Lax 1953
Lax 1953
Introduction
The principal result derived in the first part of this paper is an extension of
the Cauchy-Kowalewski theorem. I show that, for quasi-linear hyperbolic
systems of partial differential equations in two independent variables, the
analytic initial value problem has a solution which not only exists in the small,
but can be continued analytically until it reaches the boundary of the domain
of analyticity of the equations. The proof of this theorem is based on an a
priori estimate for solutions of hyperbolic equations. This estimate is only
slightly sharper than that used by Haar, [8], [lo],an important tool in the study
of hyperbolic equations in two variables, see e.g. Hadamard [ll], Schauder
[15], and Friedrichs [6].
In Section 5 of this paper the above theorem is used to prove the existence
of solutions of nomaulytic initial value problems (under various assumptions
of differentiability) by approximating the nonanalytic problem by a sequence
of analytic problems. Such a process has been used successfully to prove that
initial value problems for ordinary differential equations have a solution; I
shall sketch here this proof for ordinary differential equations because it is a
model of the proof given in section 5 for partial differential equations.
Given a system of ordinary differential equations
%
l = f,(y, , . . . , y,), i = I, 2, ,n
where the are analytic functions of all their arguments in some open set D.
f 1
If one of the endpoints of the interval (a,, , a,) is a point of singularity for one
) ) < t < a , ) has a point
of the functions y,(x), then the pointset ( y , ( t ) , . . , ~ , ~ ( t (a,
of accumulation on the boundary of D.
In particular, this corollary will enahle us to show that for large k , Zk is a t
least d / 2 M , where d denotes the distance2 of the initial point $j, , . . . , $jn from
the boundary of R and M the maximum of I j , j (i = I , 2, . . . , n) in R.
For our equations
dyk
-=
dt
ft,
+
let us take for the interval (a, , a,) the interval (to - Zk, to Zk) with the largest
possible value of Zh. Then one of the functions yr (i = 1, , n) becomes singular
at the endpoints; for, otherwise the solution could have been continued analyti-
cally t o a larger interval, which means that not the largest possible 'Z was taken:
Hence, according to the above corollary, the solution approaches the boundary
of R.
Now the functions j , , being continuous in R , are bounded in absolute value
by some constant 111; and the fZ , sinre they approximate .f, uniformly, are
bounded in absolute value for large enough k by, say, 2111. Since yr (i = 1, . . . , n)
is a solution of the system dy:/dt = j f , the bound 2M on the absolute value
of the derivative of y: lies in the interval (t, - Ik, f, +
.)'Z Therefore the devia-
tion of $(i = 1, . . . , n) in this interval from its initial value is at most 2111 Zk.
On the other hand we have shown that for each k, the solution reache? the
boundary of R as t varies in the interval (to - Zk, to +
l'). Hence a t least one
of the components yf has deviated from its initial value y, by at least the amount
d (d is defined as the distanre of the initial point $jl , . * , $jn from the boundary
of R ) . Hence
d 5 2Mlk.
This establishes the interval (to - d / 2 M , to +
d / 2 M ) ES the interval in
which all approximating problems (from a certain li on) have solutions.
Since the first derivatives of the solutions of the approximating problems
have the common bound 2 M , these solutions are equicontinuous; hence, by
Arzela's theorem, a convergent subsequence of them may be selected. And it
is a routine matter to show that the limit of this subsequence solves the original
problem.
Note that only continuity of the functions f , was required in the above proof.
It is well known that this is not sufficient to guarantee that the Picard iterations
generate a sequence which converges to a solution-certainly not if there are
several solutions. In contrast, my generalization of the above method to partial
differential equations applies only when the coefficients and initial data are a t
least Lipschitz continuous, a circumstance which guarantees that the solution
is unique and can be obtained by iteration.
As long as we have only the Haar inequality to prove convergence, the above
approximating process yields the same existence theorems as can be obtained by
Picard iteration, see Schauder [15], Cinquini-Cibrario [2],Friedrichs [6], and
Courant and Lax [4]. However, in the second part of Section 5 a more sensitive
tool is used to obtain detailed information about the way the solution depends
on the initial data. The result is couched in the language of generalized solutions;
roughly speaking, it asserts that discontinuities in the first derivatives of the
initial data are propagated only along the characteristics. This result can be
regarded as an extension of a result of Doughs [5], see also Hartman and Wintner
[la
I would like to mention that the method of appro.iimating riorianalytic
problems by analytic ones has been used by Schauder to solve nonanalytic initial
value problems for linear equations; he handled first order systems in two in-
dependent variables by this method in [ 151, second order hyperbolic equations
in any number of independent variables in [16].
