[under construction]
Existence of entropy solutions for 1D conservation laws
Basic assumptions
We will construct entropy solutions of the conservation law
\begin{equation} \tag{C}
u_t + F(u)_x = 0, \qquad u(x, 0)=g(x).
\end{equation}
where we assume that there is a constant $\delta\gt0$ such that
- $F\in C^2(\R)$,
- $F'(u)\ge 0$ for all $u\in\R$
- $F''(u)\ge \delta$ for all $u\ge 0$
For simplicity we assume that all functions are periodic in $x$ with period $\ell$:
\begin{equation}\tag{Per}
\forall x, t : \quad u(x+\ell, t) = u(x, t), \quad g(x+\ell) = g(x).
\end{equation}
In the end we want to construct an entropy solution for any bounded measurable
initial data $g\in L^\infty(\R)$, but for now we will assume that $g\in C^1(\R)$. In
particular, we will assume that
- $0 \lt g(x) \lt M$ for all $x\in\R$
- $g'(x) \leq M_1$ for all $x\in\R$
The discretization and the approximate solutions $u^N(x, t)$
We discretize the equation in the spatial direction by choosing a large integer $N$ and
setting $\Delta x = \ell/N$. Instead of considering the function $u(x, t)$ at all real
$x$, we only consider its values at the discrete set of $x$ values $x_i = i\Delta x$
($i\in\Z$). So we set
\[
u_i(t) = u(i\Delta x, t), \quad i\in\Z, 0\le t \lt T
\]
for some fixed $T\gt0$. Because of the periodicity
condition we have $u_{i+N}(t) = u_i(t)$ for all $t$ and all $i\in\Z$. Replacing the
$x$-derivative in $F(u)_x$ in the conservation law by a discrete
difference quotient, we arrive at this set of ordinary differential equations for the
functions $u_i(t)$:
\begin{equation}\tag{Disc}
\frac{du_i}{dt} + \frac{F(u_i(t)) - F(u_{i-1}(t))}{\Delta x} = 0,
\qquad
u_i(0) = g(x_i).
\end{equation}
Note that we made a choice here, namely we picked a left differential quotient instead
of a right differential quotient: i.e. we could also have picked
\[
\frac{du_i}{dt} + \frac{F(u_{i+1}(t)) - F(u_{i}(t))}{\Delta x} = 0.
\]
As we will see below, there will be one point where our construction falls apart if we
choose the right difference instead of the left difference (Disc).
\begin{equation}
\tag{4}
u^N(x, t) = u^N_i(t) \text{ for } x\in(x_i, x_{i+1}), \; t\gt0.
\end{equation}
Our strategy for proving the existence of an entropy solution to
the conservation law is to do the following:
- Verify that the solutions $u^N_i(t)$ of the discretized equations (Disc) exist for all $t\ge0$. This implies that the
step functions $u^N(x,t)$ are well defined for each $N$. The main part of the
argument here is taken care of by the theory of Ordinary Differential Equations.
- Show that the $u^N(x, t)$ satisfy the entropy condition.
- Show that there is a sequence of $N_i\to\infty$ such that the functions $u^{N_i}$
converge pointwise almost everywhere on $\R\times[0, \infty)$. The essential
ingredient here is compactness of the sequence of solutions in an appropriate
topology. We get this by using a version of Ascoli’s theorem that is adapted to
measurable functions rather than continuous functions. All compactness theorems for
functions (like Ascoli’s theorem) require some form of equicontinuity, which one can
usually prove by estimating derivatives of the functions.
- Show that the limit $u(x, t) = \lim_{i\to\infty} u^{N_i}(x,t)$ is an integral
solution of the conservation law which satisfies the entropy condition.
Existence for the discretized equation
Because of our periodicity assumption we have $u_{i+N}(t) = u_i(t)$ so that we really
only have to consider the unknown functions $u_1(t)$, …, $u_N(t)$, and so that (Disc) is a system of $N$ ordinary differential
equations. Since the function $F:\R\to\R$ is $C^1$, the theory of ODE provides us with
solution $\{u_i(t)\}_{i\in\Z}$ that is defined on some short time interval $0\leq t\lt
t_*$.
Lemma 1. $0\leq u_i(t) \leq M_0$ for all
$i\in\Z$, and $t\in [0, t_*)$.
