Box-Muller transform

1. Given r and φ independently uniformly distributed in (0,1], compute:

   <math>

   z_0 = \cos(2 \pi \varphi) \cdot \sqrt{-2 \ln r} </math>

   and

   <math>

   z_1 = \sin(2 \pi \varphi) \cdot \sqrt{-2 \ln r} </math>
2. Given x and y independently uniformly distributed in [-1,1], set r = x²+y². If r = 0 or r > 1, throw them away and try another pair (x, y). Thus, all points left after this filtering process will be uniformly distributed within a unit circle. Then, for these filtered points, compute:

   <math>

   z_0 = x \cdot \sqrt{\frac{-2 \ln r}{r}} </math>

   and

   <math>

   z_1 = y \cdot \sqrt{\frac{-2 \ln r}{r}} </math>

The second method is faster because it uses only one transcendental function[?] instead of three, even though it throws away 21% of the numbers.

External links

• Generating Gaussian Random Numbers (http://www.taygeta.com/random/gaussian)