His method relies on the following fact: In the linear case, the
Cauchy-Komalemski theorem is a result in the large in the sense that the size of
the domain of analyticity depends only on the size of the complex domain of
analyticity of the initial data, not on the magnitude of the data in this complex
domain. Schauder’s method has been adopted by Petrowski for general hyper-
bolic systems, see [13].
Schauder, in his treatment of the nonlinear case [15], relies on his fixed
point theorem. In the particular situation he deals with. the fixed point could
have been obtained by Picard’s method of i t e r a t i ~ n . ~
The theorem on the existence in the large of solutions of analytic hyperbolic
equations has this immediate corollary: The set of analytic solutions of a quasi-
linear, analytic, hyperbolic system is dense in the set of all solutions. For
general nonlinear systems the main theorem of this paper gives only a weaker
result: the set of analytic solutions is dense in the set of those solutions whose
highest derivatives occurring in the equation satisfy a Lipschitz condition. In
other words, some nonlinear equations* may possibly have some isolated solu-
tions; so far none have been found, but their occurrence is not excluded by a
general theorem, not even for single nonlinear first order equations. If it could
be shown that a single first order nonlinear equation has no isolated solutions
then the gap in Haar's proof [9] that the classical theorem on characteristic
strips is valid when no requirement beyond continuity is made on the first de-
rivatives of the solution mould be filled.
For a large class of nonlinear equations (e.g. the Monge-Amphre equation)
it has been shown that no isolated solution exists.
In the course of proving the main theorem, i.e. the existenre in the large
of an analytic solution of an analytic initial value problem, one is obliged t o
estimate successive derivatives of romposite functions. I obtained satis-
factory estimates by using as majoraiit a special power series, very similar t o
one used by Gevrey [7,12]. This majorizing series has many uses. 111 the
appendix it is used to give a simple direct proof of Cauchy's theorem on the
convergence of the formal power series solution of the analytic initial value
problem for a system of ordinary differential equations.
I am greatly indebted to R. Courant, K. 0. Friedrichs and A. Dough;
many of the ideas presented here were thrashed out in conversations with them.
Section 1
In this paper we shall study quasilinear systems of first order equations
for ~tfunctions of two independent variables. The general form of such a system
is
(1.1) A'U, + B'U, + C = 0
wbere L: stands for the column vector of n unknown functions, A' and B' are
coefficient matrices, C' is a column vector; A', B' and C' depend on s,y and U .
We shall be concerned with nondegenerate systems, i.e. systems where a t
least one free (noncharacteristic) direction exists. \I-e take this free direction
to be the direction defined by y = const. and solve (1.1) for the y derivatives
(1 2) U, = AU, + C.
From here 011 we shall deal exclusively with systems of the form (1.2).
(This means a slight loss of generality since it amounts to assuming that the
direction y = const. is free for all relevant r , y, U . )
(1.2) is called hyperbolic a t a point To , 2/o , li, if the eigenvectors of the
matrix A(r, , yo , U,) span the n-dimensional space, and the corresponding
eigenvalues are real. This is the same as saying that there exists a (real) matrix
T such that
(1 *3) T-'AT = D
is diagonal.
+
If (1.2) is hyperbolic at every point of an (n 2)-dimensional domain 3,
we say that (1.2) is hyperbolic in E. If in addition .4,C and T belong to some
NONLINEAR HYPERBOLIC EQUATIONS 235
differentiability class' (say Ck)in El we say that (1.2) is hyperbolic of that class
in E. In particular we shall speak of systems which are analytic hyperbolic in
a domain E.
The eigenvalues of the matrix A are the negatives of the slopes of the
characteristics of the system (1.2). Multiplying (1.2) by T , we transform it
into the characteristic form, i.e. a form in which the i-th equation contains only
derivatives in the i-th characteristic direction.
Linear (and semilinear) systems can be expressed in a more special normal
form by introducing new unknowns V = T-lU for these new (so-called character-
istic) quantities. The equations then read:
STis not determined uniquely by condition (1.3); therefore we should properly say: If T
can be chosen to belong to a certain differentiability class.
