Вы находитесь на странице: 1из 19

COMMUNICATIONS ON PURE AND APPLIED MATHEMATICS, VOL XVIII, 697-7 15 (1965)

Solutions in the Large for Nonlinear


Hyperbolic Systems of Equations*
JAMES GLIMM

1. Introduction
We consider the quasilinear system of equations

Here - 00 < x < + co,t 2 0, and u and f are vector-valued functions. The
function
(1.2) u ( x , t ) = const.

is always a solution of (1.1). We assume that (1.1) is hyperbolic and thatf is a


strictly nonlinear function. Our main result, Theorem 1.1, is an existence theorem
for solutions of (1.1) with prescribed Cauchy data near the constant data u(x, 0) =
const. The solution is defined for all t 2 0 and all x and is near the corresponding
constant solution (1.2). This answers in a fairly satisfactory manner the existence
portion of a problem posed by Lax [4], p. 6.
The solutions we obtain are weak. They are not differentiable or continuous
in general. Even for the simplest equations and initial data, the nonlinear nature
off forces the characteristic speeds to differ at differing x and t in such a way that
a continuous solution is in general overdetermined by its initial data and does
not exist for all t 2 0. However, a smooth solution will exist in some strip
0 5 t 2 T,even without restricting the (smooth) initial data to be near a constant,
see [l]. For t 2 T there will in general be curves in the x , t-plane across which
the solution has jump discontinuities (shocks). Presumably nothing worse can
occur. We prove the weaker statement that there is a bound on the total variation
in x (for fixed t ) of the solution v, and that the bound is independent o f t . In
particular, v is a measurable function. Apparently the concept of solutions weaker
than this (e.g., distributions) has no meaning becausefis not linear.
Our solutions are limits of difference approximations ; the difference approxi-
mation involves a random choice. This can be illustrated roughly as follows.
Suppose we have determined the difference approximation u at the points

* This work was supported in part by the National Science Foundation under Grant NSF-GP
2600, and in part by the ,4ir Force Office of Scientific Research grant AF-OSR-684-64.
Reproduction in whole or in part is permitted for any purpose of the United States Government.
697
698 JAMES GLIMM

xo - h, to and xo + h, t o . We take

u1 = u(xo - h, to) for x < xo ,


u2 = u(xo + h, to) for x > xo ,
as Cauchy data on the line t = t o . The equation (1.1) has a (weak) solution u in
some strip to 5 t 5 to +
Oh with this Cauchy data. We define
u(xo, t + eh) = u(a, t + eh) ,
where a is chosen at random in the interval [xo - h, xo +
h]. A similar differ
ence scheme, but lacking this probabilistic feature, has been considered by
Godunov [2].
Our proof requires two series of estimates, one for v and one for the first
derivatives of u. For the estimates involving u we use the sup norm on v ; for those
involving the first derivatives we use the total variation norm on v. We define as
the distance between the initial data u ( - , 0) and a constant 5 either
do = Ilu(., 0) - 511, (1 + total variation v(*, 0)) ,
or
dl = IIu(-, 0) - 511, + (total variation u ( - , 0)) .
We must assume that one of these distances do or dl is small. Assuming do is small
is less restrictive and this will be permitted only for a restricted class of equations,
The restricted class includes all 2 x 2 systems under consideration. I n particular,
it includes the equations of gas dynamics (for constant entropy and with one space
variable).
We set A(u) = af(v)/au. Then (1.1) can be written as

(1.3) vt + A(v)v, = 0 .
Let N be the order of the system. In saying that (1.1) or (1.3) is hyperbolic we
mean that, for each value of v, the matrix A(u) has N real and distinct eigenvalues
A, = L l ( V ) < 12 < * a ' <
Let r j = rj(v) and lj = l,(v) be the right and left eigenvectors of A for the eigen-
value A j . Following Lax [3], we assume that the system (1.1) or (1.3) is strictly
nonlinear in the sense that
T i VL, # 0. -
We say that a local coordinate system w l , - * , w N defined in an open set in RN
consists of Riemann invariants if
ri Vwj = 0 , i#j.
In general such a coordinate system will not exist; but, if N = 2, then it will
necessarily exist. We say (1,l) is smooth iff is smooth.
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 699

THEOREM 1.1, Let the equation (1+ 1) be hyperbolic, strictly nonlinear and smooth in a
neighborhood of v. There is a K < OC) and a 6 > 0 with thefollowing properg.
If the initial data v(x, 0 ) are given so that dl 6, then there is a weak solution v(x, t )
of (1.1) dejined for all x and all t 2 0 with initial data v(x, 0 ) such that
(1.4) IIV - vllm s K Il4.,0) - vll, Y

(1.5) total variation u ( - , t ) K (total variation v ( * , 0 ) ) , t z 0 ,

(1.6) /mmlu(x, tl) - u(x, t2)1 dx 5 K Itl - t,l (total variation v(-, 0)) .

