/2]\) of
\[
t(z)=-1+2z+\sum_{j=1}^{\infty} z^j T_j,
\]
through
\[
R=\frac{1}{w_0}.
\]
The expected Horton numbers satisfy the recursion
\[
N_k[K]=2N_{k+1}[K]+\sum_{j=1}^{K-k}T_j\,N_{k+j}[K],\qquad 1\le k<K,
\]
which is the mean counterpart of deterministic branch-count equations used in earlier Horton-law work [1511.01558].

This self-similar viewpoint has an information-theoretic analogue. For planted binary plane trees satisfying Horton law with exponent \(R\), the entropy rate is
\[
\mathscr{H}_{\infty}(R)=1-\frac{1-H(2/R)}{2-2/R},
\]
where \(H(z)\) is the binary entropy function. In that model, \(\mathscr{H}_\infty(2)=0\), \(\mathscr{H}_\infty(4)=1\), and \(\lim_{R\to\infty}\mathscr{H}_\infty(R)=1/2\), so the Horton exponent directly constrains asymptotic structural complexity [1804.06989].

## 3. Enumerative combinatorics, reductions, and refined variants

Classical analytic combinatorics links HS to Dyck-path height. If \(\overline{\mathcal B}_{n,s}\) denotes full binary trees with \(n\) internal vertices and Horton–Strahler number \(s\), and \(\overline{\mathcal D}_{n,s}\) denotes Dyck paths of length \(2n\) whose height \(h\) satisfies
\[
\left\lfloor \log_2(1+h)\right\rfloor=s,
\]
then
\[
\left|\overline{\mathcal B}_{n,s}\right|=\left|\overline{\mathcal D}_{n,s}\right|.
\]
Equivalently, the HS class \(s\) corresponds to Dyck-path heights in the interval
\[
2^s-1 \le h \le 2(2^s-1).
\]
Thus the classical HS statistic records a logarithmic bin of Dyck-path height rather than the exact height itself [2406.03025].

A recent refinement resolves this logarithmic collapse. An increasing sequence of trees \((\tau_i)_{i\ge 0}\) interpolates the complete binary trees, with \(\tau_{2^h-1}=\mathrm{cb}(h)\). The refined Horton–Strahler number is
\[
\mathcal S(T):=\max\{r:\tau_r\text{ can be embedded in }T\}.
\]
The main refined theorem states that if \(\mathcal B_{n,h}\) is the set of full binary trees with \(n\) internal vertices and \(\mathcal S(T)=h\), and \(\mathcal D_{n,h}\) is the set of Dyck paths of length \(2n\) and exact height \(h\), then
\[
|\mathcal B_{n,h}|=|\mathcal D_{n,h}| \qquad\text{for all }n,h\ge 0.
\]
The classical HS number is recovered by grouping refined values in the range
\[
2^s-1\le \mathcal S(T)\le 2(2^s-1).
\]
This refinement makes exact Dyck height, rather than only \(\lfloor \log_2(1+h)\rfloor\), visible on the tree side [2406.03025].

A complementary combinatorial interpretation comes from reduction theory. For binary trees, a reduction \(\Phi\) deletes all leaves and then repeatedly merges nodes with exactly one child until no such node remains. The register function satisfies
\[
\Reg(\Phi(t))=\Reg(t)-1
\qquad\text{whenever }\Reg(t)\ge 1,
\]
and therefore
\[
\Phi^r(t)=\square \quad\Longleftrightarrow\quad \Reg(t)=r.
\]
In this formulation, HS is exactly the maximal number of reductions that can be applied before the tree collapses to a single leaf [1602.06200].

This reduction viewpoint also induces a natural family of substructures. An \(r\)-branch is a maximal chain of nodes labeled \(r\), and for uniformly random binary trees with \(n\) internal nodes the expected number \(E_{n;r}\) of \(r\)-branches satisfies
\[
E_{n;r} = \frac{n}{4^r} +\frac{1}{6}\Bigl(1+\frac{5}{4^r}\Bigr)+\cdots,
\]
while the variance begins with
\[
V_{n;r} = \frac{4^r-1}{3\cdot 16^r}\,n+\cdots.
\]
The leading factor \(4^{-r}\) exhibits the same geometric rarity of higher-order structure that appears in Horton laws [1602.06200].

Explicit fixed-level generating functions are available in several tree classes. For binary trees and weighted unary-binary trees, a substitution such as
\[
z=\frac{u}{1+(a+2)u+u^2}
\]
yields closed forms for the generating functions \(R_p(z)\) of trees with Horton–Strahler number exactly \(p\). In those families, the average HS is logarithmic, with main term \(\log_4 n\) and a tiny periodic fluctuation arising from dyadic singular structure [2106.14782].

