I have the following problem:

Let $X$ and $Y$ be two independent gaussian distributions, with same mean $\mu$ and standard deviations $b$ and $c$ ($X \sim \mathcal{N}(\mu,b)$ and $ Y \sim \mathcal{N}(\mu,c)$).

Let $U$ and $V$ be two "gaussian" distributions with resp. means $X$ and $Y$ and same standard deviation $\sigma$ ($U \sim \mathcal{N}(X,\sigma)$ and $V \sim \mathcal{N}(Y,\sigma)$).

I want to compute the probability of overlap between $U$ and $V$, which I define as :

$P(U>X+\frac{|X−Y|}{2})+P(V<Y− \frac{|X−Y|}{2})$

More precisely, I would like to express this quantity as a function of $\mu$, $b$, $c$ and $\sigma$, or at least prove that the probability of overlap grows with $b+c$ and is inversely proportional to $\sigma$).

Maybe this could be useful (but I didn't figure out how) : $X - Y$ is gaussian with mean $0$ and s.d. $b+c$ (as a sum of 2 independent gaussians), therefore $E(|X−Y|)= \sqrt{2(b+c)/\pi}$ (general result for a centered gaussian).

The context : in my experiment, $X$ and $Y$ are sampled once, then $U$ and $V$ are sampled hundreds of times with the obtained means, and I want to know the probability to sample for $U$ a value that is "more probable" for $V$ and vice versa, ie to be in the yellow or red region :

enter image description here

(this drawing might be biased, since according to an answer here, the resulting distribution of $U$ and $V$ might in fact be gaussians with same mean $\mu$ and different standard deviations... But I am quite confused because I do not know how to define the probability of overlap in my experiment anymore...)


If the event of interest is that $U$ and $V$ are "more probable" for the alternative Normal, conditional on $X=x$ and $Y=y$, it could be written as $$\varphi((U-y)/\sigma)>\varphi((U-x)/\sigma)\quad\text{and}\quad\varphi((V-y)/\sigma)<\varphi((V-x)/\sigma)$$ where $\varphi(\cdot)$ is the standard Normal density. Expanding the density leads to $$(U-y)^2<(U-x)^2\quad\text{and}\quad(V-x)^2<(V-y)^2$$ and, if $y>x$, $$2U>x+y\quad\text{and}\quad 2V<x+y\,,$$ while, if $y<x$, $$2U<x+y\quad\text{and}\quad 2V>x+y\,,$$ Conditionally on $X=x$ and $Y=y$, since $U\sim\mathcal N(x,\sigma^2)$ and $V\sim\mathcal N(y,\sigma^2)$, this event has probability $$\left.\begin{aligned}\Phi((x-y)/2\sigma)^2 &\text{ if } ~y>x\\\\\Phi((y-x)/2\sigma)^2 &\text{ if } ~y<x\end{aligned}\right\} = \Phi(-|x-y|/2\sigma)^2$$ The final probability thus writes as $$\mathbb E[\Phi(-|x-y|/2\sigma)^2]=\int \Phi(-|z|/2\sigma)^2\,e^{-z^2/2(b^2+c^2)}\,\text dz/\sqrt{2\pi(b^2+c^2)}$$ In the event $4\sigma^2=b^2+c^2$, this integral enjoys a closed-form expression since $$e^{-z^2/2(b^2+c^2)}\propto\frac{\text d}{\text dz}\Phi(-z/2\sigma)$$


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.