- -
If there is a coordinate system w1 , * , wN deJined in a neighborhood of v which consists of
Riemann invariants (for example, ;f N = 2), and ifthe initial data v(x, 0 ) are given so that
do 5 S, then there is a weak solution v(x, t ) of (1.1) dejned for all x and all t 2 0 with
initial data u ( x , 0 ) such that (1.4) and (1.5) hold.
That v is a weak solution with initial data v(x, 0) means that v is a bounded

c
measurable function and that

[J-mm(6tu + 6 z f ( v ) ) dA! dt + +(x, O ) V ( X , 0) dx = 0

for all smooth functions 6 which are identically zero outside some bounded set.
The remaining sections of this paper are devoted to the proof of this theorem.
By simple examples it can be shown that the weak solutions of (1.1) are not
unique in general. This is due to the existence of extraneous solutions of (1.1)
which are not stable (with respect to small perturbations of initial data) and which
contain unnecessary discontinuities not of shock type. We conjecture that our
solution u of Theorem 1.1 is not of the extraneous type, [4], [5]. For general
background and for the more fully developed theory of a single equation we refer
the reader to [3], [4], [6] and to the references quoted there. The author is
grateful to P. Lax for several stimulating discussions.

2. Bounds for Solutions of the Riemann Problem


The eigenvector r j has a unique direction but its length is arbitrary. Near v
we choose its length in a smooth manner so that r , V2, = 1. Each rk defines a
vector field R, ,
Rk = r k * V .
The integrals of R, (i.e., w for which rk V w = 0 ) are called k-Riemann invariants
of (1.1). There are N - 1 independent Riemann invariants defined in a neighbor-
hood of 5; the integral curves of R, in this neighborhood are exactly the curves
along which all k-Riemann invariants are constant.
If vL and v, are any two N-vectors near v, then (1.1) has a weak solution v
with Cauchy data
VJ, x<o,
v(x, 0) =
vr, x>o.
700 JAMES GLIMM

This problem is called a Riemann problem and its solution u exists by Theorem
9.1 of [3]. Furthermore, u takes on constant values uo = uz , * * * ,u j , * ,vN = -
u7 in sectors * * , X i = { x , t : t i t < x <
Eit>, * * * . The sectors Xi-l , Xj are
consecutive. Also by [3] we have two possibilities for u in the sector between
Xj-l and Xj :
(aj) XiPl and clj have a common boundary x = f j t , also 3ij(u(tEj - 0, t ) ) >
t j > Aj(4tEj +0, t ) ) ,
(bi) between CjP1 and X I , A j ( v ( t t , 6)) = [ and all j-Riemann invariants
are constant.
I n both cases u is a function of x l t . I n case (bJ, u is continuous on and between
and X I ; in case (aj) there is a jump discontinuity. This jump discontinuity
is called a j-shock wave. The configuration (bj) is called a j-rarefaction wave,
while aj-wave refers to either. The u j are called intermediate states. They are
near v and are C2 functions of v1 and uT , see [3], p. 563. Third derivatives are
piecewise continuous (with a jump at the transitions from (ak) to (bk)).We call u
the resolution of the discontinuity v l , u, into j-waves, 1 j 2 N. We choose s
functions wi defined near v such that

(2.1) R 1. w3. = r j * V w I = 1,
and we let the quantity
El = Wj(Uj) - wi(uj-l)
denote the magnitude of thej-wave in u. I n the case (bj) of rarefaction waves the
range of u restricted to the sector between Xi-1 and X, lies on a n integral curve of
R j , and c j represents the change of parameter along this curve. Along the curve,
Ai(u) increases. Since sign RjAj = sign R j w j , wj(u) increases and c j 2 0. I n
case (aj), let u: be the point on the integral curve of Rithrough uj-l with
Wj(U?) = Wj(Vj) .
+
Then u; = u p O(E;); see [3], p. 563. We suppose uz and u, are close to each
other. Then E~ is small, and
sign [Aj(vj) - Aj(vj-l)] = sign [Aj(u!) - Aj(vj-l)] = sign c j .
Since 3Lj(uj) - A,(vjPl) < 0, we have ci < 0 also. Hold uip1 fixed and consider
u j as a function of E ~ The
. tangent dvj/daj to this curve is

rj(uj) if ci 20,
ri(uj) + O(E;) if cj <0 ,
(see [3]), and the derivative of the tangent is

-
R j ( r i ( v j ) )= ( r j V ) r j if E~ 2 0,
Ri(Tj(Uj)) +O ( 4 if cj <0.
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 70 1

By Taylor's expansion about uj-, , we have

with all the Y evaluated at vl and with E = cl, * , EN -- -


Now suppose that urn is some other state near v" and that u' (respectively u") is
the resolution of the discontinuity u l , urn (respectively urn, u,,) into j-waves with
intermediate states ui (respectively u"). Let thej-wave in u r (respectively v") have
magnitude y i (respectively Sj). Let y = y1 , - -
,y, and 6 = 8, , * * , 8 , . -
Applying (2.2,) to the case u", we have

with all the Y evaluated at u, . T o analyze u r we proceed in the order of de-


creasing j:
ui-l = ui - yjY,(vj) + 2-'y;(rj V)r, + o(y;).
By an induction we obtain
(2.4) u1 = v; = - 2 yiyi
i
+ 2 yhyi('h
hzi
' v)ri(l - 2-16hi) + 0(i6i3) Y

with all the r evaluated at u, .


