I'm trying to solve the following partial differential equations: $$ u_t + \frac{1}{3}{u_x}^3 = 0 \tag{a} $$ $$ u_t + \frac{1}{3}{u_x}^3 = -cu \tag{b} $$ with the initial value problem $$ u(x,0)=h(x)= \left\lbrace \begin{aligned} &e^{x}-1 & &\text{for}\quad x<0\\ &e^{-x}-1 & &\text{for}\quad x>0 \end{aligned} \right. $$ My idea was to set $v(x,t)=u_x(x,t)$, because then I get the transport equation in $v$ which I am able to solve: $v_t + v^2v_x =0$. But when I do this, my solution for $v$ is $$ v(x,t)= \left\lbrace \begin{aligned} &\phantom{-}ae^{x-v^2 t} & &\text{for}\quad x<0\\ &{-a}e^{-x+v^2 t} & &\text{for}\quad x>0 \end{aligned} \right. $$ Can someone help me with this equation? Is $u_x = a e^{x-u_x^2 t}$ the correct answer? Or should I maybe do something very different to solve this equation?


Both equations (a) and (b) are Hamilton-Jacobi equations. Indeed, they are derived from a canonical transformation involving a type-2 generating function $u(x,t)$ which makes vanish the new Hamiltonian $$ K = H(x,u_x) + (\partial_t + c)\, u \, . $$ Here, $H(q,p)=\frac{1}{3}p^3$ is the original Hamiltonian and $\partial_t + c$ defines the time-differentiation operator. Setting $v=u_x$, we have $$ v_t = u_{tx} = -\left(\tfrac{1}{3}{u_x}^3\right)_x - cu_x = -v^2v_{x} - cv \, . $$ Thus, we consider the first-order quasilinear PDE $v_t + v^2v_{x} = -cv$ with initial data $v(x,0) = h'(x) = \pm e^{\pm x}$ for $\pm x<0$, and we apply the method of characteristics:

  • $\frac{\text d t}{\text d s} = 1$, letting $t(0)=0$, we know $t=s$.
  • $\frac{\text d v}{\text d s} = -cv$, letting $v(0)=h'(x_0)$, we know $v=h'(x_0)e^{-ct}$.
  • $\frac{\text d x}{\text d s} = v^2$, letting $x(0)=x_0$, we know $x=\frac{1}{2c}h'(x_0)^2(1-e^{-2ct}) + x_0$.

Injecting $h'(x_0) = ve^{ct}$ in the equation of characteristics, one obtains the implicit equation $$ v = h'\!\left(x-v^2\frac{e^{2ct}-1}{2c}\right) e^{-ct}\, . $$ Along the same characteristic curves, we have

  • $\frac{\text d u}{\text d s} = \tfrac23 v^3 - c u$, letting $u(0) = h(x_0)$, we know $u = h(x_0) e^{-ct} + \frac23\! \int_0^t e^{-c(t-s)} v(s)^3 \text d s$.

Thus, we get $$ u = \left(h\!\left(x-v^2\frac{e^{2ct}-1}{2c}\right) + h'\!\left(x-v^2\frac{e^{2ct}-1}{2c}\right)^3 \frac{1-e^{-2ct}}{3c} \right) e^{-ct} \, , $$ where the link between $x_0$ and $v$ along characteristics has been used. For short times, the previous solution is valid. The method of characteristics breaks down when characteristics intersect (breaking time). We use the fact that $\frac{\text d x}{\text d x_0}$ vanishes at the breaking time $$ t_B = \inf_{x_0\in \Bbb R} \frac{-1}{2 c} \ln\left(1 + \frac{c}{h'(x_0)h''(x_0)}\right) . $$ However, it seems pointless to look further for full analytical expressions in the general case.

If $c=0$, the characteristics are straight lines $x=x_0+v^2t$ along which $v=h'(x_0)$ is constant, and along which $u = h(x_0) + \frac23 v^3 t$. A sketch of the $x$-$t$ plane is displayed below:


The breaking time becomes $t_B = \inf_{x_0} -(2h'(x_0)h''(x_0))^{-1} = 1/2$. For short times $t<t_B$, the implicit equation for $v$ reads $v = h'(x-v^2t)$, i.e. $v = \pm e^{\pm (x-v^2t)}$ if $\pm(x-v^2t)<0$. Its analytic solution $$ v(x,t) = \pm\exp\! \left(\pm x- \tfrac{1}{2}W(\pm 2t e^{\pm 2x})\right) \quad\text{for}\quad {\pm (}x-t) < 0 $$ involves the Lambert W function. The expression of $u$ is deduced from $u = h(x-v^2 t) + \frac23 v^3 t$. For larger times $t>t_B$, particular care should be taken when computing weak solutions (shock waves) since the flux $f:v\mapsto \frac{1}{3}v^3$ is nonconvex.

  • $\begingroup$ Thank you a lot for your answer! So with this method, you get a formula for $v$, but not for $u$? But if I want to compute u, for $c=0$ I could calculate $u$ with the formula for the Hamilton-Jacobi equation. Is there any way to compute $u$ directly if $c \neq 0$? $\endgroup$ – Infinite_28 Oct 25 '18 at 21:54
  • $\begingroup$ Okay, thanks. So there is no theory how to solve a non homogeneous Hamilton-Jacobi equation and get an explicit formula of the function, is there? $\endgroup$ – Infinite_28 Oct 26 '18 at 16:46
  • $\begingroup$ @Infinite_28 Maybe you were hoping that there is a Lax-Hopf formula for this non-homogeneous HJE... I don't know if such a thing exists, but at least we can find $v=u_x$ via the method of characteristics. $\endgroup$ – Harry49 Dec 11 '18 at 21:53

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.