COMMUNICATIONS ON PURE AND APPLIED MATHEMATICS, VOL.

XVII, 35-92 (1964)

Estimates Near the Boundary for Solutions

of Elliptic Partial Differential Eqiiations
Satisfying General Boundary Conditions 11*
S. AGMON A. DOUGLIS L. NIRENBERG
Hpbreui University The Crniversity o j Maryland New York Eniversity

Introduction
This paper is the second part of a study of estimates pertaining to solutions of
boundary problems for linear elliptic equations. The first part [ 11, hereinafter
referred to as ADN (and sometimes as Part I), dealt with a single equation for one
dependent variable; here we treat systems ofseveral equations for an equal number
of dependent variables. The estimates, as in ADN, are of two kinds: (a) Schauder
estimates, in which Holder norms for the solutions and their derivatives are assessed
from the values of similar norms for the given data; (b) L , estimates for the solu-
tions and their derivatives. The boundary conditions considered are of differential
type and satisfy a condition of suitability to the system of differential equations
which we call the Complementing Condition. This Complementing Condition
is necessary and sufficient in order that it be possible to estimate all the derivatives
occurring in the system, without loss of order, i.e., that inequalities of “coercive”
type be valid. It is expressed algebraically in Section 2, but also can be formulated
in the following non-algebraic way very useful in practice (see Lopatinskii [IS]).
If P is an arbitrary point on the boundary of the region 9, make a coordinate
transformation flattening the boundary in a neighborhood Q of P into a portion
of a plane, which we represent by t = 0 with the half-space t 2 0 containing the
image under the transformation of !J n 9. Then consider the problem to which
the given problem reduces when all the “leading” coefficients in the differential
equations and in the boundary conditions are fixed equal to their values at P
and the lower order coefficients dropped out. For this system with constant
coefficients, consider the homogeneous boundary value problem in the half-space
t 2 0. The Complementing Condition requires that, in this new problem, the

* Part of the work for this paper was supported by the United States Army under Contract
DA-ARO(D) 31-1244156 and by the United States Navy under Contract Nonr-285(46). Dough’
work was supported by the National Science Foundation under Grant NSF G16003 and, prior to
this, by the United States Air Force through the Air Force Office of Scientific Research of the Air
Research and Development Command, under Contract A F 49 (638)-590. Reproduction in whole
or i n part is permitted for any purpose of the United States Government.
35
36 S. AGMON, A. DOUGLIS AND L. NIRENBERG

only solution u defined for t 2 0, tending to zero as t + 00, and consisting of a

finite linear combination of exponential polynomial solutions is u 3 0.
The following is another way of formulating the Complementing Condition;
the new formulation is again in terms of the homogeneous problem with constant
coefficients introduced just above. Let x refer to coordinates in the hyperplane
t = 0. For any real vector 6 # 0 in this hyperplane, consider solutions of the
problem of the form eix’t times a function u of t ; u is thus a solution of a system of
ordinary differential equations with constant coefficients. The Complementing
Condition requires that u = 0 for every 6 # 0, if u is bounded on the semi-axis
t 2 0.
This paper is concerned only with estimates. Existence theory under the same
boundary conditions is developed with the aid of L , estimates in Hormander’s
book [12]. Peetre’s method, generalized from a single equation [21] to systems,
also should lead to an adequate existence theory.
Since the publication of Part I of this paper, much further work has been done
on elliptic boundary value problems. Some of this work is compiled in a supple-
mentary list in the Russian translation of ADN; to make this list more accessible,
we have added it to the end of our own bibliography below. We now mention
more recent papers on this subject.* Fife [ 1 11 has shown that Schauder estimates
still can be obtained when Holder continuity assumptions with respect to some of
the independent variables are dropped. Cordes [9] has derived L , estimates for
systems of fixed order. Lions and Magenes [15] have extended previous results
in various directions with the aid of interpolation theorems for mappings in
Banach spaces. Schechter (see, for instance, [22], which contains additional
references to other work) has obtained quite general results for single elliptic
equations; see also Miranda [ 171. Some authors have discussed boundary
conditions not purely local in character, i.e., not purely of differential type, Bade
[5] abstractly formulated boundary conditions (for a single equation) and
Agranovich and Dynin [3] boundary conditions (for a system) expressed with the
aid of singular integral operators. Respecting the latter type of boundary con-
dition, we refer also to the authors listed in [3] and mention Atiyah’s and Singer’s
application [4] to the problem of the index of elliptic operators on manifolds.
Many of the works mentioned above overlap with the present paper, and a few
in some respects are more general. None, however, outside of [12], [24] and 1 **
treats systems of such comprehensive type as are considered here*.
The systems of equations treated by us are those of [lo], which are permitted
to be of different orders in the different dependent variables: the maximum order
allowed of the j-th dependent variable in the i-th equation, namely, is si t j , +
where si and t j are integers. The boundary conditions, which, as noted, are of
differential type, similarly may involve varying orders of the dependent variables ;

* See also references 1 ** to 7** added in proof at the end of the Ribliography; the results of this
paper have also becn announced by Solonnikov in 7**.
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 37

the maximum order allowed of the j-th dependent variable in the h-th boundary
+
condition is rh t , , where the r,, are integers.
Our treatment follows closely that of Part I, many details as well as applica-
tions therefore being only summarily handled. Chapter I describes our formal
hypotheses concerning the differential equations and the boundary conditions.
In Chapter 11, explicit Poisson kernels are developed for systems in a half-space
with constant leading coefficients and zero lower terms. The constructions are
based on a consideration of general initial value problems for systems of ordinary
differential equations ; believed to be of independent interest, this theory is given
self-contained treatment, in Section 3. After the paper was completed we found
that Volevich [24] has also recently constructed Poisson kernels for general systems.
Chapters I11 and IV are devoted to Schauder and L , estimates, including local
L, estimates which had been left out of ADN, but omitting equations of vari-
ational form such as had been treated in Chapter 111. Chapter V consists of a
few applications and comments; the last section contains remarks on singular
integral representations.
A question was raised in ADN (p. 649) concerning the relation between two
different ways of expressing the fact that a function has a fractional derivative of
order 1 - lip belonging to L,. More specifically, the question concerned the
relation between two norms, called Jfll-l,p,L,and l J l ~ - l , p , L , (see (3.13); in ADN).
We had conjectured the two norms to be equivalent, but J. P. Kahane, in the
summer of 1960, showed us a simple example of a function constructed with the
aid of lacunary series for which, for 1 < p < 2, the first norm is infinite and
the second norm not. Eli Stein subsequently demonstrated that

lfl:-l,p*L* 6 constant - l f i l - l , p , L 9
for 1 < p < 2 and that the opposite inequality holds for p 2; see also M. >
Taibleson [23].
It is natural to ask whether a given system of elliptic equations in some domain
admits any boundary conditions of complementing type. In Section 2 we have
considered only strongly elliptic equations and shown that Dirichlet conditions
are of complementing type. A complete (unpublished) answer to this question
has been given, however, by R. Bott. We mention also that recently P. D. Lax
and R. Phillips, in connection with their work on hyperbolic symmetric systems,
have shown that when A , , j = 1, n, denote real, symmetric N by N
a ,

n
matrices, the first order operator 2
Ai(a/axi) can be elliptic only for certain
j= 1
values of R .
 CHAPTER I

FORMULATION OF BOUNDARY PROBLEM. BASIC NOTATION

AND ASSUMPTIONS

1. The Differential Equations Considered

We shall be concerned with functions defined in domains of (n l)-d’imen- +
sional Euclidean space, n 5 1, with coordinates P = (xI, * . , x,+~),and we use -
the notation
= ai a/ax,,
a= (al,. an+l).
.,
Differentiations oforder I, for instance afl - - a$$; ,where XDi = 1, are symbolized
*
i
in terms of the multi-index /? = (PI, - , pn+l) by ap or, more schematically, by
*

-
8. Monomials - Ek;; of degree 1 = Ij3I = Zj3, in the components of any
Efl * *

vector G = (El, - *, tn+& are similarly abbreviated as E”or El. By IEI we shall
*

\i=1 /
T h e systems of partial differential equations considered will be represented as

where the ,!ij(P,a), linear differential operators, are polynomials in a with

coefficients depending on P over some domain 9 in x,, *, xn+l-space. The + -
orders of these operators, as in [lo], will be assumed to depend on two systems of
-
integer weights, sl, . -,sS and t,, * -
t, attached to the equations and to the
a,

unknowns, respectively, si corresponding to the i-th equation and t j to thej-th

dependent variable u j . T h e manner of the dependence is expressed b y the
inequality
. .
(1.2) deg l d j ( P ,E ) 5 si + t j , t,j = 1, * - - ,N,
“deg” referring of course to the degree in z.
It is to be understood that lij = 0
+
ifsi t j < 0.
Adding a suitable constant to one system of weights and subtracting the
constant from the other, we readily achieve, as a normalization, the condition
(1-3) si i0.
Since, for fixed j , not all lij vanish, and since st + ti 2 0 for non-vanishing l i j ,
we thus also have
(1.4) 0 5 tj t‘,
t’ being the maximum of the t j .
38
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 39

Only elliptic systems of the form ( I . 1) will be considered. These are defined as
systems for which
(1.5) L(P, Z) E det (&(P, a)) # 0 for real E # 0,
where lij(P, E) consists of the terms in l i j ( P ,Z) which are just of the order
si + t, -
Remark. Our definition of ellipticity is a bit artificial for it is not obvious
how to determine whether a given system (1.1) is elliptic, i.e., when weights
si , ti can be found so that (1.5) holds. Recently, however, Volevich [25] has given
an equivalent formulation which is easily verifiable. To describe it consider a
general system of the form (l.l), and let aij = deglij. The determinant
A ( P , E) = det consists of sums and differences of terms of the form lli, ,
lziz, * . , lni, , Let R be the maximal degree of such terms and r the degree of A
(so that in general r R ) . Our definition of ellipticity is equivalent to the
following: r = R and A ( P , S) is elliptic, i.e., for any P, E = 0 is the only real
root of A’(P, 5 ) = 0, where A‘ is the leading part, the part of order Y , of A. That
our definition of ellipticity implies this formulation is immediate ; Volevich
gives an elegant proof of the converse using a result in linear programming.
All elliptic systems in three or more independent variables are known (cf.,
for example, ADN, pp. 631-632) to satisfy a further condition described below.
In the two variable case, however, we must additionally assume this.
SuwmmNrARY CONDITION ON L. L(P, E) is of even degree 2m (with respect
to a). For e v e v pair $ linearly independent real vectors E,E‘, the polynomial
L(P, E +7s‘)in the complex variable r has exactly m roots with positive imaginary part.
This condition is actually used only at points P of the boundary & of 9 with
c
z a tangent, and S’ the normal to 9, a t P.
When rn = 0 we may explicitly solve (1.1) for the uk in terms of the F, and
their derivatives as has been seen in [lo, p. 5061, We do not lose, therefore, in
requiring here that

(1.6) m = 4 deg (L(P, 5 ) )> 0.

Uniform ellipticity will be required in the sense that there is a positive constant
A such that
’
( *7) A-I 1E12” 2 IL(P, =)I A 1SI2”
for every real vector = ($,, - ., t,,,)
* and for every point P in the closure of
the domain considered.
To summarize, we shall always assume for our systems conditions (1.2),
(1,3), (1.4), (1.6), (1.7), and also the above Supplementary Condition on L.
Accompanying continuity requirements for different purposes will be stated in
Sections 9 and 10.
40 S . AGMON, A. DOUGLIS AND L. NIRENBERG

Remarks on elliptic systems of first order. Any elliptic system satisfying

+
(1.2) can be remodeled so that si t, 5 1 : first order elliptic systems in this
sense are the most general. This fact was first pointed out to us by Atiyah and
Singer. I t is not used in this paper, but, being of interest in itself, will be justified
here.
The remodeling procedure consists in introducing as new dependent variables,
which are added to the original dependent variables, appropriate derivatives of
the latter, and as new equations, added to the original system, the equations
which define the new variables. The procedure is well known, but it destroys
ellipticity in the ordinary sense and hence has not heretofore seemed significant
for elliptic systems. What Atiyah and Singer noticed is that this procedure does
preserve ellipticity in our sense. Consider as a simple illustration Laplace’s
equation ayu + = 0, 8, = a / a x k , which in terms of the new variables
u1 = a,u and u, = a2u is reducible to the first order system

alu1 + aaUz= 0,
alu - u1 = 0,
- u 2 = 0.

