Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Assume the $1 \times N$ vector

$\mathbf X = [X_1, X_2, \ldots , X_N]$

contains i.i.d. normal samples such that $\mathbf X$ has a multivariate normal distribution. Now assume another random variable $Y$ also has a normal distribution, independent from the samples in $\mathbf X$, such that $Y$ has a mean and variance not necessarily equal to that of $\mathbf X$.

What is the probability that $Y$ will exceed all the values in $\mathbf X$ concurrently (or jointly)? In other words what is the probability that the value of $Y$ will be the maximum value of $[X_1, X_2, \ldots , X_N, Y]$

Possible approach?

In the simplest case $N=1$, the problem reduces to

$p(Y>X_1)$ = $p(Y-X_1>0)$

which can be calculated by convolution of $p(Y)$ and $p(-X_1)$ and integrating the resulting PDF from $0$ to $\infty$. When $Y$ is also i.i.d. w.r.t the samples of $\mathbf X$, the resulting PDF is zero-mean, such that $p(Y>X_1)$ = 0.5. This makes intuitively sense, since the probability that one RV will exceed another i.i.d. RV is 0.5. The probability

$p(Y>X_1, Y>X_2, \ldots, Y>X_N )$

should therefore equal $1/(N+1)$ if $Y$ is i.i.d. w.r.t. $\mathbf X$. When $N>1$, the problem does not seem straightforward (maybe it is?).

share|improve this question
    
Maybe I'm misunderstanding your question, but if everything's independent, isn't the probability of $Y> max(X_1,\dots X_N)$ just the product of the probabilities of the independent events $Y>X_j$, $j=1, \dots N$? –  Mike Jury May 6 '13 at 19:58
1  
@Mike Jury: That variables $X_1, X_2, Y$ are independent doesn't mean events such as $Y \gt X_1$ and $Y \gt X_2$ are independent. –  Douglas Zare May 6 '13 at 20:58
    
@Douglas Zare: Ah, of course. Long day. –  Mike Jury May 6 '13 at 22:32
    
en.wikipedia.org/wiki/… –  Steve Huntsman May 7 '13 at 0:43
add comment

3 Answers

up vote 4 down vote accepted

Assume without loss of generality that each $X_i$ is standard normal and that $Y$ is normal with mean $\mu$ and variance $\sigma^2$. By definition, for every $y$, $$ P[y\gt\max(X_1,\ldots,X_N)]=\Phi(y)^N, $$ hence, by the independence of $Y$ from $(X_i)_i$, $$ P[Y\gt\max(X_1,\ldots,X_N)]=E[\Phi(Y)^N]=\int_\mathbb R\Phi(\mu+\sigma x)^N\varphi(x)\mathrm dx, $$ where $$ \varphi(x)=\frac1{\sqrt{2\pi}}\mathrm e^{-x^2/2},\qquad\Phi(x)=\int_{-\infty}^x\varphi(z)\mathrm dz. $$ Except when $\mu=0$, $\sigma^2=1$, I see no simplification. An equivalent formula is $$ P[Y\gt\max(X_1,\ldots,X_N)]=\int_\mathbb R\Phi(x)^N\varphi\left(\frac{x-\mu}\sigma\right)\frac{\mathrm dx}\sigma. $$

share|improve this answer
    
Thanks! The key is to write the question in the form $p(Y > {\rm max}(\mathbf X))$ –  JD Vlok May 7 '13 at 11:00
add comment

The case $n=1$ can be done more explicitly than you say. If $X \sim \mathcal N(\mu_X,\sigma_X^2), Y \sim \mathcal N(\mu_Y,\sigma_Y^2)$ are independent then $Y-X \sim \mathcal N(\mu_Y-\mu_X, \sigma_X^2 + \sigma_Y^2)$ so $P(Y - X \gt 0) = 1 -\Phi(\frac{\mu_X - \mu_Y}{\sigma_X^2 + \sigma_Y^2}) = \Phi(\frac{\mu_Y-\mu_X}{\sigma_X^2+\sigma_Y^2}).$

If you are interested in a particular value of $n$, then I think you should use numerical integration which works quite well. You may also be able to get some asymptotics. A lot is known about the distribution of order statistics including the maximum of $n$ IID normal distributions. For example, the mean value for the maximum of $n$ standard normals is approximately $\Phi^{-1}(\frac{n-0.375}{n+0.25}).$ See also the accepted answer to "Expectation of the maximum of gaussian random variables" and the reference in the comments.

The maximum of $n$ IID normals is not normal. However, it may be ok to approximate it as a normal distribution, and use the case $n=1$, since it doesn't start too far away from a normal distribution (it is continuous and unimodal) and you are essentially convolving it with one Gaussian afterwards. This approximation should work better closer to the mean.

share|improve this answer
    
The vector $\mathbf X$ is typically large ($N >> 1$). The $N=1$ case is merely a sanity check to see whether the problem is correctly described mathematically. –  JD Vlok May 7 '13 at 13:16
    
If you use a normal approximation for the maximum of $n$ normal random variables (and this may be a better approximation than you would expect) then it reduces to the case $n=1$. So $n=1$ is not just a sanity check. –  Douglas Zare May 7 '13 at 13:18
add comment

The probability that $Y$ will exceed all the values in $\mathbf X$ can be written as (see answer of Didier Piau above)

$p_m = p(Y > {\rm max}(\mathbf X)) = p(Y - {\rm max}(\mathbf X) > 0)$

The density of ${\rm max}(\mathbf X)$ can be obtained using (Papoulis p. 193)

$f_{\rm max}(z) = N f_X(z) F_X^{N-1}(z)$

with $f_X$ the density of $X_i$ within $\mathbf X$ and $F_X$ the associated distribution function (CDF).

The density of the difference between the RVs $Y$ and ${\rm max}(\mathbf X)$ can be obtained using (Papoulis p. 185)

$p(Y - {\rm max}(\mathbf X)) = f_z(z) = \int_{-\infty}^{\infty} f_Y(z+w) f_{max}(w)~dw$

which is simply the convolution between $f_Y = p(Y)$ and $f_{max} = p({\rm max}(\mathbf X))$. The final answer is then the area under $f_z(z)$ beyond $0$ which can be written

$p_m = 1-F_z(0)$

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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