I have a list of random variables, $X_1$, $X_2$, ..., $X_N$, associated with binary random variables $A_i$ such that $P(A_i) = \pi$ is known. I also know that, for all $i$ $$X_i|A_i\sim f(x)\\ X_i|\bar A_i\sim g(x)$$ where $f$ and $g$ are known, and thus the distribution of the $X_i$ is a mixture given by $$p(x) = \pi\cdot f(x) + (1-\pi)\cdot g(x)$$ Is there a general expression for the distribution of $Y=\sum_{i=1}^N X_i$?


1 Answer 1


If you denote by $$g^m\ast f^k$$ the convolution of $m$ $g$'s and $k$ $f$'s, meaning the density of the distribution of the sum of $m$ realisations of $X\sim g(x)$ and $k$ realisations of $X\sim f(x)$, with for instance $$g^0\ast f^1=f\,,\quad g^1\ast f^0=g\,,$$ \begin{align*}(g^m\ast f^k)(x)&=\int (g^{m-1}\ast f^k)(x-y)g(y)\,\text{d}y\\&=\int (g^{m}\ast f^{k-1})(x-y)f(y)\,\text{d}y\,,\end{align*}the density of the sum of $n$ variables from the mixture with density$$\pi g(x) + (1-\pi) f(x)$$is

$$\sum_{m=0}^n {n \choose m}\pi^m(1-\pi)^{n-m} (g^m\ast f^{n-m})(x)$$

by a straightforward binomial argument.

For instance, if $g$ is the density of a N$(0,1)$ distribution and $f$ the density of a N$(\mu,\sigma^2)$ distribution, we have that $(g^m\ast f^{n-m})$ is the density of a $$\text{N}\{(n-m)\mu,(n-m)\sigma^2+m\}$$ distribution, hence the above sum has for distribution $$\sum_{m=0}^n {n \choose m}\pi^m(1-\pi)^{n-m} \text{N}\{(n-m)\mu,(n-m)\sigma^2+m\}$$

Note: the part about the convolution is detailed in a connected answer to a Stack Exchange question about the sum of a mixture variate with a normal variate.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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