11
$\begingroup$

Let $X_1 \sim \mathcal{N}(\mu_1,\sigma_1^2)$ and $X_2 \sim \mathcal{N}(\mu_2,\sigma_2^2)$, what's the CDF of $X = \max\{X_1,X_2\}$? Both variables are assumed to be independent.

I tried the following:

\begin{equation} F_X(x) = \text{Prob}\{X<x\} = \text{Prob}\{X_1<x, X_2<x\} = \text{Prob}\{X_1<x\}\cdot \text{Prob}\{X_2<x\} = F_{X_1}(x)\cdot F_{X_2}(x) \end{equation}

Since $X_1$ and $X_2$ are normal distributed variables, their CDFs are well known:

\begin{equation} \tag{1} F_X(x) = F_{X_1}(x)\cdot F_{X_2}(x) = \bigg(1-\mathcal{Q}\Big(\frac{x-\mu_1}{\sigma_1}\Big)\bigg)\cdot \bigg(1-\mathcal{Q}\Big(\frac{x-\mu_2}{\sigma_2}\Big)\bigg) \end{equation}

where $\mathcal{Q}\big(\frac{x-\mu}{\sigma}\big) = \int_x^{\infty} \frac{1}{\sqrt{2\pi}\sigma} e^{-\frac{(x-\mu)^2}{2\sigma^2}} dx$.

However, MATLAB simulation shows that this is not correct. The MATLAB code and the plot result comparing simulation and theoretical is shown below.

S=1e6;

% Define Normal Distributed RV #1
mu_1 = 1;
sigma_1 = 1;
X_1 = sigma_1/sqrt(2)*(randn(1,S))+mu_1;

% Define Normal Distributed RV #1
mu_2 = 1;
sigma_2 = 1;
X_2 = sigma_2/sqrt(2)*(randn(1,S))+mu_2;

% Define variable X = max{X_1,X_2}
X = max(X_1,X_2);
[X_pdf, X_var] = ecdf(X); % Obtain empirical CDF (similar to function hist)

X_pdf_th = (1-qfunc((X_var-mu_1)/sigma_1)).*(1-qfunc((X_var-mu_2)/sigma_2));

plot(X_var,X_pdf); hold on;grid on;
plot(X_var,X_pdf_th,'r');  legend('Simulation CDF', 'Theory CDF'); hold off;

enter image description here

$\endgroup$
  • $\begingroup$ Why do you divide by $\sqrt{2}$ in 'X_1 = sigma_1/sqrt(2)*(randn(1,S))+mu_1' ? $\endgroup$ – d.k.o. Aug 8 '15 at 1:15
  • $\begingroup$ Remove that $\sqrt{2}$. You can also replace '1-qfunc()' in the definition of 'X_..._th' by 'normcdf(X_var, mu, sigma)'... $\endgroup$ – d.k.o. Aug 8 '15 at 1:23
  • $\begingroup$ @d.k.o. Ohh, my post deserves a negative evaluation. Thanks a lot. $\endgroup$ – lucasrodesg Aug 8 '15 at 1:35
  • $\begingroup$ @luikn can you post the answer, for future reference? Does the distribution of $X$ have a name? Can you also plot the density of $X$ ? $\endgroup$ – Basj Jan 31 '16 at 13:38
  • 1
    $\begingroup$ The pdf of the maximum is given on page 1 of the paper by Nadarajah and Kotz (2008) available here: gwern.net/docs/conscientiousness/2008-nadarajah.pdf Finding a closed form for the CDF might be a bit more tricky. In your case, just set $\rho=0$. $\endgroup$ – wolfies Jan 31 '16 at 15:16
3
$\begingroup$

This is not really an answer, but notes for future reference about the distribution of $X = \max\{X_1,X_2\}$ when $X_1, X_2 \sim \mathcal{N}(0,1)$, independent, identically distributed.

As you stated, the cumulative distribution function (CDF) is

$$F_X(x) = P(X\le x) = P(X_1\le x\ \cap X_2 \le x) = P(X_1 \le x) \cdot P(X_2 \le x) = \Phi(x)^2, $$

with $\Phi_{\mu,\sigma}(x) := \int_{-\infty}^x \frac{1}{\sqrt{2\pi}\sigma} e^{-\frac{(t-\mu)^2}{2\sigma^2}} dt$, and $\Phi := \Phi_{0,1}$.

The density of $X$ is then $$f_X(x) = 2 \Phi(x) \Phi'(x) = \sqrt{\frac{2}{\pi}} \Phi(x)\ e^{-x^2/2}.$$

The expectation of $X$ is $\int_{\mathbb{R}} x f_X(x) d x = \frac{1}{\sqrt{\pi}}$ and its variance is $\int_{\mathbb{R}} x^2 f_X(x) d x - \Big(\frac{1}{\sqrt{\pi}}\Big)^2 = 1 - \frac{1}{\pi}.$

We can note that $f_X$ is very close to the density of a normal distribution $\mathcal{N}(1/\sqrt{\pi}, \sqrt{1-1/\pi})$:

enter image description here

Note: Does this distribution has a name? Any reference welcome.

$\endgroup$
  • 1
    $\begingroup$ Thanks for your answer. Hope it helps future visitors. $\endgroup$ – lucasrodesg Apr 18 '16 at 21:50

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.