We hold urn fixed and regard E as a function of y
and 6. Then E has continuous second derivatives and piecewise continuous third
derivatives. Since E = 0 if y = 0 = 8, E = O(lyl 181). Since +
(2.5) rf(ul) = yi(um)- 2h yh('h ' V)ri + o(lr12) 9

we get from (2.2,)


'7 - ' 1 = 2: Eirf(U,)
i
+ O([lvl + ls1l2) *

If we compare this to (2.3) and (2.4), we obtain

(2.6) Ei = Yf + 6, + Wlrl + l~Il2) '

We substitute (2.6) and (2.5) in (2.2,) and get


u, - uz =
1
EiTi + hSi
EhEi(Yh V)Yi(i - 2-'6,,)

1 if h = i,
dh, =
0 ifh # i.
702 JAMES GLIMM

with all the r evaluated at u, . Comparing this to (2.3) and ( 2 . 4 ) , we get

Z:i ('i - ~i - 'i)ri = I:C - ~ - ' ( Y , + 6i)2 +


i
(ri + 6i)yi
+ 2-16: - 2-17,"] (Ti * V) Ti

2 ( E g - Yi -
i
Si)Ri = 2 7ishLRi
h<i
2 Rhl + o([lrl + lS1l3) '

I n the case where there is a coordinate system w i near v" with R j w k = 0 if


j # k , we take ( 2 . 1 ) as a determination of the length of Ri (rather than a restric-
tion on wj).In this coordinate system R , = a / a w , and [Ri, Rh] = 0 ; thus

(2.7) Ei = yi + + O([lYl + ldIl3)


'i *

We say that an i-wave from u' and a j-wave from u" are approaching if

i>j
or if
i=j and yi < O or di < O .
Let
D(Y, 4 = Z: IYiI 1691 Y
where the sum is over all pairs i a n d j for which the i-wave from u' and thej-wave
from U" are approaching. The principal result of this section is the following
refinement of (2.6) and (2.7).
2.1.
THEOREM We have
(2.8) Ei = yi + di + a y , 6)0(1)
as IyI+ 161 -+0 and ;f ( 2 . 7 ) holds we have
(2.9) ~i = + + D ( Y >6)O(lyl + 161)
~i di a

Proof: First we observe that the theorem is true in the case D = 0. In fact
by the results of [ 3 ] , the waves in u' and U" can be fitted together exactly to produce
a solution of the desired type, which by uniqueness must be v. Next we prove
the theorem with D replaced by ( y ( (61 . Fix i, y , and 6 and let
F ( C , 7) = E i ( O Y , 7 8 ) - CYi - ~ 6 ~ .
Then F(0, T) = 0 = F( G, 0 ) ,F is C2and is C3where G # 0 # T, and the bounds on
F and its derivatives are uniform in y and 6, IyI 161 0. We have F,(a, 0) = 0, + --f
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 703

thus T-'F,,(c, T) is bounded and continuous, uniformly in y and S. I n the same


way ~ - ' F ( c ,T ) is bounded and continuous. Since T-'F(O, T ) = 0, it follows that
F = a d (1) uniformly in y and 6. Therefore,
(2.10) = yi + 6, + Iyl 181 O(1).
Let C ( y , 6) be the O(1) in (2.10). Then G is continuous and has piecewise
continuous and bounded first derivatives. If (2.7) holds we must have G(0,O) = 0.
+
However this implies G = 0 ( ( y l IS()so that
(2.11) ~i = yi + 6i + IYI 161 O ( l ~ +
l ISl)
as asserted. The 0 in (2.10) and (2.1 1) is uniform in urn for urn near 5.
We suppose 0 = d,,, = * . . = 6, . 'To complete the proof we use induction
on p . For p = 0 the theorem is true; we suppose it true for p - 1,0 p - 1 5
N - 1. Let
a = s i , . . . 6,-1 ,o, . * * , 0 ,
9

p i = 0 , vi = E,(Y, A) if the i-wave in u' and the p-wave in v" approach,


pa= ei(y, A), vi = 0 otherwise.
We want to consider the interaction of the waves with magnitude v j , and the
p-wave from u". In order to do this we must take vg-' in the role of t m .We
indicate the dependence of E on v, by writing
Ei = E ~ ( Y >6, urn) .
The result of this interaction is the solution of a Riemann problem: it consists of
j-waves with magnitudes n j ,where

71j = T T i ( Y ,* - - ,0, 6,, 0, * - , u;-l) .


