Statistical Mechanics of Strahler Number via Random and Natural Language Sentences

Kumiko Tanaka-Ishii, Akira Tanaka

Introduction

For decades, cognitive scientists and linguists have observed a recurring pattern: human language processing appears to require a small, nearly constant amount of short-term memory—typically cited as 3 to 5 "chunks" or memory areas. While empirical studies have confirmed these values for parsing, a rigorous theoretical explanation for why the number is specifically 3 or 4 has remained elusive. This paper investigates whether this "magical number" is a fundamental property of sentence structure, how it scales with sentence length, and whether it is a unique feature of human language or a universal characteristic of branching trees.

Methodology: The Strahler Number as a Memory Metric

The core of this study is the Strahler number, a measure of tree complexity originally used to describe river networks. In a binary tree, the Strahler number is calculated bottom-up: leaves are assigned a value of 1; if a node's children have different values, the node takes the maximum; if both children have the same value $i$, the node takes the value $i+1$. This recursive process is illustrated in Figure 1.

Figure 1: Calculation of the Strahler number on a binary tree. Leaves start at 1, and internal nodes increment when their children have equal values.

Crucially, in computer science, the Strahler number of a computation tree equals the minimum stack depth (memory) required to evaluate that tree using shift-reduce operations. Figure 2 demonstrates how the sequence of shifts and reductions determines the peak stack depth, with the Strahler number representing the optimal memory requirement.

Figure 2: Shift-reduce operations on an expression tree. The Strahler number of the tree corresponds to the minimum number of memory slots required to process the structure.

To apply this to real-world data, the authors analyzed the Universal Dependencies (UD) version 2.8 dataset, comprising 202 corpora across 114 languages. They binarized these dependency trees using both grammar-based and distance-based heuristics to find the absolute maximum and minimum Strahler numbers possible for any given sentence. These results were then compared against null models of random binary trees ($R_2(n)$) and all possible rooted trees ($R(n)$).

Key Findings and Results

The Logarithmic "Constant"

The paper proves that the Strahler number is not a true constant but grows logarithmically with sentence length ($n$). It only appears constant (3–4) because human sentences typically fall within a specific range of lengths. Across all 114 languages, the average Strahler number for sentences is approximately 3.1 to 3.2, with a mean upper limit of 3.56 and a mean lower limit of 2.71.

Universal Geometry

Perhaps the most surprising finding is that natural language trees behave almost identically to random trees of the same size. This suggests that the memory constraints of language are not necessarily a specialized biological adaptation for speech, but rather a general property of how branching structures work. The vast majority of human sentences have a Strahler number between 2 and 4, and this consistency holds across different language families, genres, and modes (speech vs. writing).

Independence from Entropy

The authors further show that the Strahler number is uncorrelated with Shannon entropy. While entropy measures the complexity of a sequence, the Strahler number measures the complexity of the underlying hierarchy, providing a formal lower limit on the memory required for parsing.

Conclusion and Future Work

This research provides a theoretical grounding for the "magical number" 3–5 observed in psychological experiments, showing it is a mathematical consequence of tree geometry. However, the link is currently tied to the shift-reduce parsing model. Future work could refine these limits by incorporating specific linear word-order constraints (like non-projectivity) and mapping these theoretical "stack spaces" to specific biological neural mechanisms in the human brain.

Summary updated

BibTeX
@article{tanakaishii2023statistical,
  author = {Kumiko Tanaka-Ishii and Akira Tanaka},
  title = {{Statistical Mechanics of Strahler Number via Random and Natural Language Sentences}},
  journal = {Journal of Statistical Mechanics Theory and Experiment},
  year = {2023},
  doi = {10.1088/1742-5468/ad0222},
}
Figure