%chauder [15]avoids integration along the characteristics; instead, he relies on a device
of Haar [lo].
'We assumed that the lines y = const. are nowhere characteristic.
236 PETER D. LAX
S o t e that all points of the line of integration lie iii the domain of determinacy.
\Ye state now the a priori estimate for the solution of a diagonal lincnr
hyperholic system in terms of its initial values (Haar's lemma).
Lemma I : Suhject the initial values and coeffirients of the system
lye ivish to show that u ' , 3 = 1, 2 , . . . , n, is less than w(y) throughout the
domain; if this were false for some points there would be a point with smallest
y coordinate, say yo , for which one of the vi, say vi, would be as great as w.
Consider the relation (1.8) at this point. The left side is nonpositive. The right
side on the other hand is positive; for, by (i), the term M - u'(P') is positive,
and by (iii), the term Jio [y(y) - c:] dy is nonnegative. The integrand of the
remaining term is also positive; this follows from assumption (ii) and from the
fact that the integration extends over values of y less than yo for which w is
greater than ui.
This coiitradi~tion,a consequence of the assumption that (1.7) fails to
hold a t sonie point. of the domain of determinacy, proves Lemma 1.
Note that Lemma I is a literal analogue of a basic lemma on the growth
of the soliit,ioii of ordiriary differential equations (see e.g. [I]).
\Ye return now to the general quasilinear case. Here too the notions of
chartctcriatic direction and domain of determinacy can easily be defined; the
i-th chtirwteristir direction is again defined as dx/.'dg = - d ' , where d i is the
i-th clemeiit of the diagonal matrix. Since D depends on 2 , y, :And L:, we can
spcnk only of the i-th characteristic field with respect to a particular solution
IT0(x, g ) , which is to be substituted for U in di(.c, y, C:).
Similarly, S is called a domain of determinacy of the initial interval with
respert to the solution C, if S is an open set in whivh the vector I!, is defined
and satisfies the equation, and if S is a domain of determinacy of the linear
problem
then their union is again a domain of determinacy. Thus by Zorn's lemma every
domain of determinacy is contained in a maximal domain of determinacy.
We shall state now the main theorem of this paper:
Theorem I: Let
U, = a s , y, U)U, + CYX, y, U)
be an analytic hyperbolic system in a domain E of the (n +
2)-dimensional
space 2, y, U ; assume that this domain E contains the initial line x, 0, @(x),
al 5 x 2 a, , and assume that the initial values U(z, 0) = @(x)are analytic.'
Let U(x, y) denote an analytic solution of this initial value problem in some
domain of determinacy S. Assume further that this solution does not approach
the boundary of the domain of analyticity of the system, i.e. the closure of the
pointset z, y, U(z, y) with (2, y) P S contains only interior points of E , and that
the first pwtial derivatives U , remain bounded for (x,y) in S.
Conclusion: U(z, y) has no singularities on the boundary of S.
The following example shows that Theorem I is false when the requirement
of boundedness of U, is omitted from the hypothesis:
Consider the equation
u, = uu,
which is analytic-hyperbolic in the whole z, y, u-space; assign the initial values
u(x, 0) = 2.
The function uo = z/(l - y) solves the initial value problem. The charac-
teristic curves corresponding to this solution are straight lines through the point
(0,l) of the z,y-plane. Therefore the interior of any triangle S with one vertex
at (0,l) and base below the x-axis is a domain of determinacy of the intercepted
interval on the x-axis. The solution uo(z,y) remains bounded throughout such
a triangle, hence it doesn't reach the boundary of E which is the whole space;
yet u has a singularity at the boundary point (0,l).
For semilinear systems ( A independent of U ) the boundedness of U , in S
can be deduced from the other hypotheses. On the other hand for general
nonlinear systems an additional hypothesis has to be added:
(n) All derivatives of order one higher than those entering the equations
remain bounded in S.
The following example shows that without some additional assumption
such as (n) Theorem I is false for general nonlinear systems:
Consider the equation u, = 3 u;L , and assign the initial values u(z, 0) =
+ 2'. The equation is analytic in the whole x,y, u,p-space (here p = u,), that is
E is the whole four dimensional space; the initial values are analytic on the whole
x-axis.
91.e., they are analytic in a domain of the complex z-plane which contains the closed
interval ul 5 x 5 u2 .
