Ultrametric Convergence — Variance Collapse Under Coarse-Graining

Watch diversity collapse into uniformity — not by intervention, but because ultrametric geometry makes it inevitable

Tree Parameters

Results

Variance at level k
Variance at k = 0 (leaves)
Collapse Ratio
Var(X) = E[Var(X|Y)] + Var(E[X|Y])
Law of total variance: coarse-graining reduces variance because it eliminates within-group variation. The second term Var(E[X|Y]) ≤ Var(X).
In an ultrametric tree, each level k groups leaves by their ancestor at depth d−k. Variance monotonically decreases with k.

Ultrametric Tree (k = 0)

Leaf Value Distribution

Variance vs. Coarse-Graining Level

Welcome!