1
$\begingroup$

In the study of nonlinear conservation laws a lot of time I work on the two problems given bellow:

$$(1) \hspace{1cm} \begin{cases} u_t+(f_{1}(u))_x=\lambda \cdot g(u) \\[2ex] u(x,0)=h_{1}(x) \end{cases} $$

$$(2) \hspace{1cm} \begin{cases} u_t+(f_{2}(u))_x=0 \\[2ex] u(x,0)=h_{2}(x) \end{cases} $$

Here u $\in \mathbb{R}^n$ and $\lambda$ is a constant. Besides $u$, the functions $g,f_1, f_2$ could also depend of $x$ and/or $t$.

A long time ago a professor of mine told me that he read somewhere that there is a way of transforming problem $(1)$ to the problem $(2)$ for some special cases of functions $g,f_1, f_2$. And vice versa. But he can't remember where he has read that.

Obviously it would be very useful if we can switch between problems $(1)$ and $(2)$ from time to time. In one moment you work on a problem that has a source term $\lambda \cdot g(u)$, i.e. problem $(1)$. In the other moment you use change of coordinates (or something similar) and now you work with a problem with no source term and with different initial condition (and possibly different flux too), i.e. problem $(2)$.

I am pretty sure that some general transformation doesn't exist. In the problem $(1)$ source is time dependent and in the problem $(2)$ new initial data are not time dependent. But

  • maybe source could go partially into the flux and partially into the initial data or something similar.
  • Or maybe this kind of transformation exists for some functions $g,f_1, f_2, h_1, h_2$ with the special properties.
  • Or maybe it exists for some other kind of PDE problems - that don't deal with conservation laws.

The only example of transformation of this kind I know is: if $g(u)=(w(u))_x$ (assuming that function $w$ could be found), then the source term would be absorbed in the flux term. So we would get the problem $(2)$ form where $f_2 (u) = f_1 (u) - \lambda \cdot w(u)$ and $h_2(x)=h_1 (x).$

My question is: is anyone encountered some similar transformation and where (reference in the literature would be nice)? Of course you could share your ideas or answers too. This question intrigues me from the moment the professor told me that.

$\endgroup$
5
  • $\begingroup$ such a transformation from eq. 1 to eq. 2 seems unlikely to exist, at least not for a general source: equation 2 has the conservation law that $\int u dx$ is time independent; in equation 1 it is not, because of the source. $\endgroup$ Dec 7, 2018 at 15:17
  • $\begingroup$ @CarloBeenakker: I am skeptical of existence of such transformation for a general source too. But I hope it exists for some special source term. In problem (1) I could use different kinds of source terms (the one written above is the most general one I work with). $\endgroup$
    – Mark
    Dec 7, 2018 at 15:26
  • $\begingroup$ @WillieWong: In general I am pretty sure too that it is not possible. Also your simple example explains it good. But I was wondering is there some paper that studies some conservation laws type system or equation, where authors transformed problem with some kind of source term (any type of source term) to the problem with zero source term (but now they have different initial conditions and/or flux). Or vice versa. I guess that the fact that the professor said it to me that he saw something similar intrigues me. $\endgroup$
    – Mark
    Dec 7, 2018 at 19:20
  • $\begingroup$ since you are doing general change of coordinates, are you going to insist that $g$ and $f_1$, $f_2$ are only functions of the unknown, or can they also depend on the space-time coordinates $t$ and $x$? If they are allowed to depend also on $t$ and $x$ I think what you want may be possible. $\endgroup$ Dec 7, 2018 at 19:40
  • $\begingroup$ @WillieWong: They could also depend of $t$ and $x$ (in the problem I am currently working on, they depend only on $u$ - that's why I write it that way above - I probably should explain it more clearer). You could use any type of source term, initial condition and flux. $\endgroup$
    – Mark
    Dec 7, 2018 at 19:53

2 Answers 2

4
$\begingroup$

When $n=1$, you can always do this, at least near $t=0$, by solving a single inhomogeneous, linear first-order PDE; you can even arrange that $h_2 = h_1$. When $n>1$, there is a geometrical obstruction that can be computed in terms of $f_1$ and $\lambda g$. This is a classical fact in the geometry of PDE and characteristic systems. Here is a summary of the argument:

First, let's suppose that $n=1$ and that $f_1$ and $g$ are differentiable functions of $x,t,u$. Then the condition that $u(x,t)$ satisfies the equation and initial condition is that the surface graph $\bigl(u(x,t),x,t\bigr)$ in $\mathbb{R}^3$ should contain the curve $\bigl(h_1(x),x,0\bigr)$ and that the $2$-form $$ \Omega = \mathrm{d}u\wedge\mathrm{d}x -\mathrm{d}(f_1)\wedge\mathrm{d}t - \lambda\,g\ \mathrm{d}t\wedge\mathrm{d}x $$ should vanish on the surface graph. This is equivalent to saying that the surface graph is tangent to the characteristic vector field $$ X = \frac{\partial}{\partial t} + \frac{\partial f_1}{\partial u}\,\frac{\partial}{\partial x} + \left(\lambda g - \frac{\partial f_1}{\partial x}\right)\,\frac{\partial}{\partial u}. $$ Now let $\mu$ be the solution to the first-order, inhomogeneous, linear, scalar PDE $\mathrm{d}\mu(X) = \lambda g_u$, i.e., $$ \frac{\partial\mu}{\partial t} + \frac{\partial f_1}{\partial u}\,\frac{\partial\mu}{\partial x} + \left(\lambda g - \frac{\partial f_1}{\partial x}\right)\,\frac{\partial\mu}{\partial u} = \lambda\,\frac{\partial g}{\partial u} $$ such that $\mu(u,x,0) = 1$. (Such a solution always exists on a neighborhood of the plane $t=0$. With the right conditions on $f_1$ and $\lambda g$, the solution $\mu$ could exist globally on $\mathbb{R}^3$.)

