2

Has this second-order nonlinear ODE been studied before or been given a known name that I can look up for further study or investigation?

$$ y’’ \sqrt{y}+f(x)=0, $$

in single real variable $x$, with $y=y(x)$, an arbitrary function $f(x)$, and with the double primes indicating second derivative with respect to $x$?

user135626
  • 1,309
  • I am not sure if it is transformable to a known ODE, or if it is integrable in general. However, if I were to investigate it I would start by looking at simpler cases. i.e. $f\equiv 0$, and $f\equiv const$. The first is very easy. The latter a first integral is easily attainable, however a general solution a lot more difficult and you will have to consider different branches, and I think at some stage you will have to invert a cubic.. egh. I would then consider the initial value problem $y(x_0)=y_0$, $y'(x_0)=v_0$. First numerically, then maybe expand the nonlinearity about $y_0$ –  Feb 02 '22 at 03:27
  • @epiliam thanks for your comment. Apart from the simplest $f$ cases, I was thinking to investigate whether it is scale invariant or equidimensional under transformation for certain (say, polynomial) forms of $f(x)$. What do you think? I have tried transformations like $y=u^{2}$, $y=e^{u}$, $y=u^{n}$ and $y=\sin u$ to no avail. – user135626 Feb 02 '22 at 03:56
  • Even if $f(x)=const$, the problem seems to be difficult. – Claude Leibovici Feb 02 '22 at 03:59
  • @user135626 I think that is an interesting path to take. Indeed, if you were to find any coordinate transformation under which the ODE is invariant I think that would be a very interesting result as it leads straight in to Lie's theory of DEs –  Feb 02 '22 at 04:04
  • @ClaudeLeibovici indeed it is. However, I was able to write down a solution which involves the inverse of a cubic. –  Feb 02 '22 at 04:05
  • @epiliam. I agree with you for the inverse. – Claude Leibovici Feb 02 '22 at 04:12
  • @epiliam would you please advise or direct me to an brief introduction to this Lie theory for DEs with applied examples? I guess canonical/Hamiltonian form would be a special case for area-preserving transforms, right? – user135626 Feb 02 '22 at 04:16
  • @user135626 I must admit I am not an expert on the topic (I'm a numerical guy). My knowledge of the topic is because I read Dresner's "Applications of Lie's Theory of Ordinary and Partial Differential Equations" (which happens to be sitting next to me on my desk right now). I thought it was quite a good introductory text, however, having not read any other texts on the topic I don't know how it stacks up. You'll probably need a good understanding of group theory before hand. –  Feb 02 '22 at 04:25
  • @epiliam thanks. Can you please advise on how you got the cubic solution analytically? – user135626 Feb 02 '22 at 08:18
  • @epiliam Cannot we simply say that, a particular solution for $y’’ \sqrt{y} +c_{1}=0$ by inspection is simply of the form $y=A x^{4/3}$, while the general solution for the associated homogeneous eqn $y’’=0$ is of the form $y=Bx+C$, therefore the final general solution form is $y=Ax^{4/3}+Bx+C$, and then we apply initial conditions? – user135626 Feb 02 '22 at 19:08
  • @user135626 You cannot do that for non-linear problems. I'll briefly write up how to solve the problem with a constant $f$ shortly. –  Feb 03 '22 at 02:51
  • @epiliam oh, that’s right. My mistake – user135626 Feb 03 '22 at 05:54

