The Problem of Dirichlet for Quasilinear Elliptic Differential Equations with Many

Independent Variables
Author(s): J. Serrin
Source: Philosophical Transactions of the Royal Society of London. Series A, Mathematical and
Physical Sciences, Vol. 264, No. 1153 (May 8, 1969), pp. 413-496
Published by: The Royal Society
Stable URL: http://www.jstor.org/stable/73755 .
Accessed: 27/05/2014 04:47

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

The Royal Society is collaborating with JSTOR to digitize, preserve and extend access to Philosophical
Transactions of the Royal Society of London. Series A, Mathematical and Physical Sciences.

http://www.jstor.org

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 [ 413 j

TIIE PROBLEM OF DIRICHLET FOR QUASILINEAR

ELLIPTIC DIFFERENTIAL EQUATIONS WITH MANY
INDEPENDENTI' VARIABLES

BY J. SERRIN
Universityof Minnesota

(Communicatedby W. K. HIayman,F.R.S.-Received 14 May 1968)

CONTENTS
PAGE PAGE
INTRODUCTION 414 13'. Further estimates for |Dul 450
14. Summary 451
CHAPTE R 1

1. The fundamental existence theorem 417 CHAPTER III

2. The existence programme 420 15. A variation of the maximum principle 459
3. The distance function 420 16. Irregularly elliptic equations 459
4. A basic identity 422 17. Singularly elliptic equations 462
18. Non-existence theorems for
CHAPTER II singularly elliptic equations 468
5. Basic definitions and agreements 425 = 0
19. The equation _r4'D2u 471
6. Elementary maximum principles 426 CHAPTER IV
7. Global barriers 430 20. The minimal surface operator 476
8. Regularly elliptic equations 431 21. Variational equations 478
9. Boundary estimates depending on 22. Surfaces having prescribed mean
curvature 434 curvature 480
10. Boundedly non-linear equations 437 23. Surfaces having prescribed mean
11. Mixed equations 440 curvature in central projection 485
12. Interior estimates for IDul 442 24. The differential equation of constant
mean curvature 489
13. Interior estimates for IDul in
higher dimensions 447 REFERENCES 494

This paper is concerned with the existence of solutions of the Dirichlet problem for quasilinear
elliptic partial differential equations of second order, the conclusions being in the form of necessary
conditions and sufficient conditions for this problem to be solvable in a given domain with arbi-
trarily assigned smooth boundary data. A central position in the discussion is played by the concept
of global barrier functions and by certain fundamental invariants of the equation. With the help
of these invariants we are able to distinguish an important class of 'regularly elliptic' equations
which, as far as the Dirichlet problem is concerned, behave comparably to uniformly elliptic
equations. For equations which are not regularly elliptic it is necessary to impose significant
restrictions on the curvatures of the boundaries of the underlying domains in order for the
Dirichlet problem to be generally solvable; the determination of the precise form of these restric-
tions constitutes a second primary aim of the paper. By maintaining a high level of generality
throughout, we are able to treat as special examples the minimal surface equation, the equation
for surfaces having prescribed mean curvature, and a number of other non-uniformly elliptic
equations of classical interest.

VOL. 264. AX II3 (Price 192 4s.; U.S. $5.70) 54 [Publishedl 8 May I969

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
414 J. SERRIN

INTRODUCTION
Within recent years the study of the Dirichlet problem for second-order quasilinear elliptic
equations has reached a certain degree of completeness, and, in particular, methods have
been developed which reduce the existence question to one of obtaining a _prioriestimates
for the solution and its gradient. I propose here to treat this latter problem, presenting
sufficient detail to show the pre-eminent role which certain fundamental invariants of the
equation play in the decision as to whether the Dirichlet problem is or is not generally
solvable for a given equation and a given domain.
We consider, specifically, quasilinear elliptic partial differential equations of the form
Dlu =
vQ(xgu, DOu) (x, u, Du) . (1)
Here x (xl,.., x) denotes points in the underlying n-dimensional Euclidean space, Du
and D2u denote respectively the gradient vector and Hessian matrix of the dependent
variable u u (x), and v and X are respectively a given matrix and a given scalar function
of the variables indicated. The multiplication convention implicit in (1) is the natural one
which makes a scalar of the left-hand side; that is, using component notation,
-ID2u di2UX aXj

where summation from 1 to n over repeated indices is understood.

The structure of (1) is determined by the functions v (x, u,p) and M (x, u,p). We suppose
throughout that they are defined and continuously differentiable for all real values of
u and p, and for all points x in the closure of any particular domain under consideration.
Ellipticity of (1) requires that the coefficient matrix satisfy the condition

6-d6 > O (2)

for arbitrary non-vanishing real vectors 6 (61, ... 5n). Moreover, we can obviously assume
without loss of generality that d is symmetric. Beyond this, any additional structure which
will be required of the coefficient matrix and inhomogeneous term is best left to the state-
ment of individual theorems. In any event, we shall maintain a fairly high level of generality
throughout the paper, and accordingly will be able to treat as special examples the minimal
surface equation, the equation for surfaces of prescribed mean curvature, and a number
of other uniformly and non-uniformly elliptic equations of classical interest.
Roughly speaking, the conclusions are in the form of necessary conditions and sufficient
conditions for the problem of Dirichlet in a given domain to be solvable for arbitrarily
assigned smooth boundary data. The work thus constitutes a development and extension
of the theory initiated by Bernstein in a series of papers from 1907 to 1912 and applied by
him to a number of cases of great generality. Bernstein's work was subsequently sharpened
and enlarged in an important and fundamental paper of Leray which appeared in 1939. In
comparing this work with ours, we note first of all that their considerations were restricted
to the case of two independent variables and inlvolve various regularity assumptions on the
behaviour of v' andi g for large values oftp which we avoid here. Bernsteinl had also con-
cluded for certain equations (including that governing surfaces of constant mneancurvature)
that the Dirichlet problem was not generally solvable for arbitrary smooth data, even when

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 415
the domain in question is convex and arbitrarily small. While strictly true, this statement
nevertheless sidesteps the problem of determining precisely those domain for which the
Dirichlet problem is in fact generally solvable. It is one of the goals of this paper (as it was
of Leray's) to consider questions ofjust this sort.
In higher dimensions, the Dirichlet problem for non-uniformly elliptic equations has
been comparatively less studied. We mention here specifically the work of Gilbarg and of
Stampacchia, and recent papers of Trudinger and Oskolkov, and also a treatment by
Jenkins & Serrin on the minimal surface equation. Some of the relations between this work
and ours will be discussed below (see in particular ??9, 10 and 13).
The formal work of the paper is divided into four chapters. In the first one we give a
precise statement of the problem under consideration and outline the general procedures to
be followed in its study. The fundamental existence theorem upon which the further results
of the paper are based is essentially due to Ladyzhenskaya & Uraltseva, though the form
in which it has been stated does not seem to have appeared previously in the literature.
The second chapter is in many ways the central one of the paper; in it we treat in detail
the various a priori estimates required by the existence theory. The results themselves are
strongly influenced by the structure of the equation, and in particular by the behaviour of
certain fundamental invariants. Following a discussion of maximum principles and global
barrier functions, we introduce the concept of regular ellipticity. As far as the Dirichlet
problem is concerned, the class of regularity elliptic equations behaves comparably to the
class of uniformly elliptic equations. On the other hand, for equations which are not regu-
larly elliptic it is frequently necessary to impose significant and important restrictions on
the curvatures of the boundaries of the underlying domains in order to be able to solve the
Dirichlet problem for arbitrarily given smooth data. An intimation of this state of affairs
occurs already in the work of Bernstein and Leray, and in the paper by Jenkins & Serrin
noted above. Bernstein had in fact seen the importance of convexity in the solution of
Dirichlet's problem in two variables, while Leray had found a further relation between the
second derivative of the boundary curve and the inhomogeneous term M. For the minimal
surface equation in higher dimensions, on the other hand, the mean curvature of the
boundary of the underlying domain is the significant invariant. These various conditions
are here seen as special cases of the criteria developed in ??9 and 10.
In ??12 and 13 we consider a number of interior estimates for the gradient vector,
assuming that a boundary estimate is already available. This work is a clarification and
elaboration of fundamental ideas due to Bernstein; we are in addition able to extend the
results to a large class of equations in more than two variables, an extension which had not
originally been thought possible in view of the inadequate number of critical point relations
available when n > 2. Chapter II closes with a summary of various existence theorems
which are direct consequences of the earlier a priori estimates.
Chapter III deals with certain cases of the non-solvability of Dirichlet's problem for
arbitrary smooth data. It is shown that the results of chapter II are in many ways best
possible, in that when their hypotheses fail to hold one may construct boundary data for
which the Dirichlet problem cannot be solved. Similar resultsfor two independent variables
have been given by Bernsteinl,Leray, and Finn, and indeed our proof method is based on
an interesting form of the maximum principle used by these authors.
54-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
416 J. SERRIN
The paper is concluded witli an extended discussion of various examples and special cases.
These examples possess a considerable amount of independent interest, besides serving as
concrete illustrations of the general theory. It should also be pointed out that, more often
than not, the results of this chapter go beyond mere specialization of some earlier theorem
of the paper. The following proposition may be singled out as indicative of the conclusions
obtained (see also Serrin I967):
Let Q be a bounded domainin n-dimensional Euclideanspace,whoseboundary is of classC2. Then
theDirichletproblemin Qfor theequationof constantmeancurvature
[(1 + IDul2)I-DuDu] D2u = nA(1+ IDu 2) 2

has a uniquesolutionfor arbitrarysmoothboundarydata, if and only if the meancurvature

of the
boundary greaterthanor equalto [nl(n- 1)] IAI.
surfaceis everywhere
In the above example the notation DuDu stands for the dyadic matrix generated by the
gradient vector, while I as usual denotes the identity matrix.
Particular importance attaches to this proposition in view of the geometric interest of the
equation, the paucity of known exact solutions, and the number of conjectures which have
been made concerning the Dirichlet problem in this case. Because of these circumstances
I have included in the final part of the paper a fairly complete discussion of the background
of this theorem, and have also stated generalizations for the case of prescribed (non-constant)
mean curvatures and for surfaces with a single valued spherical projection. Similar theorems
are also given for the closely related differential equation which describes the free surface
of a stationary fluid under the combined action of gravity and surface tension.
Although the paper deals with both uniformly and non-uniformly elliptic equation on
an equal footing, the former case has in fact received such careful and far-reaching study
at the hands of other writers that our conclusions must be regarded for the most part as
a contribution to the theory of non-uniformly elliptic equations. In this regard, it may be
worthwhile to point out specifically the main theorems of the paper, as opposed to the
various methods and techniques of a priori estimation used to attain them. In particular, we
mention the five existence theorems of? 14, the non-existence theoremnsof ??16 and 18, and
the four theorems of ? 19 concerning the equation

-~~~~~~ d(x, u, Du)1)O (1'

As already mentioned, the examples discussed in chapter IV all have some special interest,
either for their geometric significance or because they illustrate some particular point of
the theory.
Theorems are numbered consecutively in each separate section. To avoid ambiguity in
later references to a theorem, the section number is added as a distinguishing mark. Thus
thleorem 1 of ? 8 is referred to in later sections as theorem 8-1.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 417

CHAPTER I

In this chapter is given a precise statement of the Dirichlet problem and the general pro-
cedures to be followed in its study are outlined. The last section contains an important
identity which is basic to a number of constructions later on in the paper.

1. THE FUNDAMENTAL EXISTENCE THEOREM

Let Q be a bounded domain (connected open set) in n-dimensional Euclidean space.

We shall use the notation &Qto denote the boundary of Q and Q to denote tlle closure of Q.
The Dirichlet problem for equation (1) then consists in determining a function u= u(x)
such that:
(i) u is twice continuously differentiable in Q and continuous in Q,
(ii) u takes on given continuous boundary valuesf on MQ,and
(iii) Yu 0 in Q, where the operator Y is defined by
Ylu _ " (x, u, Du) D2u-M (x, u, Du).
In order to establish the existence of solutions of the Dirichlet problem one has available the
classical technique of apriori estimation of the norms of eventual solutions. In particular, the
following important proposition holds.
Let theboundaryof Q and theassignedboundaryvaluesf be of class C3. Also let T be an arbitrary
realnumberin theclosedinterval[0, 1]. Supposethereexistsa constantM, independent of Tr,suchthat
theconditions
(i) v CC(Q))
(ii) v-Tff, on
(iii) v (x, v, Dv) D2v TI14(x, v, Dv) in Q
imply supQ ( Iv+ IDv) 'M. Thenthe Dirichletproblemfor equation(1) in Q has at least one
solutionfor thegivenboundaryvaluesf. Moreover,this solutionis twicecontinuously in
differentiable
theclosureof Q.
The proof of this theorem involves two different sets of ideas-first, an a priori estimate
concerning the equicontinuity of the gradient vector Dv, and second, a topological argu-
ment showing that solutions of (i), (ii), (iii) exist for each value of T, 0 < T < 1. For two
dimensions, such a programme was initiated and carried through by Bernstein under the
assumption that the solutions v are unique. Leray & Schauder (I934) then provided a
general topological frameworkfor the proof, avoiding Bernstein'suniqueness condition and
at the same time giving a wide scope to the allowable homotopy. Moreover, the derivation of
the requisite a priori estimates was greatly simplified by the work of Morrey and Nirenberg;
cf. Nirenberg (I952) . Thus the theorem for the case of two dimensions may be considered
classical.
The topological apparature of Leray & Schauder applies unchanged in higher dimen-
sions, but the a priori estimates concerning the equicontinuity of Dv are now a much deeper

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
418 J. SERRIN
matter. For the case of divergence structure equations these estimates were found by a
number of authors on the basis of work of De Giorgi, Nash, and Moser, and corresponding
existence theorems were stated by Ladyzhenskaya & Uraltseva (i961), Gilbarg (I963), and
others. The first derivation of the requisite a prioriestimates for the general case was given
by Ladyzhenskaya & Uraltseva (i964). The above theorem follows as a consequence of
their work and the Leray-Schauder theorem; for completeness we shall include the proof,
using a method due to Gilbarg.
Let w be an element of the Banach space Cl+'(Q), where y is a real number in the open
interval (0, 1) which will be fixed later. Consider the linear elliptic equation
a(x) D2W - b(x) (3)
where a(x) /(x, w(x), Dw (x)), b(x) - M(x, w(x), Dw (x)).
Since the coefficients a (x) and b(x) are Holder continuous in the closure of Q, Schauder's
theorem implies the existence of a unique solution W E C2+Y(Q)which takes on the given
boundary values f. This process clearly defines a mapping T of C'+7(Q) into itself, namely
w- W - Tw.
Obviously T maps bounded sets in C'+Y(Q) into bounded sets in C2+Y(.Q),whence by
Arzela's theorem and an easy argument (see the appendix at the end of this section) it
follows that T is completely continuous.
The proposition will now be proved if we can show that T has a fixed point. To this end
we apply the Leray-Schauder fixed point theorem in a simplified version due to Schaeffer.
According to that result, Twill have a fixed point if there exists a constant M', independent
ofT, such that v- Tv (4)

implies IIvjl < M'. Here T denotes an arbitrary real number in [0, 1], while the norm of
course is that in the Banach space C1+Y(i).
Let v then be a solution of (4). By the definition of the mapping Tit is evident that v Tf
on dQ and Q (x,v,Dv)D2v-= (x, v, Dv) in Q. (5)

Consequently, by the principal hypothesis of the proposition, we have

supQ(IvI+IDvI) < M. (6)
To complete the proof it must further be shown that the quantity
Dv(x) -Dv(y) (7)
1X-Y 17
is uniformly bounded for all x, y in Q and all T in [0, 1].
In view of (6), the arguments x, v, Dv in the coefficients of (5) are uniformly bounded.
Consequently, as far as the solution v is concerned, (5) may be considered uniformly elliptic
and the functionsv' and -r uniformly of class C1in their arguments. Thus by an important
theorem of Ladyzhenskaya & Uraltseva (1964), there exists a constant y (depending on X
on th1estructure of (1), an on norms for the boundary and the boundary data, but inde-
pendcnt of T) such that the quantity (7) is uniformly bounded. Making this choice of y from
the very beginning thereby completes the proof.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 419
It is of interest to examine how the smoothness of the boundary and of the boundary data
is used in the proof. To begin with, it is required in order to apply Schauder's theorem and
thus to determine the function W; and, secondly, it is needed for certain boundary estimates
in the theorem of Ladyzhenskaya & Uraltseva. Actually, Holder continuity of the second
derivatives of the boundary representations and the boundary data would suffice equally
here, but this generalization would be pedantic. In fact, Trudinger (i967) has shown that
a slightly modified version of the main proposition continues to hold without any conditions
on the boundary data beyond continuity. Since Trudinger's modification requires further
topological apparatus, and since the result as stated is both simple and adequate to our
purposes, it will be unnecessary to consider these generalizations here.
Remark.The above result remains true if we replace conditions (ii) and (iii) by the more
general conditions
(ii') v k(T)f on dQ,
(iii') v'(x, v,Dv;rT)D2v -? (x,v,Dv;rT) in Q,
where the functions '/ (X,u,p; r), 4 (x,u, p;r), and k(r) are continuously differentiable in
their arguments, satisfy 6,' > 0 for 6 + 0, and finally are such that

v(x, u,P; 1) z/(x,

v u,p), X (x, u,p; 1) zzz>(x, u,p), k(l) 1
and M(x,u,p;0)-O for IpI< 1, k(0)=0.
In order to prove this version of the result we consider the Dirichlet problem

V (x, w, Dw; r) D2 W - M (x, w, Dw; r) +X (x, w, Dw; 0) (DwDW1

Dwj2
5
/
with W = k(r)f on the boundary. This defines a mapping T from C1+r(Q) x [0, 1] into
C'+Y(Q), namely (w, r) -? W T(w, T),

and it is required to show that T(., 1) has a fixed point. To this end we observe that T is
continuous and maps bounded sets into sequentially compact sets. Moreover, if v = T(v, T)
then v satisfies (i), (ii'), and (iii'). Hence as in the proof already given, there exists a constant
M' independent of r such that 11 vll < M'. Finally, T- 0 when T 0. It follows therefore
from the Leray-Schauder theorem (see Browder I966) that T has a fixed point when
r = 1, completing the proof.

APPENDIX

We shall show here that the mapping Tw defined above is completely continuous. To
begin with, we recall that T maps bounded sets in C'+7(Q) into bounded sets in C2+7(Q).
Hence by Arzela's theorem it is evident that Tmaps bounded sets in C1+7(Q) into sequenti-
ally compact sets in C1+7(Q). It thus remains only to prove that Tis continuous.
Let w-*wo in C1+Y(Q). We must then show that Tw-? Two in the same space. Suppose
for contradiction that this were not true. Since the elements TWare sequentially compact
in C2(Q?)as well as in C'+7(Q?),it follows that there is a sequence {wJ} such that

-?wQ in C'+Y(Q2), W zzTw -?v in C2(Q?)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
420 J. SERRIN
where W + Two. On the other hand, by (3) we have
_V(x, wi, Dw.) D2 Wm. m(x, WM,Dwm)
whence letting m tend to infinity there results
&? (X,w0,Dwo)D2W -- (x, wo)Dwo).
Consequently W=- Two) and this contradiction completes the proof.

2. THE EXISTENCE PROGRAMME

The proposition which was established in the previous section requires, for its application,
the estimation of the Cl norms of solutions of a family of Dirichlet problems. We propose to
carry out this estimation in a series of steps which are described below and which are more
or less unrelated to each other. If a particular equation possesses an appropriate structure
so that all the various steps in this estimation process can be carried out, then the corre-
sponding Dirichlet problem will have at least one solution.
The existence programme, then, will be based on the following succession of steps:
(A) The maximum absolute value of the solution is estimated in Q.
(B) Assuming the preceding step, one determines an estimate for the gradient of the
solution at the boundary of Q.
(C) An estimate is then obtained for the gradient of the solution in the entire domain,
depending on the results derived in the preceding steps.
Of course, it almost goes without saying that in this programme we are obliged to consider
afamily of Dirichlet problems and that the corresponding estimates must be independent of
the homotopy parameter T. On the other hand, in presenting apriori estimates for the various
steps of the programme it will not be necessary to retain the parameter Tin evidence, so long
as the dependence of the estimates on the structure of the equation and on the given domain
and data is set out in each case.
While there is some tendency for the existence programme as described to appear
fragmentary, I think this need not be taken seriously: the subject itself is broad and simply
does not in the nature of things possess a high degree of unity. In fact, looked at from another
point of view, the requisite classification of non-linear structures in order to obtain
meaningful results has its own particular interest.
The following chapter treats in detail the a priori estimates required by steps (A), (B), and
(C) of the existence programme. Before turning to this, however, we shall derive an
important identity which will be basic to a number of later constructions. This identity is
contained in ? 4; the next section contains certain preliminary material for the derivation.

3. THE DISTANCE FUNCTION

Consider an (n - 1) -dimensional surface S of class C3 in n-dimensional Euclidean space,

hlaving no self-intersections and no boundary. We shall be interested in the function
d-d (x), defined as the distance from x to the surface S. The following lemma holds.
LEMMA1. Let the normalcurvaturesof S be boundedin absolutevallueby K. Then d is of clatssC2 at
all points whosedistancefrom S is less than 1 /K.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 421
Proof. For points x whose distance from S is less than 1/K, we define y = y (x) to be the
(unique) point on S nearest to x.
Now let x be a fixed point whose distance from S is less than 1/K. We consider a specific
coordinate frame in which the xn axis is oriented along the normal to S at y - y(x). In the
neighbourhood of the point y the surface S can then be represented in the form
Xn ' 0 (XI, * - ) Xn-1l)

where 0 is three times continuously differentiable. Moreover, for all points x near x the
following relations hold,

xi Yi- Oi{
+ (y) -2d7 I = l ..S.,n-1; xnx=s( y) +{ ~~~~d
d (8)
where qi d=/dxi, Do is the gradient vector of 0, and y = y(x). (There is a small point of
notation here-the arguments of 0, Xi, and Do are, to be absolutely precise, Yl, )Yn-l )
Relations (8) have the general form
x- f (YI, enYn- 1d) (9)

wherefis a function of class C2in its arguments. The function d(x), which can be considered
as arising by inversion of (9), will then be of class C2 at x provided the Jacobian of the
transformation is non-zero at this point.
In order to compute this Jacobian it is clearly allowable to suppose that the coordinates
xI) ...,)nxI lie along the principal directions of S at the point y. In these coordinates the
gradient Do vanishes at y and the Hessian matrix D20 is diagonal there; consequently, since
the diagonal elements are precisely the principal curvatures of S at -, an easy calculation
yields the expression n-I
I7 (1-kid)

for the Jacobian at x, where the numbers ki are the principal curvatures in question. This
completes the proof of the lemma.
In what follows we shall require formulae for the first and second derivatives of d at the
point x. It will be sufficient to evaluate these in the special coordinate system used in the
proof of lemma 1-that is, with the xn axis taken along the normal to S at y and directed
towards x, and the remaining axes lying along the principal directions of S at -y.
LEMMA 2. In thespecialcoordinate above,we haveat x
systemintroduced
Dd - (0, .0., 0, 1)

and D2d - -k1d' ' 1-k 1d' 0)di

of S at y.
wherekl, ... kn1 aretheprincipalcurvatures
Proof. Put (y, ... Yn-1, d) and write (8) in component form

Differentiating (10) with respect to Xjyields

5S VOL. 246. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
422 J. SERRIN
where the Jacobian matrix (fi, I) is diagonal at x with elements
(1 -kid ...)1lkn d,1).
Evaluating (11) at x, it follows that (61j) is also diagonal there, with elements

I1-kld' I --kn- ldn

The final row of this matrix is just Dd, proving the first part of the lemma.
If we now differentiate (Il) with respect to xmthere results
0 fif, 1, jm +fi, lk 61, j 6k, m
where i,j, m , .. ., n. Putting i = n and evaluating this identity at x yields
0 d)jMr+fnjjm(l-kjd)-l (1 -k1 d)1 (12)
(if eitherj or m equals n, the corresponding factor 1 - kj d or 1 - kmd should be suppressed).
Now the matrix (fn, m)is easily found by direct calculation to be diagonal at x with elements
(kl (I-k1d)) ...,) kn- I(I -kn- 1d)) 0).
Inserting this into (12), the lemma follows at once.
It is perhaps worth emphasizing that the principal curvatures kl, ..., kn 1 in lemma 2 are
those arising when the unit normal to S is directed towardsthe point x.

