In the existing literature [From the Schrödinger problem to the Monge–Kantorovich problem], below the Example 3.4, the author claimed that using Laplace method, we can obtain $$ \int_0^{\infty}e^{-z^p/p}\exp(\zeta z) dz\sim (2\pi(q-1))^{1/2}\zeta^{1-q/2}e^{\zeta^q/q} $$ as $\zeta\to\infty$, where $q$ is a parameter satisfying $1/p+1/q=1$. However, I don't know the detailed deviation step. Also, is it possible to calculate the asymptotic value for the integral $$ \int_0^{\infty}e^{-z^p/p}\exp(\zeta (z + cz^2)) dz, $$ where $c>0$ is some constant?
1 Answer
$\begingroup$
$\endgroup$
3
The strategy is as follows:
- put everything in the integrand in the exponent $e^{f(z)}$, with $f(z)=\zeta z-z^p/p$;
- calculate the saddle point, the (possibly complex) number $z_0$ where $f'(z)=0$; this gives $z_0=\zeta^{1/(p-1)}$;
- expand $f(z)$ to second order around $z_0$, discarding higher order terms; this gives the exponent $g(z)=f(z_0)-a(z-z_0)^2$ with $a=\tfrac{1}{2}(p-1)\zeta^{(p-2)/(p-1)}$;
- carry out the remaining Gaussian integral of $e^{g(z)}$ over $z$, along a "path of steepest descent", which gives the asymptotic expansion as $\zeta\rightarrow\infty$ of the integral $\int e^{f(z)}dz\rightarrow e^{f(z_0)}(\pi/a)^{1/2}$.
The same approach also works for the second integral, but then I am not able to give a closed-form expression for the saddle point at arbitrary $p$; for $p=2$ I arrive at $\exp[\zeta^2(2-4c\zeta)^{-1}]\bigl(2\pi/(1-2c\zeta)\bigr)^{1/2}$.
-
$\begingroup$ Isn't there a problem with convergence for $p\le 2$ given that $c>0$ and $\zeta >0$ ? $\endgroup$ Commented Jul 19, 2022 at 21:12
-
$\begingroup$ Ah, c>0, I had missed that. This p=2 answer is for c<0, otherwise one needs p>2, as you say. $\endgroup$ Commented Jul 19, 2022 at 21:45
-
$\begingroup$ Your approach works fine for $p>2$ as well. There is only one significant positive maximum, $z_{0}>0$ of the exponent, which is (asymptotically, but that is sufficient) $z_{0}\sim (2 c^{p-1}\ \zeta)^{\frac{1}{p-2}}$. Calculating the Gaussian integral around this point gives at least the leading term of the expansion. $\endgroup$ Commented Jul 20, 2022 at 9:24