NONLINEAR HYPERBOLIC EQUATIONS 239
Section 2
Dejnition: A formal series in k variables
240 PETER D. LAX
B = C
i.=O
bil...ik~fl xik,
denot,ed by
2=1 %=I
xdt1 , - ,
* tl) 3 y*(t, , * , to,
and
X,(O, - , 0) = y*(O, * * * , 0) = 0.
Let A(xl , 0 - . , x,) and B(x, , - - , x,) be power series in z, , - , xk satisfying
+ *
A = 2
iv-0
U~,...~~X;' - - . xLk,
is said to majorize up t o N another formal power series
m
B = C
1.=O
b , x . . . z k ~* ;*'. X:k
-
(denoted by A ( x ~, * . , 2,) > N B ( x , , * . , x,)) if ui,...ik 2 1 bi,... i k I for all
values of i, , . , i, which satisfy i, i,
+ * * + + -+
ik 5 N .
It is easy t o show that rules (i), (ii), and (iii) hold if we replace the sign
> by > N *
We shall prove a number of lemmas concerning the power series
Section 3
I n this section we carry out the steps outlined a t the end of Section 1 for
the proof of Theorem I. The discussion will be restricted to quasilinear systems
already in diagonal form. The proof of the general quasilinear case can be
obtained from this special case by observing that if U is a solution of a general
quasilinear equation ( 1 .a),
U, = AU, -+- B ,
then the system of 3n functions U , U , , U , satisfies a quasilinear system that
can be put in diagonal form. To see this denote T-IU, by Z , T-'U, by W ;the
first set of the 3n equations is the identity
(3.1) U, = DU, - DTZ + TW = DU, + C'.
The next 2n equations are obtained by differentiating the original system (1.2)
with respect to x and y, replacing U , and U , , wherever they occur, by T Z and
TW and multiplying through by T :
(3 -2) Z , = DZ, + C"
(3.3) W, = DW, + C"'.
C" and C"' are functions of x, y, U , Z , W . The system (3.1)-(3.3)for
U , Z , W is in diagonal form. It is clear that every solution U of the original
system (1.2) generates a solution of the enlarged system.
Equivalence of the enlarged system with the original one-or rather equiva-
lence of the corresponding initial value problems-is not needed for the present
purpose. I should like to mention that no simple direct verification of this
equivalence is known; it is, however, clear if, as in [4], only Z = T-'U, is adjoined
as a new set of unknowns.
If we succeed in proving Theorem I for diagonal systems we can immedi-
ately state-with the aid of this reduction-that the solution of a general quasi-
linear system cannot develop singularities unless either the solution reaches
the boundary of the domain of analyticity of the coefficients, or the second
partials become unbounded. But as the discussion a t the end of Section 1 shows
the second partials cannot become unbounded if the first partials remain bounded;
thus we have Theorem I in the general case.
Since from here on we shall be concerned with the special system (3.1)-(3.3)
I should like to denote 3n by m, the vector composed of U , Z and W by V , and
write the system as
(3.4) V, = DV, C; +
D now has a slightly altered meaning: it is three of the former D put together
diagonally.
We may assume without loss of generality that the coefficients D and C do
NONLINEAR HYPERBOLIC EQUATIONS 243
not depend explicitly on the independent variables z and y. For, such explicit
dependence can always be eliminated by regarding x and y as new unknowns
satisfying trivial equations.
The tot,al derivative of D with respect to z (and y) can be expressed as a
function of V alone; for, l? depends on the first n components of the V only, and
their derivatives can be expressed as functions of V. We have therefore
d
(3.5) -D = F(V).
dx
This identity will be used in the course of the proof.
Proof of Theorem I for the system (3.4) : According to the hypothesis of Theorem
I the solution remains within a closed bounded subset E' of the domain of analy-
ticity of C(V) and F(V). Therefore there exist uniform estimates for the deriva-
tives of C and F for all relevant V , of the form
k, + k, ... + k , = i, i = 0, 1, --- ,
for some (sufficiently large) constants b and R.
These inequalities can be put into concise form in terms of the notation of
section 2":
+ + qm) C b n
m
(3.6') C(v1 ql , * * * , V, 1-1
fO(Rq,).