Proof. We use a maximum principle argument: if at any time $t$ we
choose $j$ so that $u_j(t) = \max_i u_i(t)$, then $u_j(t)\geq u_{j-1}(t)$, so that
$F'\geq0$ implies $F(u_j(t)) \geq F(u_{j-1}(t))$, and so that the discretized
conservation law
(Disc) implies $u_j'(t)\leq0$. This
suggests that $\max_i u_i(t)$ is a nonincreasing function of $t$, which is what the
Lemma claims. The complete (but less transparent) argument goes like this:
Let $M_\varepsilon(t) = M_0 + \varepsilon + \varepsilon t$ where
$\varepsilon\gt0$ is a (small) constant. We show by contradiction that $u_i(t)\lt
M_\varepsilon(t)$ for all $i$ and all $t\in[0, t_*)$. If this were not true, then
there would be a smallest $s\in[0, t_*)$ and a $j\in\Z$ with $u_j(s) = \max_i u_i(s) =
M_\varepsilon(s)$. Since $M_\varepsilon(0) \gt M_0$ it cannot be true that $s=0$, so we
have $s\in(0, t_*)$. Then $u_i(s)\leq u_j(s)$ for all $i\in\Z$, and for all $t\in[0,
s]$. Using $F'(u)\geq0$ we conclude that
\[
\frac{F(u_j(s)) - F(u_{j-1}(s))}{\Delta x} \geq0, \qquad
u_j'(s) \geq \varepsilon.
\]
This contradicts the differential equations
(Disc).
We have shown that $u_i(t)\leq M_\varepsilon(t)$ for all $i$, $t$, and
for all $\varepsilon\gt0$. Letting $\varepsilon\to0$ we conclude that $u_i(t)\leq M_0$
for all $t$, $i$, as claimed.
One proves $u_i(t)\geq 0$ in the same way. ////
Consequence. The solution $\{u_i(t)\}_{i\in\Z}$ to the
discretized conservation law
(Disc) exists for all
$t\geq0$.
Proof. The theory of ordinary differential equations implies
that the solution will exist until it becomes unbounded, while
Lemma 1 shows that the solution $u_i(t)$
remains bounded. ////
The entropy condition
Lemma 2.
The solution $u_i(t)$ to
(Disc) satisfies
\[
\frac{u_i(t) - u_{i-1}(t)}{\Delta x} \leq \frac{2}{\delta t},
\]
for all $i\in\Z$, $t\gt0$.
Proof.
Consider
\[
v_i(t) = \frac{u_{i+1}(t) - u_i(t)}{\Delta x}.
\]
The differential equations for $u_i$ and $u_{i+1}$ imply
\[
v_i'(t) + \frac{F(u_{i+1}) - 2 F(u_i) + F(u_{i-1})}{(\Delta x)^2} = 0.
\]
The Taylor expansion of $F$ at $u_i$ together with $F''(u) \geq \delta$ implies
\[
F(u) \geq F(u_i) + F'(u_i) (u-u_i) + \frac{\delta}{2}(u-u_i)^2.
\]
Hence
\begin{equation}
\tag{1}
v_i'(t) + F'(u_i) \frac{v_i - v_{i-1}}{\Delta x}
\leq -\frac{\delta}{2} \bigl( v_i^2 + v_{i-1}^2\bigr)
\leq -\frac{\delta}{2}v_i^2.
\end{equation}
For any constant $K\gt0$ we note that $\max _i v_i(t) \leq K/t$ holds true for small
enough $t\gt0$. If the inequality fails for some positive $t$, then there is again a
smallest $s\gt0$ such that $\max_i v_i(s) = K/s$, and we can choose $j\in\Z$ so that
$v_j(s) = \max_i v_i(s)$. In particular $v_j(s)\geq v_{j-1}(s)$, so
(1) implies that
\begin{equation}\tag{2}
v_j'(s) \leq -\frac{\delta}{2}v_j(s)^2 = - \frac{K^2\delta}{2s^2}.
\end{equation}
On the other hand $v_j(t)\leq K/t$ for $t\leq s$ with $v_j(s) = K/s$ implies
\begin{equation}\tag{3}
v_j'(s) \geq \frac{d}{ds} \frac{K}{s} = -\frac{K}{s^2}.