Next we consider the interaction of the waves with magnitudes p and n. The
result of this interaction is u. Hence if we let trndenote the intermediate state
between the " p-waves" and the "n-waves", we have

(2.12) 'i(Y, ', 'm) = 'I(pFC, T J .


Now D ( y , A) 5 D ( y , d), and by our induction hypothesis we have
E ~ ( YA,
, Urn) + Si, 6, = yi + 6, + D ( Y >6)O(1) .
By (2.10) we have
= vi + 6i, 6, + I6,l 0(1) IYI *

For i < p , lvil (6,I = 0. For i = p , v, = 0 or else


y, =Yp + D ( y , 4 O(1) 9

Iv,I 16,l 5 D ( Y , 6)0(1) a


7 04 JAMES GLIMM

None of the ",u-waves" are approaching a "+-wave"; hence we have

(2.13)
Ei(,u, +, f l m ) = pi + +i

= yi + di + D ( y , 6)0(1) .
If (2.7) holds, we obtain
(2.14) d , u , +, flm) = yi + + W y , 4 0 ( l y l + 14) -
di

I t follows from (2.10) that

lEZ(P, r 7 flm) -' Ei(,u, +, %)I 6 In - +I O(1)


s [WB O1( l y)l + 161)
7

if (2.7) holds.
This combined with (2.12)-(2.14) completes the proof.
We suppose without loss of generality that the coordinates wi have the property

We take

as a metric in R" near 5. We assume that w,(fl) = 0 and we define IvI =


supi Iwi(u)I.This is a norm with respect to the linear structure defined by the
coordinates wi but it is not a norm in the original coordinates in general. For
intermediate states uk--l and vk of the resolution of v z ,v, iritoj-waves we have
(2.15) wi(vk,> - wi(vk--l) = &k('jk + O(lVZl + Ivrl)) 9

and
Wj(V,) - Wj(VZ> = Ej + IEl O(lV,l + 1q.l) .
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 705

If the w j are Riemann invariants for Rk , k # j , so that Rkwj E J j k , we have


(2.16)

3. The Difference Scheme


Our difference scheme depends upon a random choice. Given a bounded
neighborhood U, of the state b, we can find neighborhoods U, c U, c Ul such
that for any two states u1 and u, in U, there is a unique resolution of u l , u, into
-
j-waves with intermediate states u1 , * * ,uAVplin U, . We choose mesh lengths r
and s such that
sup { l i l j ( u ) l : u E U,, 1 5 j 2 N } < r / s .
We hold the ratio r / s fixed, so s is a function of r. Let
Y = ((n,
n) : m,n integers, m + n is even and n 2 0} ,
(3.1) A= fl
lm,n)P
[(m- l)r, (m + 1)r x (m)] ;

each factor is a horizontal line segment in the plane. Give each factor the measure
1/2r times Lebesgue measure and give A the product measure, denoted by du.
We choose a point a = {a,,,} E A . This is our random choice and the points
are our mesh points.
Suppose that our difference approximation u = u(x, t ) has been defined for
x , t = um-l,n-l and x , t = am+l,n--l and that these values of u lie in U, . Let u
be the solution of equation (1. I ) , with range u E U3 and with initial values

which is a resolution of u(u,-l,n-l), U ( U , + ~ , ~ - ~ )into j-waves. Let u(a,,,) =


u(u,,J. This defines our difference scheme. It is convenient to set
u ( x , t ) = u(x, t ) , ( m - 1)'sx s (m+ 1)r, ( n - 1)s 2 t < ns.
Then u is a solution in this rectangle. For x near the boundary ( m - 1 ) r (respec-
+
tively ( m 1)r) we have u ( x , t ) = U ( U ~ - ~ , % - ~(respectivelyu(x,
) t) =~(a,+~,~-~)),
because of the restriction on rls. Thus ifu is defined in a strip (n - 1)s t < ns,
- co < x < +a,then u is constant across the lines x = ( m - l ) r , m n even, +
where its mode of definition changes and u is a solution in the strip.
We cannot show in a simple and direct fashion that the difference approxi-
mation u can be defined for all x and t . Instead, this will be proved simultaneously
with the proof of the bounds on u, and requires that the initial data be a small
perturbation from the constant state 5.
706 JAMES GLIMM

I n place of horizontal lines, it is convenient to consider curves consisting of


line segments of the form Lm,n,m+l, n+l , Lm,n,m+l, n-l joining am, to a,,,+l, n+l and
joining a,,,, ll. to am+,, ,respectively. If the mesh index rrz increases monotonically
on such a curve, we call it an I curve. We can partially order the I curves by
saying that the larger curves lie toward larger time. Let 0 denote the unique I
curve passing through the mesh points on t = 0 and t = s. We remark that these
I curves could be replaced by a simpler class of curves. For example, the poly-
gons joining six points of the form
(- a,12s + 0) 7 am.n 9 anz+l,n+l?