0 51 52

5, 0 0 = 0;
52 0 0
the system thus is not elliptic in the ordinary sense. The characteristic deter-
minant under our definition, when the weights
to = 2, 1, = t, = I
are assigned to u, ul, u,, respectively, and
so = 0, s, = s2 = -1
to the first, second, and third equations, respectively, is, however,

so that the system is elliptic in our sense, as contended.

Let us now prove that the procedure described is applicable to a n arbitrary
system of equations of higher than first order which is elliptic in our sense. By
inductive reasoning, it suffices to show that, when thejrst derivatives are added as
new dependent variables, the resulting system has maximum order one less than
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 41

that of the original system and is elliptic. Recalling that max si = 0, we first
divide the functions uk into two classes, { u ~ }and {u,"}, the indices written as k'
taking the values of k for which t, > 1, the k" the remaining values of k; we may
suppose 1 9 k' K and K 1 5 k" + N . With ak
= alax,, we introduce as
new unknowns the first derivatives of the functions uv: apr = u ~ , and ~ , we
introduce the additional equations
(' ")k',9 ajuk'-uy,i=O, k'= l , . . . , K , j = l;..,n.
We also rewrite each operator l,,u, (summation not used) in the following way.
The terms in which uk is not differentiated are not changed. A n y other term, which
consists of a coefficient times aPajuk for somej and some p (possibly IpI = 0), is
rewritten as coefficient times This way of rewriting is clearly not unique.
In any case, the expression Xklikuk is replaced by one of the form

(1.9)i 2E' &'Uv +2

'
t
lik"&" + 2 l&,&.j,
E',j
where the liVand I,- are functions and li,,,ja differential operator of order
+
Isi t, - 1. Our new system of equations consists of (l.9)r, and (1.9)i
-
together. Weights, s;l.,j, f, , fv,j, and lk ,for the new system are assigned as follows:
.Ti = si, ik = I,, sK,i= 1 - t k ? , lv,j = t,. - 1.
We note that the new system is completely equivalent to the old and that its order
is reduced, with the original variables differentiated, in fact, no more than once.
To see that the new system is elliptic, we have only to show that, if
t = (51, . * F,+J # 0,
-9

the corresponding characteristic determinant is not zero, i.e., the following

system of linear equations for complex variables 7,, qv,j has only the null solution:

tjqk' - q k ' , j = O, k' = 1, . * *, K , j = 1, * *, n,

21& k q k +2 &k',j(E)qk',j = O, i = l , . . . , N,
C indicating summation over all k' a n d j , XI summation over the values of k for
+
which s i t , = 0, and the characteristic polynomial of the leading part
of +
the part of order xi tk, - 1. By substituting the first set of equations
into the second, we find
21l i k q k +2 &k'.j([)tjqk' = '9 i = 1, - ' . ,X.
By the construction of the new system, these equations, however, are simply the
equations

Hence, qk = 0 and, consequently, qv,j = 0. It follows that the reduced system is

elliptic, as stated.
Boundary conditions for the original system of differential equations go over,
42 S . AGMON, A. DOUGLIS AND L. NIRENBERG

usually in several possible ways, into equivalent boundary conditions for the
reduced system. I f the boundary conditions were to complement the original
system, any equivalent form of these conditions would also complement the new.
This is because 1) the new and original boundary value problems are completely
equivalent, and 2) the Complementing Conditions are both necessary and suffi-
cient for our coercive inequalities to hold (Section 11).

2. Complementing Boundary Conditions

The boundary conditions we consider refer to a regular portion r of 9.
They are expressed as

h = 1, - - a, m,

in terms of given polynomials in ,B,,(P, Z), with complex coefficients depending

on P. The orders of the boundary operators, like those of the operators in (1. 1 ) ,
depend on two systems of integer weights, in this case the system t,, . ., t ,
already attached to the dependent variables and a new system rl, * . -,r,, of which
rh pertains to the h-th condition in (2.1), h = 1, * * ., m. The exact dependence is
that expressed by the inequality

+
it being understood that Bhj = 0 when rh t j < 0.
Let Bb(P,Z) consist of the terms in Bhj(P,E) which are just of the order
rh + t j .
The boundary conditions must “complement” the differential equations in
presenting a well-posed problem for both. They will do so or not according to an
algebraic criterion involving the lij(P, E) and the Bij(P, E) which we shall
describe below.
At any point P of r, let n denote the normal, and Z # 0 any tangent, to
r (Z, in particular, is real). Denote by ~ k f ( PE , ),k = 1, . * ., m, the m roots
+
(in T ) with positive imaginary part of the characteristic equation L ( P , Z T n ) = 0.
The existence of these roots is assured by the Supplementary Condition on L.
Set rn
hf+(p,Z, 7 ) = ( T - T l ( P , E)).
h=l

+
Let ( L j k ( ( P ,E T n ) ) denote the matrix adjoint to (lij(P,E +
Tn)). The above
mentioned criterion for the boundary problem (1. l), (1.2) to be coercive is that
the following algebraic condition be satisfied.
COMPLEMENTING BOUNDARY CONDITION.For any P E I? and any real, non-zero
vector E tangent to r at P , let us regard M+(P, Z, T ) and the elements of the matrix
N
(2.3) 2 B&P, B + T n ) L y P , E + 7n)
j=1
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 43

as polynomials in the indeterminale 7. The TOWS of the latter matrix are required to be
linearly independent modulo M+(P, E, r ) , i.e.,
m N
2 C,,2 BijLjk = 0 (mod M+),
h=l j=l
o n b ;f the constants C, are all zero.
For problems in which the Complementing Boundary Condition is satisfied,
an important type of constant, called a “minor constant”, pertaining to I’ or
a suitable partition of r,
will figure in the later estimates. Let us first describe
such a constant in connection with a portion I” of I?. For P E r”, reduce the
polynomials (in r ) (2.3) modulo M+(P, E, T ) , as is described in the congruence
s m-1
(2.4)
j=1
+
2 BLj (P,E m) Ljk(P, 2 m) = dP(P,E ) T (mod+ 2
~ M+(P, E, T ) ) .
p=o
Then construct the matrix Q = (do)having m rows: h = 1, . *, m, and m N
columns: /3 = 0, . ., m - 1, k = I , . * *, N . Under the Complementing Bound-
ary Condition, the rank of Q will be m. Hence, if

M y P , El, * . ., M a p , E)
denote all the m-rowed minors of Q, not all the M a will be zero, and, in particular,
max IM’(P, E)l
j=1, ... , a
will not be zero. Also non-zero, at least when r’ is compact, is the infimum A;.,
of these quantities for all real unit vectors Z which are tangent to I“ at P, and
for all P E I”. When I“ is plane, its minor constant A,. is defined as A;,.
When I” is not plane, a certain change of variables will be known on occasion that,
in particular, makes I?’ plane; Ar, in such a case will be defined as the value
r
A;, has in the changed variables. When itself, say, is covered by a union of
subportions I?(%),and each Pz) has a minor constant A,,,) as described above,
the corresponding minor constant for I? will be defined as
A f A,. = in€ A,,,).
Except in the case in which r = I” is given plane, these definitions, of course, are
not complete; they will be made so in appropriate contexts below.

Strongly elliptic systems. As an example of an elliptic system with com-

plementing boundary conditions, we mention a strongly elliptic system which has
been treated for instance in [19]. This is a system

i = l;-*,N,

for which deg lij = ti + tt!, i.e., si = ti = tl! 0. Let l l j denote the principal part
44 S. AGMON, A . DOUGLIS AND L. NIRENBERG

of l i j ,i.e., the terms of order precisely tl + ti'. Strong ellipticity is the requirement
that, at every point P of 3,

(2.5)

--
for all complex numbers ql, -,q n and all real vectors Z = (tl, - - ., En+l), a
being a positive constant.
I n Dirichlet boundary conditions, the normal derivatives

are prescribed. (When t; = 0, u j goes unprescribed.) The prescribed values are

called the Dirichlet data in the problem. That the Dirichlet boundary conditions
are complementing is easily seen. Suppose that E is a vector tangent to the
boundary a t a fixed point Po and n the inner normal a t Po. We must show that a
solution of
& ( P o , (l/i)a)uj = 0 ,
of the form u = eix"v(n . x ) , decaying as n . x -+ co and having zero Dirichlet data
for n * x = 0, is identically zero. We observe that u ( t ) is a solution of the system of
ordinary differential equations

satisfying the boundary conditions (d/dt)qvj= 0 at t = 0 for q 5 ti - 1. Multiply

(2.7) by ijk(t), sum over k, and integrate with respect to t on the half-line t 5 0.
The resulting integral, after suitable integrations by parts, may be expressed as
the integral of a quadratic polynomial in the derivatives of the u j of orders not
greater than t j ; because of the boundary conditions, no contribution occurs from
the initial point t = 0. Now extend the domain of integration of this integral to
the entire t-axis, for this purpose defining each v i ( t ) as zero for t < 0. In terms
of the Fourier transform C , ( T ) = JecitTuj(t) dt, by Parseval's theorem, we have as a
result
-
I li,j(Po, 6

which, in view of (2.5), proves that u

+ T~)u*j(7)&(T)'dT=

E
0,

0. Dirichlet conditions are, therefore,

complementing as claimed.
 CHAPTER II

EQUATIONS WITH CONSTANT COEFFICIENTS

IN A HALF-SPACE

Estimates in general problems (Sections 9 and 10) are based on those developed
in this chapter for problems with constant coefficients in a half-space. Such
problems arise already for ordinary differential equations, with which we find it
advantageous to begin, see also [12], Chapter 10.

3. On Ordinary Differential Equations

Some questions in the constant coefficient case will be reduced by plane wave
representation to problems for ordinary differential equations. These are treated
here.
The problems we are led to consider are for systems of N linear, ordinary
differential equations with constant (complex) coefficients for N functions u =
--
( u l ( t ) , *, u N ( t ) )defined on the semi-axis t 2 0. Denoting (l/i) (d/dt), by D,
we write the system (here we use the summation convention) as
(3.1) li3(D)uj= 0, i = l,*..,N,
where the lii are polynomials in D. Let L(D) denote det (lii(D)). We assume
that L is not identically zero; thus, L is a differential operator, say of order
p , p > 0.
Let (Ljk)denote the adjoint to the matrix ( l i i ) ,so that, in particular,
(34 liiLik = 6fL.
Only the set of exponentially decaying solutions of (3.1) is pertinent to the
applications intended below. However, the set of all solutions will be analogously
discussed, if briefly, as being of interest in itself.
We shall first carry out a well known process for reducing the matrix (li,(7))
to triangular form with zeros below the diagonal, see Ince [13]. This reduction
is carried out by two kinds of elementary operations: (a) the interchange of
two rows, and (b) the addition to some row of another row multiplied by any
polynomial in 7. The new (triangular) system of equations is equivalent to - the
original, since each such operation is reversible. The triangular matrix (lii(7))
is in fact obtainable from ( l i r ( 7 ) )by matrix multiplication on the left by a square
matrix A whose entries are polynomials in 7 and whose determinant equals & l .
Thus, L(T) = det (ti) = f L ( 7 ) . I t follows, incidentally, that fLijia? = Lzk, the
a: being the elements of the matrix A.
-
Let p i = deg (Lii(-r)). Obviously, p = &.
i
45
46 S. AGMON, A. DOUGLIS AND L. NIRENBERG

What now follows, u p to Theorem 3.1, will relate specifically to the full set of
solutions of (3.1). We begin by proving that the dimension of the latter set is
- -
equal to p. Indeed, the system l i j u j = 0 being triangular, we have li, = 0 for
i > j . Hence, the obvious procedure- in constructing independent solutions of
(3.1) is first to solve the last equation lslvuN = 0 of the triangularized system for
uN obtaining pN independent solutions, then to solve the next to the last equation
for uAvv-lin terms ofuh., and so on. The ultimate result is that there are exactly
p linearly independent solutions of (3. I), as contended.
By the above procedure, plainly a unique solution of (3.1) can be found with
prescribed values of Dkuj,0 6 k < p j ,j = 1, * * N, at t = 0. The initial value
0 ,

problem thus described is an instance of

PROBLEM 1. T o find a solution of (3.1) for which
(3.3) Bmj(D)uj= a, at t = 0, a = l,-..,p,
where the a, are given numbers and the B,,(D), cc = 1, * * . , p, j = I , * * ,N, -
differential operators with constant coefficients.
I n the absence of any restrictions on the B m i ,obviously some of the conditions
listed in (3.3) might be redundant, or they might be incompatible with others of
these conditions or with (3.1). Appropriate safeguards against these eventualities
are described in
COMPLEMENTING
CONDITION
1 . The rows of the p by N matrix, with Polynomial
entries,
B Z j ( 7 ) P7()
are linearly independent modulo the polynomial L ( 7 ); i.e., the equations
2C,B,,(7)Lik(7) E 0 (mod L ( T ) ) for k = 1, * - ,N
U

imply that the constants C, are all zero.

