I know that if I am given $$f=a+b$$ then $$\sigma_\text{specific} = \sqrt{\sigma_a^2 + \sigma_b^2} \tag1$$ which is a direct result from the more general $$\sigma_\text{general} = \sqrt{\left( \frac{\partial f}{\partial a} \right)^2 \sigma_a^2 + \left( \frac{\partial f}{\partial b} \right)^2 \sigma_b^2} \tag2$$

I have noticed that Eq. 1 and Eq. 2 only agree with each other when $a \neq b$. For example, let $a=b=50 \pm 4 \implies f = a + b = 2a$. Then, $$\sigma_\text{specific} = \sqrt{4^2 + 4^2} = 5.7$$ according to Eq.1, while according to Eq.2 $$\sigma_\text{general} = \sqrt{2^2\cdot4^2 + 0} = 8$$ So, which one is correct? and what did I do wrong?


Your result from the general formula is correct (though it doesn't in general include correlations). Specifically, if you have $$ f = a + a $$ The errors on $a$ and $a$ are obviously (completely) correlated. If one of the terms was bigger than expected, so was the other.

If you have $$ f = a + b $$ with independent, uncorrelated errors on $a$ and $b$, the the overall error on $f$ is slightly smaller, as (heuristically) it is less probable to that both $a$ and $b$ were bigger/smaller than expected, as one could be bigger than expected and the other smaller.

The simple rule of adding errors in quadrature for a sum of terms assumes that the errors are not correlated. In general, for a sum of variables $$ y = \sum x_i $$ then including correlations, the variance on $y$ is $$ \sigma^2 = \sum_i \sum_j \sigma_i \sigma_j \rho_{ij} $$ where $\rho_{ij}$ is the correlation coefficient betwen $x_i$ and $x_j$. In the case of uncorrelated variables, $\rho_{ij} = \delta_{ij}$ such that $$ \sigma^2 = \sum_i \sigma_i^2 $$ If the $n$ variables have the same variances (say $\sigma_1$), we find $$ \sigma = \sqrt{n} \sigma_1 $$

On the other hand, in the case of $n$ absolutely correlated variables ($\rho_{ij} = 1$ for all $i$ and $j$) with the same variances ($\sigma_1$), we instead find $$ \sigma = n \sigma_1 $$ Note that this is a factor $\sqrt{n}$ larger than before. You considered the case of $n=2$, resulting in a factor of $\sqrt{2}$ difference (your $8/5.7 \approx \sqrt{2}$).

Note that this analysis of linear functions can be approximately extended to the general non-linear case by Taylor expanding a non-linear to first-order, resulting in your formula involving partial derivatives.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.