I need to numerically compute this rather involved integral

$I(z) = \int_{0}^{\infty} \phi\Big(\frac{K(z) - \mu(t)}{\sigma(t)} \Big) \big( \frac{1}{\sigma(t)}\big)\exp(-\lambda t)dt$

where $\phi$ is the density of a standard normal random variable,

$\mu(t) = \frac{1}{a}\big[(r_0 - b)(1 - \exp(-at)) + abt \big]$,

$\sigma^2(t) = \frac{c^2}{2a^3}\big[ 2at+4\exp(-at)-\exp(-2at)-3\big]$

$a,b,c,r_0,\lambda$ are real parameters.

Basic quadrature routines (Simpson and the one implemented in the integral Matlab function) raise numerical issues: $I(z)$ is part of a density and when it is integrated over $[-\infty,\infty]$ the error is around 2% (I need it to be much smaller) so it boils down to the accuracy of the estimate of this integral. I wonder if there is a closed form solution or a simpler form more suitable for the implementation (some series expansion?). I tried the substitution ($y = \exp(-\lambda t)$) but the accuracy is still bad. Any advice would be highly appreciated.

  • $\begingroup$ You would attract the interest of more people if you explained the nature of your numerical difficulties. $\endgroup$ – Carl Christian Dec 6 '17 at 11:56

Use exp-sinh quadrature. If you're using C++, and exp-sinh quadrature routine is provided by boost, and can evaluate your integral to arbitrary precision.


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.