# Archived Weighted least-squares fit error propagation

1. Nov 17, 2013

### mbigras

1. The problem statement, all variables and given/known data
Suppose we measure N pairs of values (xi, yi) of two variables x and y that are supposed to statisfy a linear relation y = A + Bx suppose the xi have negligible uncertainty and the yi have different uncertainties $\sigma_{i}$. We can define the weight of the ith measurement as $w_{i} = 1/\sigma_{i}$. Then the best estimates of A and B are:

$$A = \frac{\Sigma w x^{2}\Sigma w y - \Sigma w x \Sigma w x y}{\Delta}\\ B = \frac{\Sigma w \Sigma w x y - \Sigma w x \Sigma w y}{\Delta}\\ \Delta = \Sigma w \Sigma x^{2} - \left(\Sigma w x \right)^{2}$$

Use error propagation to prove that the uncertainties in the constants A and B are given by

$$\sigma_{A} = \sqrt{\frac{\Sigma w x^{2}}{\Delta}}\\ \sigma_{B} = \sqrt{\frac{\Sigma w}{\Delta}}$$

2. Relevant equations

rules for sums and differences
$$q = x \pm z\\ \delta q = \sqrt{(\delta x)^{2} + (\delta z)^{2}}\\$$
rules for products and quotients
$$\delta q = \sqrt{(\delta x/x)^{2} + (\delta z/z)^{2}}\\$$

3. The attempt at a solution
What I'm thinking is because the uncertainty in x is negligible it will be treated like a constant. I'm not sure how to deal with all these sums. I feel like I don't know how to approach this question.

Last edited: Nov 17, 2013
2. May 1, 2016

### haruspex

Try calculating E(A2) etc. The negligible uncertainty in the x allows you to treat the Δ denominators as constant.