7
$\begingroup$

In my research I have come across the following non-linear differential equation: $$x\frac{d}{d x}\left(x\frac{d y}{d x}\right)=4(y-x^{2}y+y^{3})$$ I want to find the general solution of this equation for $x\geq0$, but didn't manage to do it. I have found some facts about the solutions by investigating the asymptotics:

  1. For $x\ll1$, the solutions go like $y=Cx^2$ for some constant $C$, or diverge.
  2. For $x\gg1$, the solutions go like $y=x$, or go to $0$ (probably after oscillating, found by numerics), or diverge.

In particular, I am interested in a one of the solutions. By looking upon the numerics, I believe that there is only one solution that goes like $y=x$ for $x\gg1$, and it does not diverge at $x=0$. I didn't manage to prove this by now. I am interested in finding this solution analytically, or if impossible, just find appropriate $C$ analytically.

Thanks!

$\endgroup$
6
  • 4
    $\begingroup$ Did you try to change the independent variable $x=e^t$? $\endgroup$ Commented Jul 30, 2020 at 21:25
  • $\begingroup$ Yes, It was helpful in order to investigate the asymptotics for $x\ll1$, but for arbitrary $x$ it gives $$\frac{d^{2}y}{dt^{2}}=4\left(y-e^{2t}y+y^{3}\right)$$ which I did not find easier to solve. $\endgroup$
    – Roy
    Commented Jul 30, 2020 at 23:35
  • 1
    $\begingroup$ All I can currently do is to show that at least one solution with the properties you required exists. But since it is just a minimizer of some quite fancy energy functional (actually, even worse than that: it is actually the limit of a sequence of minimizers on finite intervals; its energy on the whole line diverges to $-\infty$), I doubt very much that anything explicit can be done here. Let me know if you are interested in that existence proof. $\endgroup$
    – fedja
    Commented Jul 31, 2020 at 1:25
  • 1
    $\begingroup$ Looking at the plot of $z=f(t,y)=4y(1-e^{2t}+y^2)$, $y$ can't be asymptotic to $e^t$ for ever. That would either make $y=y(t)$ infinitely steep or make it cross the line $\sqrt{e^{2t}-1}≈e^{t}$ and make $z<0$, which would make $y=y(t)$ oscillatory. It would oscillate between the two positive and negative regions. $\endgroup$
    – Alapan Das
    Commented Jul 31, 2020 at 3:58
  • 1
    $\begingroup$ @fedja I am interested in your existence proof. Nice if you give the details. Thanks $\endgroup$ Commented Aug 1, 2020 at 12:33

2 Answers 2

5
$\begingroup$

Mathematica's command DSolve cannot solve your ODE in closed form; so, apparently, a closed-form solution does not exist.

However, it will shown here that, in line with your conjecture, for each real $c$ your ODE has solutions $y$ such that $y(x)=(c+o(1))x^2$ as $x\to0+$.

Indeed, your ODE is \begin{equation*} x^2 y''+x y'=4(y-x^2y+y^3) \tag{1} \end{equation*} for real $x>0$. Assuming the initial conditions $y(0+)=y'(0+)=0$ and using the Taylor formulas $y'(x)=x\int_0^1 v(sx)\,ds$ and $y(x)=x^2\int_0^1 (1-s)v(sx)\,ds$ with $v:=y''$, we rewrite (1) as \begin{equation*} v(x)=\int_0^1 K(x,s)v(sx)\,ds+4x^4\Big(\int_0^1 (1-s)v(sx)\,ds\Big)^3, \tag{2} \end{equation*} where $$K(x,s):=[4(1-x^2)(1-s)-1]s.$$ For any real $c$, setting now \begin{equation*} v(x)=c+x w(x), \end{equation*} we rewrite (2) as \begin{equation*} w(x)=F(w)(x):=-2cx+\int_0^1 K(x,s)w(sx)\,ds+4x^3\Big(\int_0^1 (1-s)(c+xw(sx))\,ds\Big)^3. \tag{3} \end{equation*} Take now any real $m>0$ and $h>0$, and let \begin{equation*} W_{m,h}:=\{w\in C[0,h]\colon\|w\|\le m\}, \end{equation*} where $\|w\|:=\max_{x\in[0,h]}|w(x)|$.