\end{equation}
Together,
(2) and
(3) imply
\[
\forall s\gt 0:\;
-\frac{K}{s^2} \leq -\frac{K^2\delta}{2s^2}\quad
\text{i.e. }
K \le \frac{2}{\delta}.
\]
We arrived at this conclusion by assuming that $\max _i v_i(t) \gt K/t$ for some
$t\gt0$. Hence, if we choose $K\gt 2/\delta$ then we have shown that $v_i(t) \leq K/t$
for all $i\in\Z$ and $t\gt0$. Since this holds for all $K\gt 2/\delta$, we can let
$K\searrow 2/\delta$, and the Lemma follows. ////
A derivative estimate
The inequality in Lemma 2 is a
universal upper bound for $v_i(t)$ in the sense that the same upper
bound holds for all initial conditions $g$. We can improve the upper bound by
taking the initial condition $g$ into account. The biggest improvement appears as
$t\searrow0$, which is good because the upper bound we have right now ($v_i(t) \leq
2/\delta t$) deteriorates as $t\searrow0$.
Proof.
The mean value theorem implies that at $t=0$ we have $v_i(0) = g'(\xi_i)$ for some
$\xi_i\in (x_i, x_{i+1})$. Hence we have $v_i(0)\leq M_1$.
For $t\gt 0$ we again use
the equation (1) for
$v_j(t)$ and apply a maximum principle argument. At any time $t_1\gt 0$ there will
be some $j$ such that $v_j(t_1)=\max_i v_i(t_1)$. At this $j$ it then follows from
(1) that either $v_j(t_1)=0$ or $v_j'(t_1)\lt 0$. This
implies that $\max_i v_i(t)$ can never increase above its initial value and there never
exceeds $M_1$. ////
Compactness
We have found upper bounds for the discrete derivative $v_i(t)$, but no lower bounds.
Nevertheless, the fact that $0\leq u_i(t)\leq M_0$ combined with $v_i(t)\leq M_1$ implies
an $L^1$ type bound for $|v_i(t)|$.
Lemma 4.
There is a constant $C$ such that
\[
\sum_{i=1}^{N} \bigl\{|v_i(t)| + |u_i'(t)|\bigr\}\Delta x \leq C
\]
for all $N\in\N$, and all $t\geq0$.
Proof.
It follows from $v_i(t)\leq M_1$ that $k\mapsto u_k(t) - M_1k \Delta x$ is nonincreasing.
Using the periodicity of $k\mapsto u_k(t)$ and
$N\Delta x=\ell$, we then get
\[
\sum_{i=1}^N \{M_1 - v_i(t)\} \Delta x
=M_1\underbrace{N\Delta x}_{=\ell} + \underbrace{u_{1}(t) - u_{N+1}(t)}_{=0}
=M_1\ell.
\]
Note that $M_1-v_i(t)\geq0$ for all $i, t$, so
\[
\sum_{i=1}^N |v_i(t)| \Delta x
\leq \sum_{i=1}^N \bigl\{ M_1 + |M_1-v_i(t)|\bigr\} \Delta x
\leq 2M_1 \ell.
\]
To get the estimate for $u_i'(t)$ we observe that
\[
u_i'(t) = - \frac{F(u_i) - F(u_{i-1})}{\Delta x}
= F'(\tilde u_i) v_{i-1}(t)
\]
where $\tilde u_i$ is a number between $u_{i-1}$ and $u_i$, provided by the Mean Value
Theorem.
Since $F\in C^1(\R)$ there is a constant $C_0$ such that $0\leq F'(u) \leq C_0$ for all
$u\in [0, M_0]$. We have already found that our solutions $u_i$ are bounded by $0\leq
u_i\leq M_0$, so $0\leq F'(\tilde u_i)\leq C_0$, and thus
\[
|u_i'(t)| \leq C_0 |v_{i-1}(t)|.
\]
Multiplying with $\Delta x$ and summing over $i$ we then get the estimate for $\sum
|u_i'(t)|\Delta x$.
////
Lemma 5. For all $\xi\in(-1, 1)$ and
$\tau\in(0,1)$ one has
\[
\iint_{Q} \left| u^N(x, t) - u^N(x+\xi, t+\tau)\right|\, dx\,dt
\leq C_1 \bigl(|\xi|+\tau\bigr)
\]
where the constant $C_1$ is independent of $N\in\N$.
