A random vector $\mathbf{X}$ has a $p$-variate normal mixture distribution when with probability $q$ it is drawn from $f_1(\mathbf{X}) \sim N_m(\mathbf{\mu}_1, \mathbf{\Sigma}_1)$ and with probability $1-q$ from $f_2(\mathbf{X}) \sim N_m(\mathbf{\mu}_2, \mathbf{\Sigma}_2)$. It has probability density function $f(\mathbf{X})=qf_1(\mathbf{X})+(1-q)f_2(\mathbf{X})$

(a) How many parameters are in this distribution?

(b) Find $E(\mathbf{X})$ and $\text{Covar}(\mathbf{X})$


$\mu_1,\mu_2$ trivially both have $m$ components. For $\mathbf{\Sigma_1}, \mathbf{\Sigma_2}$ we need only count the upper triangle since they are each symmetric about the diagonal. The diagonal has $m$ components, the adjacent upper diagonal has $m-1$ components, and if you continue this counting process we have that each covariance matrix has $m+m-1+m-2 +\dots + 1=\frac{m^2-m}{2}+m$. So all together we have:


parameters in this distribution.


I feel fairly confident about the number of parameters piece of this problem but am unsure whether or not I have communicated the expectation operations correctly. How do you interpret $E(f_1(\mathbf{X}))$ if not the way I already have it? I could also use help understanding how to calculate the covariance. Is it correct to say that $\text{Covar}(f_1(\mathbf{X}))=\mathbf{\Sigma}_1$?


Writing things like "$E[f_1(X)]$" is a bit problematic; you aren't ever really plugging in the random variable into the density function. But your computation leads to the right answer, even if the notation is not standard.

Let $Z$ be the random variable satisfying $P(Z=1)=q$ and $P(Z=2)=1-q$. Then the conditional distribution of $X$ given $Z=1$ is $N_m(\mu_1, \Sigma_1)$, and the conditional distribution of $X$ given $Z=2$ is $N_m(\mu_2, \Sigma_2)$.

Then, $E[X] = E[E[X \mid Z]] = E[X \mid Z=1] P(Z=1) + E[X \mid Z=2] P(Z=2) = q\mu_1 + (1-q) \mu_2$ which is what you got.

The law of total covariance applies to each entry of the matrix $\text{Cov}(X)$, and ultimately becomes $$\text{Cov}(X) = E[\text{Cov}(X \mid Z)] + \text{Cov}(E[X \mid Z]).$$ The first term is $q\Sigma_1 + (1-q) \Sigma_2$ by a computation similar to the expectation computation above. The other term is $$\text{Cov}(E[X \mid Z]) = E[E[X \mid Z]E[X \mid Z]^\top] - E[E[X \mid Z]]E[E[X \mid Z]]^\top = q \mu_1 \mu_1^\top + (1-q) \mu_2 \mu_2^\top - (q\mu_1 + (1-q) \mu_2)(q\mu_1 + (1-q) \mu_2)^\top$$ where I am just using the formula $\text{Cov}(Y) = E[YY^\top] - E[Y]E[Y]^\top$ with $Y=E[X \mid Z]$.

  • $\begingroup$ Thank you for the detailed response. Have a much better understanding of the problem now! $\endgroup$ Mar 23 '21 at 11:38

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.