5
$\begingroup$

I need to compute the two-sided (bilateral) Laplace transform of a numerically given function $F$, $$ I(t) = \int_{-\infty}^{+\infty} {dx} \, e^{-x} \, F(x + t) ~, $$ where $F(x)$ has some sharp features, e.g., at $ x = \{ 0, x_p, \cdots \} $, has finite support, and vanishes sufficiently fast for $ x < 0 $ to prevent a divergence of the integral.

Due to the sharp features, I have to divide the integration region to some subregions, for instance, \begin{align} I(t) &= \int_{-\infty}^{-\Delta} {dx} \, e^{-x} \, F(x + t) + \int_{-\Delta}^{+\Delta} {dx} \, e^{-x} \, F(x + t) \\ &\quad + \int_{+\Delta}^{x_p - \Delta} {dx} \, e^{-x} \, F(x + t) + \int_{x_p - \Delta}^{x_p + \Delta} {dx} \, e^{-x} \, F(x + t) \\ &\quad + \int_{x_p + \Delta}^{+\infty} {dx} \, e^{-x} \, F(x + t) ~, \end{align} where $\Delta > 0$.

I'd like to know the best and most precise way to perform such an integral numerically (quadrature methods). Is there something like Gauss-Laguerre quadrature for such an integration?

I have also considered a transformation of the integration variable, $$ u = e^{-x} ~, $$ to absorb the exponential factor, $$ I(t) = \int_{0^+ = \, u(x \rightarrow +\infty)}^{+\infty = \, u(x \rightarrow -\infty)} {du} \, F(-\ln(u) + t) ~, $$ and performing the transformed integration by a Tanh-Sinh quadrature; yet I am not sure if this is the best method.

As an example, one can take $$ f(x,\omega) = e^{-\alpha (x - \omega)} \, \Theta(x - \omega) $$ where $\Theta$ denotes the Heaviside step function and $\alpha > 0$. Hence, the ‘width’ of the peak at $\omega$ can be varied by $\alpha$. The two-sided Laplace transform of $f$ will be $$ \mathcal{L}[f](\omega) := \int_{-\infty}^{+\infty} dx \, e^{-x} \, f(x, \omega) = \frac{ e^{-\omega} }{1 + \alpha} ~. $$

$\endgroup$
  • 1
    $\begingroup$ The need to integrate high oscillatory kernels arises frequently in the evaluation of Sommerfeld integrals for dyadic Green's functions. The usual practice is to do a (numerical) rational fit of $F(x)$ as a collection of poles $\sum_i 1/(x+c_i)$ and to use a closed-form expression to convert the poles into the inverse transform. $\endgroup$ – Richard Zhang Mar 26 '16 at 5:25
  • $\begingroup$ @RichardZhang : The integrand is not oscillatory here, as you can see, for instance, in the example provided. $\endgroup$ – AlQuemist Mar 26 '16 at 17:07
  • $\begingroup$ My apologies, the intention of my comment wasn't to suggest that your kernel is oscillatory, only that difficult intergands can often be treated using a semi-analytical approach of rational fitting + closed-form transformation formulas. $\endgroup$ – Richard Zhang Mar 27 '16 at 14:06
  • $\begingroup$ @RichardZhang: The integrands are only available numerically (obtained from another numerical procedure). No analytical expression exists for them. $\endgroup$ – AlQuemist Mar 27 '16 at 15:19
  • $\begingroup$ Do you agree that a numerically evaluated $F(x)$ can be approximated to arbitrary accuracy (under regularity assumptions) as a rational function, $\tilde{F}(x)=\sum_i 1/(x+c_i)$? And do you agree that the bilateral Laplace transform has a closed form expression for $\tilde{F}(x)$? $\endgroup$ – Richard Zhang Mar 27 '16 at 15:25
4
$\begingroup$

With the substitution $y=x+t$, the $t$-dependency can be factored out:

$$I(t)=\int_{-\infty}^\infty e^{-(y-t)}F(y)\;dy = e^t \int_{-\infty}^\infty e^{-y}F(y)\;dy = C\cdot e^t$$

This still leaves you with the problem of computing $C$, but it is merely a constant number and the $t$-dependency is entirely described by the factor $e^t$. The two-sided Laplace transfrom of this function and its region of convergence (if it is non-empty) should be found in a table of Laplace transforms.

Edit: On a second thought, I doubt that the two-sided Laplace transform of $I(t)$ exists even for purely imaginary $s$, because $I(t)\sim e^t$ is not a tempered distribution.

$\endgroup$

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.