This implies that the sequence $u^N(x, t)$ has a subsequence that converges in $L^1(Q)$,
and therefore has a further subsequence that converges pointwise almost everywhere.
The limit is an integral solution
We will prove
\[
\lim_{N\to\infty}
\iint \varphi_t u^N(x, t)\,dx\,dt
+ \iint \varphi_x F(u^N(x, t)) dx\,dt
+ \int g(x) \varphi(x, 0)
\,dx
=0.\tag{5}
\]
For the first term we have
\begin{align}
\iint \varphi_t u^N(x, t) \, dx\,dt
&= \sum_i \int_0^T \int_{x_i}^{x_{i+1}} \varphi_t(x, t) u_i(t) dt \\
&= -\sum_i \int_{x_i}^{x_{i+1}} \varphi(x, 0) g(x_i) dt
-\sum_i \int_0^T \int_{x_i}^{x_{i+1}} \varphi(x, t) u_i'(t) dx dt
\end{align}
Uniform continuity of $g$ implies that
\[
\sum_i \int_{x_i}^{x_{i+1}} \varphi(x, 0) g(x_i) dt
= \int_0^\ell \varphi(x, 0)g(x)dx + o(1)
\]
as $N\to\infty$, so that
\begin{equation}
\iint \varphi_t u^N(x, t) \, dx\,dt
= - \int g(x) \varphi(x, 0) dx
- \sum_i \int_0^T \int_{x_i}^{x_{i+1}} \varphi(x, t) u_i'(t) dt
+ o(1). \tag{6}
\end{equation}
For the second term in (5) we get
\begin{align*}
\iint \varphi_x F(u^N)\, dx\, dt
&= \sum_i \int_0^T \int_{x_i}^{x_{i+1}} \varphi_x(x, t) F(u_i(t))\,dx\,dt\\
&= \sum_i \int_0^T \bigl(\varphi(x_{i+1}, t) - \varphi(x_i, t)\bigr) F(u_i(t)) dt \\
&= - \sum_i \int_0^T \varphi(x_i, t) \bigl\{F(u_i(t)) - F(u_{i-1}(t))\bigr\}\, dt\\
&=-\sum_i \int_0^T \int_{x_i}^{x_{i+1}} \varphi(x_i, t) \frac{F(u_i(t)) -
F(u_{i-1}(t))}{\Delta x}\,dx\,dt \\
&=-\sum_i \int_0^T \int_{x_i}^{x_{i+1}} \varphi(x_i, t) u_i'(t) \,dx\,dt.
\tag{7}
\end{align*}
This last expression looks almost the same as the double integrals at the end
of (6); the difference is that
in (7) $\varphi$ is evaluated at $x_i$ while in (6) $\varphi$ is evaluated at $x$.
By combining
(6), and
(7) we
find that
\begin{multline}
J^N \stackrel{\rm def}=
\iint \Bigl\{\varphi_t u^N + \varphi_x F(u^N)\Bigr\}\, dx\,dt
+\int g(x) \varphi(x,0)\, dx = \\
\sum_i \int_0^T\int_{x_i}^{x_{i+1}} \bigl(\varphi(x, t) -\varphi(x_i, t)\bigr)
\frac{du_i^N}{dt} \,dx\,dt + o(1).
\end{multline}
The test function $\varphi$ is continuously differentiable and in the above integral we
always have $x_i\lt x \lt x_{i+1}$, so that $|\varphi(x,t)-\varphi(x_i,t)|\leq C
(x_{i+1} - x_i) = C \Delta x$, where $C = \sup |\varphi_x|$, and therefore
\[
\left| \int_{x_i}^{x_{i+1}} \bigl( \varphi(x, t) - \varphi(x_i, t)\bigr)
\frac{du^N_i}{dt} \,dx
\right|
\leq C (\Delta x)^2 \left|\frac{du^N_i}{dt} \right|.
\]
Hence, summing over $i=1, \dots, N$ and integrating over $t\in(0,T)$ we get
\[
|J^N| \leq C\ell \Delta x \max_i \int_0^T \left|\frac{du^N_i}{dt} \right| \,dt + o(1)
\leq C' \Delta x +o(1) = o(1).
\]