,=I
It has been assumed further that the initial data of U , U(x, 0), are analytic
in an interval that includes the interval a, 5 2 5 a, . Therefore the same is
true for the initial data of V, and this implies uniform bounds for the xde-
rivatives of V on the initial line, of the form
1
O A scalar power series majorizes a vector power series if it majorizes each of its com-
ponents.
244 PETER D. LAX
(3.10)
(3.12)
and C=N.
, C, are abbreviations for Fz<-,, Vzn-s+5
Here F,-l , Vn--i+,
As stated before, (3.12) is linear in V , :
(3.12) (VN)v= D(Viv), + B'V -L C'
NONLINEAR HYPERBOLIC EQUATIONS 245
where
B' = N F + Cv
(3.13)
C' = 5 ( ~ ) F ~ - ~+vcN~ -- ~c,T".
i=2
+ ~
Lemma I estimates the solution of linear equations of the form (3.12) in terms
of upper bounds for absolute values of the coefficients B' and C'; these bounds
were denoted by /3 and y(y).
The choice of /3 is simple:
P
Ng =
+
where g = f c, j being an upper bound for J F J over E' and c and upper bound
for 1 C, 1 over E'.
Compound derivatives enter into 6'. However, we have chosen our esti-
mates for the derivatives of F and C-see (3.7') and (3.6')-and the induction
hypothesis 3.10N-, so that Lemma IV, or rather its corollary, on derivatives of
compound functions can be invoked":
F(V(, + f , Y)) c bf0(r(y)5),
N-1
I.e.,
These last three inequalities give estimates for all ingredients of C' as given
by formula (3.13), so for C' itself we have the upper bound
nand because the inequality r(y) 2: a(y)RK, which is a hypothesis of Lemma IV, is
satisfied. This is clear from the definition of r(y) and a(y) as given by (3.11),a(y) = may,
r(y) = rev , and the inequality between a and r itself as stated in (3.9,r > aRK.
246 PETER D. LAX
Thus
P = Ng
y(y) = a(y)N!rN-l(y)K"/N = aN!rN-leNhuK"/N
and, by Lemma I,
Section 4
The passage from the analytic to the nonanalytic case is carried out in
Section 5 . I n this section I would like to point out another possible application
of the technique described in the last section for estimating successive derivatives.
Suppose namely that the coefficients and the initial data contain a parameter k,
and that their dependence on this parameter (as well as on the dependent and
independent variables) is analytic. We claim that then the solution depends
NONLINEAR HYPERBOLIC EQUATIONS 247
Section 5
In this section we shall complete the passage to the limit from the analytic
to the nonanalytic case. We start with a lourer estimate for the size of the domain
in which the solution of an analytic initial value problem exists. This estimate
is derived from our main theorem by reasoning outlined in the introduction for
ordinary differential equations,
Denote by Ss the trapezoid topped by the line y = 6 whose base is the
initial interval and whose sides slant inward with a slope K . K is an upper
bound for the characteristic slopes in E' which is a closed subset of the domain
of analyticity E of the coefficients. We assume that E' contains the initial
interval in its interior and denote by r the shortest distance of the initial line
from the boundary of El. Note that this choice of K guarantees that Sa is a
domain of determinacy for any solution that exists over it and stays within E'.
Let 6, denote the largest value of 6 for which S , carries an analytic solution
that still stays within E'. Clearly such a solution U must either have reached
the boundary of E' or have a singularity on the boundary of Sa, ;for, otherwise
it could be continued analytically. Thus from our main theorem we conclude
that one of two alternatives must hold:
(1) U has reached the boundary of E',
(2) One of the first derivatives of U is unbounded in Sa, .
We shall show that neither (1) nor (2) can hold if 6 is less than a certain
quantity 7;this quantity 7 depends only on bounds for the coefficients A , T , T-',
C and their first derivatives" in E', on the first derivative of the initial data
+(x) and on r, the distance of the initial line from the boundary of E'. The
demonstration is based on an a priori estimate for the first derivative of the
solution of a quasi-linear hyperbolic system. For, denote by I: a solution of
(5.1) U, = AU, C; +
then Z = T-'U, will satisfy t,he equation obtained from (5.1) by differentiation
with respect to x,
(5.1') Z, = DZ, C' +
where C' is quadratic in Z with first order quantities as coefficients.
'*Quantitiesthat depend only on bounds enumerated here are called first order quantities.
248 PETER D. LAX
1.e.
1 z'(P) 1 II z'(P') j + yo(a + am*),
2m Im + yo(a + am2).
