# Exact versus numerical approximation of an integral

Why does the following approximation work very well for small theta's but not for large theta's?

$$I = \int_0^\infty x (1 + \theta x)^{-2 \left(\frac{1}{\theta} + 1 \right)} \, \mathrm{d}x$$

Exact evaluation using integration by part

$$I = \frac{1}{2(\theta + 2)}$$

Approximation using Gauss-Laguerre quadrature $$I \approx \sum_{k=1}^N w_k \exp(x_k) \, x_k (1 + \theta x_k)^{-2 \left(\frac{1}{\theta} + 1 \right)}$$ where the $x_k$'s and the $w_k$'s can be found, e.g., here.

Comparison (N = 100)

• theta = 0.5806:

true = 0.1937534 and approx = 0.1937522

• theta = 15:

true = 0.02941176 and approx = 0.01662866

my R-code

#------ package ------
#---------------------

#------ f(x) ------
f <- function(x, theta)
{
x * (1 + theta * x)^(-2 * ((1 / theta) + 1))
}
#------------------

#------ xk and wk ------
x <- laguerre[, 1]
w <- laguerre[, 2]
#-----------------------

#------ approx ------
approx <- function(theta, x, w)
{
sum(w * exp(x) * f(x=x, theta=theta))
}
#--------------------

> approx(theta=0.5806, x=x, w=w)
[1] 0.1937522
> approx(theta=15, x=x, w=w)
[1] 0.01662866

-
Sorry, your integral is not showing up in the post. – Ron Gordon Apr 22 '13 at 10:35
@RonGordon: Should be fixed now. Thanks! – Marco Apr 22 '13 at 10:38
Sorry again, I still cannot see it. – Ron Gordon Apr 22 '13 at 10:40
@RonGordon: Sorry for the trouble! Hope it is ok now – Marco Apr 22 '13 at 11:14
You should recall that Gauss-Laguerre quadrature works best if the "infinite extent" portion of your integral behaves very much like $\exp(-x)$. This for instance is the reason why Gauss-Laguerre is not very useful for computing integrals like $\int_0^\infty \frac{\mathrm dx}{1+x^2}$. – J. M. Apr 22 '13 at 12:36