2 Answers2

1

Let us write the equation in the form \begin{align} y' ={}& v\\ v' ={}& \frac{f}{\sqrt{y}} \end{align} and as mentioned in the comments we'll consider the case $f\equiv const\ne 0$. We write $$ vv' = \frac{y'}{\sqrt{y}}f $$ which, due to a constant $f$, has the first integral $$ \frac{1}{2}v^2 = 2y^\frac12 f + c $$ where $c$ is a constant of integration. Let us define $$2y^\frac12 f + c = \frac12 u^2$$ from which we infer $$y=\frac{1}{4f^2}(\frac12 u^2 - c)^2\implies y'=\frac{1}{2f^2}(\frac12 u^2 - c)uu' $$ and so $$(y')^2=v^2=u^2$$ naturally leads to the two branches $$(\frac12 u^2-c)u'=\pm 2f^2.$$ Integrating yields $$\frac16 u^3 - cu = b\pm2f^2x$$ where $b$ is another constant of integration. Taking $c=0$ there is an obvious solution (well... two obvious solutions as it has two branches... or even six if we consider the complex branches of $\cdot^\frac13$) $$u=(6(b\pm2f^2x))^\frac13\implies y=\frac{1}{16f^2}(6(b\pm2f^2x))^\frac43.$$ Setting $b=0$ yields $$y=\big(\frac{9}{4}f\big)^\frac{2}{3} x^\frac{4}{3} .$$ However let's go for something a bit harder with $c\ne 0$ (in particular with $c>0$). We'll tidy things up a bit and write $p=6c$ and $q\equiv q(x)=-6(b\pm2f^2x)$. Thus we have that $$u^3-pu+q=0.$$ Okay, now cubics are gross but lets plow ahead anyway. I will assume from now on that $\big(\frac{q}{2}\big)^2\le\big(\frac{p}{3}\big)^3$. Note this implicitly assumes $p>0$ (which in turn implies $c>0$). I'll leave you to consider the other cases. There is a cubic formula but I don't like it and I prefer to write solutions in terms of trig or hyperbolic trig functions. Let us write $u=r\cos(\frac{\theta}{3})$, $r>0$. Applying the triple angle formula, an obvious choice is for $r$ to satisfy $\frac34 r^3 = pr$. Or equivalently, $$r = 2\big(\frac{p}{3}\big)^\frac12.$$ This leaves us with $$\frac{1}{4}r^3 \cos(\theta)+q=0\implies \cos(\theta) = -\frac{\frac{q}{2}}{\big(\frac{p}{3}\big)^\frac32}.$$ Note that my earlier assumption ensures that $|\cos (\theta)|\le 1$. Therefore we can write down all three solutions to the cubic via $$u=2\big(\frac{p}{3}\big)^\frac12\cos\big(\frac{\theta + 2n\pi}{3}\big)$$ where $n\in\mathbb{Z}$ and $$\theta = \arccos\Big(-\frac{\frac{q}{2}}{\big(\frac{p}{3}\big)^\frac32}\Big) = \arccos\Big(\frac{3(b\pm2f^2x)}{(2c)^\frac32}\Big).$$ Recalling our formulae for $y$ and $p$, $$y = \frac{c^2}{4f^2}\Big(4\cos^2\big(\frac{\theta + 2n\pi}{3}\big) - 1\Big)^2.$$ Hopefully I haven't made any mistakes. Note that there are implicitly 6 different branches contained in this solution.

1

For the case where $f(x)$ is a constant $a$, first switch variable and write $$-\frac {x''}{[x']^3 }\sqrt y=a$$ Now, reduction of order $p=x'$ $$\frac {p'}{p^3}=-\frac a{\sqrt y}\implies p=\pm \frac{1}{\sqrt{4 a \sqrt{y}+ c_1}}$$ $$x'=\pm \frac{1}{\sqrt{4 a \sqrt{y}+ c_1}}\implies x+c_2=\pm\frac{\left(2 a \sqrt{y}-c_1\right) \sqrt{4 a \sqrt{y}+c_1}}{6 a^2}$$ Squaring and using $z=\sqrt{y}$ $$16 a^3 z^3-12 a^2 c_1 z^2+c_1^3=36a^4(x+c_2)^2 \tag 1$$

Assuming $c_2=0$ and $z(0)=0$ imply $c_1=0$ which gives in the real domain $$z=\sqrt[3]{\frac{9 a }{4}x^2} \implies y=\Bigg[\frac{9 a }{4}\Bigg]^{\frac 23} \, x^{\frac 43}$$

Otherwise, we just need to solve the cubic equation $(1)$ in $z$ .

