Incremental calculation of weighted mean and （增量计算加权平均数）.pdf

Incremental calculation of weighted mean and variance Tony Finch fanf2@cam.ac.uk dot@dotat.at University of Cambridge Computing Service February 2009 Abstract In these notes I explain how to derive formulae for numerically stable calculation of the mean and standard deviation, which are also suitable for incremental on-line calculation. I then generalize these formulae to weighted means and standard deviations. I unpick the difficulties that arise when generalizing further to normalized weights. Finally I show that the exponentially weighted moving average is a special case of the incremental normalized weighted mean formula, and derive a formula for the exponentially weighted moving standard deviation. 1 Simple mean Straightforward translation of equation 1 into code can suffer from loss of precision because of the difference in magnitude between a sample and the sum of all samples. Equation 4 calculates the mean in a way that is more numerically stable because it avoids accumulating large sums. n 1 µn = xi (1) n i=1 n−1 1 = (x + x ) (2) n