0
$\begingroup$

This question is the copy from mat.stackexchange.com here. I requestioned here due to the very limited responses there.

Let $\phi:\mathbb{R}\mapsto\mathbb{R}$ be the standard normal density, $$\phi(x)=\frac1{\sqrt{2\pi}}e^{-\frac{x^2}{2}}, \forall x\in\mathbb{R}.$$ Given $0<\sigma\le 1$. I am working with the differential equation $$\alpha''(u) [\phi(u) + \frac 1\sigma \phi(u/\sigma)] - \alpha'(u) [u\phi(u) + \frac 1{\sigma^3} u\phi(u/\sigma)] - \alpha(u) [\phi(u) + \frac 1{\sigma^3} \phi(u/\sigma)] = \frac 1{\sigma^5} u\phi(u/\sigma).$$

I am looking for a solution of the above equation such that its $n-$th derivative is bounded for some positive integer $n$.

What I know so far is:

  1. If the boundedness condition of $n-$th derivative is ignored, then of course the ODE above has general solution. The question is, how to choose a solution satisfying the boundedness condition.

  2. For $\sigma=1$, I could take $\alpha(u)=-\frac14u$ that satisfies all conditions. I got this function because in this case the general solution for the above ODE is $\alpha(u)= -\frac 14 u + (a + b \Phi(u))/\phi(u)$. So we can choose $a=b=0$. However, for $\sigma<1$, it makes my problem much harder. Any suggestion?

$\endgroup$
2
$\begingroup$

The Liouville transformation works as follows. Take the differential operator \begin{equation} L = \frac{d^2}{d x^2} + a_1(x) \frac{d}{d x} + a_2(x), \end{equation} where $a_1 \in C_1$ and $a_2 \in C_0$. Then, defining \begin{equation} A(x) = \text{exp}\left(-\frac{1}{2} \int^x a_1(\xi)d\xi\right) \end{equation} and \begin{equation} b(x) = -\frac{1}{2} \frac{d a_1}{d x} - \frac{1}{4} a_1^2 + a_2, \end{equation} we can transform the equation $L\,y(x) = f(x)$ into \begin{equation} A(x) \left[\frac{d^2}{d x^2} + b(x)\right] z(x) = f(x), \end{equation} where $y(x) = A(x) z(x)$.

Your equation is of the form $L y = f$ when you divide by the function multiplying $\alpha''$. You can then calculate \begin{equation} A(u) = (2\pi)^{-\frac{1}{4}}\left(\sigma\phi(u) + \phi(u/\sigma)\right)^{-\frac{1}{2}}. \end{equation} Calculating the function $b(u)$ is more involved, but should pose no problem (it's a little long to post here). Dividing by $A(x)$ yields \begin{equation} \left[\frac{d^2}{d x^2} + b(x)\right] z(x) = \frac{f(x)}{A(x)} \quad(*). \end{equation}

So far, so good: $b(x)$ and $f(x)/A(x)$ are both smooth functions. However, the problem is that $b(x)$ is unbounded: it behaves as $-x^2$ for $|x|\gg1$. Note that $f(x)/A(x)$ decays superexponentially as $x \to \pm \infty$, so that's no problem.

Roughly speaking, that means that in the `far field' $|x|\gg1$, the solution $z(x)$ will behave like the solutions to the homogeneous equation \begin{equation} \frac{d^2 z}{d x^2} - x^2 z = 0, \end{equation} which in turn behave as a linear combination of $e^{\pm x^2}$. If their behaviour would not be superexponential, then we would have an exponential dichotomy. This would allow us to choose the two linearly independent solutions such that one decays as $x \to -\infty$ and the other one decays as $x \to +\infty$. Using the method of variation of parameters (or using Green's function), one could then formally express the solution to the inhomogeneous problem $(*)$ in terms of those functions, and investigate whether it would be possible for this inhomogeneous solution to be bounded.

However, for superexponential growth/decay, I'm not sure whether this works. You could always try to mimic the approach sketched above, but I'm not sure which problems you could encounter. I think that the expertise of @ChristianRemling will be very useful here, and I hope he is able to be of further assistance.

$\endgroup$
  • 1
    $\begingroup$ $-d^2/dx^2+x^2$ is the harmonic oscillator Schrodinger operator, and $0$ is in the resolvent set of this operator. This gives right away that $-z''+x^2z=-f/A$ has a solution in $L^2(\mathbb R)$. Now that's not yet exactly what you want ($b$ is not exactly $x^2$, and $L^2$ is not the same as a derivative bounded), but it seems quite likely that one can get there with more work. $\endgroup$ – Christian Remling Sep 23 '14 at 17:13
  • $\begingroup$ Thank you very much @Frits Veerman and ChristianRemling. I got much knowledge from both of you... $\endgroup$ – Jlamprong Sep 24 '14 at 5:17

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.