It really doesn't. The whole reason this technique works is that our brain generally processes effects like noise and grain in terms of macro properties, and even if none of the pixels are anywhere like the right value, it will still look "right" as long as the noise distribution is closish.