## 4. Random-tree asymptotics and fluctuation theory

For critical Galton–Watson trees with finite variance conditioned on total size \(n\), the HS number grows logarithmically:
\[
\frac{\HS(T_n)}{\log_2 n}\xrightarrow{P}\frac12.
\]
Equivalently, \(\HS(T_n)=\bigl(\tfrac12+o_P(1)\bigr)\log_2 n\). The proof combines local convergence to Kesten’s infinite tree for the lower bound with a rotationally invariant argument for the upper bound [2010.08613].

This finite-variance result extends to critical Galton–Watson trees in the domain of attraction of an \(\alpha\)-stable law. For \(\alpha\in(1,2]\), under conditioning on size \(n\),
\[
S(\tau)\sim \frac{1}{\alpha}\log_{\alpha/(\alpha-1)} n
\qquad\text{in probability}.
\]
The case \(\alpha=2\) recovers the classical \(\log_4 n\) behavior. The boundary case \(\alpha=1\), the spectrally positive Cauchy regime, is not universal: the correct scale depends on the slowly varying tail through a function \(\Upsilon\), and the paper gives explicit examples where the growth is of order \(\ln n/\ln\ln n\) or \(\ln n/\ln\ln\ln n\) [2307.05983].

A real-valued regularization removes deterministic oscillations from the integer-valued statistic. In leaf-weighted trees, the weighted Horton–Strahler number is defined recursively so that its floor equals the classical HS number, and its fractional part is one of the leaf weights. For \(GW_\alpha\)-weighted trees with leaf weights distributed as \(\mathsf{FExp}(\gamma)\), where \(\gamma=\ln\frac{\alpha}{\alpha-1}\), the weighted HS number is exactly exponential with mean \(1/\gamma\). Under size conditioning and rescaling, a suitable exponential transform of this quantity converges jointly with the tree itself to a measurable functional of the \(\alpha\)-stable Lévy tree, called the Strahler dilation; in the Brownian case \(\alpha=2\), twice the Strahler dilation has the same law as the Brownian tree height [2401.13771].

For random planar full binary trees with \(n\) leaves, fixed-order branch counts satisfy both central-limit and large-deviation principles. If \(S_{r,n}\) denotes the number of branches of order \(r\), then for each fixed \(r\ge 1\),
\[
\sqrt{n}\left(\frac{S_{r+1,n}}{n}-\frac{1}{4^r}\right)
\Rightarrow
N\!\left(0,\frac{4^r-1}{3\cdot 16^r}\right),
\]
and
\[
\sqrt{n}\left(\frac{S_{r+1,n}}{S_{r,n}}-\frac14\right)\Rightarrow N(0,4^{r-3}).
\]
The corresponding large-deviation rate function is the Legendre transform of an iterated cumulant generator \(\varphi^r\), with
\[
\varphi(\xi)=\frac{\xi}{4}+\log\!\left(\cosh\frac{\xi}{4}\right),
\]
and the law concentrates near the Horton-law value \(4^{-r}\) [1701.03213] [1907.13346].

Other probabilistic tree models yield distinct HS regimes. For Kingman’s coalescent tree, asymptotic branch densities satisfy a Smoluchowski-type ODE system, and weak root-Horton self-similarity holds:
\[
\lim_{k\to\infty}(N_k)^{1/k}=R,
\qquad 2\le R\le 4,
\]
with numerical estimate \(R\approx 3.0438279\) [1207.7108]. In uniform simple butterfly trees with \(N=2^n\) nodes,
\[
E[HS(\mathcal T_n^{B})] = \log_8 N + O(1),\qquad
\operatorname{Var}(HS(\mathcal T_n^{B})) = \frac29 \log_8 N + O(1),
\]
and a central limit theorem holds for the normalized HS. For general butterfly trees, a deterministic upper bound
\[
HS(\mathcal T^{B}) \le \lfloor \log_4 N\rfloor
\]
is sharp, and the paper reports empirical concentration near that upper support bound [2509.11384].

## 5. Algorithmic, compression-theoretic, and complexity-theoretic roles

HS is algorithmically tractable on several compressed tree representations. For a ranked tree encoded by a string straight-line program over its preorder traversal, the Strahler number can be computed in polynomial time. The proof uses a general evaluation framework for SLP-compressed trees, reduction to caterpillar trees, and a Strahler-specific refinement that tracks the deepest one or two occurrences needed to detect whether the maximum child value is unique [1504.05535].