4. A BASIC IDENTITY
Consider a bounded domain Q whose boundary is of class C3. Let d = d(x) denote the
distance from points x in Q2to the boundary surface. It follows from the resultsof the previous
section that d is twice continuously differentiable for 0 < d < do,where do= 1/K and K is
an absolute bound for the normal curvatures of dQ.
For points x in the boundary strip 0 < d < a, we define
v(x) = g(x) +h(d)
where g is twice continuously differentiable in Q, and h is continuous for 0 < d < a and
twice continuously differentiable with respect to d for 0 < d < a. We suppose that
h' > 0 and a< do)
the latter condition guaranteeing that v is of class C2 in 0 < d < a. Our purpose in what
follows is to evaluate the expression
fv-c5 (X) V, Dv) D2v-r (X, v, Dv).
The result is conveniently expressed formally as
LEMMA 1. strip0 < d < a wehave
For x in theboundary
-Tp = 57--h`1h2
_, Xh' +,QD2g -,V.
Thenotationherehas thefollowingmeaning:
frst,
' =d(x, u,p), S = ~(x,u,p)
where P=~pn?vh', p0=Dg,

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 423
andv is theinnerunitnormalvectorat the (unique)pointy(x) on iQ nearestto x; andsecond,
n-i 1.1.
=(P -PO) a(P-PO) =E1 k.di

wherek1, .k.,kn l and A1, ..,An - are respectivelytheprincipal curvaturesand principal directionsof
LQat y(x).
Proof. In order to evaluate gv at a particular point x in the boundary strip 0 < d < a it
is convenient to introduce new coordinates (denoted with tildes) so that the x- axis coincides
with the normal direction into Q at y y (e). We may further assume that the axes xl, ,xn-
are alined along the principal directions of the boundary surface at y.
According to well known orthogonal transformation formulae, we have
-v - (x, v, Dv) D2v- x, v, Dv)
where a =
and 5 is the orthogonal transformation relating the original and new coordinate systems,
that is

Using the definition of v together with lemma 2 of the previous section, we find easily that

D 2v~D2l h, h'k -hdat

D2v=D2g-(1_kkl dh~ ~~n- kl diagonal

-
Consequently atx, 'v- ZaD2g- - n-Iaik-I2 h'+ h". (13)

Reverting to the original coordinates, one has

/D2g_,QD2g, ,=z@J

and 1jQr=Aj (i 1v.

1, ...,n-1),
Snn -

Moreover, at x the arguments of v and X obviously are x, v, and Dv = Dg + h'Dd = po+ vh'.
By inserting these relations into (13), and noting that v (p -po) /h', the proof is completed.
A slightly different version of lemma 1 will also be useful. We formulate this as
LEMMA2. strip0 < d < a we havealso
For x in theboundary
+ Hh'
Yv =
h", _h +aD2g-.
h'2
of dQ at y (x),
whereH denotesthemeancurvature
n-1i .SA
X = Az ildki + (vqlv) H,

andtheremainingnotationis thesameas in lemma1.

To obtain this result we simply add the quantity k1nHh'to the fourth term on the right-
hand side of (13), and correspondingly subtract this quantity from the third term.
55-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
424 J. SERRIN
The quantity X is a linear combination of the principal curvatures of the boundary
surface. We note particularly the inequality
> ktracenI
where k stands for the minimum (signed) normal curvature of dQ at the point y(x). To
prove this result, first observe that

I >dk.>kk H>,k
id

whence ? v k.

The quantity in parentheses is simply the trace of a? in the special coordinate frame
/11, ., An-15v. Since trace is an orthogonal invariant, the required inequality is proved.
By the same argument we find n-I
1k AiQAj.

Since k 3 - K, where K is an absolute bound for the curvatures of dQ, it follows easily that
f > -Ktrace&I.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 425

CHAPTER II
This chapter treats in detail the various a prioriestimates required by steps (A), (B), and
(C) of the existence programme described in ? 2.

5. BASIC DEFINITIONS AND AGREEMENTS

We consider throughout this chapter a bounded domain Q in n-dimensional Euclidean

space, whose boundary dQis of class C3. It will be assumed further that the boundary data
under consideration is also of class C3.
Under these assumptions, it is clear that there exist C3 extensions of the boundary data
into the closure of Q, while conversely any C3 function defined in the closure of Q generates
C3 boundary data by restriction. In consequence, we maysupposewithoutloss of generalitythat
theboundary datain questionis determinedby a givenfunctionf in C3(Q)).
We let CO= cl SUPWIDf|, , =
5fl,
SUPIfsup cC2 supED2f D

where the norm 11all of a matrix is defined to be the root square of the sum of the squares of
the components of /. We also let K be an upper bound for the unsigned normal curvatures
of the boundary surface. Note that then the distance function d d(x) is of class C2 in the
boundary strip 0 < d < d0owhere do- 11K.
For simplicity in what follows it will be convenient to normalize equation (1) by requiring
trace v/ - 1. (14)
This condition will be tacitly understood in the formulation of all further results. As an
immediate consequence of (2) and (14), the eigenvalues of / must lie between zero and one.
Consequently we have 0</ <1 for J6 1. (15)

We shall be concerned throughout the paper with certain invariant functions, and it is
convenient to place their definition at a central location. In particular, we set
=(x, u,p) - X(x, u,p)/fpf (p + 0),

(X
= x,up) 5 = p V(x5 up) p,

- (x5 u5 P5PO) -(p -Po) (X5 U5p) (p -Po)

It is clear that these expressions remain unchanged under orthogonal transformations of

the underlying Euclidean space.
The invariant 6' was introduced in 1912 by Bernstein in his fundamental paper on the
calculus of variations. In particular, Bernstein considered equations in two independent
variables satisfying the condition
<-
1i.P-khI <- khINI AI -: 1), (I16)

where 4u1and /12 are positive constants and 1is a real number (actually Bernstein considered
onlly integral values of l, but this restriction is unnecessary). It is evident that 1 < 2; the
difference g 2-il was called by Bernstein the genreof the equation.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
426 J. SERRIN
We note, for example, that uniformly elliptic equations have genre zero, while the
minimal surface equation on the other hand has genre two.t In the following work we shall
not, however, restrict considerations solely to equations which have a well-defined genre.
Step (A) of the existence programme is discussed in ? 6.
In ??7 through 11 we present several estimates for step (B), which apply under differing
hypotheses concerning the structure of equation (1). Finally, in ??12 and 13 we treat
step (C) of the existence programme. The last section of the chapter includes various
existence theorems which are direct consequences of the earlier a prioriestimates.
In what follows, a solution of (1) in Q will always be understood to be a function u in
C2(Q)such thatYu 0- in Q. For convenience, we introduce one further notational conven-
tion: if any one of the variables x, u, or p in some relation is not specifically quantified we
assume that its range is, respectively, the closure of Q, the real numbers, or the real n-tuples.

6. ELEMENTARY MAXIMUM PRINCIPLES

We present in this section a number of maximum principles, all but one ultimately based
on the well-known sign contradiction method at a maximum value of the dependent vari-
able. While no claim is made for completeness, we have, nevertheless, attempted to give a
useful selection of results, with the requirements of non-linear equations particularly in
mind. Theorems 3, 4 and 6 seem to be new.
THEOREM 1. Let w(x) beafunctionin C2(Q) suchthat

Y(w)+b) 0 in Q
for all constantsb > 0. Supposealso thatu is a solutionof (1) in Q suchthatlim sup (u- o)) 0 as
oneapproaches anypointof theboundary.Thenu < w in Q.
Proof.Let v - u- o and K -SUp0V. Suppose for contradiction that K iS greater than zero.
Then according to the boundary conditions there must exist a point P in Q where v = K. If
there is more than one such point we fix P specifically so that any neighbourhood of P
contains points where v < K; this is surely possible since v < K near the boundary.
Clearly there exists a neighbourhood N of P where
0< V < K, V 4 K.

Now by hypothesis V (x, w+ b, Do) D20 - SW(x, U + b,Dw) < 0,

this inequality being valid for all x in Q and all positive real numbers b. Consequently it will
hold for any positive function b(x). Choosing b = v in particular, and restricting considera-
tion to the neighbourhood N (hence b > 0), we obtain
J(x,u,Do))D2od - (x,u,D) < 0 in N. (17)
Moreover, since u is a solution of (1),
sd(x,u,Du)D2u- ?'(x,u,Du) 0 in N. (18)
t It is easy to construct equations having any given non-negative genre g, for example,
LI+ (1 + IDuj2)l~~ (iDul2 I-DuDu)]D2u O 0

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 427
Subtracting (17) from (18) and applying the mean value theorem, it is easy to establish that
,1 (x, u, Du) D2v+ linear function of Dv > 0 in N.
But then Hopf's maximum principle implies v -K in N, which is a contradiction. This
completes the proof of the theorem.
The next four theorems in this section are consequences of specific choices of the function
i (x).
THEOREm 2. Supposethat
u9(x,u,0)>0 for lul>M.
Let u bea solutionof (1) in Q, suchthatlim sup lul< m as oneapproaches
anypointof theboundary.
ThenIuI< max (m,M) in Q.
Proof.Let o.(x) = max (m,M). Then
Y( +b) =-A6(x, max (m,M) +b, 0) < 0.
Consequently, by theorem 1, u < win Q. Replacing u with - u we obtain a similar result, and
the theorem is thereby proved.
THEOREM 3. Let Q becontainedin a ball of radiusR, andassumethat
signu.W(x,u,p) > --I1IR
for Jul > M and JpI> l(> 0). Let u be a solutionof (1) in Q, suchthat limsup Jul< m as one
pointsof theboundary.Then
approaches
itj < max (m,M) + 21R in Q.
Proof. Let 0 be the centre of the ball containing Q, and let r denote distance from 0.
Consider the function
(0(x) max (m,M) + IR(e-erIR), r
O< ? R.
By direct calculation, jDo)J ler/R and

= lr
&g(@+b) R(rR
r r2 r )

< Ie /R (- ?0 O<r<R,

where we have written r for the radius vector from 0 and have used the main hypothesis at
the last step (note that t) +b > M and IDo)? 1).
We assert that u < ( in Q. Suppose for contradiction that this were not true. Then by
theorem 1 and the fact that u < eoon the boundary of Q, the maximum of u- must occur
at r= 0. This is impossible, however, in view of the functional form of woat r = 0. Thus
the assertion holds, and
u < ct < max (m,M) + 21R in Q.
Replacing u with -u we obtain a similar result, and the theorem is proved.
COROLLARY. SupposethatA = o( i.PI)as IU|,IPI->-. Let u be a solutionof (I) in Q, such
Thenonecangiveana priori estimate
pointsof theboundary,.
thatlim sup juj ? mas one approaches
for juj in Q.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
428 J. SERRIN
The actual estimate depends on bounds for the order term. Since we shall not need this
result, the details can be omitted.
THEOREM 4. Suppose thatforsomefixeddirectionv andforall p > 0 wehave

d (x, u, p)
o(p), p = pv,
~(x,U,p)
where0 (p) is a positivecontinuousfunction
suchthat
l0 dp
Jop20 (p)

Let u bea solutionof (1) in Q, suchthatlim sup jul< m as oneapproaches anypointof theboundary.
ThenIuj? L in Q, whereL dependsonlyon m, q andthediameterof Q.
Proof.Let L denote the supporting plane of Q which has normal direction v, and which is
such that v is directed toward the side of L containing Q. Also let d denote distance from L,
and let 4 be the diameter of Q.
For any function o(d) which is twice continuously differentiable and satisfies o' > 0 in
the interval 0 < d < A,we have by an easy calculation

T(o + b) e(x,o+b,p) c)"/o)'2- (x, o)+b,p), p o/v.

Now let a and fYbe positive numbers such that
3dp

and let o) and d be parametrically related by the formulae

( = f , +m d= q dp
pPo(p) p ~p2qS(p)
where ca< p <,. We find easily that Co'= p and O" -= O/I20q(o).Consequently, using the
main hypothesis of the theorem, (o)+ b) 0. 0 Therefore by theorem 1
'

u <' ) < m++fin Q.

Since a similar result holds for - u, the proof is complete.

Remarks.A rather obvious generalization of theorem 4 can be given in which the main
hypothesis is stipulated only for sufficiently large values of Jul, and also by using sign
u.M6(x,u,p) rather than JV(x,u,p)j. We also note that a related result was proved by
Ladyzhenskaya & Uraltseva (ig6i, page 79).
THEOREM 5. Supposethat / is independentof u and that d /du ; 0. Let w(x) be a faunction
in
C2(Q) satisfyingthecondition ?)<0 in Q.
Supposealso thatu is a solutionof (1) in Q suchthatlim sup (u- w) < 0 as oneapproaches points
of theboundary.Thenu o ;in I.
This well known result is a direct consequence of theorem 1, for in view of the hypotheses
on a?andSY6we have 92(o ?Sb) ? < 0
for any positive constant be

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 429
There is one further maximum principle rather similar to theorem 5, but applying when
equation (1) has divergence structure. While this result will not be of immediate use, we
shall nevertheless include the proof because of its special interest.
We suppose particularly that (1) can be written in the form
d{Ai(x, u, Du)}/dxi B(x, u, Du)
where Ai and B are respectively of class C2and C' in their arguments. We suppose further
that the (n+ 1) by (n+ 1) matrix
I?l/p Pi dAllaPndAlla8

! AnAPI@.... dAnIPndAnI/U }.....(9

dB/dpl ... aB/Pn dB/du/
is non-negative definite (here (PI, ... ,pn) is the position variable for Du, as usual).
THEOREM structureand that thematrix(19) is non-negative
6. Supposethat (1) has divergence
definite.
Assumethatu E C2(Q) andv E C2(Q2)aresolutionsof (1) in Q, suchthatu- v < 0 ontheboundary.
Then u < v in Q.
Proof. Let K SUP(U- V) in Q, and suppose for contradiction that K > 0. For e < K put
w - max (u-v-c, 0).
Clearly w is Lipschitz continuous in Q and vanishes in some neighbourhood of the boundary.
Consequently a straightforwardprocess of integration by parts yields

{Dw A(x, u,Du) +wB(x, u,Du)} dx 0

and {Dw A (x, v, Dv) + wB(x, v, Dv)}dx 0,

and integrations taking place over the set where w > 0. If we now subtract these equations,
there results

{(Du- Dv) (A(x, u, Du) -A(x, v, Dv)) + (u-v) (B(x, u, Du) -B(x, v, Dv))}dx

-e {{B(x, u, Du) -B(x, v, Dv)}dx - 0.

Letting e tend to zero then yields

{(Du-Dv) *(A(x, u,Du) -A(x, v, Dv)) + (u-v) (B(x, u, Du) -B(x, v, Dv))} dx = 0,

the integration now extending over the set where u > v.

The integrand is non-negative by the definiteness of the matrix (19), and consequently
vanishes on the set where u > v. Now by a well-known application of the mean value
theorem, the integrand can be written in the form

(Du--Dv)3-@(Du --DV)+(DU-DV) *(- (U-V)+a(UV2

56 voL. 264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
430 J. SERRIN
where the notation has an obvious meaning and the tilde denotes evaluation at some
intermediate value of the variables. The matrix aA/ap is positive definite due to ellipticity,
and moreover the intermediate value in question lies in a compact set of the u,p space when
x is in Q. Thus there exists a positive constant Asuch that
(Du-Dv) (dA/ap)(Du-Dv) >? AIDu-Dvl2
for x in Q. Furthermore, by the Cauchy-Schwarz inequality

(Du-Dv). *u+ ?p )(u-v) > -const.jDu-Dvl . u-vl

>- 2A Du-Dvl2-const.(u-v)2.
Consequently we find IDu -Dvl ? const. (u-v)
throughout the set where u > v.
This, however, is impossible. For let P by any point where u > v and let Q be some nearest
poinltto P where u-v. (Such a point exists since u < v at the boundary of Q.) By integrating
the preceding differential inequality along the straight line from Q to P we then get
0 < u(P)- v(P) < [u(Q)-v(Q)] exp (const.IP-QQ) 0,
and the theorem is thereby proved.
Perhaps the most interesting application of theorem 6 is to first-order multiple integral
variational problems in which the integrand F(x, u, Du) is jointly convex in u and Du for
each fixed value of x. For other maximum principles for elliptic equations, see Redheffer
(i962).
7. GLOBAL BARRIERS
Let a, M be positive constants, and let N denote the boundary neighbourhood 0 < d < a.
A globalbarrieris a function v in C2(N) which satisfiesthe condition
Y(v+b)<0 in N

for all positive constants b, and has the form

v(x) =zf(x)+h(x),
where h is continuous in the closure of N and satisfies
h-O when dz=0, h=M when dm=a.
A set of global barrier functions, defined for all M > 0, will be called a globalbarrierfamily.
The following lemma shows that when step (A) of the existence programmecan be carried
out, step (B) can be reduced to the problem of determining an appropriate global barrier.
LEMMA. Let u e C2(Q?)be a solutionof the Dirichletproblemin Q.7Supposethatu < m andthat
functioncorresponding
thereexistsa globalbarrier l
to -m +cI. Then
du/atz?iL on A

~foreverydirection
n intoQ~.HeretheboundL depends
onlyontheglobalbarrier
functionin question.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 431
Proof.Let v -f+Ih be the global barrier function. Clearly u = v on dQ. Also, recalling the
definition of c0as the supremum of If I in Q, we have
u < v when d= va.
Consequently by theorem 61 u < v in N.
It now follows (since both u and v are differentiable on dQ) that
duu dv
on
An :Z:<
)n
for every direction n into Q. Setting L lDvl completes the proof.
maxa0Q

8. REGULARLY ELLIPTIC EQUATIONS

In this section we shall impose an appropriate degree of structure on the coefficient

matrix v/ in order to carry out step (B) for an arbitrarydomain Q of class C3.The definitions
and agreements of ? 5 will be maintained throughout.
We say that equation (1) is m-regularlyellipticprovided that

1+'JlK(D(JpJ) for Jul?m, p, 0, (20)

where <)(p), 0 < p < oo, is a decreasing continuous function satisfying the condition
J~ dp-2@(= 00 (21)
Jp2D(p)
Before proceeding further, it is worthwhile considering this definition in relation to several
particular examples.
1. Suppose that 8 C lpI for lpJ I>1 and that for any m > O
J_J/J?<y(m)log(1+JpJ) when Jul<m, lp1>?1. (22)

Then we can take (D(p) --+Y p ( +P)

p
and (1) is regularly elliptic for each m.
If in particular equation (1) has a genre g s< 1, then it is regularly elliptic provided that
(22) holds. An even more special case occurs for uniformly elliptic equations, these being
regularly elliptic provided that
I.<I?y(mn) |p|2log(1 +JlpI)
for Jul< m and lpI > 1.
2. It may be worth presenting a specific example of a non-uniformly elliptic equation
which is regularly elliptic, namely
(I1+ uX) uxx+ 2uxuyuxy+(I1 + U2 ) uyy 7-
-=

where x and y are independent variables. (The regular ellipticity is obvious since the genre
is zero. On the other hand, the equation is not uniformly elliptic since its coefficient matrix
has the same eigenvalues as the minimal surface equation.)
3. If an equation has genre g > 1 then it cannot be regularly elliptic, as one easily checks.
In particular, neither the minimal surface equation nor the equation for surfacesof constant
56-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
432 J. SERRIN
mean curvature are regularly elliptic. It is this feature which leads to the convexity and
curvature restrictionswhich have proved necessaryfor the solution of the Dirichlet problem
for these equations.
We may now turn to the main result ofthe section, after firstproving a simple preliminary
LEMMA. Supposethat lPoI < c. Then
o- 2C2 g7 <? 2 +2C2.
Proof. Since 2 is positive definite and has unit trace, it is clear that for arbitrary real
vectors 6 and 1 ( < (?+Cp) < 1(1621+ C .

Setting6 - p andg p thenyieldsI'p12 < p,ip. Now

- (P -PO) -'(P -PO) ppp - 2po0Sp+Po0_Po.
For pojI< c we have obviously Jp0spj <?c < cV(p1p). Hence, recalling the definition
1s1pJI
of off,it follows that 8 2cJ/ < Y < 6 +2cJ2 +c2.

The required conclusion is now an immediate consequence of Cauchy's inequality.

THEOREM 1. Let u E C2(Q) bea solutionof theDirichletproblemfor equation(1) in Q.
Assumethat Jul< m in 12. Thenif (1) is m-regularly
ellipticwe have
IDuI< L on dQ
whereL dependsonlyon co,C1,C2 m, K, andthefunction(P. (K, cO,cl, C2aredefinedin ?5.)
Proof. We shall assume to begin with that (20) holds without the added restriction
juI< m; that is, we assume +1 (IPI) (23)

for all x in the closure of -Q,all real u, and allp # 0. This assumption will be removed at the
end of the proof.
Our goal is to construct an appropriateglobal barrierfamily and then to apply the lemma
of the previous section. Let v(x) z=f(x) +h(d), 0 < d < a,

where a < do.We assume that h is twice continuously differentiablewith respect to d and that
h(0) -0, h(a) = M, h'(d) >, a, (24)
a being a positive constant whose value will be specified later. In order to evaluate $? (v+ b)
we may apply lemma 4- 1. Since J >-K, this yields
Y (v+ b) < Yh"/h'2+ Kh'+d D2f- (25)
the arguments of X, X, and F being x, v + b, and p ?-p + vh' where p0 = Df.
Now g > 1/(, and 0 necessarilytends to zero as lpj tends to infinity (since the integral (21)
diverges). Consequently there exists a constant c,, depending only on cl and (D,such that
of 8c for IpI> x1.
Now choose oc= max (cl + ?,1MK, 1) .
Sincep = p0+vh' and jpoJ? cl it is easy to check that

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 433
We are now in position to estimate the coefficients d, 9, and Y in the inequality for
ff(v+b).
To begin with, since the eigenvalues of s all lie between 0 and 1, we have
JaD2f I < 11
D2fll < C2.

Next 11 <_IPIIcel< 2h'Icj,

and finally, using the preliminary lemma and the fact that f > 8c2,

Inserting these estimates into (25) yields

(v +b) < ?h' e3 +4--

)} (c
(Fh Y(3+&p( 1P1 4C2+4K+ 8)

? Yh' 3

the last step holding since (Pis a decreasing function of lp.I

We now choose h' so that the expression in braces vanishes and so that conditions (24)
hold. In order to do this, define fi by means of the relation

aMJ p2(D(p -Cl)

(this is possible since the integral diverges). Next, let h and d be parametrically related by

ch=-,C,dp __ cd= A ,g(dp

I p2$D(p-C)' pp3,D (p -CI)
< p < /,. One easily checks that h' = p ) OGand that the quantity in braces vanishes.
where aG
Moreover, it is evident that h(0) 0, h(a) M,

where a
1 ? dp

(note that a < M/cL< do).

To summarize, we have constructed for each M> 0 a global barrier function with
a < do.By the lemma of the previous section, therefore,
du/dn< L on dQ
where L depends only on the global barrier function corresponding to M = m+co. The
latter depends, moreover, only on the quantities listed in the statement of the theorem.
Replacing u with-u in the equation leaves the construction unaltered. Hence we also have
du/dn?-L on Q

completing the proof of the theorem subject to the initial assumption (23).