This can only hold if
(5.2) m/(a + am2) Iyo
which is a lower bound for yo , i.e. a lower bound for the y-coordinates of those
I
points where 1 2 exceeds 2m; thus we have a lower bound for 6, itself in case
I 2 1 is unbounded in Sa, , i.e., alternative (2) holds.
A similar lower bound can be found for 6o if the first alternative holds. To
show this we note that an upper bound for 1 2 1 furnishes an upper bound for
I U , I since these two are related by
U, = A T 2 + 6;
therefore
I u,I Ib I 2 I +c,
where b and c are zero order constants.
R e claim that if 6 is less than
(5.3)
(
min r, ~-
a P a m 2 ' 2mb +c
then the solution stays within E'. Namely in this case the distance of the point
(2,y, U ( z , y)) from the point (x,0, U ( z , 0)) on the initial line is less than r. To
NONLINEAR HYPERBOLIC EQUATIONS 249
+
verify this we have to check that the deviation of each of the n 2 components
is less than r. This is clearly true for the first components since they are identical;
the deviation of the second components is 1 y I which, by (5.3), is less than T.
The difference of each of the last n components is less than 6 max 1 U y 1, which is
+ +
less than 6 ( b max I 2 I 6); y is constrained by (5.3) to be less than m/(a am'),
and in this region I 2 I is less than 2m. Finally, again by (5.3), 6 is less than
+
r/(2m7, c). Thus the difference of the last n components is less than
This proves that the second alternative cannot hold if 6o is less than the
quantity given by (5.3). On the other hand we have shown previously that the
first alternative cannot hold either; thus 6" must be greater than the quantity
given by (5.3).
Observe that we have not only obtained a lower bound for the size of the
domain in which a solution of an analytic initial value problem exists but also a
bound for the first derivatives of the solution in this domain, namely j 2 I 5 2m.
Let us now take a nonanalytic initial value problem
U, = AU, + 6,
U(z, 0) = +).
Assume that the coefficients A , T and C satisfy a uniform Lipschitz condition
in a domain E' of z,y,U-space; assume that E' contains the initial line in its
interior and that T is nonsingular in E'.
The intial vector @(x) is assumed to satisfy a uniform Lipschitz condition.
Cnder these conditions the initial value problem has a generalized solution.
Proof: ilpproximate the coefficients A , T , C and the initial vector QI by a
sequence of analytic ones A , , T , , C, and ip, . Choose these so that A , , T I , C,
are analytic in E' and that their, first derivatives are uniformly bounded in E',
while the first derivatives of @, are uniformly bounded over the initial interval.
Consequently the approximate initial value problems have solutions U , in a
common domain and the first derivatives of the U , will be uniformly bounded,
The difference of two solutions U , and U , can be estimated in terms of the
coefficients and of the difference of initial data. To see this we subtract the
equations satisfied by U , and U , from each other:
( V , - CJt = +
AdUJ(UL - Urn)= [ A W J - AdUm)IUm,
+ fAl(U,) - A,(Um)lUm,+ C W , )- CdU7n) + CdUJ - C,(U,).
Denote T;' (U,) ( U , - U,) by 8; V satisfies
(5.4) V , = D,(A,)V, + EV + F .
We shall not write out the exact expressions for E and F ; it suffices to know
that E is bounded in terms of the uniform upper bound for the first derivatives
250 PETER D. LAX
of A , and A,,, and in terms of the uniform upper bound for 1 U,, I and I U,, 1.
F is bounded by a constant times I A , - A , I and I 6, - C, 1; the constant
depends on the same quantities as the upper bound for E.
The Haar inequality applied to the system (5.4) shows that V is in absolute
value less than a constant times the maximum of 1 V 1 on the initial line plus a
constant times the maximum difference of the coefficients A , , A,,, and C , , C ,
on E'. This shows that as 1 tends to infinity, U , converges uniformly to a limit.
We call this limit the generalized solution of our initial value problem.
It is clear from the foregoing that the generalized solution is independent of
the particular approximating sequence. Our estimate for I U , - U , 1 shows
further that the generalized solution depends continuously in the sense of the
maximum norm on the initial data and the coefficients.
I was able to show that these generalized solutions satisfy the differential
equation almost everywhere in this sense: Write the system in characteristic
form, i.e. multiply the given equation U , = A U , +
C by T :
TU, = TAU, + TC = DTU, + C'.