The representation of the tree is decisive for complexity. When a binary tree is given as a term, the problem “is \(st(t)\ge k\)?” is complete for uniform \(\mathsf{NC}^1\). With pointer representation, the same problem is \(L\)-complete. For trees represented succinctly by a binary DAG or a tree straight-line program, the general problem is \(\mathsf{P}\)-complete. For fixed thresholds \(k\), finer classifications appear: for example, the term-representation problem is \(\mathsf{uTC}^0\)-complete for \(k\ge 4\), the DAG case lies in \(\mathsf{UL}\cap\mathsf{coUL}\), and the TSLP case is \(\mathsf{NL}\)-complete for \(k\ge 2\) [2512.19060].

Formal-language applications use HS as a bounded-branching parameter on derivation trees. In Chomsky normal form, derivation trees are binary after terminal leaves are removed, and if such a tree has height \(d\) and Horton–Strahler number \(s\), then it has at most \(d^s\) leaves and at most \(2d^s\) nodes. This size bound motivates the class \(\mathrm{CFG}(k)\) of grammars whose acyclic derivation trees all have HS at most \(k\). For every fixed \(k\ge 1\), the context-free membership problem for symmetric groups restricted to grammars from \(\mathrm{CFG}(k)\) is NP-complete, whereas the unrestricted context-free membership problem for symmetric groups is PSPACE-complete [2206.11756].

In parity games, the Strahler number is lifted from trees to attractor decompositions. The Strahler number of a game is defined from the smallest Strahler number of decomposition trees of the relevant dominions, and it coincides with both the progress-measure Strahler number and Lehtinen’s register number. If a parity game has Strahler number \(k\), then it can be solved in quasi-linear space and in time polynomial in the number of vertices \(n\) and linear in
\[
(d/2k)^k,
\]
where \(d\) is the number of priorities. In particular, if
\[
k\cdot \lg(d/k)=O(\log n),
\]
then parity games are solvable in polynomial time [2003.08627].

## 6. Generalizations beyond rooted trees and empirical interpretations

Classical HS presupposes a tree structure and therefore fails on reticular networks with loops: cyclic dependencies prevent a consistent upstream-to-downstream ordering. For weighted planar reticular networks, one generalization orders loops and tree structure separately. Edges are sorted by increasing weight, every face starts with level \(1\), and when an edge separating two distinct faces \(f_L\) and \(f_R\) is removed, the merged face receives level
\[
\lambda(f_{\text{merged}})=
\begin{cases}
\lambda(f_L)+1, & \lambda(f_L)=\lambda(f_R),\\[4pt]
\max\{\lambda(f_L),\lambda(f_R)\}, & \lambda(f_L)\neq \lambda(f_R),
\end{cases}
\]
while the removed edge gets
\[
\lambda(e_i)=\min\{\lambda(f_L),\lambda(f_R)\}.
\]
The removed edges form a reticular hierarchy on the co-tree of faces, and the remaining edges form a spanning tree on which the classical Horton–Strahler scheme is applied [1110.1413].

This planar generalization preserves the logic of equal-level merging while separating loop-forming edges from tree edges. It also supports Tokunaga-style side-branching statistics for both the tree and the co-tree. The sensitivity analysis is asymmetric: a transposition of two reticular edges changes the loop hierarchy locally by at most one level, whereas a transposition involving a reticular edge and a tree edge can cause arbitrarily large changes in the tree hierarchy [1110.1413].

In empirical syntax, HS has been proposed as a branching-complexity measure for sentence trees and as a lower bound on memory needed for bottom-up or shift-reduce processing. For binary trees with \(n\) leaves, the upper limit is
\[
\lfloor \log_2 n \rfloor + 1,
\]
while the lower limit is \(1\) if \(n=1\) and \(2\) if \(n>1\). Dependency trees are first binarized, and upper and lower limits are computed over all binarizations [2307.02697].

Using Universal Dependencies v2.8 across 202 corpora in 114 languages, the reported average HS values are about \(2.71\pm 0.60\) under grammar-based binarization and \(3.21\pm 0.74\) under a heuristic binarization based on modifier distance and left/right placement. The median is \(4\) for upper limits and \(3\) for grammar-based binarization, heuristic binarization, and lower limits; values above \(4\) are rare, the largest observed value is \(7\), and the smallest is \(1\). When random trees are sampled to match sentence-length distributions, the average HS is very close to the linguistic values, leading the paper to conclude that tree-size distribution, rather than language-specific structure, largely determines the observed range. In this setting, HS is interpreted as a rigorous lower bound on the number of stack spaces or memory areas needed for tree evaluation, connecting it to Sethi–Ullman-style register minimization and earlier psycholinguistic claims about \(3\) to \(5\) memory resources [2307.02697].

Source: https://www.emergentmind.com/topics/horton-strahler-number-hs