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
434 J. SERRIN
If (23) does not hold, we consider a new equation with coefficient matrix a defined by
(v/(x,-m,p) if u <-m,
(x, u,p) (x, u,p) if -m < u <m,
Ii(x,m,p) if u > m,
and with a similarly defined inhomogeneous term M. Clearly u is also a solution of the
equation /D2u= , and this fact together witlhthe evident relation
A

A+SW< 11IPI)) P + O,

allows us to repeat the previous proof word for word, completing the demonstration of the
theorem.
9. BOUNDARY ESTIMATES DEPENDING ON CURVATURE

In this section we consider step (B) for equations which are not regularly elliptic. In such
cases it turns out thlatthe curvatures of the boundary surfacc must be restricted in order to
obtain boundary estimates.
We shall consider the following conditions on the asymptotic behaviour of the invariants
a and r' for large values ofp, namely
'(x, u,p) -- X O)+o(1),
O(x, o-- P/IP1) (26)
and (x, u,pP) (x,u,) o-1- (1), = P/I PI, (27)
as p -> oo, where 4 0(x, a) and '0o(x,u, a) are continuously differentiable functions of their
arguments and wlhere aIor/au -> O. (28)

for (x, u) in
In these relations, as well as later in the paper, ordertermsareassumedto beuniform
any compactset, unlessotherwisestated. We emphasize that condition (28) will be tacitly
assumed whenever we deal with the asymptotic relation (27).
The matrix di% can be used to introduce a generalized mean curvature of the boundary
surface. Let y be a point on the boundary and let v be the unit inner normal at y. Also let
k ..., kn-1 and Al, ...,) An- I bc respectively the principal curvaturcs and principal directions
of Q at y. We then put A -A(y) - -1o(y, v)
n-I
and define '= (Y)) - (AiAAi)ki+ (vAv)H,

where H is the ordinary mean curvature of dQ at y. The function which arises upon
replacing v by - v in this construction will be denoted byf (y).
It is worth emphasizing that cJ#'and J are averagesof the principal curvaturest and that
both quantities are orthogonal invariants, exactly as is the case with the mean curvature.
n-I
t That is, - E aiki where
ai >, O, Za, = 1, )

with a siinilar relation holding for ,X. Conditions (*) are a direct consequence of two facts, (i) the numbers
(A1 AA1)and vAv>are the diagonal elements of the matrix A in the coordinate frame A1, ..., Anl vo,and (ii)
the matrix S% necessarily has unit trace and is non-negative definite.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 435
Moreover, if the matrix A is proportional on dQ to the Euclidean metric tensor, then
H. This condition automatically holds in two dimensions, so that X K in
this case, where K iS the curvature of the boundary curve.
To illustrate the constructionsofs0,
Q0, X andy, consider in particular the equation for
surfaces of constant mean curvature in n dimensions, namely
(1 + IDul2)Au-DuDuD2u- nA(1 + IDul2)1.
After normalizing to unit trace, we find
o-o- ~ n
0= n-i' 0n-iA;
moreover, the error terms in (26) and (27) are of degree 2. Evidently AiAAi (n- 1)
and vAP= 0, and we have therefore t = H

(this also follows from the fact that A is proportional to the Euclidean metric on an).
THEOREM 1. Let u E C2((Q)be a solutionof theDirichletproblem
for equation(1) in Q.
Assumethat IuJ? m in Q. If (26) and (27) hold,andif bothconditions
' >o(Y,f, v), X > Wo(y,f, - v) (29)
aresatisfiedat eachpointy of theboundary
surface,then
IDuI L on dQ
whereL dependsonlyon co, c1, c2,m, K, a lowerboundforthedifferences
in (29), boundsfortheerror
termsin (26) and (27), andC1normsof thefunctionsi0 and 0%
Proof.According to the argument used in the final step of the proof of theorem 8-1, we
can assume without loss of generality that the error terms in (26) and (27), as well as the
norms of c/o and W0,are uniformly bounded with respect to u. As before, we shall carry out
the proof by constructingan appropriateglobal barrierfamily and applying the lemma of ? 7.
We put v(x) =f(x) +h(d), 0 d a

where a < do. It is further assumed that h is twice continuously differentiable with respect
to d and that h(0) =0, h(a) M, h'(d) ? a, (30)

a being a positive constant which will be determined later. By lemma 4-2

Y (v + b) = h"?Hh' _ fh' ?&/D2f-Y

the arguments of /, M, and F being x, v+ b, andp -p0+ h'.

Let 0 > 0 be chosenso that t > (Sf( ) + 50
at each point of the boundary surface (this can surely be done since the terms in (29) are
continuous functions of y).
An easy estimate yields lpI >? ?- cl and

IKd(x, v+b, p) -dr/(x, Y) Jl<I(,v ,p)d0(x, 1+ sup a?{-

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
436 J. SERRIN
Moreover, if y -y(x) denotes the point on the boundary nearest x, we have

II_QO(x,v) -do(y) P)JI<- dsupjSax?|

Consequently, when a is suitably large and d is sufficiently small, the difference

dV(x,v+ b,p) -A (y)
becomes arbitrarily small. In particular, for a ? a, and d < al,
n-I
X >E (AisAi) ki +(viv) H > X-8. (31)

Here al and a, depend only on 0, cl, K, bounds for the error term 11j4-,Vol, and a bound
for the C' norm of -d0.
Similarly, for a suitably large we have
6(x,v+b,p) > leo(x,v+b,v) -8 > lo(x,f(x),iv) -0
since Te is increasing in its second argument. Since IDf ? cl we may assume further that

to(x,f(x), v) > (o4(y,f(y), v) -8 for d A a2.

This leads to the required estimate
,4(x,v+b,p) -
fpjV'(x,v+b,p)
> lP{IWO(Y,f,1') - 20} > hI'{O(y,f v) - 30} (32)
provideda > a2and d < a2.
Now choose _ max(1 2,M/al, M/a2, MK,c2/O).

Then according to (31) and (32) we have (provided a < mmn(a1,a2,do))

h
(v+ b)< h- +Hh'2+h'{-X'-0o(y,f, v) + 508}< k+'2

where at the last line we have used the definition of 0 and the fact that H < K.
We now define 1-eKd 1 MK
h(d)=MIe-K where a=-log l+a.

It is easy to check that h"+Kh' = 0 and that conditions (30) hold. Moreover,
a < min (al, a2,do), so that finally
Y(v+b)?< O for O<d<a.
The rest of the proof is almost exactly the same as that of theorem 8 1. The only significant
change is that when u is replaced by - u we must simultaneously replace d (x, u,p) with
d!/(x, - u, -p) and R (x, u,p) with -S (x, - u, -p). This accounts for the second condition
in (29), in which X has been replaced byl and 'o (y,f, v) by - o (y,f, -v).
Remark.It is clear from the above proof that a result similar to theorem 1 can be obtained
under a milder structure condition than (27), namely

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 437
In this case (29) should be replaced by the pair of inequalities
r>-V'6'-(y,f,v) and _>>'2(y,'f-V).
The rather more complicated nature of these conditions is the main reason for emphasizing
the structure (27) and the condition (29).
Even with no structure imposed on the coefficient matrix, it is still possible to obtain a
result of interest.
THEOREM 2. Let u E C2(O?) be a solution of the Dirichletproblemfor equation (1) in Q.
Assumethat ju < in in Q. Then if (27) holds, and if theminimumnormalcurvaturekof the boundary
surfacesatisfiesthe condition k > T-F'(y,f + v) (33)
at eachpoint of y of the boundary,we have
IDuj L on dQ
whereL dependsonly on co, c1, c2, m, K, a lowerboundfor the differencesin (33), a boundfor the error
termin (27), and the CI normof thefunction %.'0
The proof is exactly the same as before, except that we use the relation AX? k in place
of the more difficult inequality (31).
Comment.Tlheorem 2 includes as a special case the boundary estimates used by Gilbarg
and Stampacchia in their study of non-uniformly elliptic equations of the form
d{Ai(Du)}/dxi = 0 Ai E (C2
In fact, when this equation is written out we have X 0, so that (33) becomes precisely
the convexity condition imposed by these authors.

10. BOUNDEDLYNON-LINEAREQUATIONS
With a very slight strengthening of the asynmptoticstructure of a? and c69, the equality sign
can be allowed in conditions (29) and (33). Although at first glance there in no great gain
in this, the fact that a number of important examples can be included here makes the
stronger result well worth stating.
We say that equation (1) is (c, m)-boundedlynon-linearprovided that (26) and (27) hold and
=zw +o+1(6-~?eIPK'
I -1olI o+ <pII<_( IPI) (34)

for Jul ? m, lpol < c, pI > c, where 4)(p) is a decreasing continuous function satisfying the
condition y0 dpIp
1.+ep2q,(p)

Let us consider some particular cases.

1. Suppose that .F ? 4#(m,c) > 0 and
1L-doll+ It-eVol log (I + pI)
y (M, c) T
when p[ ? 1 +c. Then we canPtake

@()_ 4u'.?+yl,og (i,P,) p>1+

and (1) is boundedly non-linear for each mand c.

A7 VOL. 264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
438 J. SERRIN
If equation (1) has a well-defined genre g, 1 < g < 2, then it is boundedly non-linear
provided 11 -53?oll + 1W-eoI < y(m) lp1I-5 for lul < m, lIP > 1

(the lemma of ? 8 implies9 1pI2-g for largep). Conversely,if g > 2 then (1) cannot be
boundedly non-linear.
2. The case where (1) has genre two is particularlyinterestingin view of the large number
of examples of this type, including the minimal surface equation and the equation for
surfaceshaving constant mean curvature. The following lemma states a sufficient condition
for an equation of genre two to be boundedly non-linear.
LEMMA. Assumethatequation(1) has genretwoandthat
J/-Ido =--- ?( lP-)n le-lo - ?( IP-1)
asp tendsto infinity.Then(1) is boundedly
non-linear.Iffor largep theleasteigenvalue
ofa? is bounded
belowby a positivemultipleof lpl-2, thenwe can replaceo( Ipt ) by O(IpK')in theordercondition
for /-a o.
Proof.In view of the result of paragraph 1, it isenough to show that,Y > # for IP > 1+C.
Consider first the case where v/ -sdo o( pI-1). From (26) we have obviously
[alo?o+(1)] P12.
=-
Since g = 0(1) by the genre assumption, it follows that orsI
ao= 0. Moreover, 1/ 0 must
surely be non-negative definite. Consequently we have
- (P Po)(P Po) > (P-Po) (ddi- 1o)(P-Po)

Now of > ,ulfor IpI 1. HenceF>Y 21, for all sufficiently large p, and the required con-
clusion follows at once from the fact that Y is positive and continuous for c < PI1 < ?.
When the least eigenvalue of,s1 is of the order Ipt 2, we have obviously

= (p -po) >d(p-po) > const. lP-pl12

and the final part of the lemma is an immediate consequence of this relation.
3. We have remarked in the previous section that the error termsQ9/-0 and '
-WO for
the minimal surface equation and the equation for surfacesof constant mean curvature are
of degree -2. By the above lemma, therefore, these equations are boundedly non-linear.
We now turn to the main results of the section, which slightly extend the conclusions of
theorems 9 I and 912. Note that the expressionsX* and X were defined in ? 9.
THEOREM 1. Let u E C2(Q?)be a solutionof theDirichletproblem
for equation(1) in Q.
Assumethat jut < m in Q, andthat (1) is (c,, m)-boundedly
non-linear.Thenif bothconditions
X ) To(yf)
Y), X > To(ynfn v) ~(35)
holdat eachpointy of theboundary
surface,wehave
tDuj?XL on Z
whereL depends
onlyon c0,cl, c2,m,K, thefunction4)(p), and C' normsof thefunctionsd andT0

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 439
Proof. We may suppose without loss of generality that
Ila--oll + Il -1c, (I)
tol+ IPI_ ( P>1
We carry out the proof, as usual, by constructing an appropriate global barrier family and
applying the lemma of ? 7.
Choosing v in the form v(x) --f(x) +h(d), 0 < d < a,

where a < doand where h satisfies the conditions

h(O) O, 0 h(a) - M, h'(d) > cx, (36)
we find by lemma 4-2
h"l+Hh'
Y(v+b) =,? ^ h2h_h'"t + aDY X

where the arguments of a, M, and Y are x, v+b, andp = po+vh'.

As in the proof of theorem 9 1,
Jj(x, v+b,p) -A(y)jj 1 o + 2c1sup > +dsup a?70

It follows (see (31)) that

X'>) - -,\In KI d1_01 oll - const. (lpI+ d).
Similarly (see (32))
_V(x,v+b,p) ) IPI{To(y5f, v)- Le-ol-const.( Ipl-'+d)}
> h'{60(y5f, v) -2 j - jol -const.(pj-K1+d)}
provided aX> cl. (The constants in the last three lines are of course not the same, though all
depend only on cl, K, and the C1 norms of the functions q0 and '60.)
Before inserting the preceding estimates into the identity for Y (v+ b), it is convenient to
obtain a bound for din terms of lPI.To this end we assume h" < 0, a condition which will be
satisfied by the eventual choice of h. Then, for ox> cl,
d? 2hlpl1-'d? 2MIPI'-
since by the convexity of h we have h'(d)d < h(d)-h(O) < M.
With the help of the above estimates, it follows that

Y(v+b) < ! - h'{-X- o(y,f v)+2C2IPI

+const.(1 + M) IPIK-}
+V/nK{1-voII + 2 j'&-Tolj
provided a'?> cl. Using the principal hypothesis (35) and setting
c = 2 + 2c2+V\nK+ const.(1+ M)

yields finally (v+ b) < h'i +C (k{p+ P1

j )))

h'Y {(:h-3+(h72?(k-ci)}
since <I'is a decreasing function. (It should be noted here that the constant c depends only
on cl, c2,K, M, and C1norms of the functions a? and T0.
57-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
440 J. SERRIN
Now choose a max (cl, MK)
and define / by means of the relation
cM
= fI + (p
p2@dp
_
ci)
(this is always possible since the integral is divergent). The function h(d) is finally deter-
mined parametrically by the formulae

ch X
J I +p2$D(p-c) cd= p p[1 + p2$(rpc1)]

where oa< p f,. One easily checks that h' - p ao and that (36) holds with
I fl
- - dp
a
c f 0p[I + p2 (p c)]
(note that a < Mla < do). Moreover
h -=ch'[1+h'2$(h'-c)]
so that h" < 0 and Y(v+b) < 0.
The remaining part of the proof is by now standard, and consequently may be omitted.
THEOREM 2. Let u E C2(Q) be a solution of the Dirichlet roblemfor equation (1) in Q.
Assumethat Iul m in Q, and that (34) holds with the term11K-Vol omitted.Thenif the
minimumnormal curvaturek of the boundarysurface satisfies the condition
k > T le (y,f, + v) (37)
at eachpoint y of the boundarysurface, we have
Dul AL on dQ
whereL dependsonly on c0, c1,c2, m, K, thefunction42(p), and the Cl normof thefunctionc60.
The proof is exactly the same as before, except that we use the relation X ? k in place of
the more difficult estimate for S in terms of X.
The minimal surface equation is boundedly nonlinear, as we remarked earlier; more-
over, in this case X# X H, where H is the mean curvature of the boundary surface. It
is therefore evident from theorem 1 that step (B) in the existence programme can be carried
out if H>O

at each point of the boundary surface. Now step (A) is trivial for the minimal surface
equation, and step (C) follows from the fact that the derivatives of a solution satisfy the
maximum principle. Hence the Dirichletproblemfor the minimalsurfaceequationin a C3 domainQ
is solvablefor arbitrarilygiven C3 boundarydata providedthat H > 0 at each boundarypoint. This
result was first proved by Howard Jenkins and the present author, and in fact holds equally
for C2 domains and continuous boundary data (see ? 19).

11. MIXED EQUATIONS

It may happen that (1-) is regularly elliptic for certain directions of the vector p and
boundedly non-linear for other directions. As an example, consider the 'false' minimal
surface equation (1 ?u)-( u)~ 0(8

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 441
where x and y are independent variables. Putting p -(P1,P2) and recalling the normaliza-
tion to unit trace, we find easily that
_ jpI2+4p2p2
IpJ2+2
If the vector o = pl p is bounded away from both coordinate directions, and lpI ? 1, then
1
1IPI2 <, P2 IpI2

for appropriate constants ,ucand 12. Thus (38) is regularly elliptic for these directions.
When a-lies in a coordinate direction, on the other hand, then g < 1 and the condition
of regular ellipticity fails. The condition of bounded non-linearity holds, however, since
(by an easy calculation) 1
> 2+ (1 +c)2 for pi> 1 +c.
In the following theorem we present a fairly simple result indicating the type of conclusions
which can be drawn in such mixed cases.
THEOREM. Let u E C2(Q) be a solutionof theDirichletproblem for equation(1) in Q.
Assumethat UI u m in Q, that (1) is m-regularlyellipticfor all valuesof p whosedirections
o- plp/Iplie in someopenset R of theunitsphere,andfinallythat(1) is (cl, m)-boundedly
- non-linear.t
Then if(y,f,v) > X -I) o(y

at eachpointy of theboundary
surfacewherev 0 R, we have
IDuI< L on dQ
whereL dependsonlyon co,c1, c2, m, K, thefunction42(p), andC' normsof thefunctionsdo andTO.
The proof is a direct combination of the demonstrations of theorems 8 1 and 10.1. The
assumption that (D(p) is the same in both conditions (20) and (34) is used in order to
construct a single function h(d) satisfying the requirements of each proof separately.
Returning to equation (38), let R be any fixed set {flo-,j> c, 1o-21 > e}, where c > 0. Then
1 <const. forR
e _pI fr pl ---i, oe

while clearly L Q/-011/+IPI <const. for p->, +c.

Choosing (D(p) const./p, p > 1 + c, it is apparent that the hypotheses of the theorem are
satisfied. Since steps (A) and (C) are trivial for the present case (for example, use theorem 6*2
and the fact that the first derivatives of a solution of (38) satisfy a maximum principle), we
have therefore proved the following result:
TheDirichletproblemfor equation(38) in a C3 domainQ is solvablefor arbitrarilygiven C3
boundarydataprovidedthatthecurvature K of the boundary curveis non-negative in theneighbourhood
of eachboundary point whosetangentlies alongoneof thecoordinate directions.
The condition that the boundary and the data be of class C3 can be lightened to class C2
by an approximation argument; we shall, however, omit the details.
t We suppose the function <t>(p)to be the same in both conditions (20) and (34); doubtless this assumption
could be removed by a suitable modification of the proof.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
442 J. SERRIN
It is interesting to contrast this result with the corresponding theorem for the minimal
surface equation. The rather trivial reversalof the sign of the middle term in these equations,
though leaving the eigenvalues of the coefficient matrix unaltered and apparently affecting
the structure only slightly, in fact has a profound effect on the type of Dirichlet problem
which can be solved.

12. INTERIOR ESTIMATES FOR IDu

In contrast to the work of preceding sections, the derivation of interior estimates for the
gradient of a solution naturally divides into two cases, depending on whether n 2 or n > 2.
For n = 2 the results go back to Bernstein; we shall follow his ideas in outline, though not in
detail, when treating that case.
We first introduce some notation. If f is a given scalar function of the variables x, u, p
we shall write

YX= (ad/x1, ..., df/dx), 1K=i/{p / ('dp, I ..., dll/Pn)O

Also forp + Owe put = i(x, u,p) = a7jx+ IPI U (39)
where oa=r-pl pI.When IF is a vector or matrix function it is convenient to retain the same
notation, with an obvious interpretation.
We shall require two further invariants of (1), namely
l
61and =_e,*)- P#.
p + 0
*=IpI2 In

Note that g* < 1 since trace v --1. For the remainder of the section it will be supposed
that n 2. The following result then holds.
THEOREM 1. Supposethat > 0 for PI > L'.

Letu E C' (Q) n C3(Q)bea solutionof (1) in Q, suchthatIDuI < L on MQ.ThenlDuI < max (L, L')
in U.
j
Proof.Put w IDul2 and let Q' denote the open subset of Q where w + 0. We claim that
,iD2w 2(1 &*)(9+92)w+4 Dw, x eQ'
where the arguments of Q/, ?, and off* are x, u, Du and where A- =: A (x) is a continuous
function on Q'.
Before proving this identity, we note that the result of the theorem follows from it as an
immediate consequence of the strong maximum principle.
In deriving the identity, it is convenient to use standard component notation for partial
derivatives. Writing (1) in the form =

and applying the operator Uk(alxk) to both sides, there results

dEjUkUijk?I|Du|IijiJU+ <UkUklUiJ=| lDul +? ap UkUki.

Now w-uk Uk, whence wi 2uk uk, wi - 2UkUkfJA2UkiUki. (40)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 443
Inserting these expressionsinto the preceding line gives