%a+Zk+l, n+l 9 am+zk+Z. n , ( + 00, ns + 0)


would suffice if the subsequent treatment were modified slightly.

4. Bounds for the Difference Equation


We obtain our bounds on the difference solution u by means of functionals.
I
These functionals F = F(u J ) = F ( J ) are functionals of the restriction of u to an
I curve J , and they are monotone decreasing as functions of J . According to
I
Section 3, u J consists of various shock and rarefaction waves, and the functionals
F will be expressed as functions of these waves. Let u and /? be two waves in u
which cross J ; suppose u is aj-wave and B is a k-wave.. W h e n j # k , we say u
and ? !, approach (each other) if the wave belonging to the faster family (larger index)
lies to the left on J (toward x = - a).W h e n j = k, we say u and 8 , approach if
ct+ and if they are not both rarefaction waves. Let 1x1 be the absolute value of
the magnitude of a as defined in Section 2. Let
Q ( J )= 2 1.( I,!lI : a and ,!l cross J and approach},
L ( J ) = 2 {lul: ct crosses J } .
Let amSnbe a mesh point on J. We also define

i m ( J k 2 IMI ;
here the summation is over allj-waves cc crossing J , j # i, which i f j > i (respec-
tivelyj < i) lie to the left (respectively right) of u , , , , ~on J . In other words, the
sum is over j-waves, j # i, which approach a (nonexistent) i-wave a t . Let
+ KoQ(J)
Fl(J) = L(J) 9

F2(J, i, m) = l~i(~(am,m))l+ & i m ( J ) + Bo&Q(J) ,


F 2 ( J )= sup { F ( J ,i, m) : 1 5 i 5 N , m is an integer}.
Here So and KOare positive constants to be chosen later.
Fl is equivalent to the total variation norm, and will be used to obtain the
estimates (4.4). F2 dominates the sup norm deviation of u from the constant g,
and will be used in proving (4.3).
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 707

THEOREM
4.1. (A) There are positive constants 6, and K3 such that ;f
(4.1) IlU(*, 0 ) - 4 , s 6, ,
(4.2) total var. u ( - , 0 ) s 6,,
then the diference solution u can be dejined f o r all x and all t 2 0. Furthermore,
(4.3) Ilu(-, .) - 4 l c o s K3 Ilu(*, 0) - 4 l m >
(4.4) total var. u ( * , t ) K3 total var. u ( * , 0 ) ,
(4.5) / l u ( x , t,) - u ( x , t2)I dx 2 K3(lt2 - t,l + 4s) total var. u ( . , 0 ) .
(B) Suppose the w j are Riernann invariants. Then (4.2) may be replaced by
(4.6) (IEc(*, 0 ) - C((, total var. u(-,0 ) 6,. s
Proof I t is sufficient to prove the theorem using the linear structure of the
coordinates wi , I n this case 5 is the origin and
114-> -1 - 511, - 114.3 -1 IICO = SUPizt Iwa(u(x, 011 .
We choose neighborhoods U, and U, of 77, U, 3 U, 3 U, , so that if the v' and U"
of Section 2 have their range in U, , then the bounds of Section 2 hold with a
constant Kl , K , 2 1, and also

We further require that if uL and v, are in U, , then all intermediate states lie in
U, . We suppose U, contains a sphere with center 5. Let 6, be its radius. Let
6, = min (2-12N-2K-1 1 ,2-361}.

We choose later our 6, in (0, d2). Assume (4.1) and (4.2) or (4.6) with 6, replaced
by 6,. Let
= 4NK1, 60 = 26K1 I\U(', o)(I, .
We can define u up to the I curve 0 and, for x , t between the curve 0 and the line
t = 0, we have
IU(X, 91 5 2 Ilu(*, 0) Ilm .
With the help of (4.7) and (4.8) we get
L ( 0 ) 2 N total var. u ( * , 0) ,
+
F,(O) 5 L ( 0 ) K,L(0)2 5 2 L ( O )
-
5 4N total var. u(*,0) ,
6,L,,(O) + 6,K,Q(O) d04N total var. u ( * , 0)
= 2sNK, total var. u(*,0) l\u(., 0)
s llu(., 0) Ilm 9

FdO) s 3 0)IIm -
llU(*,
708 JAMES GLIMM

Suppose that u has been defined up to some I curve J1 ,, that u is in U, below J1


and that
F1(J1) 5 4N6, J

F2(J1) f 3 8 , .
This is true if we take J1 to be the curve 0. Let J2 be a n I curve which is a n
immediate successor to J 1 . Thus J2 and J1 pass through the same mesh points
except two. In fact for some (m,n) E Y we have

am.n EJ I N J2

-
9

am,n+2 E J2 J1

Now u can be defined up to J , and is in U, below J , . We shall show that Fi(J2)


Fi(J1). Then for x , t on J2 ,
Iu(x, 01 s 2F2(J2) 5 8, *

This implies that u is in U, below J, . An induction on J would then prove that


u can be defined for all x and t 2 0 and that

114. -1Ilm 5 2F2(0) 5 6 Ilu(., 0)Ilm ,


Y

total var. u ( * , t) 5 F l ( 0 ) 2 4N total var. u ( - , 0) .

