0
$\begingroup$

How to express the distribution of measurements ($X_m \sim \mathcal{D}_\mathrm{unknown}$) given that the underlying physical process being measured follows a Gaussian distribution ($X_p \sim \mathcal{N}(\mu_p, \sigma_p)$) and the measurement uncertainty follows a Gaussian distribution ($\mathcal{N}(\mu_u, \sigma_u)$) too?

Note: $\mu_u$ here is variable and takes the value of the specific instance/sample $X_p$.

Update:

Can one start in terms of PDFs as follows?

Given the PDF of the physical process: $$f_{X_p}(x) = \frac{1}{\sigma_p\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{x-\mu_p}{\sigma_p}\right)^2}$$

In order to express the PDF of the measurements affected by measurement error ($\mathcal{N}(X_p, \sigma_u)$), the idea is to use an specific sample $t$ of the physical process as the mean of the measurement uncertainty distribution. Since the specific sample $t$ follows a PDF $f_{X_p}(t)$, therefore integration is used over the whole range of $t$, and weighted by the PDF: $$f_{X_m}(x) =\int^{\infty}_{-\infty} \frac{1}{\sigma_u\sqrt{2\pi}}e^{-\frac{1}{2}\left(\frac{x-t}{\sigma_u}\right)^2}f_{X_p}(t)dt$$

Is the above formulation reasonable?

Update after David K's answer: This question can also serve to get a more intuitive understanding of the convolution of two normal distributions.

$\endgroup$
8
  • $\begingroup$ I thought measurement error was usually considered to have a distribution like $\mathcal{N}(0, \sigma_u)$, that is, it's something added to the actual value in a way that introduces uncertainty but not bias. Is there a reason to formulate it differently? $\endgroup$
    – David K
    Feb 17, 2020 at 18:22
  • $\begingroup$ The idea here is to figure out a way to express the measurements of the noise-like physical process including the measurement uncertainty. Basically, the question is how the distribution of the noisy measurements of the noise-like physical process look like? Which I think is different than adding the two distributions, do you agree? $\endgroup$
    – Omer
    Feb 18, 2020 at 10:19
  • $\begingroup$ I agree up to the point where you say it is different from adding two distributions. I think it is the same, unless you have some reason to believe that larger actual values of a particular output of the process tend to result in larger (or smaller) errors of measurement of that output. Then you need to say what the dependency is. (It's still a sum of two variables but the joint distribution is more complicated then.) $\endgroup$
    – David K
    Feb 18, 2020 at 11:44
  • $\begingroup$ I just realized you're writing "measurement uncertainty" where I would have written "measurement error." Uncertainty of measurement is a single parameter, such as a standard deviation: physics.nist.gov/cuu/Uncertainty/glossary.html $\endgroup$
    – David K
    Feb 18, 2020 at 12:28
  • $\begingroup$ If you have a source for your definitions I think you should cite it in the question. $\endgroup$
    – David K
    Feb 18, 2020 at 12:28

1 Answer 1

1
$\begingroup$

Let's step back a little and consider the mathematics, not the application.

Suppose we have a random variable $X_1 \sim \mathcal{N}(\mu_1, \sigma_1)$ and another random variable $X_2 \sim \mathcal{N}(0, \sigma_2)$. That is, we have two general Gaussian distributions, except that the second distribution has mean zero.

The density function of $X_1$ is $$ f_{X_1}(x) = \frac{1}{\sigma_1\sqrt{2\pi}}e^{-((x-\mu_1)/\sigma_1)^2/2}. $$

The density function of $X_2$ is $$ f_{X_2}(x) = \frac{1}{\sigma_2\sqrt{2\pi}}e^{-(x/\sigma_2)^2/2}. $$

The density function of the sum $X_1 + X_2$ is the convolution of the two individual density functions, which can be computed in several equivalent ways, one of which is

\begin{align} f_{X_1+X_2}(x) &= \int_{-\infty}^\infty f_{X_2}(x - t)f_{X_1}(t)\,dt \\ &= \int_{-\infty}^\infty \frac{1}{\sigma_2\sqrt{2\pi}}e^{-((x - t)/\sigma_2)^2/2} f_{X_1}(t)\,dt. \end{align}

This equation should look familiar, because the right-hand side on the last line is the integral you developed for $f_{X_m}(x)$, provided that $X_1 = X_p$ and $\sigma_2 = \sigma_u.$

In other words, the $X_m$ you are trying to describe is simply $X_p + X_u,$ where $X_u \sim \mathcal{N}(0, \sigma_u)$.

The variable $X_u$ is what I would consider the error, which has mean zero. It is the difference between the measurement and the actual value.

$\endgroup$
1
  • $\begingroup$ Thanks for your answer. This exercise unintentionally helped me get a more intuitive understanding of the convolution of distributions. :) $\endgroup$
    – Omer
    Feb 19, 2020 at 21:12

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.