Write out the equation componentwise:
(5.5) t*,(du'/di) = c: ) i = 1, 2, . - ., n.
Here d/di denotes differentiation in the i-th characteristic direction:
du'ldi = u: + d,u: .
A generalized solution as a function of x and y has bounded difference
quotients, and therefore it has bounded difference quotients as a function of the
parameter y along any characteristic curve. Therefore by a well known theorem
of real variable theory, the directional derivative du'ldi exists at almost all
points of any given characteristic curve ( and u'is absolutely continuous along
any characteristic curve). We claim that with this determination of du'/di,
(5.5) holds almost everywhere along any given characteristic curve.
Proof: Let P and Q be two points along the i-th characteristic corresponding
to a generalized solution U . Let U , denote the solution of the I-th problem in
some approximating sequence and denote by P , that point on the i-th character-
istic (corresponding to the Lth problem) through Q which has the same y-coordi-
nate as P. Clearly P , tends t o P as I 4 0 3 .
U , satisfies the analogue of (5.5) for the I-th problem; integrate the i-th
equation for the 1-th problem between P , and Q along the i-th characteristic
curve x = xl(y) corresponding to the Z-th problem:
IP,
0
dui
tfi(U1)- dy =
dY
Q
PI
c l ' dy.
cE(U) dY
(5.7)
The limits of all three terms can be easily evaluated; the first two tend to
ti;(U(&))u'(&)- t i j (U(P))u'(P);in the last, the integrand converges uniformly
to ( d l d y ) t,i ( U ) - u ' ,and thus the passage to the limit can be performed under
the integral sign. If the limit expression for (5.8) is integrated back again by
parts and then identified with the limit of the left side of (5.6), we obtain
i.e. that the limit solution satisfies the integrated form of (5.5); it must then
satisfy (5.5) almost everywhere on the characteristic.
I wish to emphasize this: The generalized solutions deserve their name
because they were obtained by functional extension, i.e. by looking at the
solution of the initial value problem as a function of the initial data and of the
coefficients and then extending the domain of this function by completion. That
these generalized solutions satisfy the differential equations almost everywhere
is merely a mildly interesting property of generalized solutions. Of course this
property may turn out to be a useful tool in studying some intrinsic properties
of generalized solutions, in particular the manner of their dependence on the
initial data, and it is for this reaaon that I decided to include it.
If the coefficients and the initial data have further differentiability proper-
ties, so will the generalized solution. We shall present a special instance of such
a result. Suppose that A , T , and C have Lipschitz continuous first derivatives in
E' and has Lipschitz continuous first derivatives in the initial interval, then the
generalized solution too will have Lipschitz continuous first derivatives and will
satisfy the differential equation everywhere. For, in this case the functions
A , , T I , C , and a, in the approximating sequence can be so chosen that their
252 PETER D. LAX
*Note added in proof: This result has been rediscovered by Hartman and Wintner [18].
NONLINEAR HYPERBOLIC EQUATIONS 253
This assures us that the above sequence of difference quotients of uk also fails
to converge. Since the U , have continuous first derivatives, this difference
quotient can be expressed according to the mean value theorem as the derivative
of U , at some point Ri between R: and R: . R , tends to R but the sequence
(U,,,,), (R,) is not convergent, contrary to our hypothesis.
Instead of operating with the first derivatives ( U , ) , it is more convenient
to operate with their characteristic linear combinations 2, = T,-'( U l ) z, since
these satisfy the differential equations (5.1') already in diagonal form. Define
the vector W ( R )to be
lu b SUP [z,(RJ - Zt(W11
~Bi1,IB'il I--
the least upper bound being taken with respect to all pairs of sequences ( R , ]
and {R;J tending to R.
The component,s of W will be denoted by wi. Our aim is to show that
W ( R ) is zero at all regular points; this conclusion will follow from an integral
inequality for the components of W which will be derived in the next paragraphs.
We start With the integral relation satisfied by the components z; of 21 ;
it is obtained by integrating the i-th equation in (5.1') along the i-th character-
istic:
The integral on the right is taken along the i-th characteristic connecting
R with the point R' on the initial interval.