  • Thanks. How did you "switch variable" in the first step? This seems to me like the key step. – user135626 Feb 03 '22 at 09:12
  • @user135626. How did I ... what ? By the way, are they any boundary conditions ? – Claude Leibovici Feb 03 '22 at 09:14
  • Sorry text got truncated by error. I was asking: How did you "switch variable" in the first step? This seems to me like the key step. – user135626 Feb 03 '22 at 09:16
  • @user135626. It is a quite standard trick (at least for me). On the search bar, type my name plus switch variables. You should find 55 entries. Cheers :-) – Claude Leibovici Feb 03 '22 at 09:37
  • Yes, I can see that you use it quite a lot, indeed, in your entries :-). Is this related to the hodograph transform? I am still confused; perhaps it is my ignorance! I would appreciate it if you would please explain how it is done or kindly refer me to a text or webpage that explains it. I tried looking in your other entries, and you seem to give it straight away, but it is not clear to me how it is done. Thanks. – user135626 Feb 03 '22 at 09:54
  • @user135626. It it simple : use the relation between $x'(y)$ and $y'(x)$ (trivial) and the next (which I wrote). Do not go beyond since the formulae start to be a nightmare. To see the derivation : https://math.stackexchange.com/questions/566507/prove-that-d2x-dy2-equals-d2y-dx2dy-dx-3 – Claude Leibovici Feb 03 '22 at 10:12
  • Many thanks. I see it now. By the way, the boundary conditions are $y(0)=0, y(x_{0})=y_{0}$. – user135626 Feb 03 '22 at 10:23
  • I think this approach when generalized is called the hodograph method in PDEs. – user135626 Feb 03 '22 at 10:23
  • I got originally the same answer by noticing that the equation was autonomous ($x$ doesn’t appear explicitly or the transform $x\rightarrow x+ const$ leaves it unchanged), which invites the substitution $u(y)=y’(x)$ and with $y$ now as independent variable. Then I realized that is basically equivalent to your method of switching the variables and leads to the same integral… Your approach, similar to hodograph transform, I think is smarter, though.. cheers – user135626 Feb 03 '22 at 21:17
  • @epiliam Thanks for this, but I am not sure I fully understand why we need first $y’(0)\ge 0$ and we need it equal zero if $y(0)=0$? Are you saying that, to keep the square root argument nonnegative, we can keep $y$ non-decreasing over our domain ($x=0$ to $x=x0>0$) by making its slope always positive? I recognized from Picard’s theorem that a partial derivative in $y$ would give $1/\sqrt{y}$ term that is singular at 0. But this theorem only gives sufficient (not necessary) condition for a unique solution to exist. So your point may better help me in this, I can understand it. Thanks – user135626 Feb 04 '22 at 08:17
  • @user135626 I ended up deleting my comment as I realised it didn't apply to the two point boundary value problem. To check if that is well posed you end up with a very complicated set of conditions on $y_0$, $x_0$. However if you are interested in a unique real solution in a neighbourhood of the origin, then the initial value problem $y(0)=0$, $y'(0)=v_0$ is ill posed for all $v_0 \ne 0$. Try going through all the cases your self and figuring out why this is. –  Feb 06 '22 at 01:10
  • @epiliam I think you are right, and physically (from the physics related to the problem) it makes sense, but I am trying to prove it mathematically. Here is my attempt: according to Picard's theorem, if we have an equation in the form $y'(x)=F(x,y)$, then if we can show that $F, \partial F/\partial y$ are continuous at the initial point (here $x=0, y(0)=0$) then it is a well-posed problem. So, if I multiply both side of $y''=-a/\sqrt{y}$ by $y'$ and then integrate we have $y'=F(y)=\pm\sqrt{c-4a\sqrt{y}}$, where both $F$ and $\partial F/\partial y$ are indeed continuous at $(x,y)=(0,0)$ ,right? – user135626 Feb 06 '22 at 09:41
  • Does that make sense? I would appreciate it if you can let me know what you think of this method, or if you have advice on other methods to use to conclude that $y'=0$ is needed for a unique solution of the initial value problem. Thank you. – user135626 Feb 06 '22 at 09:44