We show that Fl(J2) 5 Fl(J1). Let E~ denote thej-wave in u crossing J2 N J1


and let y j (respectively Sj) denote the j-wave in u, if any, crossing J1 J2 left
(respectively right) of . Let
-
D = 2 IYjI IhI 7

where the sum is over the approaching ys and 6s. This agrees with the termi-
nology of Section 2, and by (2.8) we have

l ~ jSl lyjl + 1891 + 3

(4.9)
L(J2) 5 L(J1) + NKlD.
Let u be a k-wave which crosses J1 n J , and approaches eg . If k # j or if u is a
shock wave, then tl approaches y j and dj and we use the estimate (4.9). If 01 is a
j-rarefaction wave, then c j is a shock wave and cc approaches y i and S, if they are
shocks. Let ly;l = lyjl if y j is a shock wave, let it be zero otherwise; we define
(ST(similarly. By (2.8) we have in this case
lEll 5 lvil + lS,l + KID >

and so we see that if 1 . 1 ~ E J occurs in Q(J2),then it is dominated by (terms in


Q(J1)) + 1x1 KID.
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 709

Since D occurs in Q(J1)but not in Q(J2),we have


Q(JJ 4 Q ( J i ) + NL(Ji)KiD - D 3

Fi(J2) Fi(J1) + (NK1D + 4N2 d,KoIC,D - KoD)


5 FI(J1) + K0D(2t2 + 2-2 - 1)
5 Fl(J1) - 2t1K0D5 Fl(J1).
s
We show that F,(J,) F,(Jl). Consider an m' # m. Evidently e j lies on the
same side of (with respect to J,) as both y , and 6, do (with respect to J 1 ) .
Thus l e j l occurs in Lim,(J2)if and only if I y j l +
16J occurs in Lirn.(J1).Using
(4.9) as above, we see that
Liwt(Jd + KoQ(JZ) 5 Lirn*(Ji) + KOQ(J1)
and
(4.10) F,(J, ,i, m') 5 F,(Jl, i, m') .
Now we consider F,(J, , i, m). First suppose that F( crosses J , on both sides of
(and so ci is a rarefaction wave). Let b, (respectively b,) be the left (re-
spectively right) edge of ti at the point of crossing J , . Since w ichanges monotoni-
cally along J , between b, and b, we have
(4.11) l~i(~(~rn,n+~))l5 lwi(u(bl))l
for either 1 = 1 or 1 = 2. Let c = b, for this choice ofl. If ci crosses J , on one
side of only, let c = .
+
We compare F,(J, , i, m ) to F 2 ( J 1 i,, m a) where u = f l . We choose
G = + 1 if ei crosses on both sides of am,n+2 and if 1 = 2, or if E~ crosses to the left
of am,n+2 ; we choose a = - 1 otherwise. Let uJ, denote the restriction of u to
J i . Then
IIUJ,(' > -)llm s *)IL5 5 9 2 ( 0 )
2 lbJ1(-,
- 24 114.> 0 ) I l m
5
Using this and (2.15) we see that
(4.12) Iwi(u(c))l ~wi(uturn+u,n+l))~ + 25Kl Ilu(*, o)Ilca 2 IEkl

and
2 lEkl 5 2 ( I r k ! + Idkl) + NKID >

where the summation is over all .sk crossing J , between c and n+l . For such a
k, E, is not approaching an i-wave at urn,n+a and so Iek[ does not occur in L,,(J,).
However yk and 6, are approaching a (nonexistent) i-wave at um+a,n+l and so
the terms lykl + Idk]occur in Lirn+,,(J1).By (4.9) we have
Lim(J2) S Lim+u(JJ + NKlD - 2 (Irk1 + IN) >
where the range of summation is as before. I t follows that
Lirn(J2) + KoQ(Jz) 2 Lim+n(Jl)+ KoQ(J1) - 2 (Irk1 IdA) - 2 - 1 K ~ D-
710 JAMES GLIMM

Combining this with (4.11) and (4.12), we have

From the definition of 6, it follows that

This shows that u can be defined by induction for all x and all t 2 0 and that (4.3)
and (4.4) are satisfied.
If the w j are Riemann invariants, we change our definition of KOand 6, . Let

We use (4.6)-(4.8) and obtain as before

We suppose inductively that u is defined and in U, below J1 and that

As before, we need only show that Fiis nonincreasing. We replace (4.9) by

and we obtain
~ . - . . - . -
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 71 1