The pertinence of this condition is readily understood, in fact, from the
following theorem, in which our main results concerning the full set of solutions
of (3.1) are summarized.
THEOREM 3.1. The following statements are all equivalent to one another:
(a) Complementing Condition 1 holds.
(b) Problem 1 has a solution for arbitrary a,.
(c) Any solution of Problem 1 is unique: u j = 0 is the only solution with all a, = 0.
Theorem 3.1 can be proved by means analogous to those used below in
demonstrating Theorem 3.2, the latter pertaining to the set of exponentially
decaying solutions of (3.1). We now turn to this.
Assume that the polynomial L(T)has exactly pf complex roots with 9 m 7 0; >
then factor L ( T ) as
L(7) = L+(T) ' L - ( T ) ,
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 47

where the roots of L+(T)are those of L for which $m 7 > 0. Also factor the
-
diagonal elements of the triangularized matrix above as L t t ( ~ ) = 1: ( T ) C ( T )where
,
c1
the roots of I:(.) are just those of for which Y m T 0. Then >
(3.4) L+(T) = f l: (T),
e

and p+ = 2 p,: p+ being the degree of 1;.

1

Our first result concerning the set of exponentially decaying solutions of (3.1)
is that its dimension is p+. This is deduced with the help of the following

LEMMA.Let P ( D ) be a di$erential operator with constant coefiients f o r a single

variable u. I f f ( t ) dies down exponentially as t co, then the equation
---f

P(D)u =f
has a solution also 4i.g down exponentially.
Postponing the proof of the lemma, we now show how to obtain p+ independent
solutions of (3.1) dying down exponentially at CO. Consider the triangular system
-
l z j u j = 0 equivalent to (3.1). The last equation has pa$ linearly independent,
exponentially decaying solutions which are obtainable, namely, as the solutions
of 1.:~ = 0. Choose one of these, and regard the next to the last equation
*

1,v- = -I,- 1 , N U N
l,.V-IUA- 1

as an equation for u ~ - By ~ . the lemma, this has a solution dying down exponen-
tially, and to it can be added any linear combination of the p i - linearly independ-
ent solutions ofli-,u = 0. There are thus, in all, p i + P $ - ~ linearly independent
,,
vectors (uLY- us) satisfying the final two equations of the triangularized system.
Continuing in this way, we ultimately find Cp: = p+ independent solutions ofthe
whole triangularized system that die down exponentially. Furthermore, we may
determine a unique solution with prescribed values of the derivatives
(3.5) Dku,, k = O , l , . . * , p ~ - l , j = l , * - . , N ,
at the origin. (No prescription is made when p; = 0.) This proves our contention
as to dimensionality.
Proof of the Lemma: Factoring P, we easily reduce the lemma to the case of a
first order operator
du
- au =f,
dt
+
In this, a can be assumed to be real; for, when a = u + i/l the transformation
v = ei% reduces the equation to
dv
+
- uv = eioff.
dt
48 S. AGMON, A . DOUGLIS AND L. NIRENBERG

Suppose l f ( t ) l 5 K c b f ,6 > 0. If a 2 b, the solution

U = Catkf (T) dT

decays exponentially, while if a < 6, the solution

W

u = -e-.'L earf(T) dT

decays exponentially. Exhibiting these solutions proves the contention.

The problem of prescribing initial values of the derivatives (3.5) is an instance of
PROBLEM 2. To find an exponentially decaying solution of (3.1) for which
(3.6) R,,(D)u, = a, at t = 0, (3= l,-.*,p+,
where the a, are given numbers and the B,,(D), (3 = 1, - * - , p+, j = 1, . * , N , -
differential operators with constant coefficients.
Appropriate safeguards against redundancy or incompatibility of conditions
(3.6) are here furnished by
COMPLEMENTING 2. The rows o f the p + by N matrix
CONDITION
B,j(T)Lj'(T)

are linearb independent modulo the polynomial L+(T).

This condition is satisfied, as will be seen below, when the quantities prescribed
for t = 0 are the derivatives (3.5). Its appropriateness is evident from
THEOREM 3.2. The following statements are all equivalent to one another:
(a) Complementing Condition 2 holds.
(b) Problem 2 has a solution for arbitrary a,.
(c) Any solution of Problem 2 is unique: u j = 0 is the only exponentially decaying
solution with all a, = 0.
Preliminary to proving Theorem 3.2 we observe that the Complementing
Conditions are satisfied relative to equations (3.1) if- and only if they also
are satisfied relative to the triangularized equations lipj= 0. Indeed, if the
( ~ ) linearly dependent mod L+(T),
rows of B o j ( ~ ) . z j k are so also are the rows of
r ) BOj(T)Ljji(T)af(T).Conversely, if the rows of B,Lik are linearlj-
B ,,i( ~ ) Ljk ( -=
dependent mod L+(T),so also are the rows of B,tik, the elements of the matrix
A-l being again polynomials in T .
Let us also verify that the particular boundary conditions mentioned above in
connection with Problem 2, which consist in prescribing the values of the deriv-
atives (3.5) a t t = 0, satisfy Complementing Condition 2. (An analogous state-
ment regarding Problem 1 is similarly proved.) By the preceding paragraph, it is
sufficient to consider the triangular system lipj = 0. The matrix ( e i i ) also is
triangular, i.e., zero below the diagonal. Now suppose Complementing Condition
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 49

2 not to be satisfied. From the form of the present boundary conditions, it easily
follows that there exist N polynomials pj(.), the degree of pj being less than p,?
and not all the pj being zero, such that
f l j ( ~ ) L j k (=~ )0 (mod L+(T)) for k = 1, -- I , N.
Since (Ljk) is triangular, we thus have, in particular,
p l ( ~ ) L l 1 ( ~ )0 (mod L+(T)).
However, i,,& = n
i
<,j so that tl1 = nTjj.
jtl
It follows from (3.4) that

or
p1 fl lj -= 0 (mod l : ) .
j# 1

Since each 1)- has its roots in the lower half-plane, this is possible only if
p1 = 0 (mod l : ) ,
and, since the degree ofp, is less than p: = degree of L:, only if!, = 0. Similarly
we find, in turn, that p z , p 3 , * * . ,p., all vanish.
N'e shall now give the proof of Theorem 3.2; that of Theorem 3.1 is just the
same, Lf and p+, wherever used, simply being replaced by L and p .
Proof of Theorem 3.2: (i) Equivalence of existence and uniqueness, i.e.,
of (b) and (c) : The exponentially decaying solutions of (3.1) constitute, as noted,
a p+-dimensional space. We map this space into Cp', the space of vectors a =
(ul,. . . , u p - ) with p+ (complex valued) components, by associating with any
exponentially decaying solution u = (ul, * * ' , u N ) the vector a(.) whose com-
ponents are B C p , 0 = 1, . . * , p+. This mapping being a homomorphism
\\ill be onto if and only if the statement u ( u ) = 0 implies u = 0. Thus, a u exists
to any a if and only if such a u is necessarily unique. This is to say that (b) and (c)
are equivalent, as asserted.
(ii) Sufficiency of the Complementing Conditions ((a) implies (b)): If L f
has the form L+ = TP' +U ~ T ~ + - ' +
. * * , we introduce p+ polynomials
Ll(T) = Tfl + UITP '+ ' ' ' + up, b = 0, ' ' ' , pf - 1,
analogous to those of equations (1.5) in ADN. These polynomials have the
property that

(3.7)

where y is any rectifiable Jordan contour in the complex plane enclosing the
roots of L+(T).
We also write
P+ 1
(3.8) ~ )2
B O j ( 7 ) L j k (= qt'Tp (mod L+).
il=O
50 S. AGMON, A . DOUGLIS AND L. NIRENBERG

The matrix Q = (& is to be regarded as having pf rows: (T = 1, . . , p+, and

p f N columns: /3 = 0, . * , p+ - 1, k = 1, . , N . The Complementing Con-
dition asserts that its rank is p+. Given arbitrary numbers a,, , we shall construct
an exponentially decaying solution of (3.1) satisfying (3.6) at t = 0. Because Q
has rank p+, there exist numbers ckp (not necessarily uniquely determined) such
that (the summation convention is used here)

(3.9) Ckaq:P = a,.

We now set, for j = 1, * .- ,N,

(3.10)

where y is a contour enclosing the roots of L+(7). Since these roots lie in the open
upper half-plane, we may also choose y to lie in that half-plane. The functions
u j therefore decay exponentially as t co; they also clearly satisfy (3.1). Now
--f

at t = 0,

by (3.8). Here summation is o v e r j and k from 1 to N, and over E, from 0 to

N - 1 . By (3.7) and (3.9) we thus have
B,uj = cksqfp = a, for t = 0.

Hence the functions uj given in (3.10) have all the desired properties.
We note that, although there was a great deal of arbitrariness in choosing the
ckB to satisfy (3.9), the functions uj are independent of the particular choice. This
is due to Theorem 3.2, which guarantees uniqueness when solutions of Problem 2
exist for arbitrary a,. An easy consequence of this is the following
COROLLARY
TO THEOREM3.2. r f the Complementing Conditions are satirJied, and
if ckp are constants satisfying
cksqEP = 0 f o r (T = 1, * * , p+, then the polynomials
k,P

f o r j = I;.., N.
A direct algebraic proof of this was rather lengthy.
(iii) Necessity of the Complementing Condition ((c) implies (a)) : Suppose
that the rank of the matrix Q is less than p+. We shall construct a non-zero,
exponentially decaying solution of (3.1) with B , p j = 0 at t = 0, (T = 1, . * * , p+,
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 51

contradicting uniqueness. We may suppose that (3.1) is in triangular form:

l,p, = 0. As we observed earlier there is a system of boundary operators Boiu,
at t = 0, for which existence and uniqueness hold, and which satisfies the
Complementing Condition (such a system might consist, namely, of the deriv-
atives (3.5) at t = 0). T h e corresponding matrix 4 then has rank p+ and, there-
fore, contains a row which is linearly independent (mod L + ( T )of
) the rows of Q;
suppose this is the a,-th row. Then, if we write
p+-1
B,p,jLjk= q t t (mod
~ ~ L+(T)),
,6=0

-
the row gtf is independent of the rows q!p, a = 1, * * , p+, which correspond
to Q according to (3.8). Hence there are constants ckB such that

(the summation convention has been employed). If we substitute these constants

into the expression (3. lo), the resulting functions u j decay exponentially, satisfy
(3.1),and satisfy, at t = 0, the homogeneous boundary conditions
B o p i = 0.
These functions are not identically zero, however, because of the additional
boundary condition BoOjuj= 1. Thus, condition (c) ofTheorem (3.2) is violated,
this fact proving that (c) implies (a), completing our proof of Theorem 3.2.

4. Explicit Solution of Boundary Problems for

Homogeneous Equations
In this section, we shall be concerned with functions defined in a half-space
given, say, by x n L 1 > 0. Usually, it will be convenient to write t = x , + ~ and to
denote a point of the space by P = ( x , t ) , x here abbreviating (xl, . * . , x n ) . With
+
5 = (El, * . * , tn),the symbol (6,T ) will denote a vector with n 1 components;
8, Mill denote differentiation with respect to t , a, = (al,
* .* , a,)
differentiation
with respect to x I , * . . , x , , and 3, as before, differentiation with respect to
-
xl, * . , x, , t, while or a!, , 1 = 1
11,represents higher differentiations with
respect to the x , . If P = ( x , t ) and Q = ( y ,T), we represent the Euclidean
distance between P and Q as
11

IP - &I = 1(. -rl2 + ( t - T)')''', IX -y12 = C(xz -Jyi)'.

11.

