Statistical Properties of the King Wen Sequence: An Anti-Habituation Structure That Does Not Improve Neural Network Training

Researcher Augustin Chan has published a rigorous analysis of the King Wen sequence, a 3000-year-old ordering of 64 hexagrams from the ancient Chinese text I-Ching. Using Monte Carlo permutation analysis against 100,000 random baselines, the study found the sequence exhibits four statistically significant properties: higher-than-random transition distance (98.2nd percentile), negative lag-1 autocorrelation, yang-balanced groups of four, and asymmetric distance patterns. These properties superficially resemble principles from modern machine learning concepts like curriculum learning and curiosity-driven exploration.

Motivated by this resemblance, Chan tested whether the sequence could actually benefit neural network training through three experiments: learning rate schedule modulation, curriculum ordering, and seed sensitivity analysis. The tests were conducted across two hardware platforms (NVIDIA RTX 2060 with PyTorch and Apple Silicon with MLX) with uniformly negative results. King Wen learning rate modulation degraded performance at all tested amplitudes, and as a curriculum ordering, it was either the worst non-sequential ordering or within noise. A 30-seed sweep confirmed that only King Wen's degradation exceeded natural seed variance.

The paper explains that the sequence's high variance—the very property that makes it statistically distinctive—actually destabilizes gradient-based optimization. This demonstrates that anti-habituation in a fixed combinatorial sequence is not equivalent to effective training dynamics. The research serves as an important cautionary tale about applying mathematically interesting patterns to practical AI training without empirical validation.