Since $|K(x,s)|$ is convex in $x^2$, for $x\in[0,1]$ we have \begin{equation*} \int_0^1 |K(x,s)|\,ds\le\max\Big(\int_0^1 |K(0,s)|\,ds,\int_0^1 |K(1,s)|\,ds\Big) =\max(19/48,1/2)=1/2. \end{equation*} So, for any $w\in W_{m,h}$ \begin{equation*} \|F(w)\|\le2|c|h+m/2+h^3(|c|+hm)^3/2\le m \tag{4} \end{equation*} if $h>0$ is small enough -- which will be assumed in what follows. So, $F$ maps $W_{m,h}$ into $W_{m,h}$. Moreover, if for some $w$ and $u$ in $W_{m,h}$ we have $\|w-u\|\le t$ for some real $t\ge0$, then similarly to (4) we get \begin{equation*} \|F(w)-F(u)\|\le t/2+3h^3(|c|+hm)^2ht/2\le\tfrac23\,t \end{equation*} if $h>0$ is small enough. So, $F$ is a contractive map of $W_{m,h}$ into $W_{m,h}$, and hence $F$ has a fixed point. That is, equation (3) has a solution $w\in W_{m,h}$.

Thus, ODE (1) has a solution $y$ with $y''(x)=c+x w(x)$ for $x\in[0,h]$ and $w\in W_{m,h}$, and with $y(0+)=y'(0+)=0$. So, $y(x)=cx^2/2+O(x^3)$ as $x\to0+$. $\Box$

$\endgroup$
3
  • 1
    $\begingroup$ wouldn't the small-$x$ behavior already follow upon linearization of the ODE? the linearized ODE has solution $c J_2(2x)=cx^2/2+{\cal O}(x^3)$. $\endgroup$ Commented Jul 31, 2020 at 11:15
  • $\begingroup$ Thank you losif, nice to see the proof of the conjecture! Nevertheless, it does not give new details about the eqution, details that I still hope we can find. $\endgroup$
    – Roy
    Commented Jul 31, 2020 at 12:02
  • 1
    $\begingroup$ @CarloBeenakker : The linearization needs to be justified. The answer here may be viewed, in part, as such a justification. (However, the method -- which is rather standard -- should work, not just for this ODE, but in a general setting, without even an implicit association with linearization. In view of the Schauder fixed point theorem, one does not even need the right-hand-side function $F$ to be contractive.) $\endgroup$ Commented Jul 31, 2020 at 13:08
3
$\begingroup$

Is this too naive? To obtain the large-$x$ behavior, I take a series expansion of $y(x)$ in powers of $1/x$, $$y(x)=c x + a_0 + a_1/x + a_2/x^2+ a_3/x^3 +\cdots,$$ substitute into $$R(x)=x\frac{d}{d x}\left(x\frac{d y}{d x}\right)-4(y-x^{2}y+y^{3}),$$ expand $R(x)$ in powers of $1/x$ and demand that the leading order terms vanish. It follows that $c=1$, $a_{n}=0$ for any even $n$, while for odd $n$ I find $$a_1=-\frac{3}{8},\;\;a_3=-\frac{9}{128},a_5=-\frac{99}{1024},\;\;a_7=-\frac{11637}{32768},\;\;a_9=-\frac{627669}{262144},$$ so coefficients of the form $a_{2n+1}=-b_n/2^{2n+1}$ with integer $b_n$.

$\endgroup$
3
  • $\begingroup$ I believe that this is indeed true for solutions that does not go to $0$ or diverge at infinity. This is the way I found the asymptotics for $x\gg1$ I wrote about - but this is not the only solution, there are vanishing and diverging solutions too. Nevertheless, I do not see how to use it to prove that the solution that does not vanish or diverge at infinity is unique, and how to find the constant $C$ of the $x\ll1%$ asymptotics of this solution, or show it exists. $\endgroup$
    – Roy
    Commented Jul 31, 2020 at 12:22
  • 1
    $\begingroup$ Good question. It could happen that the convergence radius of the series in $1/x$ is zero ... so one would need a statement about the asymptotics of the $a_n $ to settle that. $\endgroup$ Commented Jul 31, 2020 at 14:26
  • $\begingroup$ @MichaelEngelhardt : I'd guess the radius of convergence of the series expansion is $0$ in this case. Of course, we could use just an asymptotic expansion for $y(x)$, provided that we have the matching asymptotic expansions for $y'(x)$ and $y''(x)$. But of course, we'd need to show that a solution $y$ admitting such asymptotic expansions exists. $\endgroup$ Commented Jul 31, 2020 at 15:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.