T h e scalar product of t\vo real vectors x and-y will be denoted by x . y = 2x,.~,.

1

This section is restricted to systems of partial differential equations of the form

(1.1) with (complex) constant coefficients such that l , , = lij. Writing l t j ( E , T )
in place of / z j ( P ;6,T), we thus assume this polynomial, if not identically zero,
52 S. AGMON, A. DOUGLIS AND L. NIRENBERG

to be of the degree
(4.1) deg ( l i j ( t , 7 ) ) = S, + tj,
where si and t j are integers. (If si ti + <
0, we assume l i j 3 0.) I n accordance
with (1.3), (1.4), and (1.6), we also assume
(44 s; 0,
(4.3) ti >= 0,
and

(4.4) m = 4 deg (UE,7 ) ) > 0,

where L(5, T ) = det (lij([, T)).
According to the Condition on L, m is an integer, and corresponding to any real
vector 5 # 0 there are m roots ~k+(t),k = 1, * * . , m, of the equation L ( [ , T ) = 0
with positive imaginary part. Defining

(4.5)

we note that the a,(E) are homogeneous functions of tiof degree fl which are
analytic for real 5 # 0. M - ( t , T ) is analogously defined.
Let (Ljk((E,T ) ) designate the adjoint to the characteristic matrix ( l i j ( f ,T ) ) .
Thus, in particular,
N
(4.6) 2 lij(6, T ) L j k ( E ,
j= 1
T) = dfL(5, T ) ? i,k= l,**.,N,

6: being Kronecker’s delta. If 1.j’”is not identically zero,

(4.7) deg (Li”(t,T ) ) = 2m - sk - t i .

The boundary conditions considered in this section are assumed to have
constant coefficients and to be such that B,bj BAj. Writing B h j ( t ,T ) in place of
B , J P ; 5, T ) , we thus assume that the polynomial B J L j (T[ ,) , if not identically zero,
is of the degree

(4.8) deg (Bhj(5, 7)) = + tj 9

=
rh denoting an integer. ( B h j 0 if rh t j < 0.) +
For each fixed vector 5, let us regard the elements of the matrix
.I-
z B h j ( f , T)Ljk(6,T )
j= 1

as polynomials in the indeterminate T . The Complementing Boundary Condition

requires that, for each fixed 5 # 0, the rows of this matrix be linearly independent,
modulo M + ( t , T ) , over the scalar field.
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 53

In this section, we shall solve the following

BOUNDARY PROBLEM.Let + j ( x ) , j = 1, * * . , m, be C” functions of compact
support in x1 , . . ’ , x,-space. Under the foregoing assumptions, we .reek C“ functions
U,(.Y, t ) in the halfspace t 2 0 which satisfr the system of homogeneous partial dzyerential
equations
N
(4.10a) 2 lij(a,
j= 1
9 at)uj(x, t) =0

and the boundary conditions

I
(4. I Ob)

The Complementing Boundary Conditions and the Condition on L will be

seen to be both necessary and sufficient, when the other assumptions (4.1), (4.4),
and (4.8) apply, for this boundary problem to have a unique solution decaying to
zero as t co. ---f

4.1. Explicit solution of the boundary problem. A solution of the

above boundary problem will be obtained from convolutions of the $ j with
“Poisson kernels” (see also [24])
Kjh(x, t ) j = l , . - . , N , h = 1,-.-,m.
The construction of the Poisson kernels, like that given in ADN, Section 2 for
single equations, is based on Fritz John’s resolution [14, Chapter I ]

of an arbitrary differentiable function + ( x ) with compact support into plane

71

waves, in which formula q is any positive integer of the same parity as n, A, = 2 a;,
i
and do), denotes the element of area on the unit sphere If1 = 1. When n = 1,
r
* * do): is to be interpreted as 2 , the formula then reducing to
Jt’=] E=fl

An important role will be played by the functions

-1
( 2 4 “ M! k-1

-1
qz;M ) = < - 1% (4i) for M = 0,
(27ri)

( - 1)-”( --M - 1) !
(27fi)
ZM for M < 0,
54 S . AGMON, A . DOUGLIS AND L. NIRENBERG

in which M is an integer and z an ordinary complex variable and the principal

branch of the logarithm is taken with the z-plane slit along the negative real axis.
Obviously, G ( ( x --y) * 6; q ) essentially enters (4.1 1). We note the formula

(4.12) (i)lG(=; M ) = C ( 2 ; Ad- u)

valid for any integer M and any positive integer u.

Let
s’ = -min sk, t ‘ = max t 5 , r = max r h , p = -min ( r A ,0),
(4.13) ,
r =r+p, r; = rh p, +
P = r‘ t‘ + + s’.
We set
g(z) = G ( z ; 2m + q + P).
I n arriving at the Poisson kernels, an intermediate step, as in the related
method of Fourier integrals, is to solve (4.10a) under certain plane wave boundary
conditions. What is required, specifically, is rn auxiliary solutions ( R i ( x , t ; t),
R ~ ( xt ;, E ) , * * * , Rj(x, t ; t ) )h, = 1, * * . , m, of the differential equations
N
(4.14) 2 /,(a,,
j=1
a,)Rj(x,t ; 6) = 0, i = 1, * * * , N , h = 1, . , m,
depending continuously on parameters 6 = (El, * + * ~ tn) for 151 = 1, such that
x
(4.15) 2 B h j ( a z ,at)Rj,(s,t ; 5)
j= 1 Lo = g(2rn+t’+s’+ri)
(x -[)d,.,

These m auxiliary solutions are constructed as follows. First, we introduce

h,h’= l , . . . , m .

g’&(X, f; 5, T) (X ‘ 6
= L j k ( a z , af)g(sk+s’tl.+p) + Tt)

= Ljk((5, 7)g(2nL+s‘t ( ’ - f i f P ) (x ’ 5+ Tt);

these are of class Cr+*+‘9for t > 0, 4 m T > 0. For them, we have by (4.6)
N

and by (4.7) and (4.8)

If, for each fixed vector E , we regard M+([,T ) , B,,(t, T), and LJk((5,T ) as
polynomials in T, and if we write
n. m-l
z B h j ( E ,7)Lik((E,T) 2 1 &(E)T‘ (mod M+(E, T)),
j=1 1=0
h = l;**,rn, k= l;.*,N,
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 55

the last condition above leads to

s
(4.17) phj(a,,
j= 1
at)gji (x * [)Y1qtz(E)d
1-0
(mod M + ( f ,T ) ) .

This observation can be regarded as the first step in the construction of the R6.
According to the Complementing Boundary Condition,
rank ( # ( f ) , 9 3 . , #([)) =m
for real E # 0. Hence, if f o is any fixed vector # 0, we can determine coefficients
ck1,(E), continuous in a suitable neighborhood of t o ,such that

(4.18)

This is the second consideration that enters into the construction of the Rj.
As the third consideration, we introduce m polynomials (in T ) associsted with
analogous to those of equation (1.5) in ADN, as follows:
(4.19) i%fT(l,T ) = T1 + Ul(E)T"-' + * * +a m
1=0,1;*.,m- 1.
These polynomials have the property that

(4.20)

where y is any rectifiable Jordan contour in the complex plane enclosing all the
roots T ; ( f ) for 161 = 1. (This is noted in ADN, Section 1 .)
Let us now cover the unit n-dimensional sphere R: 1E1 = 1 with a finite
number of domains, say S,<, /I= I , * * * , b, within each of which equations (4.18)
have a (non-unique) solution c ~ / ~ , ~ ( continuously
[) dependent upon f . Let
(&([), t2(E),. * , [,([)) denote a continuous partition of unity for subordinate a
to the indicated covering. Thus, each (,([) is continuous on R, Cf 2 0, vanishes
b
outside SP,and 2 C8 = 1.
p= I
The desired auxiliary solutions are defined as
b
w, t; E) = 2 tp(S)q&,
p=1
t; 0,
where Ri,p(x, t ; f ) , defined just for [ E S, , is given by

j = l;.., N, h = l ; * * , r n , = l,*.-,b.
Here, y is a simple closed curve in the upper complex 7-plane enclosing the zeros
(in T ) ofM+(E,7 ) for 161 = 1. The Risatisfy conditions (4.14) by (4.16), and they
56 S. AGMON, A. DOUGLIS AND L. NIRENBERG

satisfy (4.15) by (4.17), (4.20), and (4.18). By the Corollary to Theorem 3.2,
R{,@(x,t ; 6) for given [, despite appearances, is independent of /3; hence, Ri
is unique.
In terms of the m auxiliary solutions (4.21), we can now give the solution of
the problem formulated above.

THEOREM 4.1. The boundary problem formulated on p . 53 can be solved if the

Comflementing Condition is satisjed by the.formula

in which the Poisson kernels Ki,(x, t ) are given by

(A more convenient formula for calculation will be given below.) Here, q is a

positive integer, entering the definition of g ( z ) , which is of the parity of n and for
present purposes also satisfies the inequality q > r +
t'.
Proof: By (4.14), equations (4.10a) hold for t > 0. It will be seen below that
u j €CODfor t 2 0. By (4.15), (4.13), and the resolution (4.11),

(4.24)

=+ h W

Except for remarks on the differentiability of u j at the plane t = 0, this concludes

our proof of Theorem 4.1.
The Poisson kernels defined by (4.23) can be represented as

for t > 0,
where
N m-1
Njh(E, 7) = 2
k=lI=O
Ljk(6, T)ck[h(t)ML-[-l(t, .)-

(The dependence of the ckrhupon @ is not displayed in this argument, because, as

noted, the integral over y in (4.25) is independent of p.) To justify (4.25), we note
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 57

from (4.23) and (4.21) that

(4.26) K,,(x, t ) = APt4)’2Kjh,q(”, t ) ,
where

From the definition of gJk we have

hence, from the definitions ofg(z) and N j h and from (4.12),

for t > 0.
Formula (4.25) is an immediate consequence of this and (4.26).
It is easily seen that K j h , e and all its derivatives up to those of order q + +
r,
t i - 1 are continuous in the closed half-space t 2 0 (the same is true of derivatives
+ +
of order q T,, t i if n )= 2). From this observation, (4.22) and (4.26), u j can be
proved to be infinitely differentiable in the closed half-space t 2 0 by the same
argument as given in ADN, p. 637, in the case of a single equation. This settles
the last remaining contention of Theorem 4.1.
Remark 1. If q is of the parity of n, 0 s 6 q - 1, and n q - rh 2 0, +
+,
and if the are compactly supported functions belonging to Cn+qPra,we can write

as+tiA(n+q)12 K
2 jh,q(X --y, t ) h ( Y ) dr

as a combination of terms of the form

s a*+tj+r”Kjh& --y, -
t ) ijy+“-‘”+,(JJ)4,

these being continuous for t 2 0. Hence, u, E Cs+tjfor t 2 0. This conclusion

holds, in particular, if q = 1 + s or 2 +
s, according to the parity of n, and
s )= max (0, r - n - 1). In sum, therefore, u j E W t j for t 2 0 if
s 2 max (0, r - n - 1) and c$h
4.2. Estimates relating to the Poisson kernels. The above representa-
tions lead to useful estimates, relating to the Poisson kernels, which are expressible
in terms of a “characteristic constant”
E =n + N + t’ + 2 + A + b + A-1,
IrhI
58 S. AGMON, A. DOUGLIS AND L. NIRENBERG

where b is a bound for the coefficients in equations (4.10a) and conditions (4.10b),
A the minor constant for the boundary plane t = 0 (Section 2), and A the
ellipticity constant (Section 1). The most basic of these estimates is that of
LEMMA 4.1. The kernels Kjh,q(x,t ) are of class C“ in the half-space t 2 0, except
at the origin, and satisfi
(4.28) 5 c(s, E ) (IxIz f t2)(TA+t’+q-s)’2
lasKjh,ql (1 + llog (1.12 + t2)1), s 2 0.
Furthermore, i f s +
rh + t j q + 1, then asKjil13,is homogeneous of degree rh t i + +
q - s, and the logarithmic term in the inequality may be omitted. Here, C(s, E ) is a
constant de@ending only on s, q, and B.
The proof is as that of ADN, Appendix 1.
From (4.26), we have

COROLLARY 4.1. The Poisson kernels K j h ( x ,t ) are of class C” f o r t 2 0, except

at the origin, and satisfi

(4.28)’ Ia8KjhI 5 C(s, E)(IxI2 + t2)(rhftj-‘n--s)’2(1 + Ilog 1(1.2 + t2)1), s 2 0.

s > rh + ti - n, then 8 K j h is homogeneous of degree Th + t , - n - s, and the
logarithm in (4.28)’ can be omitted.