Then $\bar\Omega = \mathrm{e}^{-\mu}\Omega$ will be closed (i.e., $\mathrm{d}\bar\Omega = 0$) and then it is easy to see that it can be written in the form $$ \bar\Omega = \mathrm{d}v\wedge\mathrm{d}x -\mathrm{d}f_2\wedge\mathrm{d}t $$ for some functions $v$ and $f_2$ on $\mathbb{R}^3$ with $v(u,x,0) = u$. Since $\mathrm{d}v\wedge\mathrm{d}x\wedge\mathrm{d}t = \bar\Omega\wedge\mathrm{d}t = \mathrm{e}^{-\mu}\mathrm{d}u\wedge\mathrm{d}x\wedge\mathrm{d}t\not=0$, it follows that $v,x,t$ are (local) coordinates on $\mathbb{R}^3$ (and, given reasonable assumptions on $f_1$ and $g$, will actually be global coordinates on $\mathbb{R}^3$). Writing $f_2$ as a function of $v,x,t$ we now have $$ \bar\Omega = \mathrm{d}v\wedge\mathrm{d}x -\mathrm{d}\bigl(f_2(v,x,t)\bigr)\wedge\mathrm{d}t $$ It follows that the coordinate transformation $(u,x,t)\mapsto\bigl(v(u,x,t),x,t\bigr)$ takes the first sytem to the second system (with $h_2 = h_1$), as desired.

What happens when $n>1$ is that the $2$-form $\Omega$ on $\mathbb{R}^3$ is replaced by $n$ $2$-forms on $\mathbb{R}^{n+2}$. In order to get to a system of the second kind, you would need to find $n$ linearly independent closed $2$-forms that have the same linear span as the given $n$ $2$-forms on $\mathbb{R}^{n+2}$. However, when $n>1$, there are `integrability conditions' on the $2$-forms in order to be able to do this.

For example, when $n=2$, you are asking when a pair of $2$-forms on $\mathbb{R}^4$ has the same linear span as a pair of closed $2$-forms, and there are well-known 'curvature' obstructions to this. Most of the time, the space of local, closed $2$-forms in the linear span of the two given $2$-forms is a finite dimensional space; in fact, nearly always, it's the zero dimensional space.

$\endgroup$
1
  • $\begingroup$ Thank you for the answer. I didn't expect this angle of explanation (geometry of PDE), but from it I clearly see the problems that could happen if transformations of this kind exist. And I like this explanation. I guess I expected (deep down) that there is some easy change of variables (which I have never encountered) that could work in the case of systems $(1)$ and $(2)$. And of course the common sense says no. Thanks again. $\endgroup$
    – Mark
    May 22, 2020 at 16:24
2
$\begingroup$

Here's one possible solution, but this may or may not be what your professor had in mind.


Since $\lambda$ is a constant, we can ignore it by absorbing it into $g$.

Assume $u$ is scalar (takes value in $\mathbb{R}^1$).

Assume that $g$ is a function of $u$ only.

Let $G(s; u_0)$ be the solution to $\partial_s G = g(G)$ with initial data $G(0;u_0) = u_0$.

Suppose for every fixed $s$, the two functions $f(s;j)$ and $F(s;k)$ are related by $$(\partial_k F)(s;G(s;j)) = \partial_jf(s;j). \tag{*}$$

Then if $v(t,x)$ is a solution to the homogeneous equation

$$ \partial_t v + \partial_x [f(t;v)] = 0 $$

then the function $u(t,x) = G(t; v(t,x))$ satisfies

$$ \partial_t u + \partial_x [F(t; u)] = g(u) $$

You even have that $u(0,x) = v(0,x)$.

In particular, you can definitely convert from (1) to (2) in this setting, just by solving two ODEs (for $G$ and for $f$).


When $u$ is not a scalar (has multiple components), the replacement for (*) may not be integrable, giving an obstruction to the sort of transformation you seek using this naive method. (Roughly, the equation is $ \partial f = (\partial G)^{-1} \cdot [(\partial F)\circ G] \cdot (\partial G)$.)

$\endgroup$
1
  • $\begingroup$ In the case of equations, the addition of $t$ in the fluxes solved the problem very elegantly. For systems, when $u$ has multiple components such as 2 or 3 this construction gets complicated (it would depend of the concrete system in question and of course of solving $(*)$). Still, this is the first example I've seen with transformation between $(1)$ and $(2)$. Thanks for that. Also I talked with my professor yesterday. He said to me that he didn't have this in mind. :( Who knows on what he was thinking of. :) $\endgroup$
    – Mark
    Dec 9, 2018 at 12:58

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.