We consider F, . Formula (4.10) follows as before. We replace (4.12) and


(4.13) by
IWi(4 5 l~i(~(~m+#,n+l))l + 21K, Ilu(., O)I/2,2 lEkl
and
F,(J, > i, 4 5 FZ(J1, i, m +4
+ 2OKl Ilu(., O)Il2,Z l%l - 2 + 60 (Irk1 Idkl) - 2-, 6oKoD
5 U J , , i, m + a) + NDK,(~OK~ IM., O)II: - 23 so)
+ z: + 14cl)(210KlIlu(., 0) :1 -
(Irk1 60) *

I t follows that, if (4.6) holds, p can be defined for all x and t 2 0 when the wi are
Riemann invariants, and in this case (4.3) and (4.4) are satisfied.
We prove (4.5). Choose 6, < Y4N-l 6, and so that the estimates (4.3) and
(4.4) remain valid uniformly in v if the v of Theorem 4.1 is allowed to vary in a
disk of radius 2,N 6, . This is possible because Kl and U, can be chosen uniformly
in G as 7 varies in a small neighborhood. We assume (4.1) and (4.2) or (4.6)
without replacing 6, by 6, now. We suppose without loss of generality that
t, > t , . Let
to = sup { t : t 5 t, , t = m>,

S = [(t, - to)/s] + 1.
Then SS 5 t, - t, + 2s. For any x , u(x, t i ) is determined entirely by the Cauchy
data
U(Y>to) )
y E [x - Sr, x + Sr] .
This Cauchy data satisfy (4.1) and (4.2) or (4.6) with 6, replaced by 6, and 5
replaced by u ( x , to). By the part of the theorem already proved, we get
Iu(x, t,) - u ( x , t,)I 5 K, sup {lu(y, to) - u ( x , t0)I: x - Sr 5 y 5 x + Sr)
for some constant K , . Hence
J u ( x ,t,) - u(x, tl)l 5 K, I
total var. (u(*,to) [ x - Sr, x + Sr]) ,
1
where u B is the restriction of u to the set B. Thus the integral of Iu(x, t,) -
u ( x , tl)I over x in one mesh interval [mr, m +
2 r ] in the line t = to is bounded by
I
2rK, total var. ( u ( - , to) [(m- S)r, (m + 2 + S ) r ] ).
It follows that
m
[
J-m
lu(x, t,) - u ( x , t,)l dx 5 2rK2(2S + 2) total var. u ( * , to)
+
- 8N(r/s)K2(2S 2)s total var. u ( - , 0)
I
2 &(It2 - t,] + 4s) total var. u ( * , 0) .
This completes the proof of Theorem 4.1.
712 JAMES GLIMM

5. Convergence of the Approximate Solutions


Let y ( x ) be the initial value of our desired solution u of (1.1). We assume y
is bounded and has bounded total variation. For each mesh length r let y , be
defined by the equation
y,(x) = d m r ) , (m - 1)' 2 x < (m + 1 ) r , m even.
Then yT is bounded and has bounded total variation with the same bounds as y
and y , + y at points of continuity of y. Hence y , -+yl a.e. and y , + y in I,,
on bounded intervals. Let 5 = y ( 0 ); 6, is the 6 of Theorem 1.1. We assume
the hypothesis of Theorem 1.1 with this 6 and we conclude from Theorem 4.1
that u is defined for all t 2 0. The difference scheme, and thus u, depend on the
mesh length r and the choice of a in A . We express this by writing u = u,, . Let
4 be a C2 function with compact support, and let
b ( r > a>4) = rJIrn(+tUra + +zf(ura)) dx dt + J m --m +(%, o)yr(x) dx .
Since u is a (weak) solution of the equation in each horizontal strip ns <t <
+
( n l)s, we can compute
m o o

(5.1) b ( r , a, 2
4) = n=1 + ( x , n s)(u (x,ns) - u ( x , ns - 0 ) ) dx .

Our measure space A depends on r ; however there is a n obvious isomorphism


A = Tl-[O, 11
of A with an infinite product of copies oftheinterval [0, 11given by theisomorphism
am, + 2-'( (first component am, J r - I - m + 1)
in each factor. By means of this isomorphism we regard the functions b as defined
on a fixed probability space. In the next lemma we choose a null set N in A ,
independently of r. This means that, under the above isomorphism, N is mapped
onto a null set in II[O, 11 which is independent of r. This result can be regarded
as the weak law of large numbers in the present context.

a
LEMMA
EA -5.1. There is a null set N c A and a sequence ri --f 0 such that for any
N a n d any test function 4, we have b(r, , a, 4) -+ 0.
Proof: Let b ( r , a, p, n ) denote a summand in (5.1). We regard this summand
as a function of a (i.e., a random variable). We assert that, for any continuous p,

(5.2) llb(r ,a, p, n ) 11 5 Kr llpllm(total variation y).


Now suppose p has compact support and is piecewise linear and is linear in each
of the triangles bounded by

((m - l ) r , ns) , ((m + I)r, n s ) , (mr, ( n + 1)s) , m + n even.


NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 713

We consider from now on only r of the form r = 2-k. I n this case our piecewise
linear as has the required form for all smaller z also, and we assert further that

b(r, .>as, 4 1w,.>91 n2) z 12


(5.3)
(5.4) IlNr, as) 1:
*>

The inequality (5.2) follows from the inequality


- 0
9 n1 1

- r Ijp/lm(total variation
5 u(*,m - 0))

and from (4.4).The orthogonality in (5.3) is with respect to &(A). To prove this
suppose nl < n2 and let 2,dd be the measure space product with a factor corre-
sponding to m2,n2 omitted. Let
u ( x , n) = u ( x , ns) - u ( x , ns - 0) .
The inner product of b(r, p, ril) and b(r, *, q ~ n2)
1, , is a sum of terms of the form

However p is constant on [(m2- l ) r , (m2 + 1)r] X {ng)and

c =C V ( X nls) 4%
nl) dx

is independent of amBn,. Hence (5.5) is equal to

By our definitions,

=jja(amBnz, ns - 0) - u ( x , ns - 0 ) dx durnln2
=0 ,

which proves (5.3). Also


714 JAMES GLIMM

Since p has compact support, there are O(r-l) nonzero terms in this sum. Thus
(5.6) follows from (5.2). I n the same manner we obtain
(5.7) llb(r, -,V)II 5 const. llpllm .
For each piecewise linear # there is a sequence r, -+ 0 such that b ( r , , -+ 0 a.e. a, 4)
By the diagonal process we can achieve this for a sequence dl, # 2 , * * of the
4. We choose the sequence to be dense; let N be the null subset of A such that
b(ri,-,+J-+O on ANN as i+ to, k = 1,2,.--.
We apply (5.2) with q5 replaced by # - #k and conclude that b ( r , , *, (5) 0 --f

on A N N for any 4. This proves Lemma 5.1.


For any a in A N N let u i = u,~, . By Theorem 4.1, ui is bounded and has
bounded variation on horizontal lines, uniformly in i. By Helly's theorem a
subsequence of the ui converges in L, on bounded intervals of any given horizontal
line. By the diagonal process we can achieve the same result for a countable
number of horizontal lines, located at (say) rational times t = p/q. For a n arbi-
trary time t, we have (by (4.5))

l U " ~ , P / d - .j(.,P/dI dx +Jx1dM I+,P/d - Ui(% 41 dx


- const.
5 (It - ( p / q )I + Ti + rj) +J x l 5 M lu"%P/d - U j ( % P / d l dx *

Thus the subsequence ui(j)converges (uniformly for bounded t ) on the intervals


1x1 M of any horizontal line.2 Therefore,
limf(ui(j)) =f( v ) (in L, on bounded sets)
j
and so

/J=:m#tu + 4,f(v> dx dt +s_mn+(% O)W(") dx


= lim b ) , =0 .
' ( ~ ~ (a,~ #)
j-0

Hence v is a weak solution of equation (1.1). The bounds on u in Theorem 1.1


follow from those on u given by Theorem 4.1.
We remark that since f is nonlinear, it need not be continuous with respect
to weak limits. Thus the technique from linear equations in which one first takes
a weak L, limit of approximate solutions and then proves that the limit has the
required derivatives does not seem to be applicable here. I n fact, it appears that
such a weak limit need not be a weak solution. Our approximate solutions
converge in the strong L, topology on bounded sets.
This compactness argument was used by Oleinik [ 6 ] . I am grateful to her for finding an error
in my original proof at this point.
NONLINEAR HYPERBOLIC SYSTEMS OF EQUATIONS 715

Bibliography
[ 11 Friedrichs, K. O., Nonlinear hyperbolic dzyerential equations f o r functions of two independent variables,
Amer. J. Math., Vol. 70, 1948, pp. 555-589.
[2] Godunov, S. K., Bounds on the discrepancy of approximate solutions constructed f o r the equations of gas
dynamics, J . Computational Math. and Math. Phys., Vol. 1, 1961, pp. 623-637. (In Russian.)
[3] Lax, P. D., Hyperbolic systems of conservation laws, ZZ, Comm. Pure Appl. Math., Vol. 10, 1957,
pp. 537-566.
[4] Lax, P. D., Nonlinear hyperbolic systems of conservation laws, in Nonlinear Problemr, University of Wis-
consin Press, Madison, 1963, pp. 1-12.
[5] Oleinik, 0. A., On the uniqueness of the generalized solution of the Cauchy j r o b l m for a nonlinear
system of equations occurring in mechanics, Uspehi Mat. Nauk, Vol. 12, 1957, pp. 169-176. (In
Russian.)
[6] Oleinik, 0.A., Discontinuous solutions of nonlinear dtyerential equations, Uspehi Mat. Nauk, Vol. 12,
1957, pp. 3-73. (In Russian.)
[7] Glimm, J. and Lax, P. D., Decay of solutions of systems of hyperbolic conservation laws, to appear.

Received March, 1965.

Вам также может понравиться