The remaining estimates refer to combinations of the kernels such as occur in

the boundary conditions. They will be applied in Theorem 4.2 below to give a
generalization of the maximum principle.

4.2. For h, h’ = 1,
COROLLARY * *- , m, we have
(4.29) 12Bhj(a)Kj,(X, t)I 5 c ( E ) t ( l x I 2+ t2)(TA’-
i
rh-n-’1)r2 (1 + 11% (I+ + t2)1).
Hence, in particular,
(4.29)’

(4.30)

Proof: The Poisson kernels possess the following “reproducing” property:

for all C” functions $h with compact support,

XI/[
h‘ j
a hi( --Y2 t)1t=04%’(Y) dr = $ h ( X ) ’
Hence,
2 [ B h l ( a ) K j h , (t)]t=
x , o= 0 for x # 0, h, h’ = 1, ) m.
j
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 59

This, (4.28)', and the theorem of the mean sufficeto prove (4.29). Inequality
(4.30) follows in a similar way from the fact, a consequence of (4.15), that
t ) ] t = o = 0 for x # 0, h # h'.
[I:B,,j(a)Kjh,,p(~,
i
Remark 2. Let 1, = max (0, rh), and for 1 2 1, and h = 1, - - - , m, let A(.)
be functions of class Cz+n-*h'+2for a11 h' = 1, * , m such that --
ak+h(x) = O(lxl-l-rh-k(l + (logI x l l ) ) , k = 0, . . , max (1 - rh,)
+

h'

for large 1x1. Let %-*A denote a particular differential operator with respect to x
of order 1 - rh , and for t > 0 define

Then uh may be continued as a continuous function into the closed half-space

t 2 0, and uh(x, 0) = al-rA+h(x).
Proof: The integral converges absolutely, because of (4.28)' and because of
the conditions imposed on the rjhh; a,
can be replaced by and partial integra- ay,
tions can be carried out, without contributions from infinity, to transfer a',-'h from
Kjh,to . The remainder of the proof is like that of ADN, p. 640, using Remark 1
and (4.29).
The next result estimates combinations of derivatives of the solution such as
are to be found in the boundary conditions. It is a kind of generalization of the
maximum principle. In it, we make use of the following norm defined for any
non-negative integer j:
[f(x)lj = 1.u.b. IDBfl.2
IBl=i
THEOREM 4.2. Consider the solution u j ( x , t ) giuen by (4.22), and let 1 be an integer
suchthat 1 2 r h , h = 1, - , m. Thn for any i = 1, * * * , m and any d&rentiation
*

with respect to x of order 1 - rh , we have

l20 i
t)l
sup la~-rh~Bhi(a)ui(xr c(f,E)z
j
[4j11-rj.

Proof By (4.22), it suffices to show for any fixed h , j that

sup 1
120
a?hx 1
B h i ( a)Kih'(X, ' )* I 5 '(', [4hl lk*h*

In the case h = h', we transfer %-'b to $*(x) and apply (4.29)'. In the case
h # h', we make the substitution (4.26) with q n - 1 rh, 2 0, transfer a + +
suitable number of A to +h' in whole or in part, and use (4.30). Details are as in
ADN, p. 641.
4.3. Boundary problems not well posed. In this section, we shall show
by examples that, if the Complementing Boundary Condition, or the Condition
60 S. AGMON, A. DOUGLIS AND L. NIRENBERG

--
on L, is violated, then there are solutions (ul, * , uN) of the homogeneous system
(4.10) ($h = 0) such that for any positive integer c the derivatives ofuioforders
up to a + t , , inclusive, are bounded, while for at least one index k, has
Holder coefficients (with any exponent u < 1) as large as desired in any half
neighborhood of the boundary t = 0.
Suppose first that the Condition on L holds, and consider homogeneous
boundary conditions
(4.31)

such that the Complementing Condition fails for some value of 5. For this value
o f f , which we keep fixed throughout the following reasoning, the discussion a t the
end of Section 3 assures us that there exist constants ckl such that the functions

are not all identically zero, where y is a contour enclosing the zeros (in T ) of
M+(E,T ) . These functions satisfy the differential equations (4.10a) by (4.1) and
the boundary conditions (4.31) by (4.8), (4.20), and (4.25). Not all of these
functions reduce for any fixed value of x to a polynomial in t. Hence, in particular,
for any positive integer a and for at least one value of k , 1 5 k 5 N, there is a
constant ak for which
(4.33) ck 0 ,- a;ttkvk(O, 0 ) # 0.
= a0.t. ~ ~ (ak)
For A > 0, define
(4.34) Uj(X, t ; A ) = 1 - O - t . 'Uj(AX, At), j = I , . . . , N,
From (4.1) and (4.8),it is clear that the u j satisfy both the differential equations
(4.10a) and the boundary conditions (4.31). It also is clear that the derivatives of
uk with respect to x, t of orders up to a +
t, are bounded uniformly as il -+ 00.
Since
ay+'.u(O, ak/A; A ) - a;+'*u(O, 0;A )
= (A/a~)"[a;"'Uk(O, ak) - a ~ + " U k ( o y o)]
= (A/uk)"ck 2

the Holder coefficients for however, tend to infinity with il.The example
of the u j thus shows that the Complementing Boundary Conditions are necessary,
if the desired estimates are to hold, as contended.
Suppose now that the Condition on L is violated; this is possible only for
n = 1. Define also in this case
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 61

where T: (E), * * . , ~ & ( tare

) the ~ ( 6 roots
) with positive imaginary part of
L(6, 7). There exists in this case a unit vector E = f l such that r ( E ) 2 m +
1.
Since, for such E, M+(E, T ) is of degree 2 m +
1, whereas the number of boundary
operators is m, it follows that there exists a solution uj of (4.10a) of the form (4.33)
which satisfies any given homogeneous boundary conditions (4.31). The functions
uj(x, t ; A), defined from these ui by (4.34), again violate the estimates in question.
Thus, the Condition on L, like the Complementing Boundary Condition, is seen
to be necessary as well as sufficient in order that the desired Holder estimates
hold for the boundary problems considered.

5. Estimates for Singular Integrals of Certain Types

I n this section, we shall give estimates for singular integrals of the form

(5.1) u(x, t ) =P(x -y, t l f b ) 4, t > 0,

with kernels of the specific type

(5.2) K(x, t ) = an+q+m+Mih,@(X, t ) .

(Integration is over the entire n-dimensional x-space.) The estimates are immediate
consequences of results obtained in ADN, Section 3, for convolutions (5.1) in
which

(5.3)

+
where P = ( x , t ) and IPI = (1xI2 t2)lf2. The function Q ( x , t ) is supposed to be
+
continuous on the half-sphere t >= 0, lxI2 t2 = 1, and at least to satisfy a uniform
Holder condition at points of intersection of this half-sphere with the plane t = 0.
In addition, it is required that

(5.4)
I,=l R(x, 0)dw, = 0,

where dw, denotes the element of surface area on the unit sphere 1x1 = 1. For
n = 1, the latter assumption takes the form Q(x, 0) = -Q(-x, 0).
I t is easily seen that the kernels given by (5.2) possess the above properties.
First, Lemma 4.1 shows that these kernels are of the requisite form (5.3). Secondly,
their behavior on the unit sphere is adequately controlled by (4.28). It remains
only to prove (5.4) for these kernels. Since by construction z\ l f j r ( a ) K f k ,=
Q 0,
j.
we see by applying P ( a ) and summing over i that L(8)Kj,,, = 0. This fact
enables us to deduce (5.4) from the corollary of ADN, p. 645.
We shall have to state estimates which are concerned with several norms or
semi-norms off(x) and u(P)already used in ADN. The first are those of Holder
62 S. AGMON, A. DOUGLIS AND L. NIRENBERG

type, namely
[f], = 1.u.b.
If(.) -fCv)I
2#2/ 1% --Yl"
and

with 0 < cc < 1. The norm off ( x ) in L, , 1 < p < CO, we denote by I f I L , and
for fixed t > 0, the L,-norm of u ( x , t ) , regarded as a function of x , by IuIL,,,.
For any non-negative integerj and again for 1 < p < CO, we also define

Finally, iff ( x ) represents the boundary values of a function v(x, t ) defined for
t2 0, i.e., f ( x ) = v(x, 0 ) , and if I U ~ ~ , is~ , finite, we define
If Ij-1,p,Lp = g.1.b. IUljJ, 9

the greatest lower bound being taken over all functions v as described.
The indicated character of the kernels (5.2) and Theorems 3.1, 3.2, and 3.3
of ADN enable us to infer fundamental inequalities which govern the corre-
sponding singular integrals. These are stated below in a theorem analogous to
Theorem 3.4 of ADN.
THEOREM 5.1. Let u ( x , t ) be the convolution (5.1) with kernel (5.2).
(a) Iff is in L, for somejnite p > 1 and [f], < cc for some positive u < 1, then
[ u ] , isjnite and, in fact,
[ u l a S C(E, q, Lfla
(b) Iff is in L , for somejnitep > 1, then so is u(x, t ) for every t, and
14Lp,t5 C(EY Q ,P) If IL,.
(c) Iff is in L , for somejnite p > 1 and I fJ1-llp,L, isjnite, then also is
Jinite, and
lull,&, s w,4 , P ) Ifll-l,*,L,'

6. A Representation Formula in the Inhomogeneous

Boundary Problem
In this section, we shall obtain an integral representation for a C" solution
-
1, * * , N, with compact support in the half-space t 2 0, of an
(uj(x, t ) ) , j =
inhomogeneous boundary problem
N

h = l,..*,m.
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 63

The construction is based on the fundamental solution r ( P - P ) , P = ( x , t ) ,

P = ( f , i), of the elliptic equation L ( ~ ) =
u 0 with singularity at P = P. We note
the estimates
(6.2) IPr(P)I 2 const. IP12nr-n-1-a(1 + [log [PI I) for a 2 0,
in which the logarithmic term is omissible, if a > 2m - n - 1. (The constant
depends on s, m, n, and the ellipticity constant for L.)
The first step in finding the desired representation of uj is to construct explicitly
a vector ( u j ( x , t ) ) satisfying the stipulated differential equations, if not the
boundary conditions. For this purpose we assume the& extended in the manner of
+
ADN, p. 652, to all (n 1)-space as functions with compact support of class CR+
-
for R sufficiently large. Letf,,,(P), i = 1, * , N, denote the extended functions.
We then define
n N

+
3 acting on P, and the integration being over all (n 1)-space, see [lo, p. 5191.
We easily see that v j E CR+tj-2and also that the v j satisfy the stipulated differential
equations
n:
2 lij(i3)vj(P)=fi(P) for t > 0.
j=1

Remark. If 0 < a < 1, b 2 2m, andf, E Cb-'it', then of course

~ i i ( 8 )c~b -~2 m + t j + a ,
and it is easily seen with the aid of a classical result (see Theorem 3.1, ADN)
that u j E Cb+'>+' and

[Uilb+lj+a 5 * 2i [ f R , i 1 b - s i + a const ' 2i [&lb-si+rx.

Since the f R , i have compact support, for sufficiently large [PI the integral
defining u j in (6.3) will not be singular, and integrations by parts, as well as
differentiations with respect to P, will be possible on a scale only restricted by R.
For sufficiently large lPl, we thus have, in particular,

1 = aa+tj / r ( P - P)$22"-*i-tfR,i(P)dP
- (- 1)2m-tiaa+b

= /i32na+aI'(P
s- $zm-tir(p - P)$--StfR,.(p)

P)8-.ifR,,(P)dP
df)

for a 2 0; hence, by (6.2),

(6.4) ]aa+'jujl const. IPl-n-l-a for a 2 0.
64 S. AGMON, A. DOUGLIS AND L. NIRENBERG

for a 2 0 and sufficiently large IPI, and for the boundary quantities

we have for large 1x1

(6.6) I ae-mgh(x)1 s const. IxI-~--~-' for a 2 0.
The above estimates will be of aid in justifying representations we shall shortly
describe, not of the u j , indeed, but of derivatives of the uj of suitable orders.
Again let
1, = max (0, rh).

