The One-Nat Threshold

Five systems, one universal crossover

Across wildly different systems, a pattern repeats: detection becomes possible when the log-likelihood ratio (LLR) per unit of evidence crosses approximately 1 nat. Below this threshold the signal drowns in noise; above it, structure becomes extractable. Drag the sliders to watch each system cross the line.

1. Community Detection (Stochastic Block Model)

Two communities of size n/2. Edges within communities have probability a/n, across have b/n. The signal-to-noise ratio c = (a-b)/sqrt(2(a+b)) controls detectability. At c = 1, the phase transition occurs and communities become invisible.

1.00

2. Signal in Gaussian Noise

Detect a constant signal μ buried in Gaussian noise with std σ. The LLR per sample is μ²/(2σ²). The crossover happens at SNR = μ/σ = √2 ≈ 1.41, where LLR per sample hits 1 nat.

1.41

3. Biased Coin

A coin has bias p = 0.5 + ε. Each flip gives DKL(p || 0.5) nats of evidence. How many flips until the total LLR reaches 1? For small ε, you need roughly 1/(2ε²) flips.

0.100

4. Election Fraud Detection

In an election with K constituencies, a small bias δ favors one of 2 candidates. The LLR per constituency is approximately (1 - ln 2)δ². Detection requires K ≈ 1/[(1 - ln 2)δ²] constituencies.

0.50

5. Neutral Drift (Kimura)

In a population of size N, selection coefficient s is nearly neutral when Ns ≈ 1. The LLR per fixation event is 2s/(e2Ns - 1) - ln(2Ns/(e2Ns - 1)). Over many fixations, the signal accumulates.

1.00