2ij Wij
~~~Dul ~~~~~uj
-4?ij UikUjk + Il8
(-D ?jij
+ 1w1
2w (y
-
- ~~u(41) (41
ap i-ij

Let us further temporarily introduce the classical (p, q, r, s, t) notation for the partial
derivatives of functions of two independent variables; that is, we put
p = u1, q-U2, r-u11, S = U12, t- 22

Then clearly pr+qs = lw1,

ps+qt 2w2- (42)

Ar+2Bs+Ct - D,)
the last line simply being equation (1) in the classical notation. Setting
E-Aq2- 2Bpq+ Cp2
and solving the linear system (42) for r, s, t, we find that
r- q2D/E? linear function of Dw,
s =pqD/E+?linear function of Dw,
t p2D/E+ linear function of Dw,
valid for p2 + q2 = 0. From these results, it follows that
aj U DE/E+ linearfunction of Dw
and i A (r2+S2)+ 2B(rs+st) +C(S2+t2)
_1ijUkiUk
- (r+t) (Ar+2Bs+Ct)+(A+C) (S2-rt)
wD2/E+ linear function of Dw.
Inserting the preceding expressions into (41) and rearranging terms proves the required
identity (note here that D =s2 and E (.p2+q2) (1I *)).
For the special case v -((p) the result of theorem 1 was obtained by Bernstein (I9IO,
page 129; see also Leray I939, page 255).
Theorem 1 gives an important method for dominating the gradient of a solution in terms
of the gradient at the boundary. Bernstein noted (I9I2, pages 459-461) that the result
could be further refined by a change of dependent variables, though in order to carry out
the construction in this case it is also necessary to have an a prioriestimate for the modulus
of the solution itself. Because Bernstein'sproof contains a seriousinaccuracy (i.e. formula (7)
on page 459 applies only if the coefficients of the second derivatives are independent of
(x, y, z), while on the other hand in the application of this formula on page 460 the coeffi-
cients of the transformed equation definitely involve the variable z), it is worthwhile
presenting a complete discussion here. There are two essentially different situations to
consider, roughly corresponding to the case when (1) is boundedly non-linear and when it
is regularly elliptic. The results may be compared with those of Leray (I939, page 263).
THEOREM2. Supposethatn 2 andthatfor someconstanti>O0the
invariantfunction&6satisfies
theconditions so-<(l1 ). g 1 ()3

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
444 J. SERRIN
largevaluesofp, say IpI> 1. Assumethatas p tendsto infinity
at leastforall sufficiently
6xI, Gun (P 0 ) (44)
and that (27) holds with the remaindera - ' -W satisfying

- ?(g/IPI(45)
X, SU, IPI SP
,-V
Let u c C1(Q) n C3(Q) be a solutionof (1) satisfyingIuI< m in Q and IDul L on dQ. Then
IDuIAM in Q
whereM dependsonlyonpu,m, 1, L, boundsfortheordertermslistedabove,andon theC1normof the
function Te.
Remarks.Conditions of the type (43), (44), (45) are consistent with the assumption that
equation (1) is boundedly non-linear, and indeed one of the main applications of theorem 2
will be to the equation for surfaces having prescribed mean curvature.
Although the hypotheses appear somewhat complicated at first sight, the reader will
notice that they amount roughly to the statement that differentiation of the invariants
e and R with respect to x and u does not alter their order, while differentiation with respect
to p operates as if the functions were rational. Thus conditions (44) and (45) are in fact
'natural' assumptions in the sense of Ladyzhenskaya & Uraltseva. This being the case, it
is apparent that the crucial additional requirements for theorem 2 are the existence of an
interior estimate for Iu and the condition p - < (1 -,u) J.
Proofof theorem2. Since IuI< m, we may assume without loss of generality that the order
terms in (44) and (45), as well as the norm of the function W0,are bounded independently
of u.
Now introduce a new dependent variable uiby means of the transformation
u-~9 t(U) (99'(U) > 0).

One checks without difficulty that uisatisfies the equation

-gD2ii (46)
with z=zs and R ' (+ ),
where the arguments of 4, -, and 6' are x, u, p - p'(iu)p and where
(0- "1/'2

In order to apply the result of theorem 1 to the new equation, we note that
W= z +jIpI, g * =g*

Moreover, recalling that p depends on -, an easy calculation yields

IsIweI
fi, w rAX+
r
IatPI
a a a
ThuX+ (4P6)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 445
where a is an abbreviation for the differential operator written out in full on the preceding
line. The expression 9+-2 will be non-negative provided that

( (p
) + IPl) + ( + IPI ) (47)

We shall show that (47) holds for all sufficiently large values ofp when

f (u) (l/a) logu

and a is chosen to be a suitably large constant. For this choice of p it is easy to see that
U a constant.
Now M?= -6x+ lP1e-a lplp %ewhence (44) and the first inequality of (43) imply
-(C+ (I 1-)a) lpIoff<, M < C(l +a) Ip2(48)

for all p sufficientlylarge, say lPI > a,; here we use the letter C as a generic bound depending
only on the order terms in (44) and (45) and the norm of the function Te. Similarly, from
(45) and the second inequality of (43),

= A ++ cr dro
X ;
> 0f.x+0 a -C(&?I + I +a) e (49)

> 0 andp- del0/dp 0. From (48)

provided IPl a2; here we have used the facts that d?el0/du
and (49) we obtain 6(c+ae/jpD) +-1)C)?
(pa2_(I+2a

and also, since Meg*=pj-26+201 lp I

((1 +#) a-C) 'p < M*

5 < (2pa + (I +a) C) e

We now write V + agl lp I=o + (,I + a/ lp) and estimate the two parts separately: thus

leo> - C Mt+ p > aC)

- rp-,
for IPI>- a%3
Let us assume a final choice of a satisfying a > C. Using the estimates of the previous
paragraph, the function I defined by (47) then obeys
I> e[(1-e*) (,ua2_(I+2a+1,c1) C) +*(a-C) ((1 +p)a-C)-(2aC+ (1+a) C2)]
provided aP max (2l Z2,a3, 1) L'. Assuming u < 1, as is certainly allowable, it is easy
to see that

Now choose a (necessarily greater than C) so that the quantity in square brackets vanishes.
We have thus shown that 9+ 2 ?> 0 provided 1p > L'/min q/. Therefore by theorem 1
LmmaxT(L,eL')/min'--9
and hence jDuj? maiDx max
I (L, L') / (L, L') exp (2am),
max

completing the proof of theorem 2.:

58 Vou. 264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
446 J. SERRIN
Remark.The hypothesis & ->,cc was used during the proof only to estimate the term
aTho/dxin inequality (49). Consequently, this condition can be dropped from (43) if 6%is
independent of x.
The following result is the analogue of theorem 2 for the case of regular ellipticity. Again
the crucial assumption is an a prioriestimate for Iu
THEOREM 3. Supposethatn - 2 andthatforsomeconstant e * satisfies
#u> 0 theinvariantfunction
thecondition * <I -. (50)

Assumefurtherthat ,G e O(of) (51)

and Gu p, u t = (?( pI) (52)
Let u E C1(L?)n C3(Q) be a solutionof (1) satisfyingIuI< m in Q and IDulI L on 3Q. Then
IDuI< M in Q,
whereM dependsonlyonp, m, L, andboundsfor theordertermslistedabove.
Proof.As in the preceding proof we can assume without loss of generality that the order
terms are bounded independently of u. Moreover, as in that proof, it will be enough to show
that (47) holds for an appropriate choice of the function p (iu) and for suitably large
values of p.
Now assume w > 0. Then
w > C(IpI-'+w), 1&e <C(1+? lpI)
for IpI > a,, where C is a generic bound depending only on the order terms in (51) and (52).
Consequently, since O
SZ Ipj , and &1pJN_ Np)IpIN+l

we have S + )> -1+ 0,)C)

+ _ (I+ () ( IPI
and 1e*I < (20+ ( I-1+oi)

Since also W+sIP <(C+60l (IPIE2)

we obtain the following estimate for I:

I > e[(j_e) (0)1/99+0)2_(1j+60)( IPI-1+6o)C) (C+6o) (2o+ ( IPj-1+o) C)]
e [6'19/' -A(o?2+6+ pl-1)]

where in the last step we have used (50) and put

A= (2+C)2/P;
here IpI > max (a1,Ia2, 1) .
We now choose 9(-u)according to the inverse relation
m1
r<P dt
Since u varies in the range (- m, m) the integral is clearly well defined and q'(ul) >- 0.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 447
Setting n-I one finds
p-rm
799 (1-e2AVf) /2A, w (e2AR 1)l

and ' '- 2A((02 + N)

Substituting this into the preceding inequality for I yields then
I A
Af' (?2+0- lpI-1)> Aflg(e2A3-lpI-1)
Therefore I > 0 provided
IPI >
max (061, 1)) -L.
06251, e2A(2m+
The rest is similar to the proof of theorem 2. Indeed with the help of theorem 1 we find

IDul < max 99max(L,L') < 2max(al, 2, e45( ?m),L)

and the demonstration of theorem 3 is completed.

Comments. By following the steps of the proof it is not hard to show that the theorem
remains true if the order conditions on Muand 5uare weakened to
_> ?o(e)5 9u o (9)
au

Indeed, with this change the final inequality for I takes the form
I>Ape(e2AVfo(l));
thus again I > 0 for sufficiently large p, and the required conclusion follows immediately.
Note finally that condition (50) is satisfied for any equation with a well-defined genre
g > 0, or, more generally,wheneverthe functione satisfiese i Jpt2-0 J for 0 > 0.

13. INTERIOR ESTIMATES FOR IDul IN HIGHER DIMENSIONS

The estimates of the preceding section apply only in the case of two dimensions. While it
would be highly desirable to have fully correspondingresults for higher dimensions, we see
no way to obtain such conclusions. Nevertheless, it is possible to state analogous theorems
for certain special classes of equations.
In this section we consider equations for which the coefficient matrix has the representa-
tion vQ(x,, u, ,b) = (X, u,p) a?' (p) + C i (x, u, )PP,
u,p) (54)

where 9 and Wiare scalar functions and dW'is a positive definite matrix. It can be assume(d
without loss of generality that trace =w' 1, (55)

if necessarily by multiplying 9 by an appropriate factor. Evidently W > 0 since za/is positive

definite.
A large number of classical non-linear partial differential equations fall into the class of
equations defined by (54). In particular, this representation is obviously available when-
ever v is independent of x and u. The representation also holds for the Euler-Lagrange
equation associated with regular variational problems of the form

u, jD)uj)dx = 0, F E C3.
S&JF(x, (56)
58-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
448 J. SERRIN
Indeed, a short calculation shows that in this case
11 G
i+Gn+l+G '

where G - 2 pI2F"'/nF'and the primes denote differentiation with respect to p 2* Thus

d' I/n, =l1/(1+G), and 91=G/(1+G) pl2
for the Euler-Lagrange equation in question.
Now putting y =el p +0
and recalling the definition (39), we have the following result.
THEOREM 1. Supposethat (54) holdsandthat
?+/y2?O
for IpIL'
Letu E C'((Q)n C3(Q)beasolutionof(1) inQ,suchthatIDuI< Lon Mi. ThenIDuI< max (L,L')
in Q.
Proof.Put w = IDul2 and let Q' be the open subset of Q where w # 0. We claim that
s/D2w > ;S(9w+922) w+iX Dw, x E Q'.
where the arguments of /, C, and Y are x, u, Du, and where .IV'- (x) is a continuous
function on Q.
To prove this inequality, we may assume without loss of generality that =1 by dividing
both sides of (1) by W.Of course the normalization traceQ/ =1 is then lost, but we shall not
in fact require this for the proof.
With the new normalization in mind, we now apply the differential operator ukd/ax,kto
both sides of (1), as in the proof of theorem 12 1. This yields easily

i k Uijk +1 gDuI S*lUi Uija+ pUkUklUij =

|DU| -
+1 UkUkl.

Using (40) it follows that

l wijWDu|
W uui+ X +linear function of Dw.
Now we have ZZ Uijk
Uj(57)

(introduce new coordinates so that (uzj)is diagonal; then

-
(a, uji)2 u(i)2 < (Edvi) (Xi'i U2) =ju- u
which proves (57) since both sides are invariant under orthogonal transformations).
Furthermore d'. ui = i uij + linear function of Dw
a j j?+linear function of Dw. (58)
Using (57) and (58) to eliminate /sM%Ujk from the identity for dsVjwijthen gives
2 Du~IX +linear function of Dw, (59)
and the required inequality follows at once.
The rest of the proof is an immediate consequence of Hopf's maximum principle, as in
tnhecase of theorem 12*1.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 449
In order to obtain a result corresponding to theorem 2 of the preceding section it is
necessary to specialize still further. We shall in fact suppose that (54) holds in the form
/ (x, u,p) - (x,u,p) a' (O.)+ (x,u,p)pp (60)
where o = p/lIp. Such a representation applies in particular to regular variational problems
(56), as we have noted earlier.
THEOREM 2. Assume(60) holds.Supposen > 2 butotherwiselet thehypotheses
of theorem
2 of the
precedingsectionbesatisfied.Thentheconclusion
of thattheorem
alsoholds.
Proof. We introduce the transformation
u =(-)
exactly as before. Then iuobeys the equation n7D2u= a with
IQ/ + 99, pp and + =$), Xk'
where the arguments of 9, w1, , and y are x, u,p = R'(ii), and where
CL) - gl/912

Evidently c1 has the form (54), so that theorem 1 can be applied.

For this purpose, we observe that

Now forming the trace of (60) there results

I= 'V+Ipl2g
while also cr -= ('o?) +p pJ21.
O from the last two relations and solving for 9 yields
Eliminating Ip12

,1-crJd
1 -cr0da'-

whence recalling the definition of the invariant q, we find

Thus, with regard to the transformedequation,

Y+Y2= (-?'Sq to) O-q
Io-S') 292

This quantity will be non-negative provided ?> 0, that is, provided the expression (47)
in the proof of theorem 12-2 is non-negative. This being the case, the rest of the demonstra-
tion can be carried over word for word from the proof of theorem 12-2.
Theorem 2 applies particularly to boundedly non-linear equations whose coefficient
matrix satisfies (60). For the case of regularly elliptic equations we have the following
corresponding result.
THEOREM3. Assume(60) holds.Supposen> 2 butotherwise
let thehypotheses
of theorem
3 of the
precedingsectionbesatisfied.Thentheconclusion
of thattheorem
alsoholds. -
This is proved in exactly the same way as theorem 2, except that in the final step we make
use of the proof of theorem 12*3.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
450 J. SERRIN

13'. FURTHER ESTIMATES FOR |Dul

Here we present two additional interior estimates for the gradient of a solution. As was
the case in ? 13, these also apply only to certain special equations.
THEOREM4. Let A- A(x, u,p) be the minimumeigenvalueof the coefficientmatrix 4, andsuppose

5 _-I1112/4A> O for lpI > L'.

Letu E C1(Q) n C3(Q) bea solutionof (1) in Q, suchthatIDu I < L on MQ.ThenIDuI < max (L, L')
in Q.
Proof. The identity (41) used in the proof of theorem 12 1 holds equally in any number of
dimenslons. Since .ii UikUjk> AllD2ujj
2,

it follows that D2ull2+ DuI ( 1-f11DI D2ull)+

sz/D2w A 11 (x) Dw

1l /4A)w+ 9
> (1W-qk11 (x) *Dw
using Cauchy's inequality, where 9 = (x) is continuous in Q. The conclusion of the
theorem is now an immediate consequence of Hopf's maximum principle.
THEOREM 5. Supposethat -5 _
O(A), ?U5VP O(A/lp )

and A4,Ox= O(AP

Ifp2), 5 45 - = O(Alp
IP)-
Let u E C1(Q) n C(3(Q)be a solutionof (1) satisfyingIuI< m in Q and IDuI L on MQ.Then

IDul AM in Q
for theordertermslistedabove.
whereM dependsonlyon m, L, andbounds
This result is essentially due to Ladyzhenskaya & Uraltseva (I96I, page 45); we have
retained in the formulation only those relations actually used by Ladyzhenskaya &
Uraltseva in their proof, which accounts for the difference between their result and the one
stated here. Since the proof of theorem 5 shows a close unity with previous results, it is
worth including for completeness.
We make the usual change of dependent variable
U -(u)

and apply the result of theorem 4 to the transformed equation ?7D2u = X. Then as in the
proof of theorem 12 3 we find

4A 1S11= s(W+IPI g 4A11Sfl2

-X 02)
(Nlgg (1 +O0)(1 + lp ICA pIPI--(1+ ,N2IP12)CA
IP
g'-_A(&)2 +
AIpi2[Co)'/ +1IPI-1)] (IIPI->-max (a,, 1))

where A 20+1 and C is a generic bound depending only on the order terms in the
theorem (it is assumed that ol/qY> 2 a condition which will in fact hold for the final
choice of p).

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 451
This much being shown, the rest of the argument can be carried over almost word for
word from the proof of theorem 12-3. (In their proof Ladyzhenskaya & Uraltseva choose
p' (u) -const. exp (-um)
rather than the function (53) used here; since both choices accomplish the same ultimate
purpose it is a matter of taste which one is finally adopted.)
As in theorem 12-3, the order conditions on /u and Mucan be weakened to
sz o(A), JU> _-o(AIpI2)
without affecting the conclusion.
The parallel developments of theorems 1 to 3 of ? 12, theorems 1 to 3 of ? 13, and theorems
4 and 5 of the present section, has probably not escaped the reader's notice. A further
connexion arises from the fact that the following well-known result can be considered as
a consequence of either theorem 1 of ? 12, theorem 1 of ? 13, or theorem 4 of the present
section.
Supposethat sd =s1(p) and?i --A (u,p) wheredV4/du) 0. Let u E C1(L) n C3(Q) bea solution
of (1) in Q, suchthat IDuI< L on dL. ThenIDuj ? L in Q.
Indeed, using theorem 12-1 and supposing n = 2 we have
_-2 = (1*)-1 U+_2 > 0;

using theorem 13-1 we have ++y2 > 0;

and using theorem 4 of the present section
- 11 11214A-IpjI
-1 u > 0.
Still another proof of this result may be given by using theorem 605 together with a
consideration of the difference quotients of u.
Remarks.Other estimates for IDu j have been given by Hartman & Stampacchia (I966)
and by Trudinger in his Stanford dissertation. These authors considered the case when
equation (1) can be written in divergence form,

a{Ai(x, u, Du)} B(x, U,Du).

Since we shall not require these resultsin the sequel, the detailed conclusions can be omitted
here. On the other hand, it may be remarked that the condition l 0 (A)is an important
hypothesis in this work. Consequently the results apply to non-uniformly elliptic equations
only when e 00, a feature which evidently restricts their usefulness and indicates the
need for theorems of the type derived in the two preceding sections.
Further estimates for the gradient of solutions of divergence form equations have been
given by Oskolkov (I967), though again the structure which he treats has the property
that Wo= 0 whenever the equation is not regularly elliptic.

14. SUMMARY
In chapter I we outlined a general existence programme for the Dirichlet problem, based
upon a series of aprioriestimates. According to the discussion of ? 2, if a particular equation
possessesappropriate structureso that one can carry out the estimates required at each step
of the programme, then the corresponding Dirichlet problem will have at least one solution.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
452 J. SERRIN
In the present section we shall illustrate this theory by giving a representative selection
of existence theorems based directly on the aprioriestimates developed earlier in the chapter.
We retain the notations and conventions of ? 5, except that in the interest of generality the
theorems are formulated for boundaries and boundary data which are only of class 02. The
first result applies to a class of equations which are regularly elliptic.
THEOREMI. Let Q be a bounded domainin n-dimensional Euclideanspacewhoseboundaryis of
class C2. Assumeeithern = 2 or that (60) holds.
Supposethereexistsa constantp > 0 suchthattheinvariantfunction
g = pnip satisfiesthecondition

plpl < <(1 -) JpJ2

at leastforall sufficiently thatas p tendsto infinity
largevaluesofp. Assumefurther

and Rwa> eu)t = O(91/pI)I

andfinally suppose-4 and tf satisfy thehypotheses of eithertheorem6-2 or 6 4.
Then the Dirichletproblemfor equation(1) in Q is solvablefor arbitrarilygiven boundarydata
of class C2.
Proof.We assume first that the domain and the boundary data are of class C3,as in ? 5, and
that the functions -s(I(x, u,p) and (x, u,p) are of class C2. These restrictionswill be removed
at the end of the proof.
In order to apply the fundamental theorem (? 1), consider a function v e C2(n) such that
ds (x, v, Dv) D2v = r4(x, v, Dv) in Q (61)
=
and v rf on dQ.
Evidently lvl < coon dQ; therefore by theorem 6-2 or 6-4, whichever may be applicable,
it is clear that vI < m' in Q
where m' is a constant independent of r. Moreover,
+g1- = ?(IPI-1)5

the order term being independent of r. Thus by theorem 8-I

IDvIAL' on dQ
where L' is independent of r.
It remains to derive an interior estimate for lDvl. We first observe that the coefficients
in (61), considered as functions of x, are in Cl(Q). Thus, since the boundary data is of
class C3, it follows from the Schauder estimates that v is in C2+A(Q) for any A < 1. This in
turn implies that the coefficients of (61) are in C0+A(Q2).Using linear theory once more, we
find that v EC3+A(Q)
The hypotheses of theorem 12 3 or of theorem 13 3 are now satisfied by v and by equa-
tion (61), with the order terms independent of r. Hence
IDvlI M' in Q
where M' is independent of r.
We have thus established that tvf+ Dvf < m'+M' in Q; therefore by the fundamental
existence theorem, the Dirichiet problem for equation (1) in Q2is solvable for arbitrary
boundarydata of class 03.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 453
The differentiability assumptions on the domain, the data, and the functions a? and X
can be removed by a standard approximation and compactness argument. To this end, we
observe first that the numbers m' and M' above depend only on C2norms of the domain and
the data, and on the structure stated in the hypothesis of the theorem. This being the case,
we can approximate the given problem 'uniformly' by new problems possessing appro-
priate differentiability conditions; the new problems then have solutions which, by the
above argument, satisfy Jul+ Du <m' + M' +1, say.
According to the theorem of Ladyzhenskaya & Uraltseva these solutions are then uniformly
bounded in C1+7.Finally, by linear theory, the solutions are uniformly bounded in C2-7 on
compact subsets of Q. Applying Arzela's theorem, we can choose a subsequence of these
solutions which converges together with its second derivatives to a solution of the original
problem. This completes the proof of theorem 1.
It is not hard to see that the solution which has been constructed is of class Cl (Q), and
even (using the theorem of Ladyzhenskaya & Uraltseva) of class C1+7((Q)for an appropriate
value of y.
The above result includes a main existence theorem of Bernstein (i9i2, page 464) and
Leray (I939, page 281), who required not only n = 2 but also various additional restrictions
on the behaviour of the functions a/ and M for large values ofp.
THEOREM 2. Supposethehypotheses oftheorem1 hold,exceptforthefinalcondition 9 and4.
concerning
ThentheDirichletproblem for equation(1) in Q is solvable
for arbitrarydataof classC2 providedthe
diameterof Q is sufficiently
small (dependingonly on thestructureof theequationand on a boundfor
sup f 1).
Proof.As in the proof of theorem 1, we can assume that the domain and the data are of
class C3, and that the functions v' and M are of class C2.
Let g(u) be a non-decreasing C2function such that
-2-c0 if u < -2-c0,
(u) U. if -1 c0?u I+co, (c0= supQf )
( 2+co if u? 2+co.
We consider a new equation with coefficient matrix
A~~~~~
and with a similarly defined inhomogeneous term S. It will be enough to show that the new
equation has a solution u satisfying
IuIA1+co, uzzf on Q
for this solution will evidently also solve the Dirichlet problem for the original equation.
Consider a function v E C2(Q) such that
JID2v=ZZX and v =-rf on dAQ.
In order to establish the existence of a solution ofthe new Dirichlet problem it will be enough
a -
to show that vj 1+co v in Q2,
for in all other respects the previous proof can be applied unchanged to the present situation.
59 VOL. 264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
454 J. SERRIN
Since -0(e), there exist positive constants a and C such that

/lIl < C for pI ag, u 2

2+co.
Consequently e s< C for pj .
Now, by a simple modification of the proof of theorem 6-4, we find that

Ivl < so+-log in Q

provided A< 1/xC, where S is the diameter of Q. In fact, if

i-e-c
cC
then Iv 1 + co, and the proof is complete.
Remarks.In both theorems 1 and 2 the conditions on I and 6' can be weakened

?u>-o(e) and 6u=o(e)

according to the remark at the end of ? 12. Furthermore, it almost goes without saying that
the final hypothesis of theorem 1 can be dispensed with entirely in any situation where the
solutions of (61) are known to be uniformly bounded in absolute value.
Finally, we note that the size of the domain Q required in the conclusion of theorem 2
will be independent of the given boundary conditions whenever the order relation - 0 (6 )
is uniform over all values of u. This holds in particular for the important case when the
coefficients a? and X are independent of U.
THEOREM 3. Let Q be a boundeddomain in n-dimensionalEuclideanspace, whoseboundaryis of
class C2. Assumeeithern 2 or that (60) holds.
Supposethereexists a constantji> 0 such that the invariantfunction e satisfies

at leastfor all sufficientlylargep. Assumefurther that as p tends to infinity

ex) Gun 'p 0(g)

\

and that the asymptoticformulae (26) and (27) hold with the remainder$B l -W satisfying

,)X, -U,IP ? I
?(elIP1
Finally, supposethat X satisfies the hypothesisof eithertheorem6 2 or 6 3.
Then the Dirichlet problemfor equation(1) in Q is solvablefor given C2 boundaryvaluesf, provided
both conditions X >-'6o(y,f,v), X > 60(y,f-v) (62)

are satisfiedat eachpoint y of the boundarysurface, whereX andVfare thegeneralizedmeancurvatures

definedin ? 9. If --d 0 -o( /ep ) the equalitysign can be allowedin (62).
Proof. A simuplifiedversion of the proof can be given if it is assumed that (62) continues to
