Bias, Correlation, and a Tiny Proof

A compact derivation I keep forgetting. Clean statement, clean assumptions.

the setup

When analyzing stream ciphers, we often need to understand the relationship between bias and correlation. Here’s the key insight I always have to re-derive.

the statement

For random variables $X$ and $Y$ over ${0,1}$:

\[\Pr[X \oplus Y = 1] = \frac{1}{2} + \varepsilon\]

where $\varepsilon$ is the correlation bias.

why it matters

This simple relationship is the foundation of differential-linear cryptanalysis. The bias propagates through rounds, and understanding how it compounds is key to breaking ciphers.

The key insight is that for independent biases $\varepsilon_1, \varepsilon_2$:

\[\varepsilon_{total} \approx 2 \cdot \varepsilon_1 \cdot \varepsilon_2\]

quick notes

  • Bias compounds multiplicatively (mostly)
  • Independence assumptions matter more than we think
  • Always check your experimental correlations against theory
written at 2:30 am