I'll star by saying that I am not really familiar with the field of PDEs so this questions may be trivial or ill-possed in that case please let me know.

I am in search of some existence (Global) result regarding a system of first order PDEs with many state-variables and a non-homogeneity that is non-linear in the solutions, i.e.

\begin{align} \begin{cases} \partial_t \mathbf u+\sum_{n=1}^{K} A_n(t)\partial_{x_n} \mathbf u=\mathbf b(t,\mathbf u,x_1,\cdots,x_K)\\ \mathbf u(0,x_1,\cdots,x_K)=u_0\in\mathbb R^N, \end{cases} \end{align}

where the matrices $A_n, n=1,\cdots, K$ are diagonal, and $\mathbf b$ is Lipschitz continuous in $\mathbf u$ uniformly in $t,x_1,\cdots,x_n$.

In Bressan's "Conservations laws" book the case with two variables is considered, namely $t$ and $x$. I have found some articles online considering the case with many state variables but in that case the non-homogeneity was linear in the $\mathbf u$ .

As far as I know the method of the characteristics applies only the one-dimensional case, so I am a little bit lost.

Is there some result I can use to show the existence (and uniqueness if possible) of a solution for this system?

Thanks in advance.

  • 4
    $\begingroup$ I think the standard local existence result for your system would require higher regularity for $\mathbf{b}$. Your equation is automatically symmetric hyperbolic, and so Kato's result (1975, ARMA) can apply, provided that $\mathbf{b}$ is continuously differentiable $\lfloor K/2\rfloor + 1$ times. $\endgroup$ Dec 1, 2020 at 17:49
  • $\begingroup$ Thanks @WillieWong, well asking for more regularity on $b$ is not an issue, what worries me is the local character of the result. Still do you have the complete name of Kato's article/book? I don't find it $\endgroup$
    – Chaos
    Dec 1, 2020 at 17:53
  • 1
    $\begingroup$ The method of characteristics for first order scalar PDEs approaches by finding a foliation of the domain by hypersurfaces on which the solution is constant. Even though $A$ is diagonal, the corresponding foliations are different for different components of $\mathbf{u}$, and thus in general they are not compatible and cannot be reassembled to a coupled equation. $\endgroup$ Dec 1, 2020 at 20:15
  • 1
    $\begingroup$ Kato's paper: link.springer.com/article/10.1007/BF00280740. For the "global" problem: if you let me know more precisely what can be assumed about $\mathbf{b}$, I may be able to help you with it. But there is unlikely any general thing you can say about it; results about global existence of hyperbolic systems typically are very structure dependent. $\endgroup$ Dec 1, 2020 at 20:21
  • 3
    $\begingroup$ Wait, hang-on: your $A_n$ are actually diagonal, not just block diagonal? And they are real-valued, right? I seem to have been overthinking your problem because I thought it is more general. So what you have is a coupled system of transport equations; then you don't need Kato. $\endgroup$ Dec 2, 2020 at 16:22

1 Answer 1


Okay, so I would write your equations instead in the following form: $$ \partial_t u_i + v_i(t) \cdot \nabla u_i = b_i (t, \vec{x}, \vec{u}) $$ This is a system of transport equations and so can actually be solved by using a variation of the Picard-Lindelof argument.

(I am implicitly assuming that your function $b$ is suitably nice in certain ways, which are implied by what you said is allowed in this comment.)


Consider the iteration scheme $$ \vec{w} \mapsto \vec{w}' $$ where $\vec{v}'$ solves the linear inhomogeneous equation $$ \partial_t w'_i + v_i(t) \cdot \nabla w'_i = b_i(t,\vec{x}, \vec{w})$$ with the fixed, given, initial data $\mathring{\vec{w}}:\mathbb{R}^K\to \mathbb{R}^N$.

Given a continuous function $\vec{w}: [-t_0, t_0] \times\mathbb{R}^K \to \mathbb{R}^N$, the function $\vec{w}'$ can be solved component wise by integrating along the integral curves of $\partial_t + v_i(t) \cdot \nabla$ on $[-t_0, t_0]\times \mathbb{R}^K$.

It thus suffices to show that this iteration scheme converges.


Suppose $\vec{z}$ and $\vec{w}$ are given functions that are uniformly bounded by some constant $M$. Then there exists (under what I hope are reasonable assumptions on $b_i$) some constant $M'$ such that $b_i(t,\vec{x},\vec{z})$ and $b_i(t,\vec{x},\vec{w})$ are uniformly bounded by $M'$. Thus you have that

$$ | \vec{z}'|_\infty, |\vec{w}'|_\infty \leq |\mathring{\vec{w}}|_\infty + M' t_0 $$

Choose $M = 2 |\mathring{\vec{w}}|_\infty$ and there exists some $t_0 > 0$ such that the iteration mapping maps the (closed) ball of radius $M$ in $C^0([-t_0, t_0]\times \mathbb{R}^K,\mathbb{R}^N)$ to itself.


For the differences you find that $$ \partial_t (z'_i - w'_i) + v_i(t)\cdot \nabla (z'_i - w'_i) = b_i(t,\vec{x}, \vec{z}) - b_i(t,\vec{x},\vec{w}) $$ Using the Lipschitz continuity in the final slot of $b_i$ you get that $$ \big|\partial_t (z'_i - w'_i) + v_i(t)\cdot \nabla (z'_i - w'_i) \big| \leq M'' |\vec{z} - \vec{w}| $$ for some $M''$. Integrating you get that if $t_0$ is chosen sufficiently small, the iteration mapping is a contraction mapping.


Applying Banach fixed-point theorem you get local existence.


Having proven local existence, you can upgrade this to global in the same way you argue for ODEs: by showing that the sup norm of the functions involved does not blow-up. This follows from the global Lipschitz property of the functions $b_i$, which guarantees that $|b_i(t,\vec{x},\vec{w})| \leq M''' (1 + |\vec{w}|)$

So a standard Gronwall's inequality argument well tell you that the sup-norm of the solution cannot grow faster than $A e^{2 M''' |t|}$ for some constant $A$.

  • $\begingroup$ thank you so much Willie! This is really helpful $\endgroup$
    – Chaos
    Dec 2, 2020 at 17:08
  • $\begingroup$ +1 Willie Wong and @Chaos: I like the analysis of the posed problem you did, up to get the final, optimal solution. In sum, really a nice piece of real research mathematics. $\endgroup$ Dec 2, 2020 at 21:33
  • $\begingroup$ Ey Willie sorry for bothering you again, in the last step, how can I apply the Gronwall inequality? I am not familiar with the application to PDEs. $\endgroup$
    – Chaos
    Dec 3, 2020 at 13:37
  • 2
    $\begingroup$ THe differential inequality, after integration, gives $$|\vec{w}(t)|_{\infty} \leq |\vec{w}(0)|_{\infty} + \int_0^t M'''(1 + |\vec{w}(s)|_{\infty} ~ds$$ Add 1 to both side and apply Gronwall to the quantity $1 + |\vec{w}(t)|_\infty$. $\endgroup$ Dec 3, 2020 at 15:04
  • $\begingroup$ Why do we end up integrating wrt to $t$ alone, what about the $\vec{x}$? Sorry if this is a stupid question, I know very little about this stuff. $\endgroup$
    – Chaos
    Dec 4, 2020 at 9:04

Your Answer

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

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