hold when T0 and f are replaced by TWoand -rf, 0 ? X 1. Since a large number of special
examples have this property (cf. ??19-22), it seems worthwhile to carry out the proof first
ulnder this additional assumption.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 455
As in the proof of theorem 1, we shall suppose to begin with that the domain and the data
are of class C3, and the functions a and X are of class C2.
Now let v E C2(Q) be a solution of (61) in Q, such that v Tfon 3Q. We wish to show that
jvI+ IDvI<--M in Q, where M is independent of T. By theorem 6-2 or 6-3 we have lvl < m'
in Q. Moreover, using the special assumption above, the required estimate for Dvl follows
immediately from theorem 9-1 and either theorem 12'2 or 13'2.
Next suppose that v -do ?o(g/elpI) and
-o( ,) /A>J0(y,f,-v). (63)
Here we must apply theorem 10-1 rather than theorem 9-1. To this end, it will be sufficient
to show that + Ile lol
11 K1-O--doll + IPI 1)

as p-oo, uniformly for lpol in any compact set. This in turn follows directly from the
hypotheses of the theorem provided
eF > const. of (const. > 0)
for lpoI < c and all sufficiently large values of p. (Here c denotes an arbitrary fixed real
number, 0 < c <oo.)
Now by integrating the condition p &p (1-,u) e it is clear that
const. Ip1II
for all sufficiently large p. But then, as in the proof of the lemma in ? 10,

and v >e+o(e) for poI ?c.

Consequently Y 12e for IpoI < c and all sufficiently large values ofp (it almost goes without

saying that we are here considering only a compact set of values of x and u, namely x E Q
and IuI < m').
In order to remove the differentiability assumptions on the domain, the data, and the
functions a/ and X, we can apply the approximation and compactness methods used in the
proof of theorem 1. In so doing, however, an additional point enters the argument when
condition (62) is replaced by the sharper condition (63). Indeed, for the approximating
problems in the latter case we can only assert that
> - 'o0(y,f, v) 6, 0(y,f,v)-6
o (64)
though of course 6 can be chosen arbitrarily small.
Consequently, theorem 10'1 cannot be applied directly to the approximating problems.t
The proof of theorem 10'1, however, continues to hold after a slight modification to take
into account the weakened condition (64). To see this, we observe that the principal
estimate for Y (v + b) in this proof must be replaced in the present circumstances by

Y (v+b) < k'> +c + +@(k -c

t A somewhat simnilarpoint occurs in the use of theorems 6'2 or 6'3 to obtain a maximum bound for the
solutions of the approximating problems. The difficulty here is more apparent than real, however, since one
can obviously choose the new problems so that the hypotheses uG > 0) or sign u . h-lR
- continue toshold
for large values of u.
59-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
456 J. SERRIN
(note that j -/,ufor sufficiently large values of p). lThe remaining part of the proof of
theorem 101] then carrics over essentially unchanged provided that
T0 dp - > cM, (65)

where (for the purposc at hiand)

$D(p) --const./p, M-co0+-m'+1.

Inequality (65) is obvious, however, provided c is sufficiently small. Tlnhusif the approxi-
mating problems are sufficiently ncar to the given problem the conclusion of theorem 10 1
continues to hold. T lhe argument of tlheorem 1 can tlhen be carried over word for word to
the present case, and tlleorem 3 is thus proved under the initial assumption on the structure
of condition (62).
In order to carry out the proof without particular assumptions concerning condition (62),
we must resort to a more refined hiomotopy than used in (61). To this end, let k(T) be a
function of class C2 wlliclh vanishes for T < 0 and cquals one forTX 4. Also let

l-kl (T) -= k(T- ) k2 - k2(T) -k(T )

and 0(i(p)- k(p- 1).

We now put, for 0 < TX 1,

(x, u,P; T) [0(p) +{1 - 0(p)}kj (x, k1u+ (1 k1)f;p)

and consider tllc family of equations

S (x, v, Dv) D 2v - (x, v, Dv; T) (66)

together with the boundary condition v - k(T)f. Note in particular tllat the assumptions
stated in the concluding remark of' ? 1 are fulfilled by this hormotopy.
Now let v e C2(Q) be a solution of (66) in Q, such that v = k(T)fon 9Q. In order to apply
tlle concluding remark of ? 1, we wish to slhow that IvI+ IDvI ? M in Q, wlhere M is
independent of T. It is casy to see that the hypothesis of theorcn 6(2 (or thcorem 6.3)
continues to hold for equation (66) when 3- < T < 1. Moreover, f' (x, U, 0; T) 0 for
0 s TT -J.Tlhus theorem 6-2 remains (is) applicablc in tllis range. Hence a uniform bound
for vi follows at oncc, say IvI < m' in Q, where m' is independent of.
It is, furthermorc, clear thatWo(x, U, 0'; T) ' Co(x, k1u-+ (1 -kl)f, o). Hence e, J, and
c0o(y,f, I v) are invariant under the given homotopy, and conditions (62) continue to
hiold at cach stage. Consequently, under tlhe usual smootlhncss assumptions on the domain,
the data, and the coefficients, theorem 9-1 can be used to bound the gradient of v on dQ
independently of T. Finally, by tllcorem 12 1 or 13-1 wc obtain the required estimate for
IDvl in i?. TIhllereforcthe existence theorem given in the concluding remark of ? 1 can be
applied, and the theorem is proved in thlc case where (62) applies and where the domain,
the1data, and the functions .V and @ are suitably smoothl.
As in the first part of the proof; it remains to consider the c ase whecreequality is allowed
in (6)2) and to carry out the usual approximation and comnpactness arguments. But this can
be done exactly as before, and thecdemonstration of thecorem 3 is thauscompleted.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 457
THEOREM 4. Supposethe hypothesesof theorem3 hold, with the exceptionof thefinal conditioncon-
cerningX. Then the conclusionalso holds, providedthe diameterof Q is sufficientlysmall (depending
only on the structureof the equationand a boundfor sup if ).
Proof. The argument of theorem 2 applies until the final paragraph, with obvious
modifications for the alternate homotopy used above. For the completion of the proof, we
then replace the final part of the argument of theorem 2 with the following reasoning.
Since (62eo +, and g < const. lp I1- it is evident from the hypotheses of theorem 3 that
l1e < C for IPI->, JUl< 2+co
where a and C are appropriate positive constants. Consequently, since o > ,u, we have
lXl/gAC'p I for lpI>
where C' = C/,. Now, by the proof of theorem 64 ,
-2 C/
?vI< co+ 1(I ) in Q

provided A ? 1/2'c2C', where 5is the diameter of Q. In fact, if

( 2-cC' <' aC

1/2jC'2 >
aC'C1,
then Iv I 1 +co, and the proof is complete.
The following result is a variant of theorem 3 in which it is not necessary to introduce the
asymptotic formula (26).
THEOREM 5. Let Q be a boundeddomain in n-dimensionalEuclidean space, whose boundaryis
of class C2. Assumeeithern 2 or that (60) holds.
Supposethereexists a constant, > 0 such that
*~~~ 9pg, < ( ),t)
for all sufficientlylargep. Assumefurther that as p tends to infinity
gx eu) ep= 0(g)
and that the asymptoticformula (27) holds with the remainderM = -W satisfying

,q X,vMUIPI P= ?(9/ IPI)

Finally, supposethat a obeys the hypothesisof either theorem6-2 or 6-3.
Then the Dirichletproblemfor equation(1) in Q is solvablefor given C2 boundaryvaluesf, provided
that k > +o
T (y, f, J v) (67)
at eachpoint y of the boundary,wherek k(y) denotesthe minimumnormalcurvatureof the boundary
surfaceat y.
If the minimumeigenvalueofQv is boundedbelow by a positive multipleof IpI-2 as p becomeslarge,
then the equality sign can be allowed in (67).
Proof. The first part of the theorem is proved in the same way as the first part of theorem 3,
with the exception that theorem 9-2 rather than theorem 9 l is used to bound iDvl.
Next suppose the minimum eigenvalue of/ is bounded below by a positive multiple of
pK-2 and that (67) is replaced by
k > T 6o(Y,f, A?i8). -(68)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
458 J. SERRIN
Here we must apply theorem 10-2 rather than theorem 9-2. To this cnd, it will be sufficient
to show that -1

as p - oo, uniformly for IpoI in any compact set. This in turn follows directly from the
Ilypotheses of the theorem provided
.>- const. dD (const. > 0)
for jpol< c and all sufficiently large values oflp.
To prove this, lct A -A(x, u,p) denotc thc mninirnurncigenvaluc of , arid suppose

A1i0/llpl2 for IpI --

where /to is a positive constant. 'then obviously

for pj - c and IpI max ((Y,'2c). It follows that

2 (Ano+82)

when J' I (u0- I-8c2); on the otlher hiand, wlien ( l (p l- 8c2) wC iave by the lemma of 8

1 4c2
C

and the samc inequality lholds. This being slhown, the second part of the theorem now
follows from the same argument used to prove the second part of tlheorem 3.
Remark.Th-e h-ypothesis requiring ' to satisfy the conditions of eitlher theorem 6-2 or 6-3
can obviously be replaced with various other conditions wlhiclhprovide a priori bounds for
the magnitude of a solution. With this remark taken into account, theorem 5 includes as
a corollary the result stated on page 282 of' Leray's memoir on tlle Dirichlet problem.
A final example of interest concerns the equation
(.+ U2l) XX U U x+(. + U2 ) 88y = CU ( |-U2 d_-U?J

(c -- const. > 0) whiclh governs tlhc frec surface of a stationary fluid under tlhe combined
action of gravity and surface tension. Since
CU(Il+ p2), ? p 12,

2 2 pl2+p,-
A 2 1P2
IC(+2)- I +
and 6'0 -,- cu - -1-cu
it is easily checked that tlhe hypotheses of theorem 5 are satisfied (thleorem 6-2 at thle final
stcp). Tllus the Dirichlet problem is solvable for given C2 boundary valucsf provided
K C If I
at each point of the boundary, wliere K denotes the ordinary curvature of the bounding
curve.
Whlile various furthecrresults arc suggested by the foregoing summary, these lmay be left
to the reader flb formnulation. In chapter IV we shlall consider a numbrerof special equations
whoxe discussion requires more detailed individual analysis.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 459

CHAPTER III

In this chapter we shall treat various cases of the non-solvability of Dirichlet's problem
for smooth data. It will be shown that the resultsof chapter II are in many ways best possible,
in that, when their principal hypotheses fail to hold, one may construct boundary data for
which the Dirichlet problem cannot be solved.
We shall assume throughout that (1) is normalized by the condition trace /1. Also,
l
the invariants and e will retain the meanings assigned in ? 5.

15. A VARIATION OF THE MAXIMUM PRINCIPLE

The results of the chapter depend on an interesting form of the maximum principle, in
which the hypothesis u < w is required only on a subset of the boundary. For the purposes
of stating this result, as well as for later use, we introduce some simple terminology. Let r be
a domain in n-dimensional space with boundary dI. If P is a point of dT'and N is an
n-dimensional neighbourhood of P, the set N' = d n N is called a boundaryneighbourhood
of P. A subset of is called openif it contains a boundary neighbourhood of each of its points.
r

THEOREM 1. Let F be a bounded domainin n-dimensionalEuclideanspace,whoseboundary

contains
an opensubsety of class Cl. Let w(x) beafunctionin C(F) n C2(I) satisfying
Y(w+b) < O in 1
for all positiveconstantsb, andsuchthat dw/dn=-oo at eachpointof y (heren denotesthe normal
directioninto IF.)
Suppose that u E c (F) n C2( + y) is a solution of (1) in 1 such that u < w on d -y. Then
u < w in IF.
The underlying idea of theorem 1 was known and used by Bernstein (I9IO, page 243;
I9I2, pages 465-469). More recently, Leray (I939) and Finn (I965) have applied the
result of theorem 1 in essentially the form in which it is stated here. For completeness, we
give the simple proof.
Suppose for contradiction that the conclusion is not true. Then by theorem 6 1 and the
fact that u < w on d17-y, the maximum of the function u-w must occur at some point P
on y. This is impossible, however, since at P we have
_ _ ~dii _ _

(u-0) dn -n

so that u- w takes values in r which are larger than its value at P.

16. IRREGULARLY ELLIPTIC EQUATIONS

For the purposes of classifying equations for which the Dirichlet problem is not always
solvable, it is convenient to introduce some further terminology. We say that equation (1)
is irregularlyellipticprovided that
1W1yooas p->oo~, uniformly in u (69)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
460 J. SERRIN
and I/rl - T(lpl) for Jul> M, lpl )?l, (70)
where M and I are positive constants and T(p) is a positive continuous function satisfying
the condition ? dp (71)
p~iT(p)

Evidently the classes of regularly elliptic and irregularly elliptic equations are disjoint,
though obviously there are equations which belong to neither class.
To illustrate the concept of irregular ellipticity, suppose that e ,ui pl when 1p 1.
Then (1) is irregularly elliptic provided that

I- Ile
/6 (log lpI)1'--0 for large u and p, (72)
where 6 is a positive constant. In particular, a uniformly elliptic equation is irregularly
elliptic whenever2 > lpl (log lpl)1+O0 for large u andp.
1ec1
Condition (72) should be compared with the corresponding relation (22), which guarantees
regular ellipticity.
The following theorem relates the concept of irregular ellipticity with the non-solvability
of Dirichlet's problem.
1. Let Q bea bounded
THEOREM Euclideanspace,andsupposethatequation
domainin n-dimensional
elliptic.ThenthereexistsC??boundary
(1) is irregularly datasuchthattheDirichletproblemfor
equation
in
(1) Q has no solution.
As a consequence of theorem 1, we see that for equations having a well-defined genre the
Dirichlet problem is generally not well set when the degree of the inhomogeneous term a
is greater than max (1, 2 -g). Conversely, when the degree is at most this large the results
of ? 14 become applicable.
Proofof theorem1. We may obviously assume without loss of generality that (69) and (70)
hold with the absolute value signs removed from T.
Now let P be any fixed boundary point of Q at which there is an internally touching
sphere Z. We shall show that there exists a boundary neighbourhood N' of P and a con-
stant c such that if u is a solution of the Dirichlet problem for equation (1) in Q, then
u <max(m,M)+c atP (73)
where m denotes the supremum of u on dQ - N' and M is defined by (7Q0).
Assuming this result for the moment, in order to prove theorem 1 it then sufficesto choose
smooth boundary valuesf such thatf_ 0 on Q- N' andf> c+M at P.
Turning to the proof of (75), let 2a be the radius of 2, let b be the diameter of Q, and let r
denote distance from P. Consider the function
i(x)-=M+lh(r), a < r <,
where h is twice continuously differentiable for a < r < A,and
hz'(a)=-c-o, lz(&) 0, h'(r) < -1. (74)
An easy calculation yields
sSff+b)<g(x,o@ b,Da) (h"/lz'2)
-gW(x,aZ+b, Dw)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 461
where b is an arbitrary positive constant. Now using (70) we obtain
?1(6o+b) < ehh'{"/lh'3+?T(lh'l)}
(note here that +b > M and jDoj I= h'Il?1).
As a preparatory step for the application of theorem 1541, we shall choose h so that the
preceding expression is non-positive. To this end, put

C -max (1, (Q-a)/f p )

p3TP(p)!
and let a be a constant defined by

J _-a if C= I

- l
aoc if C= (-a)/ 3T2,.

Clearly a 1. Now let h and r be parametrically related by

AC P dp ra+ . dp
h-= C td() r = a + C 0SP-
a ~p2'T(p)' ~ p3pT(p)
where a < p < oo. One easily checks that h' -p < -I and that (74) holds. Moreover
hA" C-1
ht+Tf( htl) cC Tf(jhtl) _>O
so that ? (a)+b) < 0.
We may now apply theorem 1541 to the domain F - Q n {r > a} and the open boundary
set y = Q n{r = a}. Putting N' =Q n {r < a}
then yields at once
u < max (m,M) + C d p- M* (75)
J1p2'T(p)'
in r, where m is the supremum of u on dQ - N'.
Although (75) supplies a bound for u on the set Q n {r a}, a further argument is needed
to obtain an estimate for u at P. Let e be an arbitrary real number between 0 and a, and
let s denote distance from the centreof L. Consider the function
w(x)-mzm*+h(s), a?<s?<2a-e,

where h is twice continuously differentiable for a < s < 2a - , and

h'(2a-e) - oo, h(a) 0, h'(s) > l*.
(No confusion should result from the fact that h(r) in the earlier part of the proof and h(s) in
the present part represent different functions.) An easy calculation now gives

Y(o+b) < 6'(x,o+b,D6) ,2+--(x,o+b,Do)

_ a ~12

t Note that the integral -converges in virtue of condition (71).

6o VOL. 264 A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
462 J. SERRIN
Since 6 -- oo as p -> oo there exists a positive constant 11 11(a) such that

6'e?2a-1 for IpKl1.

With the help of (70), it now follows that
h
(o+ b) < 6h t f (hf)
provided 1* max (1,11).
We now choose h so that the expression in braces in non-positive. The details are
essentially the same as in the first part of the proof (replace I by 1*, b-a by a- , and r- a by
2a -e -s) and consequently may be omitted.
The proof is completed by applying theorem 15-1 to the domain

F* - {r < a} n {s < 2a-c}

and the open boundary set y* {r < a} n {s = 2a-6}. Indeed, according to (75), we have
u < m* on the remaining boundary set {r a} n {s < 2a-e}. Hence by theorem 15.1 there
results
u < m*+C* p C = max 2 a p

in F*. Letting c -- 0 and using the continuity of u yields the same inequality at P. This
establishes (73), and thus completes the demonstration of the theorem.
Bernstein (I9I2, pages 465-469) proved for the case of two independent variables that if
equation (1) has a well defined integer valued genre g < 2, and if, moreover:
(i) sl and 7 are analytic and have certain special expansion properties for large
values of p,
(ii) =/ - (x,p) and d/ du > 0,
(iii) /le ? const. lpI for large p (const. > 0),
then there exist convex domains Q and smooth boundary data f for which the Dirichlet
problem is not solvable. The cases g = 0, 1 of Bernstein's theorem are easily seen to follow
from our results, for in fact theorem 1 shows that under the hypothesis
(iii') Ml > const. IPI?, e > const. 1/p1
alone, the Dirichlet problem is not generally solvable for any domain, no matter how it is
chosen. The case g -2 of Bernstein's theorem falls more properly under the work of
succeeding sections, and is in fact a special case of the following theorem 18 3.

17. SINGULARLY ELLIPTIC EQUATIONS

In the preceding section, attention was given to the effect of a large inhomogeneous term
on the solvability of Dirichlet's problem. Here we shall consider a second type of non-
regularly elliptic equation, in which the regular ellipticity now fails due to the structure of
the coefficient mnatrixv rather than due to the forcing effect of the inhomogeneous term X.
We say that equation (1) is singularly elltStic provided that

1/ ?NP(IpI) for 1p1>1 (76)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 463
where T(p) is a positive continuous function satisfying the condition

fco_dp
p2YT(p)
In parallel with the comment in the previous section, it is evident that the class of singularly
elliptic equations is disjoint from the class of regularly elliptic equations, though again
there are equations which belong to neither class. If equation (I) has a well-defined genre g,
then clearly a necessary and sufficient for singular ellipticity is that g > 1.
In this section, as in the earlier ?? 9 and 10, it is the curvatures of the boundary surface
which play a crucial role, and it is the failure of the recurrent condition (29), (35), (62)
which leads to non-existence. It will be convenient to defer the main results until a series
of lemmas have been proved. From ? 12 we recall the definition e*= /IPI2.
LEMMA 1. Let u be a solution of the Dirichlet problemfor equation (1) in a boundeddomain Q.
Supposethat (1) is singularly elliptic. Assumealso that
g*(x up) l ,u

for somepositive constantp, and that

M4(x,u,p) >0 for u>M.

Let P be a point of dQ and let m be thesupremumof u on the boundaryset J)Qn {r > a}, wherer denotes
distancefrom P.
Thenfor any given e > 0 we have

u <?max (m, M) +c on Q n {r-a}

for all sufficientlysmall values of a (dependingonly on c-, the diameterof Q, and the structureof the
equation).
Proof. As in the previous section, we consider a comparison function of the form
@(x) - M+h(r), a< r A,
where h is twice continuously differentiable for a < r < A, and
h'(a) -oo, h(a) O, h'(r < o.
h"f h'
An easy calculation gives ? (o + b) h-2+ (1 _ e *)
r

where b is an arbitrary positive constant and the arguments of Y,g and * are x, c9+ b, and
co
Dwo.Hence, using the conditions 0 and 9'* < 1 -,c, we obtain

Y(w+b) < e h2+ rh.

Now define 77(p) -

p-2 if <np
0 1 and if p> 1, where NP(p) is the given
H(p)--t(p)

upper bound for 1 /6 in the definition of singular ellipticity. Clearly

dp? <
JOp2-jp) <o
60-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
464 J. SERRIN

For 0 < C <0o, put x(2) T= pH -

p3"(p).
Evidently x(o) is monotonically decreasing and has range (0,00). Thus there exists a well-
defined inverse function 4 (fi), also monotonically decreasing and with range (0, oo). We
have, furthermore, co CI

l(l
?(dfl - X(a) da <co,
o o

as is easily seen by an application of Fubini's theorem.

We may now define
h(r) ( log -) dt.

One checks easily that hl" r ,h') Do)I= h';

h'3 r
thus (wo+b) < 0 since e < 1/.
This much being shown, it follows by applying theorem 15d1 to the domain Q n {r > a}
that u < max (m, M) + h(a) in Q n {r > a}.

Since h(a) = f (/iog -) dt - aJ q(logt) dt (<oo),

we find by L'Hopital's rule

lim h(a) -lim AC(,ulog ((F/a))0-
a >-o a-
and the lemma is proved.
This proof is closely related to one given by Nitsche (I965) in his discussion of the
maximum principle for equations of positive genre.
LEMMA 2. Let (1) be singularly elliptic, and supposethat the asymptoticformula (26) holds. Then

rdo/ =o

Vioreover,correspondingto any compactset of valuesx, u thereexists a positive constant,usuch that

&e
<-a

Proof. We first show that oar - 0. Indeed, since g pStp we find from (26) and (76)
that
0-?S I--p- +o(l) < lp +o(1).
lpl,1p2~
Since the first term on the right-hand side is integrable, there must be a sequence of values
of lp Ifor which it tends to zero. Keeping r fixed and lettingp tend to infinity on this sequence
then proves the first part of the lemma.
This being shohwn, we now observe that e o-0+ o (1), whence 5 *0 as h oo .
SinCe 6' < 1 on any compact set of values of x, u, p, the proof is completed.
LEMMA3. Let u be a solutionof theDirichletproblemforequation(1) in a domain?2whoseboundary
is of class 02.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 465
Supposethat (1) is singularlyelliptic,andthattheasymptoticformulae
(26) and (27) holdwith
theerrortermsbounded independently of u. Let P be a pointof aQ where
X < Wo(y,f, v) (77)
andlet m' be thesupremum of u on theset Q n {r a}.
Thenthereexistsa constant60, depending onlyonf (P), onthestructure
of theequation,andona lower
boundforthedifference '
in (77), suchthatforfixed < 6othecondition
f(P) - 5c' - m' ?f(P) - 6'

cannotoccurwhena < aO;hereao depends on thesamequantitiesas 60andalso on thedomainandone'.

Proof.To distinguish the values ofyr, f and v at P = (y) we shall henceforth write f(y)
f and P. Then in view of (77) there exists a positive constant 0 such that
(y-)<1eo(y,f- - S0 (78)
It is geometrically evident (and easy to prove using a modified Dupin indicatrix) that
there exists a quadric surface S tangent to aQ at P such that
(i) S has a unique parallel projection onto the tangent plane at P,
(ii) the point set S n {r < a} is contained in the closure of Q for all sufficientlysmall values
of a, and
(iii) the generalized mean curvature Js of the surface S satisfies the condition
is (Y-)<1 > Y)- -0 (79)
Our purpose is to apply theorem 15d1somewhat as in the second part of the proof of
theorem 1 6 1, but with the surface S playing the part of the internal tangent sphere Z. To
this end, let S+ be the half space of points x on the same side of S as the normal vector v at P.
For points x in S- we define d- d(x) to be the distance from x to S.
Now consider the point set
rf--{r< a} n {d> e}, e small.
Clearly, if a is suitably small (say a < a1) then the set 17is contained in Q. We may assume,
moreover, that d(x) is of class C2 in r, if necessary by making al somewhat smaller. For
convenience the geometric situation is shown in figure 1, the domain 1 being indicated by
diagonal shading.

FIGURE 1

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
466 J. SERRIN
For x contained in F, and a < a,, we define
@(x) -m'+h(d), 6c<d<a,
where h is twice continuously differentiable for 6 < d < a and

h'(6) -oo, h(a) > , h'(d) < or, (80)

a being a positive constant whose value will be specified later. From lemma 4 1 one finds
easily that Y (o+b) = eh-/I'2-fhi'-M, x E r, (81)
where the arguments of a, R, and 6' are x, w+b, and p vih'. (It almost goes without
saying that the normal vector P, and the principal directions and principal curvatures
which appear in the formula for J , are calculated at the point y = y(x) on S nearest to x.)
As in the proofs of chapter II, it will be necessary to estimate the various terms in the
preceding inequality. To begin with, using the definition of l and the asymptotic formula
(27), we obtain
- (x,(t)+b,p) = h'e(x,t+b p)
-h'6o0(x,wf + b, -v)-o (l)} ?< hI{o9(x,m', -v)o(lP)},
since Wois increasing in its second argument and h' is negative (the expression o(l) tends