Our immediate aim is to get an upper bound for w'(R);
wz((R) was defined
as the least upper bound of
lim sup [z;(R,)- d ( R { ) ] ,
I-CC
(5.10)
+ [[ cl'(R,)d q - I'""c:'(R;)
II
dq;
is an upper bound for the upper limit of either side of (5.10); k here is a constant
(a first order quantity). According to what we said before, this is also an upper
bound for wi(R):
(5.11)
NONLINEAR HYPERBOLIC EQUATIONS 255
Appendix
In this appendix we shall apply Lemma IV to demonstrate the existence of
a solution of an analytic system of ordinary differential equations. Let the
system be
There exist numbers M and p such that I a ~ " ' " " " I 5 M p i ' + i ' f ' * * ffor
in
all i and j. Let R be any number greater than p , say R = 2p. Clearly there
exists a constant b (for R = 2p we can take b = (9/4)"M) such that
Y,(4 c aqf0(r4.
Furthermore we have chosen T 1a RK. Thus by the corollary of Lemma IV
we have
BIBLiOGRAPHY
[I] R. Bellman, The boundedwss of solutwns of linear differential equatwns, Duke Mathe-
matical Journal, Volume 14, 1947, pp. 83-97.
[2] M. Cinquini-Cibrario, Un h e m a di esistenza e di unicitd per un sistema di equuzwni dle
derivate parzidi, Annali di Matematica (4), Volume 24, 1945, pp. 157-175.
[3] R. Courant and D. Hilbert, Methoden der mutlaematischen Physik, Volume 2, Springer
Berlin, 1937.
[4] R. Courant and P. D. Lax, Nonlinear partial differential equations with two independent
variables, Communications on Pure and Applied Mathematics, Volume 2, 1949,
pp. 255-273.
[5] A. Douglis, Existence theoremsfor hyperbolic systems, Communications on Pure and Applied
Methematica, Volume 5, 1952, pp. 119-154.
[6] K. 0. Friedrichs, Nonlinear hyperbolic differential equationsfor functh-s of two independent
variables, American Journal of Mathematics, Volume 70, 1948, pp. 555-589.
[7] M. Gevrey, Sur la nature analytique des solutwna dea t!quationS aux derivtes partklles,
Annales de 1'Ecole Normale, Volume 35, 1918, pp. 129-190.
258 PETER D. LAX
[8] A. Haar, Sur l’unicite des solutions des equations aux derivees partielles, C . R. des seances de
I’Academie des Sciences, Volume 187, 1928, p. 23.
[9] A. Haar, Zur, Characteristiken theorie, Acta Litterarum ac Scientiarum, Volume 4, 1928,
pp. 103-114.
[lo] A. Haar, Uber Eindeutigkeit u n d Analyzitat der Losungen partieller Differenzialgleichungen,
Atti del Congress0 Internasionale dei Matamatici, Bologna, Volume 3, 1928,
pp. 5-10.
[ l l ] J. Hadamard, Observations sur la note prdcddente, C . R. des seances de I’Academie des
Sciences, Volume 187, 1928, p. 23.
[I21 J. Hadamard, Les fonctions de classe supdrieure duns l’dquation de Volterra, Journal d’Ana-
lyse Mathematique, Volume 1, 1950, pp. 1-10.
[13] I. Petrowsky, Uber das Cauchysches Problem fur Systeme von partiellen Differenzialglei-
chungen, Rec. MathBmatique, Volume 2, NO. 44, 1937, pp. 815-868.
[14] M. Rose, O n systems of two non-linear partial differential equations in two independent
variables, Master Thesis, New York University, 1950.
[15] J. Schauder, Cauchysches Problem f u r partielle Differenzialgleichungen erster Ordnung.
Anwendung einiger sich auf die Absolutbetrage der Losungen bezeihenden Abschat-
zungen, Commentarii Math. Helv., Volume 9, 1937, pp. 263-283.
[ 161 J. Schauder, Das Anfangswertproblem einer quasilinearen hyperbolischen Differenzial-
gleichung zweiter Ordnung in beliebiger Anzahl von unabhangigen Veranderlichen,
Fundamenta Mathematicae, Volume 24, 1935, pp. 213-246.
[17] H. Lewy, Generalized integrals and diferential equations, Transactions of t,he American
Mathematical Society, Volume 43, 1938, pp. 437-464.
[18] P. Hartman and A. Wintner, O n hyperbolic partial diferential equations, American Journal
of Mathematics, Volume 74, 1952, pp. 834-864.