THEOREM
6.1. If the vector ( u j ( x , t ) ) is a C" solution of (6.1) with compact support
in the half-space t 2 0, then the representationformulas

are valid for t > 0. Here, v j and gh are dejned by (6.3) and (6.5) with f R , i a suJicienth
smooth extension of f i to the whole (n +
1)-dimensional x, t-space.
The representation formula is obtained from the lemma and the uniqueness
theorem (Theorem 6.2) below by the argument used for Theorem 4.1 of A D N .

LEMMA.Let h, =f, - g, , and let I be any integer 5 I,. Then the integrals

4p,ht ) = Jal+'iK,,(x -y, t)h,Cy) 4, t > 0,

converge absolutely. Furthermore,
(6.8) lljv,l(x,t)l&,,t constant independent o f t ,
and

(6.9) l'jp,l(x> t)ll,~,< ~ 0 .

TheJirst derivatives aIjp,lare absolutely integrable on every plane t = const. > 0, their
integrals converging unifrmly in any interval 0 < E t T. s
Proof: The integrals converge absolutely by (4.28)', (6.6), and the fact that
f, has compact support. By (4.26),

If 1
Ifp,l(x,
s --y, t)h,(y) 4.
t ) = a2+tlA,(n+q)'2KjP,P(~

- r , is even, choose q such that n + q - 1 + r , > 0 (q also must be even).

 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 65

Integrating by parts, we are then able to transform the right member above into

1aZ+tlA,c"+9-l+'.)/2K.
39.9( x --y, t) .A Cl-T~.'/2hp, (Y )

there being no contributions from infinity in view of (4.28) and (6.6). Since,
dY,

in accordance with (6.6), a1--'9hp belongs to L , , estimate (6.8) now follows

from Theorem 5.l(b) in the case in which 1 - r p is even. Estimate (6.9) we
similarly obtain in that case from Theorem 5.1 (c), once we have verified that
~ i 3 - r ~ h p ~ l - l , 2 ,<
L 2co,or, in other words, that aL-rghp(x) is the restriction to the
plane t = 0 of a function Vp(x,t ) with square-integrable first derivatives in the
half-space t > 0. Such a function is given, however, by the formula
VJx, t ) = [ ( t ) a ' - w z , ( X ) ,
where [ ( t j is in C" for t 0, ((0) = 1, and [ ( t j = 0 for t 2 1.
The final statement of the lemma is easily proved, in the case in which I - r p
is even, from the previous formula and the estimates (4.28) and (6.6) already
used.
The case in which 1 - r p is odd, say 1 - r p = 2k +
1, is treated by the same
arguments as above applied to the formula

This completes the proof of the lemma.

The uniqueness theorem referred to above remains to be justified. Its proof
will depend on Fourier transformation and for this reason will be formulated in
terms of D, = (1/i)az , D, = ( l / i ) a t , and vj = itJui rather than of an, a, , and
u,. In these terms, the differential equations take the form

(6.14) 2 li,(Dx,D,)v, = 0 for t > 0, i = l , . . . , N,

j
and the boundary conditions
(6.15) 2 B,,(Dx, D,)vj = 0 for t = 0, h = I,---,m.
j

THEOREM -
6.2. Let v j , j = 1, * , N , be functions of class C1l for t 2 0, where
l j = t j + max (si, rh), and sntisfy the homogeneous dzxerential equations (6.14) and the
i.h
homogeneous boundary conditions (6.15). Assume that vi and its derivatives up to order 1,
are absoluteb integrable on each plane t = const. > 0, the integrals converging uniformly
with respect to 1 in every jinite interval 0 < E t 5 R. Assume also that v j and its
derivatives of orders up to ljPossess u n ; f r m b bounded L , n o r m on the planes t = const. 2 0.
FinalCy, suppose that
r
(6.16) Jt > O 2i Iv,(x, t)I2dx dt < co.
Then v j = 0.
66 S. AGMON, A. DOUGLIS AND L. NIRENBERG

Proof: For t > 0, the Fourier transforms

D,(E,
s
t ) = ( 2 7 ~ ) - 4ePizEuj(x,
~ t) dx

are continuous in (5, t) and possess continuous derivatives with respect to t of

orders up to li. From (6.14) we have
(6.17)

Intending to apply Theorem 3.2, we next prove that the 6, exponentially decay
as t + CO. First, the 0, all satisfy the ordinary differential equation

qt,D*)4(5,t ) = 0
obtained by applying L k i ( t ,D t ) on the left of the members of (6.17) and summing
over i. Any solution of the latter equation is a linear combination, with coefficients
which are polynomials in t , of exponentials e”‘, where p is a root of the algebraic
equation L ( l , p / i ) = 0. Secondly, by condition (6.16), j”ID,([, t ) I 2 d t < co for
almost all 5, and for these evidently 9%’~ p < 0. Since, by ellipticity, W e p # 0
for any 5, we coiiclude by continuity that 3h p < 0 for every 6,or, in other words,
that the 6,(6, t) decay exponentially, as asserted.
Now we shall show that
(6.18) Z B , , ( t , D,)D,(E,t ) = O for t = 0.
j

Theorem 3.2 will then prove, in view of (6.17) and the exponential decay of the
D , , that if([,t ) = 0 and thus that u j ( x , t ) = 0, as contended. To prove (6.18),
it suffices to show that

for every 4 in C” of compact support. The integral on the left, however, by

Parseval’s formula, is equal to

lE j
$(X)Bh,(R , D,)& t) 4
4
and by (6.15) the limit of this integral is zero, as t 0, because of the assumed
uniform boundedness of the L , norms involved. This completes the proof of the
theorem.
 CHAPTER 111

THE SCHAUDER ESTIMATES

7. Notation
Let 9 be a k-dimensional domain (not necessarily bounded), and denote by
and 3 its boundary and closure, respectively. C ' ( 9 ) will denote the space of
functions possessing continuous derivatives up to order 1 2 0 in 9 ; the meaning
of C ' ( 3 ) will be similar. Forfin CZ(9),we define the pseudonorms

[fll = Vl," = 1.u.b. rm

where the least upper bound is taken over all derivatives of order 1, as well as
over the domain 9,and we define the norm

-
The subclass of those functions in C l ( 9 ) whose derivatives of order 1 satisfy a
uniform Holder condition of order GI, 0 < GI < 1, is denoted by CZ+.(9). For
f E C 2 + ' ( 9 ) ,we define the pseudonorm

-
where the 1.u.b. is over P # Q in 9 and all derivatives of order 1, and the norm

= IA+ rfil+,.
ifiz+*
Let C = C,: +
1 . ~ 1 ~ t2 < R2, t 2 0, be a half-sphere in x,t-space with
A'= (xl, . * , xn). Let (I = uR denote its planar boundary IxI2 < R2, t = 0.
For the linear spaces Cl(X), C z ( u )and
, CZta(z),Cl+'(a), 0 < CI < 1, we introduce
seminorms such as
[f]a,l = [fJzl = 1.u.b. d$tl la"f(P)I _S Ra+' IJIF
for any integer a with a + +
12 0. If a 1 < 0, we define [ f ] a ,= l 0. Here,
the least upper bound is taken over all derivatives of order 1 and all points P
in C, d, denotes the distance from P to the spherical part of the boundary of C,
1x12 + t2 = K 2 . Correspondingly, we also define
1
Ifla,r = .z= a
I
[fla,j,

Ifla.r+. = Ifla,i + [fla,i+a.

Norms for functions in Ca(o)are defined in the same way.

67
68 S. AGMON, A. DOUGLIS AND L. NIRENBERG

I t should be pointed out that Lemma 5.1 is stated incorrectly in ADN. The
condition stated in the lemma in the words, “There is a 6 > 0 such that every
point P E D is the extremity of a segment of length 6 lying entirely in D”,should
be replaced by the cone condition, viz., “There is a finite right spherical half-cone
such that every point P E D is the vertex of a congruent half-cone that lies wholly
in D”.

8. The Schauder Estimates for Systems of Equations with

Constant Coefficients
In this section, we shall still consider boundary problems

for equations with constant coefficients in a half-space of the type introduced in

Chapter 11.
As before, let 1, = max (0, r J ; let 1 be an integer 2 I,. Throughout this
section, we shall assume that u , ( x , t ) E C ‘ O + ~and
J + fi(x,
~ for t 2 0 and
t ) E C1--sd+3
that qh(x) E C‘-Q+lXfor all x , GI being a positive number less than one.
The basic estimate is for functions of compact support. From it, the general
inequalities will be deduced.
THEOREM 8.1. In addition to the preceding hypotheses, assume each u , has compact
support. Then u, E C1+tj+r,and
s
(8.2) [%11+t,+u 5 const. !2
=l
[
[filc-s,i3 + 2 [pl*l,-r,+”/’I
7tL

h= 1

where the constant depends only on I, a, and E. ( E was defined before Lemma 4.1 .)

Proof: We shall consider only the case in which u, E C“, the whole theorem
being deducible from that case by a procedure of approximation. Formula (6.7),
after ( 1 - /,)-fold differentiation, gives us
?it

(8.3) 8+lJU,(X, t) = 8+tJUI(X, t) + 2 I , p , l ( x ,t ) ,

p=l
the integrals Z 3 , , , l ( x , t ) being those defined in the Lemma of Section 6 and v , the
function (6.3). By the Remark made in Section 6,

(8.4)’

the constant depending only on 1, u,

.v
(8.4)”
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 69

From formulas given (in the two cases distinguished) in the proof of the Lemma
in Section 6,

IjlD.I=
san+qfr*ft*Kj,,q(x

Applying Theorem 5.1 (a), we thus deduce

--y, -
t ) a"-'.(q?,(y) - g J y ) ) 4.

['js.lla 5 C(E, (IJ')[Ts - gp]Z-r,+a

N
- const.
5 ([vs~l-r~tx +2
i=l 1
[.~~t-s,+a >

the latter inequality following because of (8.4)". This, (8.3), and (8.4)' suffice to
prove (8.2).
COROLLARY.The representation (6.7) of Theorem 6.1 is valid when u j ~C'o+'j+~,
provided the u j have compact support in the half-space t 2 0.
The above estimates for solutions with compact support in a half-space now
are applied to the study of solutions defined just in a hemisphere. After this, we
shall see how to modify estimate (8.2) to apply to solutions in a half-space that are
not of compact support, but are regulated less severely at infinity.
THEOREM8.2. L e t u j E C ~ ~ + ' 0J +<~ u, < 1, beasolutionof(8.1) in the half-sphere
X.= Z,. Assume thatfor an integer 1 with 1 5 1,
N m

Then u j E C'+'j+, and

N
(8.5) 5
i ~ i I t ~ - t ~ , ~ + t ~ const.
-+a (K + 2 Idr-tk,d,
k=l
where the constant depends onb on I, a, and E.
Proof: A condition for deriving (8.5) is the prior knowledge that u j E C ' + ' ~ + ~ ,
--
j = 1, * , N . By assumption, u j E C 1 ~ + f j +hence,
z; (8.5) is deducible, in partic-
ular, for 1 = 1,. We shall now show how, accepting inequality (8.5) for 1 = l,,
we can prove u i ~ C ' o + ~ ' - ~ Inequality
j+~. (8.5) can then be demonstrated for
1 = I, + 1 and the whole procedure repeated as many times as required.
Inequality (8.5) with I = I, is applicable to x-difference quotients for the u j ,
as in ADN, p. 661, paragraph (b), with the result that a,u, E C'o+'i+'. Knowing
this, we shall seek to express the pure t-derivatives v j = in terms of
derivatives of the uk of the form i3zo+fk-1azuk, to this end applying 8fo-Q to the i-th
differential equation in (8.1). Thereby we have
2 &,(a,, a,)ui = ap$,
j
where
lij(&
7 ) = 7*o-*ilii(tJ 7 ) .
70 S. AGMON, A. DOUGLIS AND L. NIRENBERG

(The degree of Aii is I, + t,.) From the first sentence in this paragraph,

2j A,,(O, = a;o-s$ + 1(Aij(O, a,) - nij(a,, a,))uj E C1+%,

j
the derivatives ofeach ui entering the right member above being of order I , + tj
and containing a,. Moreover,
nij(0, T) = C i j T z o t f j ,
and, incidentally,
Iij(0, T ) = E i j T s i + t f ,
where cij is a constant. Hence,
2 ciivi E Cl+a,
j
while det (cij) # 0 because, for r # 0,
0 # det (li,(O, T)) = det (C~,T"+~I) = T~~ det (cti).
Hence, v j E C1+', and u, E CZo+lftjfa,as contended.
From the foregoing remarks, in justifying inequality (8.5) we can validly
assume u j E Cz+tj+a,j = 1, * * , N, and proceed as in the proof of Theorem 6.2
in ADN, pp. 660-662, but omit the rather intricate details.
Now that estimates have been obtained for solutions of (8.1) defined just in
a hemisphere, we are in a position to generalize Theorem 8.1, as noted above, to a
class of solutions without compact support.
THEOREM 8.3. Let u j ( x , t ) E Czo+tj+a,0 < u < 1, satisfy (8.1) in the half-space
t 2 0.Assume that for a j x e d integer 1 2 1, both the quantities
N m

and
N
-
M , = lim 2 max
R-(l+tj+z) luj[
R-00 j=1 ZR
arejnite. Then u j E C1+tJ+a,
and

[ui1z+tj+a S const. ( K O + Mo).

