I'm trying to find the closed-form CDF and PDF of $Y = \max(X_1, ..., X_n)$ where $X_i \sim \mathcal{N}(\mu_i, \sigma^2)$.

My thought process so far: $$ \begin{align*} F_Y(y) &= \mathbb{P}(\max(X_1, ..., X_n) \leq y) \\ &= \prod_{i=1}^n F_{X_i}(y) \\ \implies f_Y(y) &= \frac{\partial}{\partial y} \prod_{i=1}^n F_{X_i}(y) \end{align*} $$ I'm not sure how to proceed from here, or whether this is even the correct approach. Any help is much appreciated.

Edit: The linked question pertains to IID random variables, whereas I am more interested in INID random variables.


Problems like this, where you want to differentiate the product of a bunch of functions that depend on your variable of interest, can be dealt with by logarithmic differentiation. Let $\Phi$ and $\phi$ denote the CDF and PDF of the standard normal distribution (respectively). Since the normal random variables in your question have the same variance you get:

$$\prod_{i=1}^n F_i(y) = \prod_{i=1}^n \Phi \Big( \frac{y-\mu_i}{\sigma} \Big) = \exp \Bigg( \sum_{i=1}^n \ln \Phi \Big( \frac{y-\mu_i}{\sigma} \Big) \Bigg).$$

Differentiating with respect to $y$ and applying the chain rule gives:

$$\begin{equation} \begin{aligned} f_Y(y) = \frac{d F_Y}{dy}(y) &= \Bigg( \frac{1}{\sigma} \sum_{i=1}^n \frac{\phi ( (y-\mu_i)/\sigma ) }{\Phi ( (y-\mu_i)/\sigma )} \Bigg) \exp \Bigg( \sum_{i=1}^n \ln \Phi \Big( \frac{y-\mu_i}{\sigma} \Big) \Bigg) \\[6pt] &= \Bigg( \frac{1}{\sigma} \sum_{i=1}^n \frac{\phi ( (y-\mu_i)/\sigma ) }{\Phi ( (y-\mu_i)/\sigma )} \Bigg) \Bigg( \prod_{i=1}^n \Phi \Big( \frac{y-\mu_i}{\sigma} \Big) \Bigg). \\[6pt] \end{aligned} \end{equation}$$

In the special case where $\mu = \mu_1 = \cdots = \mu_n$ this reduces to the well-known formula:

$$f_Y(y) = \frac{n}{\sigma} \cdot \phi \Big( \frac{y-\mu}{\sigma} \Big) \cdot \Phi \Big( \frac{y-\mu}{\sigma} \Big)^{n-1}.$$

  • $\begingroup$ Thanks! This is very helpful. Do you happen to know if this can be approximated by some well-known distribution in the case where n is large? I need to store this resulting distribution of the maximum in some compact way, e.g. with two parameters. $\endgroup$ – cosine180 Dec 10 '18 at 9:55
  • 1
    $\begingroup$ Have a look at the extreme value distributions - these are asymptotic distributions for maximums when $n$ is large. $\endgroup$ – Ben - Reinstate Monica Dec 10 '18 at 22:18
  • 1
    $\begingroup$ As far as I can tell, these only apply to the IID case though, don't they? $\endgroup$ – cosine180 Dec 11 '18 at 5:19

We can express each $F_{X_i}$ in terms of an error function, and when differentiating via the product rule to get $f_Y$ we can re-obtain from each of those error functions in turn a Gaussian function. It depends what you mean by closed form -- if you leave the $\mu_i$ as symbolic parameters then you will have explicit sums and products over $i$. It will be in suitable form to compute numerically for given parameters if you have access to elementary and error functions.


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.