uniformly to zero as a tends to infinity).
We shall assume from here on thatfJf(P) < m'+ 56'. Thus to complete the proof it
must be shown that m' >f(P) - c' provided that-e' < c0and a is sufficiently small.
Under the conditionf < m'+ 5c', we assert that
W (X, m', v) :>- wo (X, V-) O (e ) .

This is obvious iff < m', since Wois increasing. On the other hand, when m' <f < m' + 56'
the argument m' differs fromf at most by 5e', and the same result holds (it will be assumed
that 60< 1). Since, moreover,

teo(xJ5 - P) -> W(y, f5 - -) -O (a)

we have finally
--(x, w+b,p) < h'{eo0(yJ, -v) -o(i) - 0(a) - 0(e')}.
Similarly, from the asymptotic formula (26),

l ki =-s (y) +o(i) +0(a) (82)

where we have also made use of the fact that vAv 0 (see lemma 2).
Substituting the preceding two estimates into (81), and using (78) and (79), now yields
2 (?l + b) < 4h"/Ih'2+h'{4t9 o(1) 0 (a)-O (6')}.

We choose e0so that the term 0 (e') is less than 0 when e' < c0. Then for acsufficiently large
and a suffciently small, say ac> a0 and a ? a2, we have

?@b A'(Iz"/h'2)+ Oh'.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 467
Now let a3 be defined by the condition
a yO dp
3 a aoP31F
(p)
(we may of course assume ao >, 1 so that the integral converges). Then for values of a
satisfying the condition
a < min (a,, a2, a3),
we define h(d) parametrically by the formulae

d= 6 p

where c <, p < oo and where c satisfies the relation

a=09 )} 3dp (83)

p3NT(p),
Evidently c > ao and conditions (80) hold. Moreover, as one easily checks, Y(w +b) < 0
inE.
This being shown, a direct application of theorem 15 1 now implies

u < m'+ I p p
0ap 2T(p)

in 1. Letting e-> 0 yields the same inequality in the limit set {r < a} n {d > 0}. Consequently,
since P is a boundary point of this set, it follows that

f (P)<MI -J0
j'00 dp
2tlP
p T~(p)

with a defined by (83) and a < min (a,, a2,a3).

Clearly a tends to infinity as a tends to zero. Therefore, by choosing a even smaller if
necessary, we can makef(P) < m'+ e', completing the proof.
Remarks.1. A similar result can be proved for the generalized mean curvature X'. In this
case the hypothesis (77) must be replaced by
X01 < -leo(y, f v)

and the final condition should correspondingly be changed to

f (P) + c' < m" < f (P) + 5c'

where m" denotes the infimum of u on the set Q n {r a}.
2. If 'ec is independent of u, then there is clearly no need to introduce the restriction
f < m' + 5e' in the proof. The conclusion of lemma 3 can then be simplified to the following
statement:
Foranygiven 6 > 0 wehave f (P) ? m+n c

for all sufficiently

small valuesof a (depending
onlyon e, Q~,thestructure
of theequation,anda lower
boundfor thedifferencein (77)).

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 468 J. SERRIN

18. NON-EXISTENCE THEOREMS FOR SINGULARLY ELLIPTIC EQUATIONS

The lemmas of the preceding section can be used to derive several non-existence theorems
for singularly elliptic equations. The simplest case occurs when 'o is independent of u. We
phrase the result for the invariant f, though obviously a similar result holds for X.
THEOREM1. Supposeequation(1) is singularly elliptic and that the asymptoticformulae (26) and
(27) hold. Assume,furthermore,thlatTOis independentof u and that a > 0 for u > 0.
Let Q be a boundeddomainwhoseboundaryis of class C2. If the (geometric)condition
J ->-
leo v
(y, P) (84)
fails at a singlepoint y of the boundarysurface,thenthereexists smoothboundarydata havingarbitrarily
small absolutevaluefor which no solutionof the Dirichletproblemis possible.
Proof. Let P be a point where (84) fails, and let e and a be fixed positive constants (with
C < 1). Consider smooth boundary valuesfsuch thatf- 0 on 3Q n {r > a} whilef -3e at P.
We shall show that if a is chosen s-ufficcientlysmall (depend'ing on e) then the corresponding
Dirichlet problem has no solution.
Suppose in fact that a solution u did exist for the given data f. Since X > 0 for non-
negative values of its second argument, it follows from theorem 6-2 that
u < 3e i n Q..
We now wish to apply lemma 1. Since the hypothesis * < 1-t, may not hold, it is necessary
to add some comments. In particular, we observe that the comparison function ( + b in the
proof of lemma 1 need take values only in the closed interval [0, 3]. According to lemma 2,
however, there exists a constant ,u such that g * > 1 -,u when x E Q and u c [0, 3]. The proof
of lemma 1 therefore carries over unchanged to the present situation. This being the case,
it follows that if a is sufficiently small, say a < a(e), then
u< e on Q n {r a}.
By the same argument, we are entitled to apply lemma 3 in the present circumstances,
even though (26) and (27) are not assumed to hold with the error terms bounded indepen-
dently of u. Accordingly, taking into account the remark at the end of the previous section,
we have u(P) A m' + e A 2e

when a is suitably small, say a < ao (e). In summary, if a < min (a (e), ao(e)) then the Dirichlet
problem can have no solution for the given boundary valuesf. Since e can be chosen arbi-
trarily small, this completes the proof. (For the special case of the minimal surface equation
theorem 1 is due to Jenkins & Serrin (I968); cf. also Bernstein (i912) and Finn (I965). The
above proof is a generalization of the work of these authors.)
The result of theorem 1 strikingly illustrates the priimary importance of the geometry of
the boundary surface in the study of Dirichlet's problem for singularly elliptic equations.
When 6'odepends on u, the results are not quite as sharp, since the boundary data itself
enters into the solvability conditions X$2 >-T6(y, f, v) anld X/ > T6'(y,f, -v) . N!evertheless,
these conditions are essenltially best possible, in the sense that if eitheroneof themfails globally
for given boundaryvaluesf, thenthe corresponding Dirichletproblemis badlyposed. The precise result
is as follows.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 469
THEOREM 2. Supposethat equation (1) is singularly elliptic and that the asymptoticformulae
(26) and (27) hold. Assumefurthermorethat a > Ofor u >. 0.
Let Q be a boundeddomainwhoseboundaryis of class C2, and considerarbitrarycontinuousboundary
dataf such that (y,f, -v) f 0 (85)

at eachpoint y of the boundarysurface. Then thereexists smoothboundarydata, whosedifferencefromf

has arbitrarilysmall absolutevalue,for which no solution of the Dirichlet problem is possible.
Proof. Let P be a point of OQ where f takes its greatest value, and let e and a be fixed
positive numbers with e < 1. Consider smooth boundary data g such that

g-fjI <?e on OQ n {r>a} (86)

and g-f+ 3e at P.
We shall show that ife is suitably small, and a is chosen appropriately, then the corresponding
Dirichlet problem cannot have a solution.
Suppose in fact that a solution u did exist corresponding to the given data g. Then by
lemma 1, if a is sufficiently small, say a < a (e), we have
u < m+e on Q n {r =a} (87)
where m is the supremuim of g on the boundary set OQ n {r > a}. (The hypothesis I1 - ,
of lemma 1 is treated exactly as in the proof of theorem 1. We note also that if m < 0, then
(87) must be replaced by u < e; the rest of the proof carries through unchanged, however.)
Since g(P) > f (P), we have also

f < wo(y,g,-v) at P
because Te is increasing in its second argument. Consequently by lemma 3 there exists
a constant e0 such that the condition

g(P)-5e' <m' <g(P)-6' (m'-supu on QinJ{r=a})

cannot occur when 6' < e0 and a < ao(e'). Moreover, it is clear that we can choose e0 and
ao(e') to be independent of e.
Now suppose e < 60 and let a =
min (a(6) a (6), al (6))

where a, (e) is defined by the continuity condition

If(y) -f(P) I < 6 if y E OQn {r < al (e)}.

We assert that the Dirichlet problem for g cannot then have a solution.
Indeed, since a < a(e) we have from (87) that m' < m+e. Thus using (86) and the fact
thatf takes its maximum value at P, we obtain

m' < m+ e < f(P) + 26 - g(P) -6. (88)

Ont the other hand, by considering points y in the set dAQn {r-a}, it is clear that
m' ~g(y) ?f(y)-e
> f(P) -2e g(P) -Se, (89)

6I Nt~~~~~~~~~~~~~~~~~OL.
264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
470 J. SERRIN
the first inequality of the second line following since a < a, (e). From (88) and (89),

g(P)- 56 m' < g(P)-e.

But this is impossible according to the earlier part of the proof, since e is less than co and
a < ao(e). This completes the demonstration of theorem 2.
The following corollary generalizes theorem 1.
COROLLARY. Sutppose thatequation(1) is singularlyellipticandthattheasymptoticformulae (26)
and (27) hold.AssumefurthermorethatX > Oforall sufficiently largevaluesof u.
Let Q be a boundeddomainwhoseboundary is of class C2, andsupposethereis somepointy of the
boundary surfacewhere f < lim O(y, u P)
U->00

Thenthereexistssmoothboundary
datafor whichnosolutionof theDirichletproblemis possible.
Proof. From the proof of theorem 2, one sees that its conclusion remains true even if (85)
holds only at a single point P on the boundary wheref takes its greatest value. Consequently,
to prove thel corollary, it is enough to start with sufficiently large constantboundary valuesf
and then to apply theorem 2.
While the global failure of the condition-J > le(y,f, ) indicates that the Dirichlet
-

problem is badly posed, it remains of interest to construct explicit boundary data such that
9 >o (y,f,-v) -c

and yet for which no solution of the Dirichlet problem exists. We have the following
particular result in this direction, complementing theorem 2.
THEOREM 3. Supposethatequation(1) is singularlyellipticandthattheasymptoticformulae(26)
and (27) hold.Assumefurthermorethat thefunctionV (x, u,p), 1pf > 1, is boundedbelowfor u > 0
and boundedabovefor u < 0.t
Let e > 0 begiven. Thenthereexistsa smoothlyboundeddomainQ andsmoothboundarydataf,
suchthat a >- f(y,f,-v)6e and X (90)

at eachpointy of theboundary
surface,yet theDirichletproblemhas nosolution.
Proof. Let y be a fixed point in the underlying Euclidean space, say the origin of coordi-
nates for definiteness, and let v be a fixed direction at y, say P - (1, 0, .0., 0). We can suppose
without loss of generality that y and P satisfy the condition
eo(Y,O, -)v t(Y- 0, v),
le
since if the opposite inequality holds one can replace u by -u
Now putk= (y,O-)1

and let S denote the quadric surface.

X1 = 2;k(p?2+ v + X2)-
22the right of S will be denoted by
As in the proof of lemma 3, the domain to We consider
the set {r <a} n 8

t It is in any case bounded fior x, u in compact sets.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 471
where r denotes distance fromy, and we let Q be the domain obtained from it by just barely
rounding off the edges.
For this domain it is evident that
k at the boundary point y
and X -a-' on the boundary set {r a}.
On the rounded-off portions of the boundary X and f can be assumed to be arbitrarily
large, say greater than 2a' for definiteness. Next, we introduce smooth boundary dataf
such thatf _ 0 on the part of the boundary disjoint from S, 0 <f < 3e' on the part of the
boundary coincident with S, andf z 3e' at -y.
Clearly if a and 6' are sufficiently small, then (90) holds at each point of ag, while
moreover f ( -, 36', - i) - 6.

We now assert that if 6' and a are still smaller if necessary, then the corresponding Dirichlet
problem cannot have a solution. Suppose in fact that solution u did exist for the domain Q
and the dataf described above. By hypothesis there exists a constant R > 0 such that
W(x,u,p) -1/R for JpJ> 1, u > 0.
Assuming a < R, theorem 6-3 then implies
u<3+2R in Q, (e'_1).

Since u is bounded above, the argument of lemma 3 may be applied. In particular, if 6' is
sufficiently small and a < ao(e'), then the condition
f (y)-56' < m'< f (y)-' (91)
cannot occur, where m' denotes the supremum of u on the part of the boundary of Q which
is disjoint from S. But with this final choice of 6' and a the proof is completed, for by our
construction f (y) 3e' and m' -- 0

and these values do satisfy (91).

Comments. The domain Q constructed above is evidently convex if either 0 (y, 0, -v) or
- e ( y, 0, v) positive, while otherwise it is necessarily non-convex.
is
In addition to the results noted at the end of? 16, Bernstein (i9i2, page 469) derived one
further theorem for the case n 2, g -2, namely if a and V are analytic and have certain
expansion propertiesfor largep, and if (in our notation) the function le?vanishes, then there
are nonconvex domains for which the Dirichlet problem is not solvable. This result is
clearly a special case of theorem 3.

19. THE EQUATION VD2u _ 0

In this section we intend to specialize our earlier results to the important case when
equation (1) has the form a?(x, u, Du) D2u =0. (1')

This specialization is particularly interesting in view of the relatively complete state of the
theory when v6 0, as well as the simple nature of the conclusions which are obtained. We
6I -2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
472 J. SERRIN
note in particular that the Euler-Lagrange equation associated with the regular variational
problem
e JF(Du) dx - O, F E C3,

has the form (1'); in addition the representation (54) clearly holds, so that the following
theorems are applicable in this case for all. dimensions.
Theorems 1 and 2 below are concerned with the existence of solutions of the Dirichlet
problem for arbitrarily given boundary data, while theorem 3 deals with the question of
nonexistence.
THEOREM 1. Let Q be a boundeddomain in n-dimensionalEuclidean space, whose boundaryis
of class C2.
Assumeeithern 2 or that (54) holds. Supposemoreoverthat thereexists a positive constant,u such
that the invariantfunction e - psQp satisfiesthe condition

&:# IP
for all sufficientlylarge values of p.
Then the Dirichletproblemfor equation(1') in Q is solvablefor arbitrarilygiven boundarydata of
class C2. The solutionis uniqueifQ1 is independentof u.
This is proved in exactly the same way as theorem 14-1, except that in obtaining an
interior estimate for lDvl one can now apply either theorem 12d1 or 13?1 (Note that
theorem 6 2 can be used at the first stage to obtain a bound for IvIin Q.) Uniqueness follows
from theorem 6-5.
THEOREM 2. Let Q be a boundeddomain in n-dimensionalEuclidean space, whose boundary is
of class C2. Assume that both representations(26) and (54) hold.
Then the Dirichlet problemfor equation (1') in Q is solvablefor arbitrarilygiven C2 boundary
valuesprovided J. > 0 (92)
at eachpoint of the boundarysurface, wherethe invariantsXY andX are definedin ? 9.
This is proved in exactly the same way as the first part of theorem 14.3, except that in
obtaining an interior estimate for lDvl one can now apply either theorem 12 1 or 13-1.
The equality sign can be allowed in (92) provided also that

0,Q/00- O , d- vo _ 0(6'llPl),
and e > ,u for some positive constant ,u and all sufficiently large values of p. Indeed, it is
precisely these conditions which are needed to complete the proof of the second part of
theorem 14'3.
THEOREM 2'. Let Q be a boundedstrictly convexdomainin n-dimensionalEuclideanspace, whose
boundaryis of class C2. Assume eithern - 2 or that the representation(54) holds.
Then the Dirichlet problemfor equation (1') in Q is solvablefor arbitrarilygiven boundarydata
of class C2.
This is proved in exactly the same way as the first part of theorem 14s5, except that the
interior estimate for IDvl is obtained fromreither theoremn 12' 1 or 13'1.
Theorem 2' holds also for convex (not.necessarily strictly convex) domains Q~provided
that the minimumn eigenvalue of 5/ is bounded below by a positive multiple of lpj2 as

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 473
p becomes large. To see this, we note that s is then bounded below by a positive constant,
whence the last part of the proof of theorem 14-5 is immediately applicable.
Remarks.The case n = 2 of theorem 2' has been considered by a number of authors. The
result is essentially due to Bernstein (I9IO, I9I2), though in fact he never quite considered
the situation in full generality. The most complete treatment is due to Nirenberg (I952).
With the help of the a prioriestimates of Ladyzhenskaya & Uraltseva (cf. ? 1), the result for
n > 2 and a = S(p) can be obtained by exactly the hyperplane barrier methods employed
by Bernstein and Nirenberg; cf. also the work of Gilbarg (I963) and Stampacchia (I963)
for the case when (1') has divergence structure or arises from a variational problem. The
result of theorem 2' seems to be new when / has the representation (54), as does also the
result when S2is convex but not strictly convex.
Turning to the question of non-existence of solutions, we have first the following
immediate corollary of theorem 18 1.
THEOREM 3. Supposeequation(1') is singularlyellipticaccordingto thedefinition
givenin ? 17.
Assumefurthermorethattheasymptotic
formula (26) holds.
Let Q bea boundeddomainof classC2.If eithercondition
' O or O
< f<
holdsat a singlepointof theboundarysurface,thenthereexistssmoothboundary datahavingarbitrarily
small absolutevaluefor whichno solutionof theDirichletproblemis possible.
When n = 2 the asymptotic condition (26) can be dispensedwith, and the conclusion made
even sharper. The result is as follows.
THEOREM 3'. Assume n = 2. Let equation (1') be singularly elliptic, and supposethat
of* < 1 -s
for somepositiveconstant/ andfor all u with IuI < 1.
If Q is a bounded
non-convexdomain,thlenthereexistssmoothboundary datahavingarbitrarilysmall
absolutevaluesuchthatno solutionof theDirichletproblemis possible.
Proof.According to a theorem of Leja & Wilkosz (cf. Valentine I964, theorem 4 8) there
must be some boundary point P of Q where there is an internally touching arc of a circle
with negative curvature A. Now let u be a solution of the Dirichlet problem in Q, with
boundary valuesf, and let m' be the supremum of u on the set S2n {r a}, where r denotes
distance from P. Then the following assertion holds:
For anygivene > 0, we have f(P) S m'+ e
smallvaluesof a (depending
for all sufficiently onlyone, Q, k, andthestructure of theequation).
Assuming this for the moment, the proof can then be completed as in the case of theorem
18X1,with the above result taking the place of lemma 3.
To prove the assertion, we follow the argument used for the proof of lemma 3. The only
slight changes to take into account are that the internally touching circular arc here replaces
the quladricsurface S, while instead of (82) we must use the relation

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
474 J. SERRIN
(since trace ? = 1). Now vPv = * <? 1 by hypothesis, whence recalling that k is
-#
negative we obtain f <1 ,k + O(a).

The rest of the proof of lemma 3 carries over unchanged (and indeed in somewhat simpler
form) to the present case. This completes the demonstration of theorem 3'. (A corresponding
result can be given for higher dimensions, but involves the concept of boundary points
where there is an internally touching spherical element with negative curvature. This and
other similar generalizations are left to the reader.)
The preceding results can occasionally be combined in such a way as to produce necessary
and sufficient conditions for the solvability of Dirichlet's problem. To illustrate this, we may
state the following
THEOREM 4. Let Q be a boundeddomainin n-dimensional Euclideanspace,whoseboundary is of
class C2.Assumeeithern 2 or thattherepresentation (54) holds.
Supposefurtherthat (1') is singularlyelliptic,andthat theasymptoticformulae(26) holdswiththe
errorterm /-!o o(6/jpj).

Assumefinallythatthereexistsa positiveconstant/ suchthate > pfor all sufficientlylargevaluesof p.

ThentheDirichletproblem for equation(1') in Q is solvablefor arbitrarilygivenboundary dataof
classC2 if andonlyif >
at eachpointof theboundary surface.
This result is a combination of theorem 2 and theorem 3, with lemma 17-2 and the remark
following the proof of theorem 2 being taken into account.
COROLLARY. Let Q be a bounded domainin n-dimensional Euclideanspace,whoseboundary is of
class C2. Assumeeithern 2 or that (54) holds.
Supposefurthermore genreg, 1 < g <s 2, and thattherepresentation
that(1') hasa well-defined (26)
holds with i-V'0=o(p 1-g)

ThentheDirichletproblem for equation(1') in Q is solvablefor arbitrarilygivenboundarydata of

class C2 if andonlyif X , 0 'ateachpointof theboundary surface.
The minimal surface equation obviously satisfies the hypotheses of this corollary, with
- X- = Hin this case. Thus, as a special case, we obtain a recent result of Howard Jenkins
and the author (I968) for the minimal surface equation in higher dimensions. In fact,
the following more general result holds.
The Dirichletproblemfor theminimalsurfaceequationin a domainof classC2is solvable/for arbi-
trarilygivencontinuous boundaryvaluesif andonlyif H > 0 at eachpointof theboundary. Thesolution
is uniqueif it exists.
Proof. The necessity of the condition H > 0 is obvious from what has gone before.
Now suppose that H > 0 at each point of the boundary of S, and let continuous boundary
valuesfbe given. To prove the existence of a correspondinlg solution of the Dirichlet problem,
consider new boundary valuesfm of class 02 such1that

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 475
Each of the approximating problems has a unique solution ur according to the preceding
corollary. Hence by the maximum principle (theorem 6.5) the sequence {ur} tends uni-
formly in Q to a limit function u(x). Clearly u on MQ; it remains only to show that u is a
solution of the minimal surface equation in Q.
To this end, we apply a recently announced result of E. De Giorgi, namely that the gradi-
ent of a solution of the minimnalsurface equation can be bounded in terms of the maximum
of the solution and the distance to the boundary (cf. Bombieri, De Giorgi & Miranda
I969). Using this result and the Schauder theory, it follows that the functions ur are uni-
formly bounded in C3on any compact subset of Q. Consequently by Arzela's theorem and an
easy argument we see that u must be of class C2 in Q and that ur -? u in C2on any compact
subset. Since the functions umare solutions of the minimal surface equation in Q it is clear
that u is also a solution in Q, completing the proof.
We note that the same technique applies when the boundary of Q is not of class C2 but
can be uniformly approximated by C2 boundaries which satisfy the condition H ? 0 at
each point.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
476 J. SERRIN

CHAPTER IV
In this chapter we shall give an extended discussion of various examples and special cases
in order to provide concrete illustrations of the general theory developed earlier. Sections
22 through 24 are devoted to the case of surfaces having prescribed mean curvature, where
the results have special geometric significance beyond their simple interest as examples.

20. THE MINIMAL SURFACE OPERATOR

Here we shall consider various equations arising when the left-hand side of (1) has the
form [(1 + IDuI2)I-DuDu] D2u

as in the case of the minimal surface equation. We note that the coefficient matrix ,', after
normalization to unit trace, is (1+ P12)Ip
n+(n-1) IpI2
and that the corresponding invariants XYand X are given by the simple relation (cf. ? 9)

where H devotes the ordinary mean curvature of the boundary surface in question.
Perhaps the most important case concerns surfaces of constant mean curvature. As is well
known, a smooth surface in (n+ 1)-dimensional Euclidean space with the non-parametric
representation z u(x) has mean curvature A provided that the function u u(x) satisfies

[(1+lDul2)I-DuDu]D2u nA(1+lDul2)-2.
The corresponding Dirichlet problem in a given domain Q has the geometric content that
the surface z- u(x) is defined over Q and takes on given ordinate values on the boundary
set MQ.In particular, when n 2 the surface z u(x, y) spans a given simple closed curve
in the three dimensional (x, y, z)-space.
As already remarked in the introduction of the paper, the following result holds:
Let Q be a boundeddomainin n-dimensionalEuclideanspace, whoseboundaryis of class C2. Then
the Dirichlet problem in Q for surfaces of constantmean curvatureis solvablefor arbitrarilygiven
boundarydata of class C2, if and only if the meancurvatureH of the boundarysurface OQsatisfiesthe
condition
H _n A (93)

at eachpoint of the boundary.The solutionis uniqueif it exists.

Proof. We first show that (93) is necessary. Suppose in fact that (93) fails at a single point.
If At > 0, then theorem 18 1 shows that there exists boundary data having arbitrarily small
absolute value for which no solution of the Dirichlet problem is possible. If A < 0 the same
proof applies, if we merely replace u with -u.
The proof of sufficiency is more difficult. We shall give here an argument applicable only
in two dimensions. The proof for higher dimensions will be deferred until ? 22, where in
fact we shall derive an even stronger result.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR- ELLIPTIC DIFFERENTIAL EQUATIONS 477
Suppose then that n - 2. In this case (93) implies that Q is contained in a ball of radius
1/2 1A.1 The hypotheses of theorem 6 3 are then satisfied, and consequently the solvability
of the Dirichlet problem is guaranteed by theorem 14a3. The uniqueness of solutions
follows from theorem 6.5, whatever the dimension. This completes the demonstration of the