The proof of this theorem is like that of Theorem 6.3 in ADN.
As an immediate corollary, we have the following (see also Miranda [ 181) :
THEOREM OF LIOUVILLE TYPE. Let u j ( x , t ) E CZo+'j+',0 < 01 < 1, satisfy (8.1)
in the hay-space t 2 0. Assume that the functions f i and cpn are polynomials of degrees not
exceedingk-siandk-rr,,respectively,i= l,..., N,h= l;~~,rn,wherek~l,,
and assume also that
N
lim
- 2 R--(k+fj+a) max lujl = 0.
R-w j=1 =R

Then ui is a polynomial of degree at most k + ti.

 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 71

9. Equations with Variabie Coefficients

The results just obtained for constant coefficient problems in a half-space or a
hemisphere will be applied here to the study of solutions of a system of differential
equations (1.1) with variable coefficients in a general domain 9,the solutions ( u j )
satisfying Complementing Boundary Conditions, also with variable coefficients,
on a portion I' of 9. Our goal, achieved in Theorem 9.3, is to estimate norms of
the type IujlL+tj+aover arbitrary subdomains J&' o f 9 abutting r. An important
step towards this goal is to be able, as in Theorem 9.2, to bound norms of the type
Iujlt,-tj,l+tj+zon a hemisphere, the flat face for the hemisphere being I'. This step,
in its turn, depends on a prior result, Theorem 9.1, which concerns the solutions
of problems with variable coefficients in a half-space with I' the entire boundary
plane.
In all these results, whatsoever 9 and I', the hypotheses of Sections 1 and 2
always are presupposed. Suitable assumptions also are made concerning the
behavior of Fi , + h , and the coefficients in

and

where p and u denote multi-indices indicative of the precise differentiations

involved. For a fixed integer 1 2 1, = max (0, rh) and a fixed positive number cc
+
less than one, Fi and the coefficients a i j , p , IpI = 0, * * * ,si --
t , j = 1, * , N,

are assumed in particular to belong to ClPsi" in 3)and t$h and the coefficients
bhj,o, lul = 0, + -
, rh t j , j = 1, * , N, to C ' - ' A + ~on I?. All these quantities
furthermore are assumed to have finite norms of the pertinent type; let k denote
an upper bound, in particular, for these norms of the coefficients. The resulting
estimates of the u j will involve constants, denoted C,, C,, or C,, which depend on the
latter coefficient norm bound k, the uniform ellipticity constant A (Section l),
the minor constant A (Section 2 ) , the number of dimensions n, the number of
dependent variables N, the Holder exponent a, and the selected integer 1. (The
third constant C,, which appears in Theorem 9.3 regarding a general domain 9,
also will depend on the shape of the domain.) The constants do not depend,
however, on the solution ( u l , * * * , u N ) .
THEOREM 9.1. Let ul, . * . , us satisfi (1.1) in the half-space D + : t > 0 and
(2.1) on its boundary t = 0, u j belonging to C'tlj+a in D+ with Iujll+f,+a < CO,
-
j = 1, . * , N . Assume the semi-norms

[Fill-8,+z > [+hIl--rh+~

and the norms
laij,pll-si+a > Ibhj,oll-rh+a
72 S. AGMON, A. DOUGLIS AND L. NIRENBERG

alsojnite for all possible values of their indices. Then

9.2. Let ul, . - ,uN satisfy ( 1 . 1 ) in the interior of the hemisphere

THEOREM
C = C,, R 2 1, and (2.1) on the planarportion ofthe boundary, u, belonging to C'0 tt*' a
in C, j = 1, --
, N . Assume that, f o r all values of their indices, the norms
IFh+s,,l-s,ta 9 Ihlt'tr,,,l-rh+a Y

and
~ - s , + a,
latj,p18,+tJ-~p~, Ibhj,oir~+l,-lol,Z-r~+a

arejinite. Then u , E C ' + ~ Jin+C,

~ and
N m N
~ut~t'-t,,l+t,+a5 cZ( L:1 I F i l t * + s , , l - s , + a
2=
+h2= l I+hlt,+r,,,l-r,,+a + L: ~ u ~ t * - t ~ . o ) ,
k=l

j = I,... J N.
The proofs of these theorems, the first relying on Theorem 8.2 and the interior
estimates of [lo], are so close to those of Theorems 7.1 and 7.2 in ADN, pp.
664-666, as not to require extended discussion. I n demonstrating Theorem 9.1,
it is well, noting the index J for which [u,] l+tj+cc is greatest, to consider, first, the
possibility [ u J ] l+tJ+a 5 luJl l+lJ+a , which is trivial, and, secondly, its opposite.
In connection with the second possibility, a quantity analogous to A in ADN,
equation (7.6), p. 665, is useful, defined here as
-
A = l a l u ~ ( P o )- ~ ' ~ J ( Q ) I Y
P o - &I"
where Po = (0, to) and Q = (y, T ) , 0 < to 5 T, are such points, and 8 such a
differentiation, that
A^ > a luJll+tJ+a
A similar remark applies to the proof of Theorem 9.2.
The third theorem is concerned with a general ( n 1)-dimensional domain +
9, which may be infinite, a portion r of its n-dimensional boundary 9 (I? may
be the entire boundary), and a possibly infinite subdomain d of 9 such that, in
a
the n-dimensional sense, 2 n is in the interior of With cc and 1 fixed as r.
described a t the beginning of this section, it is assumed that there is a positive
number d such that each point P in a' within a distance d of & has a neighborhood
t
U p with the following properties: (a) n & c I?; (b) Uz, contains the sphere
about P of radius d/2; (c) the set 0, n G is the homeomorphic image of the
closure of an (n +
1)-dimensional hemisphere Z'R(P),R(P)5 1-the boundary
subset o,, n 9 corresponding to the flat face F p of the hemisphere-under a
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 73

mapping T p which, together with its inverse, is of class CLf'+', where 1 =

max ( -si , --zh , t j ) . Over their respective domains, both T , and its inverse are
assumed to have finite I I norms bounded by a constant K independent of P.
Let A p denote the minor constant for 0, n 6
(Section 2) that pertains to the
transformation T p , and define A here as

A = A,- = inf A p ;
P
it is assumed that
A > 0.
It is important to see how the coefficients in (1.1) and (2.1) and the uniform
ellipticity constant behave after the mapping T f J : = i ( x ) . An arbitrary linear
combination of differentiation operators, say c ( a ) = c~,,~,,... $82 . . (summa-
tion signs will be omitted in this argument), is transformed into

- + lower order differentiations,

L A

;(a) = 2jl,j2,... *

where jj =
ax.
= T' ai, and each tjl,j,,... is a linear combination of the ci,, *,...
axi
with coefficients that are products of the %,/ax,. Correspondingly, c(E) = c*(Z),
A

where S = Ed.Hence, the coefficients in the transformed differential

equations and transformed boundary conditions are of the respective continuity
classes required above, and, moreover, the norms

(9.la)
and

(9.lb)
a , + - a , + --
for IpI = 0, . * si t,, IuI = 0, * Y t , , i , j = 1, * * * , N, h = 1, * , m,are
bounded uniformly with respect to P,the accent always referring to the pertinent
A

transformed quantity. In conformity with previous notation, uniform bounds for

the norms of types (9.lb) will be denoted by k and for the norms (9.la) by
[r$,JEr+a,
l j J F 8 i + aand respectively. (These bounds of course, will depend on K . )
Secondly, the characteristic determinant for the system (1. l), which we may denote
by L(Q, Z) at an arbitrary point Q in the domain of T p ,is invariant under T , in
A *

the sense that L( Q, E) = L( Q, E). O n the other hand, since both T, and its inverse
possess first derivatives which are subject to a uniform bound independent of P,
there is a constant p such that p-l IZI 2 151 2 p !El, and the constant ofuniform
ellipticity for the transformed system of equations is seen to be Ap2"'.
The above remarks enable us to reduce the present problem for a general
domain to that for a hemisphere treated in Theorem 9.2 above. Using, in fact,
the same methods as in ADN, p. 668, we thus obtain
74 S. AGMON, A. DOUGLIS AND L. NIRENBERG

9.3. Under the assumptions above, k t ul, * * * ,uN s a h h (1.1) in 9

THEOREM
in 9 V r, we can conclude that u j E Cz+tj+uin 2,and
r. r f u j E Cr~-ttj+a
and (2.1) on

the constant C, depending on k, A , A, n, N, M, I, K, d.

Remark 1. I n case 9 is bounded and r
= 9, we may take .d = 9 and
replace each term lukl? in the above estimate by

I.;/ dV*

Remark 2. In case 9 is bounded and I’ = &, and in case the solution

(ul, * , uLv) is unique when u j E C l ~ + ~ jwe
+ ~ may
, take A? = 9 and also omit
C lukl: on the right of the above estimate. However, a new constant not explicitly
known then replaces C.,
These remarks follow just as in ADN, pp. 668-669.
 CHAPTER IV

ESTIMATES OF INTEGRAL NORMS

10. Lp Estimates
Global, local interior, and local boundary estimates in L , will be given under
appropriate hypotheses, the results for general problems (Subsection 10.2)
ensuing from those (Subsection 10.1) for equations with constant coefficients and
solutions with compact support. In addition (Subsection 10.3), L , theory will be
applied to generalize previously stated estimates concerned with Holder con-
tinuity. We follow ADN except in deriving local boundary estimates.
As in ADN, Section 14, we use the following notation:

By H j , L , is meant the Banach space obtained from the C“ functions in 9 by

completing with respect to the norm 11 and Hi- I l p , L , denotes the class of
functions 4 which serve as boundary values of functions u belonging to Hi.L, in
53. This class is normed by

and

g.1.b. in both cases being taken over all functions u in Hj,L,that equal 4 on the
boundary.
Norm symbols frequently will be abbreviated as I [4] 1\1111 I , etc., the index

L, being omitted.

10.1. Equations with constant coe5cients, solutions with compact

support. The two theorems of this subsection are concerned with systems of
equations of the form shown in (8.1) and, as indicated, only with solutions of
compact support.

THEOREM 10.1. (Interior estimates.) If the uj have compact support, and the
derivatives &uj belong to L , ,p > I , then

(10.1)

75
76 S. AGMON, A. DOUGLIS AND L. NIRENBERG

Furthermore, $1 is apositive integer, and ;f I[z l a k ( a ) t & ] l l - s i , L , isJnite, SO is I[u~]~~+~~,L,,

and k

(10.2) l[ujllZ+tl,L, 5 2i k
iik(a)Ukl~Z-si,L,*

The indicated constants depend only on p, n, and A (Section 1).

Inequality (10.1) is proved from Theorem 3.2, ADN, applied to the repre-
sentation
atjui = ayr
* qqUi)= atqr * 2 mikUk)
i.k
- &(a2rri--ti r 2 a-8'liku,)
i,k
= a w * i2, k a-siiikuk + const. 2 a-Silikuk,
i,k

where r again denotes the fundamental solution of the elliptic equation L ( a), = 0.
Inequality ( 10.2) then follows by a difference quotient procedure.
The next theorem is concerned with boundary problems in a half-space of the
form (8.1), or A'
4 j U j =A(& t ) , i = l,.-.,N, t>0,
j=1
(10.3) N I

with constant coefficients and with finite characteristic constant E (in particular,
+
A > 0, see Sections 2 and 4). Let 1, = max (0, rh I ) , and let 1 be an integer
2 4-
THEOREM 10.2. Assume the u j vanish for (PI 2 1 and belong to H z , t J , L ,in the
half-space t 2 0 . Then

(10.4) IbjII z+t, 5 const. c Ilf, 11

i
+c I1
Il+h II/ - r h - l / P >

with constant depending on I, p , and E.

