I am given a population $P$ that is equally divided into subsets $A$ and $B$. I know that a property $H$ of the population $P$ is normally distributed with mean $\mu_1$ and variance $\sigma_1^2$ for subset $A$ and $\mu_2$ and variance $\sigma_2^2$ for subset $B$. The task at hand is to find the exact median value of $H$ for the entire population $P$.

This is a Gaussian mixture with equal weights. I know that for a Gaussian distribution, the mean and the median are equal. I am trying to set up an equation based on the given data that is going to allow me to determine the median. I believe that the density of the Gaussian mixture is:

$$ f(h) = 0.5f_A(h) + 0.5f_B(h) $$

where $f_A$ is the Gaussian for subset $A$ and $f_B$ the Gaussian of subset $B$.Now the standard way to handle this problem for distributions of continuous random variables is to solve for $m$ the equation:

$$ \int_m^{\infty} f(h)dh = \frac{1}{2} $$

This does not look like a good strategy for this problem (for computational reasons -- although I may be totally off). If anyone could suggest a nicer, more elegant approach, I would deeply appreciate it!

  • $\begingroup$ Is this a homework problem? If so, can you read and add the self-study tag? $\endgroup$ – Andrew M Oct 3 '16 at 23:35
  • $\begingroup$ @DilipSarwate I am stating in the beginning that the subpopulations are of the same size. $\endgroup$ – Orest Xherija Oct 3 '16 at 23:45
  • $\begingroup$ @AndrewM No, I encountered a version of the problem in machine learning context and was trying to see you approach it rigorously. Should I still add the tag, since it is, literally, self-study? $\endgroup$ – Orest Xherija Oct 3 '16 at 23:47
  • 1
    $\begingroup$ Well, you know that $F_A[x]+F_B[x]=1$, where $x$ is the mixture median and $F_A,F_B$ are the cumulative distribution functions (CDFs) of the mixture components. For a numerical solution, this is not too bad of a strategy, given that you know $F_k(x)=\Phi[(x-\mu_k)/\sigma_k]$ for $k=A,B$, where $\Phi$ is the CDF of the standard normal distribution. $\endgroup$ – GeoMatt22 Oct 4 '16 at 1:45
  • $\begingroup$ @GeoMatt22 Could you please elaborate a bit on how we get $F_A[x] + F_B[x] = 1$? The rest of your argument is perfectly clear! $\endgroup$ – Orest Xherija Oct 4 '16 at 2:12

Let $m$ denote the median of the mixture distribution whose CDF is $\frac 12F_A(x) + \frac 12 F_B(x)$. Then, \begin{align} \frac 12F_A(m) + \frac 12 F_B(m) &= \frac 12 \tag{1}\\ &\Downarrow\\ F_A(m) + F_B(m) &= 1\\ &\Downarrow\\ \Phi\left(\frac{m-\mu_A}{\sigma_A}\right) + \Phi\left(\frac{m-\mu_B}{\sigma_B}\right) &= 1\\ &\Downarrow\\ \frac{m-\mu_A}{\sigma_A} + \frac{m-\mu_B}{\sigma_B} &= 0 \tag{2} \end{align} where the last implication follows from the fact that $\Phi(x)+\Phi(-x)=1$. In short, the "computational reasons" that are deterring the OP are not really worrisome at all: solving $(2)$ for $m$ is trivial. Of course, if the mixture weights are $p$ and $1-p$ where $p \neq \frac 12$, that is, the mixture distribution is $$p\cdot F_A(x) + (1-p)\cdot F_B(x), \quad p \neq \frac 12,$$ then we need to solve $$p\cdot \Phi\left(\frac{m-\mu_A}{\sigma_A}\right) + (1-p)\cdot\Phi\left(\frac{m-\mu_B}{\sigma_B}\right) = \frac 12. \tag{3}$$ This will likely need numerical solution for $m$: at least, no straightforward analytical solution to (3) springs to my mind.


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.