theorem, subject of course to the deferred existence proof in higher dimensions.
As we have already remarked, when n -2 condition (93) implies that Q is convex and
contained in a ball of radius 1/2 IAI. In higher dimensions no such simple conclusions are
available, and (93) can in fact be satisfied even by domains which are not simply connected.
For example, the torus generated by revolving the disk (x - a)2+y2 = b2 about the y axis
satisfies the condition whenever b < 1/3 1AI and a _> 2b+ 3 1AIb2/(I-3 JAIb).
The minimal surface operator also appears in the equation of Laplace and Gauss
governing the free surface of a stationary fluid under the combined action of gravity and
surface tension. The following result holds:
Let Q be a bounded domainin n-dimensional Euclideanspace,whoseboundary is of class C2. Then
theDirichletproblemin Qfor theGauss-Laplace equation
[(c + IDul2)I-DuDu] D2U= cu(l + IDu2) c> 0 (94)
is uniquelysolvableforgivenC2 boundary valuesfprovidedthat
If I < (n -1)H/c (95)
at eachpointof theboundary.Moreover,condition(95) is bestpossible.
Proof. Existence follows directly from theorem 14-3 (with theorem 6-2 used at the final
step.) By saying that (95) is best possible we mean that, given any domain Q and any
positive number c, then the following statements hold:
(A) If H < 0 at a single point of the boundary there exists smooth boundary data g
such that lgl ? c and yet for which no solution of tlle Dirichlet problem for (94) exists.
(B) If H > 0 at each point of the boundary there exists smooth boundary data g such
that IgI< (n- 1) H/c+? and yet for which no solution of the Dirichlet problem exists.
The first assertion follows by settingf - 0 and using theorem 182 (together with the fact
that its conclusion remnainstrue even if (85) holds only at a single point P of the boundary
wheref takes its greatest value).
In case (B), let P be an arbitrary point on the boundary and let smooth boundary values
f be assigned such that f (n -1)H/c+ 6 at P
and f : min{f(P), (n-1) Hlc+ 6} elsewhere.
The conclusion then follows exactly as in case (A), completing the proof of the theorem.
COROLLARY. Let Q bea ballof radiusR. ThentheDirichlet problemaboveis solvableforarbitrarily
givenC2 boundary valuesf satisfyingtheconditionIf I < (n- 1)/cR. Conversely,if If I > (n-1)/cR
at a singlepoint of the boundary,thentheDirichletproblemis not wellset.
It would be interesting to know similarly whether the failure of condition (94) at a single
point implies the existence of nearby data for which the Dirichlet problem cannot be solved.
Because of their general interest we shall consider two further equations,
[(1 + lDuel2)
I-DuDul D2u =nA(i + IDuI2)AO (96)
and [(1 + IDuI2) I--DuDu] D2u =cu(1 + lDuI2)AO (97)
62 VOL. 264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
478 J. SERRIN
where A, c, and 0 are constants, c > 0. The case 0 = 3 has already been discussed in the
foregoing work. We may therefore restrict the following treatment to values 0 + 3.
When 0 > 3 it is evident that both equations are irregularly elliptic; accordingly the
Dirichlet problem is not well posed in that case, whatever the domain.
When 0 < 3 a maximum principle is available for both equations (theorems 6 3 and 6 2),
and %- 0. Also when 0 < 2 both equations are boundedly non-linear. Finally theorems
12*1and 13-1 are applicable, with

and n(?I2) I 0-' > 0, Y C n > 0,

respectively. The argument of theorem 14 3 thereforemay be applied to prove the existence
of solutions of the Dirichlet problem for (96) and (97) when 0 < 3.
On the other hand, according to theorem 18 1 the Dirichlet problem for 0 < 3 cannot
generally be solved if H < 0 at some point of the boundary. A slight modification of the
proof of the antecedent lemina 17-3 in fact allows this result to be strengthened: when
2 < 0 < 3 the Dirichlet problem cannot always be solved if H 0 at some point on the
boundary (for this modification alone, we make the tacit assumption that the boundary of
Q is of class C3 in the neighbourhood of points where H 0). Combining the preceding
results, we obtain the following theorem.
Let Q be a boundeddomainin n dimensionalEuclideanspace, whoseboundaryis of class C2. Thenthe
Dirichletproblemfor equations(96) and (97) in Q is solvablefor arbitrarilygiven C2 boundaryvalues
underthefollowing circumstances:
(i) when0 <_ 2, if and only if H > 0 at eachpoint of the boundary,and
(ii) when 2 < 0 < 3, if and only if H > 0 at eachpoint of the boundary.
When0 > 3 the Dirichletproblemis notgenerallysolvable, whateverthe domain.
In a short paper published in 1910 Bernstein listed both equations (96) and (97) as
examples (though with n equal to 2 and 0 restricted to integervalues), and showed that the
corresponding Dirichlet problems are solvable for arbitrary (analytic) boundary data in an
arbitrary strictly convex (analytic) domain if and only if 0 < 2.
In the same paper Bernstein introduced a further pair of equations, namely
+2uxu uX+ (I + U2Y)UYY=( + U2+U2Y)20
(I + U2)UXX
and (I+ u2)u+ 2uxuyuy+(I + u2) uyy = u+ u2+ U2)-1
with the parameter 0 once more being restricted to integer values. A recent note of mine
(I967, page 1834) called attention to the n-dimensional generalizations of these equations,
and gave precise conditions for the solvability of the correspondingDirichlet problems. The
interested reader should be able to supply the necessaryjustification of those resultswithout
undue difficulty.
21. VARIATIONAL
EQTUATIONS

Both the equation for surfaces having constant mean curvature and the equation for
fluid interfaces under the action of gravity anld surface tension can be obtained from an
appropriate variational problem. It is the purpose of this section to present several further

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 479
variational equations of interest, including higher dimensional generalizations of the
problems introduced by Bernstein in his memoir on the calculus of variations (19I2).
We begin with the example
d|N/(1+ Xj2+ Duj2) dx = 0,

for which the corresponding Euler-Lagrange equation has the form

[(1 + Ix12+ Dul2)I-DuDu] D2u x Du.
By theorems 14 3 and 18 1 the Dirichlet problem in a bounded domain Q with Q2 boundary
is solvable for arbitraryboundary values of class C2 if and only if H 0 at each point of the
boundary. By theorem 6-5 the solution is unique if it exists.
Consider next the problem
J/(1+u2+Dul2) dx- 0,

whose Euler-Lagrange equation has the form

[(1 +u2+ Dul2)I-DuDu] D2u - u(I +u2+2jDuj2).
Again by theorems 14-3 and 18 1 the Dirichlet problem is solvable for arbitrarily given
boundary values if and only if H ? 0 at each point of the boundary (note in particular that
an a priori bound for IuIis given by theorem 6 2). Uniqueness of the solution follows from
theorem 6 6.
In both preceding cases the slight modification of the integrand from thlearea expression
V(1 + IDu 2) has no effect on the type of Dirichlet problem which can be solved. A rather
different situation arises for the variational problem

&f0(x)v(1 + Duj2)dx-0

where 0 (x) is a given positive function of class C2. The corresponding Euler-Lagrange
equation is Du
[(I + Dul2)I-DuDu] D2u=D (I+ Du 2)

thus while the left-hand side involves only the minimal surface operator, the right-hand side
yields the relation 1 o--Do
?o n-1-I
Hence the geometric conditions required for the solvability of Dirichlet's problem neces-
sarily differ from those for the minimal surface equation. In particular, by theorem 14*3the
Dirichlet problem in a bounded domain Q with C2 boundary is solvable for arbitrary
boundary values of class C2 provided that
1 v-Dq
H>1 q (98)
at each point of the boundary. It can be shown that this condition is also necessary for the
solvability of the Dirichlet problem for arbitrary data, though we shall nlotdo so here.
The special case b V(1+ xl2) was remarked by Bernstein (I9X2, page 479). If for
definiteness we take Q to be convex and to contain the origin, then(98) is always satisfied.
62-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
480 J. SERRIN
When Q is a ball of radius R whose centre lies at a distance S from the origin, (98) becomes
1> 1 S2_R2--X12
R > n - I 2R(1+ IX12)
that is 2(n -1) - (2n-1) Ix12 > S2-R2.
The niinimum value of IxIencountered on the surface of the ball is JR- S. Thus we obtain
the following solvability condition for the Dirichlet problem in this special case,

(S-R) R-S ;

in particular the problem is solvable whenever

(1) Q contains the origin, or
(2) the radius of Q is less than 2(n- 1), or
(3) the centre of Q lies in a ball of radius 2J{n (n- 1)} about the origin.
This result serves to emphasize the delicacy of the situation when any but the simplest
equations are treated. We observe specifically thiateven if the Dirichlet problem is solvable
for a given ball Z, it need not be solvable for all balls contained in X, nor even for all smaller
concentric balls!
22. SURFACES HAVING PRESCRIBED MEAN CURVATURE
In this section we shall generalize the result of ? 20 to include cases where the surface
z - u(x) has a prescribed mean curvature A(x) of class Cl.
An example will make clear that beyond the simple geometric condition (93), some
additional requirements on the behaviour of A (x) are necessary in order for the Dirichlet
problem to be solvable. Consider in particular a smooth function A(x) which is identically
2 when Ixj< 1 and vanishes identically when -x- - 1. Then (93) is satisfied if Q is the unit
ball, but according to a result of Bernstein (igio, page 243) no solution of the Dirichlet
problem can possibly exist, no matter what boundary data is assigned.
In the following result we shall give a family of sufficient conditions, depending on a
parameter a, such that the Dirichlet problem is solvable for a given domain Q whenever
A(x) satisfies any one of the conditions in the family.
For points y on the boundary of Q, let k, (y), ..., k1,, (y) be the principal curvaturesof the
boundary surface,listed in decreasing order of size. Let d = d(x) denote as usual the distance
to the boundary of Q, and let Q, be the (open) subset of Q consisting of points x such that
(1) the closed ball of radius d(x) centred at x touches the boundary of Q at a single point
y-y(x), and
(2) d(x) < llkl(y).
For x in Q we can now define

-
H(x = nn-I1 -k; (y d(x) 'ua
.
so, ~~~xXQiuQ.
Clearly H(x) agrees with the ordinary mean curvature of the boundary surface for points
x on AQ Moreover, although the fact is unnecessary for our present purposes, we remark

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 481
that wnen the boundary is of class C3 the function 11(x) can be interpreted as the mean
curvature of the level surface d = constant passing through the point x.
The maximum value of d(x) in Q will be called the internal radius of Q, and will be denoted
by d1. We can now state the main result of the section:
Let Q be a boundeddomainin n-dimensional Euclideanspace,whoseboundaryis of class C2. Let
A(x) beof classC1in theclosureof Q, andsupposethatthereis someconstanta, 0 < a < 1/d1, suchthat

nIA (x)I< (I-ad)H(x) + a1 in Q (99)

ThentheDirichletproblemin Qfor theequation
[(1 + IDul2)I-DuDu] D2u -nA(x) (1 + JDuj2)2
valuesif andonlyif
is solvablefor arbitrarilygivenC2 boundary
n
H > n-I IA(y)

at eachpointy of theboundary.Thesolutionis uniqueif it exists.

The proof is somewhat different in the three cases 0 < a < l/dl, a 1/dl, and a 0, and
accordingly we shall consider these situations separately.

Case1. 0 < a < /ld1

Supposing that the boundary of Q is of class C', we assert that
Jul< supIfI+l/a in Q. (100)
Assuming this for the moment, the result then follows exactly as in the proof of theorem 14*3,
with (100) taking the place of the final hypothesis on the coefficient X.
It is therefore sufficient to prove (100). Let
h(d) V- {2ad- (ad)2}1a, 0 < d < 1/a.
Since 0 < h < i/a the assertion will follow provided we can establish the relation
lu(x)I < sup If I+h(d) d --d(x), (101)
for all xin Q.
Suppose for contradiction that (101) fails at some point where u > 0. Since (101) neces-
sarily holds on the boundary of Q, it follows that there exists a positive constant K and a
point x in Q such that u(x) < sup If I+h(d)-+K in Q (102)
with equality holding at x. There are now two cases to consider, depending on whether x
is in the set Ql or in its complement.
I. -xE Q,. According to the work of ? 3, one sees easily that relations (8) hold and d(x) is
of class C2 in Q,. Hence, letting v(x) denote the function on the right-hand side of (102), we
find by direct computaion (as in the proof of lemma 4 1) that
[(+ lDvl2)I-DvDv]D2v h"_ h' ki(y)
3
(
(1+lDvl2)2 ~ ~ (1+h' 2) k'2 1-k () d(x)
2)~ - -ff ad(In1 h')1()nA

for x in Q1 With the help of a standard maximum principle argument this imnplies u _v
in Q1 which is impossible since u < v on dLL

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
482 J. SERRIN
II. -x0 Q?. Let y be any point where the closed ball of radius d(x) centred at x touches the
-
boundary of Q. Then for any point x on the closed line segment joining to x we have
d(x) |Ix-Yl.
Since h' > 0, it follows from (102) that Du = 0 at x. The level set u = constant - u passing
through x is consequently a surface of class C2 in the neighbourhood of -x. This being the
case, there exists a closed ball Z tangent to this surface at x, such that u ? u in Z. But then
by (102) it is clear that d > d(x) in Z, and a fortiori d > d(x) on
X22.
If the centre of Z is xo and its radius o-, then according to the geometric meaning of d(x)
the ball of radius d (x) + uo-centred at xo must be contained in Qi. This implies first of all that
xo is on the extension of the line segment joining y to x, and in turn that x is in Q,. This is
contrary to assumption, and accordingly we have proved that (101) must hold whenever
u > 0. Since a similar argument applies to points where u < 0, this completes the proof of
(101).

Case2. a 1/Id,
Let a be an arbitrary constant, 112d, < a < l/dl. If A satisfies (99) for a a, then by
what has already been shown in case 1,
Jul supIfI+2d, in Q
(provided as always that the boundary of Q is of class C0). The solvability of the Dirichlet
problem when a I/d, can now be established as in the proof of theorem 14-3, since each
approximating problem obviously can be chosen so that (99) holds for some a < 1/d,.

Case 3. a 0
As in the proof of case 1, it is sufficient to establish an a prioriestimate for u when the
boundary of Q is of class C3. To this end, we put
v(x) sup If I+h(d), O< d < d1,
ecdi
where h(d) (1ecd) d = d(x),
C

and c nsup IA(x) + 1. Now for x eQ,

[(1+lDvI2)I-DvDv]D2v h" h' ( 1)H(x)
(I+IDvl) (1+h2)- (+h'2)2
h"l h'
+- ~ nA (x)
(-l +h-'2) -2+ (I +h'2)W ()

? nA(x) + (1 A I2) (1- +

? nA.(x) + + =nA(x)

the second to last step holding by virtue of the fact that h' > 1. This being shown, the proof
of case 1 carries over with only minor changes to the present situation. Thus we have
lu(x) I? sup If l+h(d) < sup If l + (ecd1 l)/c
in Q2. This completes the proof of case 3, and consequently also of the main result.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 483
The necessity of the condition H ? (n/n- 1) IAl follows directly from theorem 18 1 in
case A does not change sign. If A takes on both positive and negative values the argument is
somewhat more complicated but follows the same basic lines (we shall ornit the details).
Finally, the uniqueness of the solution is an immediate consequence of theorem 6-5.
The result of case 3 is worth stating as a separate corollary because of its special interest.
Let Q be a boundeddomainin n-dimensionalEuclideanspace, whoseboundaryis of class C2. Then
the Dirichletproblemin Qfor surfaceshavingprescribedmeancurvatureA (x) is solvablefor arbitrarily
given C2 boundaryvaluesprovidedthat

nl() 1H(x)

at eachpoint x in the closureof Q. The solution is unique if it exists.

Since H(x) obviously takes its minimum value on the boundary of Q, this corollary clearly
includes as a special case the theorem stated in ? 20. The results of ? 20 are thus completely
proved.
The case a - 1 /d, of the main result is also of individual interest, and yields the sufficient
condition:
n
H -nl A (y), yon dQ
n-i

and A (x)l + 11--in1--d H(x), x in Q.

nd, \ n~ ,
By dropping the final term on the right-hand side we get the sirnple, though not especially
strong, sufficient condition IA(x) I < I /ndl.
It is unfortunately not particularly easy to give an explicit formula for H(x) for any but
the simplest domains. On the other hand, for a given domain Q it is always possible to
employ in (99) the function H(x) corresponding to any domain Q containing Q. A case of
particular importance occurs when Q is a ball of radius R. If r denotes distance from the
centre of Q, then d(x) R- r, H(x) - l/r

a I1-aR an
and (I -ad) H(x) +_ r +

We have proved the following result:

Let Q be a domainof class C2 in n-dimensionalEuclideanspace, containedin a ball of radiusR. Let
A (x) be of class C' in theclosureof Q, and supposethat thereis someconstanta, 0 < a < 1/R, suchthat

IA(x)I <a+ n-in 1-aR in Qi. (99')

r

Then the Dirichlet problem in Q for surfaces having prescribedmean curvatureA (x) is solvablefor
arbitrarilygivenboundaryvaluesof classC2 if andonlyif

at eachpoint y of the boundary.The solution is uniqueif it exists.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 484 J. SERRIN
It is interesting to observe that condition(99') is bestpossible, in the sense that when A(x) is
subject only to the restrictions
n___. __aR n-
A (x) a ?n1 i-aR+ in Q, IA(x) n H on a
it may not be possible to solve the corresponding Dirichlet problem. To prove this, suppose
that Q is itself a ball of radius R, and consider a smooth non-negative function A(x) such that
n-i 1
At(x)<% Rwhen r=R

n-I I-aR
and A(x) - Ia+ r +6 when S<r<R-S
n r
where S will be determined later. Using the identity
[(I + Dul 2)I-DuDu] D2udU ( Du
(I + IDul 2)32 \(1 + JDuJ2) 2J

we obtain, for any eventual solution u of the Diriclhlet problem in Q,

f I-Du? -
nA(x) dx < -- ds <_?Rn-I
i91+IDuH2)-dsOnR
where con
denotes the area of the unit sphere in n dimensions. Also

nA(x) dx >f (n - I) ( aR+ an +?) d

> 6ln[(J+6nh) (R &)n(S+ n

Combining the preceding two inequalities and letting S 0 yields

6O)Rn-1 > 60) n -) Rn-1

i1 +6CR

which is impossible. Hence a contradiction occurs for all sufficiently small values of S,
proving our result.
The value of the preceding result lies in the fact that it does not require a specific calcula-
tion of the function H(x) for the given domain Q. We conclude the section with another
theorem having the same property.
Let Q be a boundeddomainin n-dimensionalEuclideanspace, whoseboundaryis of class C2. Let A(x)
be of class C1 in the closureof Q, and supposethat

<n 2
DAI _n-I A in Q. (103)
Then the Dirichlet problem in Q for surfaces havingprescribedmean curvatureAl(x) is solvablefor
arbitrarilygiven C2 boundaryvalues if and only if

H n-
n-iiq A(Y)
A(y) 1 (104)

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 485
Proof.We shall show that (103) and (104) imply
I
| A(x)I <_ n_ H(x), x E-Q, (105)

which will obviously complete the proof (see case 3 of the main result).
The rate of change of H(x) with respect to d in a direction normal to the boundary is
clearly given by dH(x) 1 E ki(y) 2

dd n-I 1-ki(y) d(x))'

Consequently, letting a prime denote summation over all values of i for which ki(y) is
positive, we have dH(x) I / ki(y) \ 2
dd n- I -ki(y)d(x)

(n- 1) 2 (1 I -k (y) d(x) H(X) 2

since H(x) > 0 in the present case because of (104). Thus for x in Q,

&dd n H(x) _'n-I( n H(x)) (106)

If we now integrate both (103) and (106) along normals to the boundary, and make use of
(104) as an initial condition, it follows that

A(x) I _< n- H(x), x E-Q, u MQ.

Since (105) obviously holds at the remaining points of Q, this completes the proof.

23. SURFACES OF PRESCRIBED MEAN CURVATURE IN CENTRAL PROJECTION

Let Z be a smooth surface possessing a one-to-one central projection onto a domain Q of
the unit sphere in (n+ 1) dimensions. If the radial values of Z are assigned on the boundary
of Q, and if the mean curvature of Z is prescribed as a function over Q, we say that v is a
solution of the Dirichlet problem for surfaces of prescribed mean curvature (in central
projection). We shall adopt the sign convention that the mean curvature of Z is based on
the normal to the surface on the same side as the extended radius vector. The following
result then holds.
Let Q be a domainon the unit spherein (n+ 1) dimensions,whoseclosureis containedin an open
hemisphere andwhoseboundaryis of class C2. Let A(x) be a non-negative functionof class C1 defined
overthe closureof Q. ThentheDirichletproblemin Qfor surfaceshavingprescribedmeancurvature
A(x) (in centralprojection)is solvablefor arbitrarilygivenradialvaluesf of class C2 assignedon the
boundary of Q, providedthat
g> n- A(Y)f (y)

at eachpointy of the boundary,whereHg denotesthegeodesicmeancurvature surface

of the boundary
(withrespectto theunderlyingunitsphere).Thesolutionis uniqueif it exists.
63 VOL. 264. A.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
486 J. SERRIN
The assigned radial valuesf are of course necessarily positive, so that (just as in the
parallel projection problem studied in the previous section) the mean curvature of the
boundary must be non-negative. In particular, when n = 2 the restriction of Q to a hemi-
sphere is already a consequence of the solvability condition.
If A(x) is identically zero the solution is a minimal surface in central projection. For this
case the solvability condition becomes simply
Hg > 0 (107)

at each point of the boundary. The case n = 2 of this result is due to Rado (1932), his proof
resting on the general solution of Plateau's problem for Jordan boundary curves. We note
thatsincetheassignedboundary valuesdo notappearin (107), thisconditionis alsonecessary.
For non-constant curvatures it is remarkable that no restrictions need be placed on
A(x) beyond the simple condition that it does not take negative values. In this respect the
Dirichlet problem in central projection is much simpler than the Dirichlet problem in
parallel projection.
In order to prove the result we must first introduce an appropriate analytical structure
for the problem. Let 2 be a surface with one-to-one central projection onto the domain Q.
If P is a point of Z whose projection on the unit sphere is the point x, we write v = v(x) for
the radial distance of P (that is, the distance from P to the centre of the sphere). We next
introduce a system of local coordinates over Q. To this end it can be supposed without loss
of generality that Q lies in the upper hemisphere of the unit sphere, and it is convenient to
use stereographic coordinates based on projection from the South pole onto the tangent
space at the North pole.
With these coordinates, we may in what follows consider x to be a Cartesian n-tuple
(x, ..., xxn)in the tangent space, and we may similarly interpret Q as a domain in this
space.
LEMMA.The surfaceZ has meancurvatureA (x) providedthat v - v(x) satisfiesthe equation

I-DvDv] D2v - anAv(a2v2+IDvI2)l

[(a2v2+ IDvI2)
+ -ax Dv{(n-2) a2v2+(n-) +(n+1) IDvI2} (108)
a2v{na2v2
IDvI2}+
for x in Q, wherea 4
4/(4+ IxI2). Moreover,wehave

H H xv
g a 2

whereH denotestheordinarymeancurvature of theboundary as a surfacein thetangent

of Q considered
spaceendowedwithits Euclideanmetric.
This lemma is proved at the end of the section. Granting the result, in order to prove the
first part of the theorem we must show that the Dirichlet problem for equation (108) in Q
has a positive solution v for arbitrarily given positive boundary dataf of class C2, provided
that
H> a (1lA(y)f-2y.v) (109)
at each point y of the boundary.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 487
The invariant 6' for (108) is given by

t0(x, v, ) = a (n_ ljA(x) v+ 1p- x)

> 0 since A (x) > 0. Using (109), it is therefore clear that the solvability condi-
and dW60/dv
tions (62) in theorem 14-3 are satisfied. Consequently, were it not for the singularity in
equation (108) when v = 0, theorem 14-3 could now be applied to obtain the existence of
a solution of the given Dirichlet problem.
This being the case, it is natural to look for some artifice which avoids the fact that (108)
is singular when v 0. To this end, let K be a small positive constant, and consider the
modified coefficient matrix (a2g(v)+ JD2V ) I-DvDv,
wv2 for vl > K,
where g (v) 2
t 2K2 for V
v <2K
and v2 Ig(v) < K2 otherwise. Let (108) be written in the form 1v = 0, and let fv 0 be
the corresponding equation with modified coefficient matrix. By theorem 14-3 the Dirichlet
problem for /v 0 has a solution v corresponding to the given boundary valuesf. We assert
that v is positive, and, moreover, that if K iS chosen suitably small, then v satisfies the given
equation S-v 0 as required.
To prove this, consider a family of comparison functions of the form
w(x) -h(r2), r= IX1
where h, h', h"> 0. Clearly (since x Dw > 0 and g(w) > w2)

a2g(w) + Dw12 / D)a2w ax Dw-(n+

> (n-1) (2-a )h'-anAh(a h +4 h _)F_ (n+1) 2h.
By assumption the closure of Q lies in the open upper hemisphere of the unit sphere, and
correspondingly we have r < 2 there. Thus there exists a constant e > 0 such that
ar2 < 2-e in Q.
Hence Yw > 0 in Q if e(n -1) h' ? (n+1)h+nAh(h2+16h'2)k
This inequality, as well as the conditions h, h', h" > 0, is satisfied by the family of functions

