3
$\begingroup$

Does anyone know if I have a chance to get a closed form for the following integral (and if yes, what would be the trick)?

$$\int \left\{1-\Phi(x)\right\}^k \Phi(x)^{n-k} \varphi\left(\frac{x-\mu}{\sigma}\right) dx,$$

where $\Phi$ and $\varphi$ are the normal cumulative distribution and density functions, $k \leq n$ are integers, and $\mu$ and $\sigma >0$ are real numbers.

I had a look at this page https://en.wikipedia.org/wiki/List_of_integrals_of_Gaussian_functions, but I'm not sure it contains what I need.

$\endgroup$
1
  • 1
    $\begingroup$ $k=0$ or $n-k=0$ can yield analytical solutions: see stats.stackexchange.com/questions/61080. (See Dilip Sarwate's solution for a suggestive interpretation.) Nothing else does in general. You can of course expand $(1-\Phi)^k$ to reduce the integral to an alternating sum (for $k\gt 0$, $\mu=0$, $\sigma=1$). Would that count as a "closed form"? $\endgroup$ – whuber Oct 8 '17 at 14:21
2
$\begingroup$

One better reference for Gaussian integrals is "A table of normal integrals" by D. B. Owen, in Commun. Statist. Simulat. Computa. B9(4), 389-419, 1980. That reference includes a few very special cases of your integral, but nothing close to its generality. One very special case which yields simply is when $\mu=0, \sigma=1$ when the substitution $u=\Phi(x)$ reduces the integral to $$ \int_0^1 u^{n-k} (1-u)^k \; du = \frac{k(n-k)B(k,n-k)}{n(1+n)} $$ where $B$ is the beta function. That could at least serve as a check on numerical solutions.

$\endgroup$
2
  • 1
    $\begingroup$ +1 Good point about the Beta. You don't need to limit yourself to the definite integral: if one accepts the Beta CDF as a "closed form" (which seems reasonable to me), then the indefinite integral is also available in closed form. Conceivably, for small $|\mu/\sigma|$, this could serve as the beginning of a good approximation, too. $\endgroup$ – whuber Oct 8 '17 at 18:40
  • 1
    $\begingroup$ Many thanks for this. I will use it as a reference. $\endgroup$ – user79097 Oct 9 '17 at 10:09
2
$\begingroup$

Let $X_1, X_2, \ldots, X_n, Y$ be independent normal random variables with $X_i \sim N(0,1)$ for $i = 1,2,\ldots, n$, and $Y \sim N(\mu, \sigma^2)$. Then, consider the probability that $k$ specified $X_i$ are larger than $Y$ and the remaining $n-k$ of the $X_i$ are smaller than $Y$. As a specific example, suppose that we wish to find the probability of the event $$\mathcal A =\big(X_1 > Y, X_2 > Y, \ldots, X_k > Y, X_{k+1} < Y, X_{k+2} < Y, \ldots, X_ n < Y\big).$$ Following the argument used in this answer of mine, we have that the conditional probability of $\mathcal A$ given that $Y = x$ is \begin{align}P(\mathcal A \mid Y = y) &= P(X_1>x, \ldots, X_k > x, X_{k+1} < y, \ldots, X_n < y \mid Y =x)\\&= P(X_1>x, \ldots, X_k > x, X_{k+1} < y, \ldots, X_n < y)\\ &= \prod_{i=1}^k P(X_i > y)\prod_{i=k+1}^n P(X_i < y)\\ &=[1-\Phi(x)]^{k}[\Phi(x)]^{n-k} \end{align} and so $$P(\mathcal A) = \int_{-\infty}^\infty [1-\Phi(x)]^{k}[\Phi(x)]^{n-k} \phi\left(\frac{x-\mu}{\sigma}\right)\,\mathrm dx.$$ There is no closed-form expression for this probability and there are no simple approximations that can be used except in special cases such as $k=0$ (cf. @whuber's comment).

$\endgroup$
0

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.