Weighted sum of a series of values, weighted based on the scale of the error.

184 Views Asked by At

I need to find the value and error of a weighted sum of values all with their own respective error.

Say I have three data points:

$ x_1 = 7 \pm 1$
$ x_2 = 9 \pm 0.8$
$ x_3 = 10 \pm 0.7$

I need to find the mean of these values, weighted by the magnitude of their errors so that values with greater error have less weighting.

I have tried a couple of approaches but have gotten unreasonable answers. Searching the web, there seems to be a lot of noise and cannot find anything specifically that I am looking for.

My initial approach was to calculate in this manner:

$\bar{x} = \frac{\sum_i x_iw_i}{\sum_i w_i}$

For my example this gives:

$\bar{x} = \frac{7*(1)+ 9*(0.8) + 10*(0.7)}{1+0.8+0.7} = 8.48$

I know this is obviously incorrect as it gives greater weighting to the values with larger error. I'm struggling to work out a mathematically proper fix. Can I just take the inverse of the weighting?

Errors I can propagate trivially once I know how to calculate $\bar{x}$

1

There are 1 best solutions below

0
On BEST ANSWER

According to Chapter 7, "Weighted Averages" in An Introduction to Error Analysis, Second Edition, by John R. Taylor, the appropriate procedure is as follows.

Let's say your three values are $x_i \pm \sigma_i$ for $i=1,2,3$. Compute weights $w_i$ by $$w_i = \frac{1}{\sigma_i^2}$$ Then your best estimate is the weighted average $$x_{wav} = \frac{\sum w_i x_i}{\sum w_i}$$ The idea is that these weights minimize the variance of $x_{wav}$.