Wasserstein

Distribution shape, transport cost, self-similarity
distributionaldim distribution space4 metrics

What It Measures

How the distribution of values differs from uniform, and how stable that distribution is across the signal.

Bins the data into a 32-bin histogram and treats it as a probability distribution. Then asks three questions: how far is this distribution from uniform (optimal transport cost)? How concentrated is it (peak height)? And does the first half of the signal look like the second half (self-similarity)?

Metrics

concentration

The peak bin height times the number of bins. 1.0 means uniform (De Bruijn scores exactly 1.0 — its construction guarantees every byte pattern appears equally). Above 1.0 means the distribution has a spike. Collatz gap lengths (31.8), Rainfall (31.5), and Forest fire (29.4) are the most concentrated signals in the atlas — their heavy-tailed distributions pile most of their mass into the lowest bin.

dist_from_uniform

Earth mover's distance from the uniform distribution: the minimum amount of "dirt" you'd need to move to make the histogram flat. Collatz gap lengths (0.48) and Rainfall (0.48) are farthest from uniform. Neural net pruned weights (0.46) are close behind — pruning creates a spike at zero. De Bruijn scores near 0 (already uniform).

entropy

Shannon entropy of the 32-bin histogram. De Bruijn, circle map quasiperiodic, and phyllotaxis all score 5.0 (near the maximum of 5 bits — flat distribution). Collatz gap lengths scores 0.05 (almost all mass in one bin). This is the classical measure of distributional spread, here computed on the Wasserstein embedding.

self_similarity

One minus the earth mover's distance between the first-half and second-half histograms. 1.0 means the distribution is perfectly stable over time (logistic period-4, logistic period-2, De Bruijn). Hilbert walk scores 0.60 (its deterministic sweep creates different distributions in the first and second halves). This catches nonstationarity that entropy and concentration miss: a signal can have high entropy overall but low self_similarity if its distribution drifts.

Atlas Rankings

concentration
SourceDomainValue
Collatz Gap Lengthsnumber_theory31.8133
Rainfall (ORD Hourly)climate31.4687
Forest Fireexotic29.4376
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
De Bruijn Sequencenumber_theory1.0000
dist_from_uniform
SourceDomainValue
Collatz Gap Lengthsnumber_theory0.4842
Rainfall (ORD Hourly)climate0.4830
Neural Net (Pruned 90%)binary0.4643
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
De Bruijn Sequencenumber_theory0.0000
entropy
SourceDomainValue
De Bruijn Sequencenumber_theory5.0000
Circle Map Quasiperiodicchaos4.9996
Phyllotaxisbio4.9996
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Collatz Gap Lengthsnumber_theory0.0517
self_similarity
SourceDomainValue
Logistic r=3.5 (Period-4)chaos1.0000
Logistic r=3.2 (Period-2)chaos1.0000
De Bruijn Sequencenumber_theory1.0000
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Hilbert Walkexotic0.5974

When It Lights Up

Wasserstein self_similarity is the distributional lens's nonstationarity detector. Signals that change character midstream — sensor drift, regime switches, concatenated recordings — score low on self_similarity while potentially scoring high on all other distributional metrics. In the atlas, Wasserstein's concentration axis separates the heavy-tailed cluster (Collatz, rainfall, forest fire) from the uniform-distribution cluster (PRNGs, De Bruijn), while self_similarity provides an orthogonal axis that catches temporal instability invisible to any single-histogram metric.

Open in Atlas
← 2-adicFisher Information →