h o
aexp (n(+ ) (r2 -4), 0 < a < K1

where K1is a suitably small constant depending only on , n, and the supremum of A in Q.
We can now prove the assertion. First, we have v > 0 according to theorem 6-2. Then,
using the family of comparison functions above, we find by an easy application of the
maximum principle of Eberhard Hopf (I927, page 152) that

v)min (K1, K2) exp (4 e(n-1) } inQ

where K2 ( > 0) is the infimum of the given boundary values. Choosing K to be the constant
on the right-hand side of the last inequality completes the proof of the assertion. Hence we
63-2

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
488 J. SERRIN
have establishied the existence of a positive solution of (108) taking on any given positive
twice continuously differentiable boundary valuesf.
Uniqueness of the solution is essentially a consequence of a theorem of Alexandrov (I956).
In fact, a stronger result holds: if Z and Z both possess one-to-one central projections onto
a domain Q, if the radial valuesfofZ andJof2ron the boundary of Qsatisfyf < f, and if the
mean curvatures A (x) and A (x) finally satisfy the condition A (x) > A (x), then the radial
values of Z never exceed the radial values of ?. For completeness we include a proof of this
result which is independent of the work of Alexandrov (though it uses the same basic idea).
Let v(x) and v(x) be the radial values in question, and suppose for contradiction that v(x)
exceeds v-(x)at some point x in Q. Then there exists a constant -a> 1 and a point x in Q such
that v(x) < iv(x) in Q, v(x) = cv(x).
Let u(x) o-i(x). Using the fact that -vis a solution of (108), we find that
[(a2u2-JjDuj2)I-DuDu] D2u = an(A/-) u (a2u2+jDuj2)
+ Iax *Du{(n -2) a2u2+(n-1) IDul2}+a2u{na2u2+(n+1) jDuJ2}.
Since A/oa< A we conclude from the maximum principle of Eberhard Hopf that u -v in Q,
which contradicts the given boundary conditions. This completes the proof of the theorem,
subject of course to the deferred lemma.

Proofof thelemma
It is well known (and easily verified using rectangular Cartesian coordinates) that the
local mean curvature A of a surface satisfies the variational equation
3(dA)+nA&(dV) = 0,
variations in the volume V being considered positive when they occur in the direction of the
normal defining A.
For a surface Z defined by the function v(x) as in the lemma, we have
1 an_
dV -vn+ 1d6w(x)z-n vn+1dx
n+1 n+1
where dw is the element of area on the unit sphere, and a -4/(4+ IX12); see figure 2.
In the same way, using the fact that stereographic projection is conformal, it is not hard
to obtain dA= vnV{1+ fDvI2/a2v2} d@(x) = (av)n-'J{a2v2+ JDvJ2}dx.

nan
If we put F(x, v,p) = (av)n-',{a2v2+ lpI2}, G(x, v) =+ vn+
it follows from variational calculus that
div {Fp(x, V,Dv)} Fv+ AG
wrhich upon expansion yields equation (108).
To prove the second part of the lemma, note that since stereographic projection is con-
formal the relation between H and 1Hgmust necessarily be linear,
I4Hg c+,5 * .

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 489
To determine the coefficients we can thus resort to special cases where the boundary of Q
is a plane or a sphere.
Considering first the boundaries x *v = 0 we find that fimust be directed along a ray from
the origin, that is f yx where y is a scalar. Taking next the case x *v = constant and evalu-
ating at the point x rv, we get
sin_l2
H= O, Hg , =
=Cos 1O2'r, v yr

whence y =. Finally, putting Ixj r R, we find at any point on this sphere

H~~1/r,
H l/ CosO _
4-r 2 ~
4H=sinO 4r ' -
2

so that - = '(4+r 2), completing the proof of the lemma.

N r dx.

S
2
FIGURE

Remark. The above results, as well as those of the preceding section, can be generalized
to the case where the mean curvature A is allowed to depend on v and Dv in addition to x.
We shall consider these matters in more detail in another paper to appear later.

24. THE DIFFERENTIAL OF CONSTANT

EQUATION MEANCURVATURE
In 1760 Lagrange published a new method-essentially the variational technique still
used for determining the differential equation to be satisfied by functions which minimize
an integral expression under appropriate boundary conditions and side conditions (Euler's
earlier method had been based on geometrical considerations and was moreover not
particularly suited to problems involving functions of more than one independent variable).
As an example illustrating the advantages of his general. procedure Lagran.ge considered
the problem of minimizing the area integral ;

f fv' +u2+u2}dxdy
63-3

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
490 J. SERRIN
of a surface z u(x, y) while keeping the volume integral

f udxdy
fixed. He was correspondingly led to the differential equation

~~(;{' constant -2A (10

UX a<e ++_(;'
+
UY)+ A<28} u~2})- (110)

which upon expansion yields the well-known equation for surfaces of constant mean
curvature. This is thc first partial differential equation to have been obtained by the
methods of the calculus of variations, and is justly famous on this account as well as for its
obvious geometric interest.
In the final part of his memoir, Lagrange noted that spheres (or more precisely their
upper and lower hemispheres) generate a four-parameter family of exact solutions of
equation (110), but beyond this he did not pursue the problem from the point of view either
of the calculus of variations or of the differential equation itself. Further analysis of
Lagrange's equation awaited the later interpretation of A as the mean curvature of the
surface z u(x, y). This interpretation in particular established that the upper and lower
parts of right circular cylinders form a new family of exact solutions.
While the minimal surface equation (A 0) shortly became the object of serious investi-
gation, owing to the transformation formulae initiated by Monge and Legendre, no
corresponding analytical techniques were found for dealing with the case of non-zero
curvature. Accordingly it was not until 1841 that a further family of exact solutions was
discovered by Delaunay. Delaunay proposed to determine the surface of revolution having
constant mean curvature, and found that the generating curve must be such as would be
traced out by the focus of a conic section, if the conic section itself rolls without slipping
along the axis in question. t The particular case of a rolling parabola accounts for the catenary
generating the minimal surface of revolution. Although no further exact solutions have been
discovered since Delaunay's surfaces of revolution, a result of Bonnet (i853) should be
noted for its related interest. If we consider a surface S of constant mean curvature A, where
A may be taken positive without loss of generality, then the two parallel surfaces at
distances 1/2A and 1/A in the direction of the preferred normal vector to S have, respec-
tively, constant Gaussian curvature 4A2 and constant negative mean curvature - A.
A proof can be given on the basis of the formulae in ? 4. In fact, one easily establishes that
the mean and Gaussian curvatures A and K of the parallel surfaces at distance d from S are
given by A-Kd I
K
1-2Ad+Kd2, 1-2Ad+Kd2'
where K is the Gaussian curvature of S, and the result follows at once.+
The fact that there are few known exact solutions is of course neither unduly surprising
nor seriously disturbing, though it naturally indicates the importance of general theoretical
investigations. From the beginning of the eighteenth century mathematicians were aware
t Euler had found the differential equations corresponding to Delaunay's solutions as early as 1732,
though he did not attempt their integration. The generating curves are illustrated by Minkowski (i906,
page 582).
+ See also W. Blaschke, Di/ferentialgeometrie, Bd. 1, p. 119. Berlin: Springer (I930).

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 491
of the necessity to formulate and solve problems in the calculus of variations corresponding
to definite boundary conditions. In fact, already in 1785 Meusnier had clearly stated the
Plateau problem: among all surfaces passing through a given closed curve not lying in a
plane, to find that one whose area is least. Similarly, an early Dirichlet type problem for
equation (110) was proposed by Gergonne (i 8I6): Find the surface of minimum area among
all those which are bounded by the sides of a square, and which include between themselves
and the square a given volume. In view of the difficulty of such problems it is not to be
wondered at that few attempted their study until well into this century.
In a series of remarkable papers appearing between the years 1907 and 1912 Bernstein
reopened the Dirichlet problem for elliptic partial differential equations of the type (1) in
two independent variables. While these papers were concerned for the most part with a
fairly general situation, the example of surfaces of constant mean curvature still figures
prominently and was undoubtedly an important stimulus to Bernstein's thought. At the
beginning of his work Bernstein evidently considered equation (110) as unexceptional, for
he states (I907, page 1027) that the corresponding Dirichlet problem in a circular domain
is solvable for arbitrarily given continuous boundary data. By 1910, however, he had
realized the various difficulties in the problem in particular, that the inhomogeneous term
on the right-hand side grows too rapidly in IDuI for such a conclusion to be true.
At this time (I9I0, page 243) he obtained the remarkable theorem that if a domainr Q
properly contains a circle of radius 1/IA I then no solution of (1 1 0) can possibly exist over Q,
irrespective of the boundary data which might be assigned.t On page 244 Bernstein sug-
gested a procedure for determining when a solution exists, and applied this to the particular
case of an elliptical domain with zero boundary data. He writes furthermnore: 'II serait, sans
doute, inte'ressant d'indiquer la marche systematique a suivre pour poivoir decider dans
chaque cas particulier, par un nombre fini d'essais, si le probleme est possible, ou non. Je
n'ais pas l'intention de m'occupier ici de cette question.' Apart from the result discussed in
the following paragraph, in fact, Bernstein never returned to this question.
Equation (110) appears again in a note in the ComptesRendus, Paris (I9I0, page 515)
where it is stated (among other things) that the Dirichlet problem is not generally solvable
for convex domains. This result is of course a consequence of the theorem noted in the
previous paragraph; Bernstein's general argument, however, is directed to the case when
the domain may be arbitrarily small as well as convex, as is apparent from the later published
proof of the result (I912, pages 465-469). This work is of considerable interest for the non-
solvability of Dirichlet's problem, even though it is not sufficiently delicate to make a
distinction between those convex domains for which a solution of Dirichlet's problem for
equation (110) is always possible and those for which it is not (cf. ? 20). Bernstein did not in
fact realize that such a distinction might exist, for in a footnote to the theorem on page 465
he remarks: 'Il n'est pas douteux, d'ailleurs, qu'en completent convenablement la demon-
stration on trouveriat que tous les contours jouissent de la meme propriete.'
A similar situation occurs in his treatment (19I2, page 479) of the variational problem

8 fv1(1+ [xl2)JA/(1+jDuj 2) dx -0, n-z2.

t Other proofs and generalizations of this fact have been given by Heinz (I955), Chern (I965), and
Finn (1r965).

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
492 J. SERRIN
He states there that this extremum problem does not, in general, admit a solution even for
convex domains (compare ? 21 of the present work). The adjoining remark is prophetic:
'Ce problerme, qui difflre si peu du problerme de Plateau, rend compte, il me semble, des
difficultes qu'on a du recontrer en cherchant des dernonstrations directes de l'existence d'une
surface minima pasant par un contour donne". The example which Bernstein notes at the
foot of page 483 provides still another instance where a more delicate treatment is
required (see ?20).
To conclude the discussion of Bernstein's work, attention should be drawn to the final
paragraphs of the memoir of 1912. That the result stated there, 'Si le probleme de Dirichlet
est toujours possible pour toute fonction analytique sur un contour convex, il est egalement
possible si la fonction est supposee seulement continue sur le contour', cannot be correct
follows from an example of Finn (1954, page 399). The problem is of course that of obtaining
interior estimates for the gradient of a solution in terms of a maximum bound, but indepen-
dent of the particular boundary data. While such estimates are available for certain equa-
tions in two independent variables (Finn I963; Serrin I963) and for most uniformly elliptic
equations (cf. Ladyzhenskaya & Uraltseva I964), there are scarcely any non-uniformly
elliptic equations in higher dimensions for which such conclusions are known (in this
regard, E. De Giorgi has recently shown the existence of such an interior estimate for the
minimal surface equation in higher dimensions; see ? 19).
During the period between Bernstein's memoir of 1912 and Leray's fundamental paper
of 1939, I am not aware of any contributions to the problem of Dirichlet for surfaces of
constant mean curvature. We have already commented in the introduction on the relation
between Leray's work and Bernstein's. It does not seem to have been noticed before, how-
ever, that Leray's results can be applied to the Dirichlet problem for equation (110),
provided a certain degree of care is taken. In particular, the following result can be obtained.
Let Q be a boundeddomainin two-dimensional Euclideanspace,whoseboundaryis of class C5.
ThentheDirichletproblemforequation(11 O)in Q is solvableforarbitrarilygivenboundary valuesof
class C5 providedthat K 2 JA at eachpoint of the boundary,where K denotes the of the
curvature
boundary curve.
To see this, one may use corollary II on page 282 of Leray's paper (a straightforward but
somewhat tedious calculation verifies the hypotheses) together with the argument given in
? 20 of the present paper.
In a review article which appeared in 1961, Akhiezer & Petrovsky attribute to Bernstein
the result that the Dirichlet problem for surfaces of constant mean curvature is solvable in
any region which is entirely inside a circle of radius 1/JAI. This statement seems to have
resulted from reading more into some of Bernstein's remarks than was actually intended.
The question of uniqueness of surfaces having prescribed mean curvature is included in
the important work of Alexandrov (1956-58) on surfaces in the large defined by functions
of their principal curvatures. In addition to considering various Dirichlet-type problems in
parallel and central projection, Alexandlrov also obtained general results concerning the
uniqueness of comnpactsurfaces. These latter results imply in particular that a compact non-
self-inrtersecting surface of constant mean curvature is a sphere, generalizing earlier work of
Licbmlann (I900) who required that the surfacebe convex. While this result applies equally

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 493
in any number of dimensions, a complementary theorem of Hopf (195I) states that a simply
connected compact surface of constant mean curvature in three dimensions is a sphere. The
literature on these and related questions is fairly extensive, and the interested reader is
referred to Alexandrov's paper for further details.
In another recent paper, Finn (I965) has obtained the following result: Let u u(x, y)
be a solutionof (i10) definedoverthe circulardomainX2+y2 < 1/A2. Then
u -const. +{(I/A2) -X2 -y2}.

This is the natural analogue for equation (110) of Bernstein's celebrated theorem that a
minimal surface z =u (x, y) defined over the entire plane must itself be a plane. There are
finally some conclusions of a differential geometric character concerning surfaces of con-
stant mean curvature (e.g. the lines of curvature on such a surface form an isothermal net,
etc.) for which we refer the reader to standard treaties on classical differential geometry
or the theory of surfaces.
The discussion would not be complete without mention of the important series of papers
of Heinz (I954) and Werner (1957, I960) concerning the existence of surfaces of constant
mean curvature A in three-dimensional Euclidean space with arbitrary Jordan curves as
assigned boundary. The surfaces in question are defined parametrically, and consequently
(as in the problem of Plateau for minimal surfaces) may have self-intersections, multiple
coverings, and branch points at which the local representation is singular. These matters
aside, the elegant sufficient condition for solvability due to Werner is that theJordan curves
in question should be contained in a ball of radius R where 1/R > 2 1Al. It is remarkable
that this condition almost exactly parallels the solvability condition K > 2 AI for the
Dirichlet problem for arbitrary assigned data.

This work has been in progress over a number of years and accordingly I have discussed
it with a great many of my friends. To all of these, and particularly to D. E. Edmunds and
J. C. C. Nitsche, I express my heartiest thanks and gratitude. The manuscript was written
at the University of Sussex during a sabbatical leave of absence from the University of
Minnesota. I am most grateful to both these institutions for the opportunity which they
provided.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 494 J. SERRIN

REFERENCES

Alexandrov, A. D. I956 Uniqueness theorems for surfaces in the large. VestnikLenin. Univ. 11, 5-17.
Bernstein, S. I907 M[ethode generale pour la resolution du probleme de Dirichlet. C. R. Acad. Sci.
Paris 144, 1025-1027.
Bernstein, S. 1910 Conditions necessaires et suffisantes pour la possibilite du probleme de Dirichlet.
C. R. Acad.Sci. Paris 150, 514-515.
Bernstein, S. i910 Sur la generalisation du probleme de Dirichlet, IT. Math. Ann. 69, 82-136.
Bernstein, S. ijio Sur les surfaces definies au moyen de leur courbure moyenne et totale. Ann. Sci.
Lcole Norm. Sup, 27, 233-256.
Bernstein, S. 1912 Sur les equations du calcul des variations. Ann. Sci. Ecole Norm. Sup. 29, 431-485.
Bombieri, E., De Giorgi,IE. & Miranda, M. I969 Una maggiorazioni a priori relativa alle iper-
superfici minimali non parametriche. Arch. Rat. Mech. Anal. 32.
Browder, F. i966 Problemes Non-lineaires (Seminar de MatheimatiquesSupeJrieures).University of
Montreal Press.
Finn, R. I965 Remarks relevant to minimal surfaces and to surfaces of constant mean curvature.
J. d'Anal. Math. 14, 139-160.
Gilbarg, D. I963 Boundary value problems for nonlinear elliptic equations in n variables, pp. 151-
159. Nonlinearproblems.Madison: University of Wisconsin Press.
Hartman, P. & Stampacchia, G. i966 On some nonlinear elliptic differential functional equations.
Acta Math. 115, 271-310.
Hopf, E. 1927 Elementare Bemerkungen uiber die Losungen partieller Differentialgleichungen
zweiter Ordnung von Elliptischen Typen. Sber. Preuss. Akad. Wiss. 19, 147-152.
Jenkins, H. & Serrin, J. I968 The Dirichlet problem for the minimal surface equation in higher
dimensions. J. Reine Ang. Math. 229, 170-187.
Ladyzhenskaya, 0. A. & Uraltseva, N. N. I96I Quasilinear elliptic equations and variational
problems with many independent variables. Uspekhi Mat. Nauk. 16, 19-92. Translated in Russ.
Math. Surveys16, 17-91.
Ladyzhenskaya, 0. A. & Uraltseva, N. N. I964 On Holder continuity of solutions and their deriva-
tives of linear and quasilinear elliptic and parabolic equations. Trudy SteklovInst. Leningrad73,
172-220.
Ladyzhenskaya, 0. A. & Uraltseva, N. N. I964 Linear and quasilinearelliptic equations.Moscow.
Leray, J. 1939 Discussion d'un probleme de Dirichlet. J. Math. Pures Appl. 18, 249-284.
Leray,J. & Schauder,J. 1934 Topologie et equations fonctionelles. Ann. Sci. ]2coleNorm.Sup.51, 45-78.
Nirenberg, L. 1952 On nonlinear elliptic partial differential eqations and H6lder continuity. Comm.
PureAppl. Math. 6, 103-156.
Nitsche, J. C. C. I965 On differential equations of mode 2. Proc. Am. Math. Soc. 16, 902-908.
Oskolkov, A. P. i967 A priori estimates for the first derivatives-of solutions of the Dirichlet problem
for non-uniformly elliptic quasilinear equations. Trudy Mat. Inst. Steklov (Akad. Nauk. SSSR)
102, 105-127.
Rado, T. I932 Contributions to the theory of minimal surfaces. Acta Szeged 6, 1-20. Cf. also The
problemof plateau. pp. 35-37. Berlin: Springer (I932).
Redheffer, R. i962 An extension of certain maximum principles. Monatsh. Math. 66, 32-42. Cf. also
J. ReineAng. Math.211, 70-77.
Schaeffer, H. I955 Uber die Methode der a priori Schranken. Math. Ann. 129, 415-416.
Schauder, J. 1934 Qber lineare elliptische Differentialgleichungen zweiter Ordnung. Math. Z. 38,
257-282. Cf. also Courant, R. & Hilbert, D. Methodsof mathematicalphysics, vol. 2, pp. 331-346.
New York: Interscience (1962).
Serrin, J. I967 The Dirichlet problem for quasilinear equations with many independent variables.
Proc. Nat. Acad. Sci. 58, 1829-1835.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
 QUASILINEAR ELLIPTIC DIFFERENTIAL EQUATIONS 495
Serrin, J. (To appear). The Dirichlet-Alexandrov problem for surfaces having constant mean
curvature.
Stampacchia, G. I963 On some regular multiple integral problems in the calculus of variations.
Comm.PureAppi. Math. 16, 383-421.
Trudinger, N. I966 Quasilinear elliptic partial differential equations in n variables. Dissertation,
Stanford University. Cf. also Bull. Am. Math. Soc. 73, 410-413.
Trudinger, N. I967 On the Dirichlet problem for quasilinear uniformly elliptic equations in
n variables. Arch. Rat. Mech. Anal. 27, 108-119.
Valentine, F. A. I964 Convexsets. New York: McGraw-Hill.

SUPPLEMENTARY BIBLIOGRAPHY FOR SECTION 24

Akhiezer, N. I. & Petrovski, I. G. I96I S. N. Bernstein's contribution to the theory of partial
differential equations. UspekhiMat. Nauk16, 5-20.
Alexandrov, A. D. I956-58 Uniqueness theorems for surfaces in the large. VestnikLenin. Univ. 11
(no. 19), 5-17; 12 (no. 7), 15-44; 13 (no. 7), 14-26; 13 (no. 19), 5-8. Translated in Am. Math.
Soc. Translations(Series 2) 21, 341-403, 412-416.
Bernstein, S. 1907 Methode generale pour la resultion du probleme de Dirichlet. C. R. Acad. Sci.
Paris 144, 1025-1027.
Bernstein, S. I910 Conditions necessaires et sufficisantes pour la possibilite du probleme de Dirichlet.
C. R. Acad.Sci. Paris 150, 514-515.
Bernstein, S. I910 Sur les surfaces definies au moyen de leur courbure moyenne et totale. Ann. Sci.
tcole Norm.Sup.27, 233-256.
Bernstein, S. i91i Sur les equations du calcul des variations. Ann.Sci. Icole Norm.Sup.29, 431-485.
Bonnet, 0. 1853 Sur une propriete de maximum relative 'a la sphere. Nouv. Ann. Math. 12, 433-
438.
Chern, S. I965 On the curvatures of a piece of hypersurface in Euclidean space. Abh. Math. Sem.
Univ.Hamburg29, 77-91.
Delaunay, Ch. i84i Sur la surface de revolution dont la courbure moyenne est constante. J. Math.
PuresAppl.6, 309-315. Cf. also Sturm, J. ibid., 315-320.
Euler, L. 1738 Problematis isoperimetrici in latissimo sensu accepti solutio generalis. Comm. acad.
sci. Petrop.6 (1732/3), 123-155. Operaomnia25, 13-40 (especially p. 30).
Finn, R. 1954 On equations of minimal surface type. Ann. Math. 60, 397-416.
Finn, R. I963 New estimates for equations of minimal surface type. Arch. Rat. Mech. Anal. 14,
337-375.
Finn, R. I965 Remarks relevant to minimal surfaces and to surfaces of constant mean curvature.
J. d'Anal.Math. 14, 139-160.
Gergonne, J. D. I8i6 Problemes de geometrie, dependant de la methode des variations. Ann.Math.
PuresAppl.7, 156.
Heinz, E. 1954 Uber die Existenz einer Flache konstanter mittlerer Kriimmung bei vorgegebener
Berandung. Math. Ann. 127, 258-287.
Heinz, E. 1955 Uber Flachen mit eineindeutiger Projection auf eine Ebene, deren Kruimmung
durch Ungleichungen eingeschrankt sind. Math. Ann. 129, 451-454.
Hopf, H. I95I Uber Flachen mit einer Relation zwischen den Hauptkrtimmungen. Math. Nachr. 4,
232-249.
Lagrange, J. L. 1760 Essai d'une nouvelle methode pour determiner les maxima et les minima des
formulae integrales ind6finies. Misc. Taur. 2, 173-195. Oeuvresde Lagrange, vol. 1, pp. 333-362.
Leray, J. I939 Discussion d'un probleme de Dirichlet. J. Math. PuresAppl. 18, 249-284.
Liebmann, H. I900 Uber die Verbiegung der geschlossenen Flachen positiver Krtummung. Math.
Ann. 53, 81-112.

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions
496 .1. SERRIN
Meusnier, M. 1785 Memoire sur la courburedes surfaces. Mem.DiversSavans(Etrangers) 10, 477-5 10.
Minkowski, H. I906 Kapillaritat. Ency. Math. Wiss. V 1, 559-613.
Serrin, J. I963 A priori estimates for solutions of the minimal surface equation. Arch.Rat. Mech.
Anal. 14, 376-383; 28, 149-154.
Werner, H. 1957 Das Problem von Douglas fuir Flachen konstanter mittlerer Kruimmung. Mllath.
Ann.133, 303-319.
Werner, H. I960 The existence of surfaces of constant mean curvature with arbitrary assigned
Jordan curves as assigned boundary. Proc.Am. Math. Soc.11, 63-70.

Thispaperis dedicated
to the memoryof
SergeiNatanovichBernstein
1880o1968

This content downloaded from 81.194.27.158 on Tue, 27 May 2014 04:47:56 AM

All use subject to JSTOR Terms and Conditions