I n proving this theorem, it suffices to consider solutions u i which are infinitely
differentiable. The representation formula (6.7) applies to such solutions, and
then an argument can be given like that for Theorem 14.1, ADN, p. 701.

10.2. Equations with variable coefficients. T h e above results for com-

pactly supported solutions of constant coefficient problems in a half-space will
be applied here to solutions of general systems of differential equations (1.1) in
arbitrary, bounded, sufficiently regular domains 9.Local interior, global, and
local boundary estimates, in this order, will be given-the boundary estimates being
for solutions u j that satisfy Complementing Boundary Conditions (2.1) on the
boundary 3. Preliminary estimates (Theorems 10.3 and 10.4) still deal with
solutions having compact support.
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 77

Our first result is a n interior estimate for a solution u j of (1.1) that vanishes
outside a sufficiently small sphere of radius r. For a given integer 1 2 0, it is
assumed that, in the notation of Section 9,
pi E C"-'i, a f j , +E C1-'i;
k will denote an upper bound for laij,pll-si. It also is assumed that IIFil[l-si< 00.

THEOREM 10.3. Positive constants rl and Kl exist such that, if r 5 rl and the
-
l l ~ ~ ,l j( ~=, 1 , * * , N, are j n i t e , then I(ujI(z + t j also is Jinite for j = 1, * , N, and
(10.5)

The constants rl and Kl depend on n, N, t' (Section 4), A (Section l), b (just before Lemma
4.1), p, k, and I, and also on the modulus of continuity of the leading coeficients in the lii.
The next result has to do with equations (1.1) defined in a hemisphere C,,
Complementing Boundary Conditions (2.1) being satisfied on the flat part of the
boundary of the hemisphere. A solution u j is considered that is zero outside a
smaller hemisphere E,,, 0 < r < R. With 1, = max (0, r, +
1) it is assumed that,
for a given integer 1 1 I,,
a i j , +E C'-"i, bhjs,E 6 + h ;
k in this case denotes a n upper bound for [ a i j , p [ f ! s Iba,iuJiI$
i, the indices taking all
values. I t also is assumed that llFillz-siand l ~ C $ h l ( l - , . - l ~ p are all finite.
10.4. Positive constants r , and K , exist such that, if r 5 r , and the
THEOREM
are alljnite, then I I u ~ I I ~ + ~ , also i s j n i t e f o r j = 1, * * * , N , and
quantities IIuj

K , and r2 depend on p , k, I, E (Section 4), and the modulus of continuity o f t h e leading

coej5cients of the l i j .
The proof of this theorem is parallel to that of Theorem 15.1, ADN, pp.
702-703.
The preceding results enable us to derive, in boundary problems over arbitrary
domains, L, estimates for solutions not of compact support. We first consider
global estimates in, for simplicity, a finite domain 9. Let u j , j = 1, * ,N, denote -
a solution of equations (1.1) in 9 satisfying Complementing Boundary Conditions
(2.1) on the boundary&. With l,.= max (0, rh + I ) , let 1 denote a fixed integer
2 I,. We assume the boundary 9 to be of class C' in the following sense: 6 is
- - +
coverable by a finite number of ( ( n 1)-dimensional) open sets Upsuch that each
intersection Ufin 9 is the image, under a 1-1 mapping Tp, of the closure of a n
+
(n 1)-dimensional - hemisphere CRD ' R8 1, the flat face of the hemisphere
corresponding to Up n 9 ; each Tpand its inverse have continuous derivatives of
+
orders up to 1 t' bounded by a constant K.
78 S. AGMON, A. DOUGLIS AND L. NIRENBERG

In the coordinate system appropriate to each hemisphere C R s , a minor

constant Ap is determined by the definition of Section 2. As minor constant for
&, we define
A = A 6 = min As;
B
according to the remarks in Section 2,
A > 0.
Respecting the coefficients in the differential equations and the boundary
conditions, we assume
-
aii,p E Cz-'@), bhj,o E C'-'h(&);
and [bhj,olE,.h. For each i, we
k will denote a bound for the norms (aij,plF-8i
assume I I F J - = IIF~~
~ ~I I F ~ to be finite. Defining ~ ( ~ h ~ ~ ~ - $ =
. h -I lI ~+p~ I I ~ ~ ~as
.-,,~
the maximum of the corresponding norms over the flat faces of the C,,, we
assume these norms, for all h, too to be finite.

THEOREM 10.5. A constant K exists such that, if\\~~ll~,+~,

isjnite f o r j = 1, - * * ,N,
then 11 u j Ill+fj also i s j n f t e , and

K is dependent on p, 1, K , k, E (defined befre Lemma 4. l), the domain 9,and the modulus
of continuity of the leading coeflcients in the lij.

Remark. The term C ) ( u ~ on ( ( ~the right may be replaced by X ~ ~ u j \ ~ oor

,I,l
may be omitted altogether when the solution in question is the only one for which
u j has L,-derivatives of orders up to I, +
t j , j = 1, --
* , N . In either case, K

may have to be replaced by another constant.

The proof of this theorem, depending on a partition of unity, is like that of
Theorem 15.2, ADN, pp. 704-705, and thus need not be given.
COROLLARY TO THEOREM 10.5. Under the conditions of Theorem 10.5, if p
exceeds the dimension n + 1, so that a = 1 - ( n + I)/! > 0, then Iujlz-l+fj+x is
majorized by \lui11 l+tj,L,, and hence is majorized by the right side of (10.6).
Now we shall derive "local" boundary estimates applicable to subregions of
i? abutting 9. The estimates are formulated here for a hemisphere E R whose flat
face FR is part of 6.New norms are used, which we now introduce. Let
a = ( n + 1)(1 - l/p), and let q be a non-negative integer. For v E Hi,L,,(CR),
define

For a function 4 defined o n F R , which is the restriction toFR ofsome u E Hi,L,(CR),

 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 79

taking the g.1.b. for all w E HI,L,(CR)that equal 4 on FR.

It is worth noting that
- -
+
II I1j - 1 / p , o 5 max (1 Y Rjtqfa) II d II j-l/,,
where -
1I d 1I ,-l,tl = g.1.b. IIw 1
I j,L,,
W

an alternative norm for 4. The g.1.b. in the last definition again is taken for all
w E Hj,L,(CR)that equal 4 on FR.
Norms of the first kind above are subject to a basic inequality stated as follows:

LEMMA.If h > 0 and q > 0, there is a constant K depending only on h, n, p, q

such that, for 0 5 j < h,
(10.8) ja~41, s K(iah41,)"(Rqll4lL1(zRJ1-" + KR* I I 4 L 1 ( Z R b
where
0 = (j + a)/(h + a).
COROLLARY.For any E > 0 and j < h, a constant Kl of the sane &scription as
K exists such that
(10.8)' l18ullv IK1&jahuIla + K,(1 + E - " / ( ' - ~ ) ) RI ~~ U I I L ~ ~ ~ ~ ~ -
Proof of the Lemma: The inequality in question is based on the existence
of a constant k such that, in any hemisphere i2 of radius p ,
(10.9) llwL,(nl I14IZTp~,+ k
5 k( llah~IlL,~~da l141Ll~nl~-~-z.

This fact is deducible by a dimensional argument from the case p = 1, for which
reference is made to Remark 5 in Nirenberg [20],p. 126.
To deduce (10.8) from (IO.9), let d be fixed such that 0 < d < R and
I

( 10.10) II
dU+j+% I P j U J'X(CR-d, h4 liaju Il Q '

Cover CR-d by a finite number of hemispheres Ri of radius d/2 and contained in

CR-dIP;no more than a fixed number, independent of d and depending only on
the dimension n + 1, of the Qi are permitted to have non-empty intersection.
Applying (10.9) in each Q, and raising to the p-th power gives us

Next sum over i and apply Holder's inequality, the existence of a limitation on the
overlapping of the Ri , and the inequality C cp 2 (C c i ) P for ci 5 0 and p 2 1,
80 S. AGMON, A . DOUGLIS AND L. NIRENBERG

to obtain

Taking the P-th roots of both sides and multiplying by dq+3fa now gives us

dQ+i+aIIaju I l q , ( ~ ~S
- ~const-
) ( ~ a h ~ l l g ) " (lRl ~
g l l ~ , ~ z ~ const.
) ) ~ - ~R4 l + l ~ l l ~ ~ ( ~ ~ ) ~
the desired inequality (10.8) following from this and (10.10).
We now formulate an L , estimate over E R . As before, let 1 2 max (0, r, 11, +
and with this 1 assume the coefficients in the differential equations and the
boundary conditions to satisfy the same continuity requirements as in Theorem
10.5, but now only in X R or FR ,respectively. For each i and h, assume the norms
- -
II akFiI t'+Si, I l hII l-q-l,,,r+V$ = 0, * - - 1 - si,
9

to be finite. We then shall prove the following result:

THEOREM 10.6 (Local boundary estimate). Under the foregoing assumptions, ti

constant C exists such that, f o r j = 1, * * , N, -

1-8, -
(10.1 1) IIaz+t%llt'-t, Ic 2
i
2 lIakFillst+t~+ c 2h II+hllz-rh-l/ll.f~+7A
k=O
+ c 2j II%llL,(SR),
where C is a constant independent of u i, but inuoluing R, p, I, E, the modulus of continuity
o f the leading coe$icients in the lii , and bounds f o r certain derivatives of the other coeficients
in the ldjand the Bhj, in XR.
Proof: I n what follows, for any function a norm of integral index will refer
to the restriction of the function to C, and a norm of fractional index to the
restriction of the function to F R . C, , k = 1, 2, --
* , will denote constants of the

same description as was given C in the statement of the theorem above.

Let 53 denote any C" region containing C, such that the boundary of 9
contains F,, and consider any function u belonging to Hi+ti(L@)with support in
ZR-.e for some positive E . Theorem 10.5 applied to 53 provides the following
estimate :
2j II az+'~~jIIL,(~,)5 GCi II2j l i i u j IIz+~,,L,
(10.12)
+ GZ: IIZ h j
Bhjvj II z--T,,-l/o L, + C12 IIuj
j
II0,L;
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 81
-
Let the largest of the quantities ll#+tjujllt,-tj be that with the index j = 1,
and let d be such that 0 < d R and <
(10.13) dz+t'+" IIaz+stlul[ I L , ( n R- d ) L_ 4 I(a2+'1u1II

Let < be a C" function in ZIR-dsuch that 5 = 1 for 1x1 5 R - d, 5 = 0 for

1x1 2 R - 4 2 , and 1851 2 const./dj, the latter constant depending of course o n j .
We shall make the substitution
v j = 5ui

in (10.12). Note that, the bhj, denoting known functions,

-- +
(C, might be replaced by C,(1 + R + * R"'A), where C, is independent of
R.) If w, is any function in Hj,L(,xK,equalling 4h on F , , we have furthermore
I-r,
II ~+hlll-rh-l/i3,Lp III5WhII I-rh,L, <
= c3
I;-0
2 dkfrh-t ~~akWA1~Lp(C~-~,r)'

From the foregoing, therefore,

< C4d-l-t'-z
I I I B h j ( b j ) II t-r,,-l/v,L, = j x=o
j

Both these results are used in (10.12) (with u j = 5u3), and then (10.8)' is applied.
Multiplying the result by dl+t'+uwe are thus led to inequalities of the form
82 S. AGMON, A. DOUGLIS AND L. NIRENBERG

The first member of the latter inequality is estimated by (10.13) and the last
summation in the third member by N ~ 8 t t l u llt,-t,.
l Choosing E = 1/4 C, we now
deduce the desired inequality (10.1 1) f o r j = 1 ; by the maximizing property of
this value of the index, the inequality immediately follows for all values ofj.
10.3. Interior Schauder estimates improved. In an arbitrary domain
9, consider a system of differential equations of the form (1.1) with FiE Ca-',
and a,j,pE C*-*1 (Section 9), where 0 < u < 1 ; k will denote a bound for the
norms lai,,pla-s, over 9.The burden of the following theorem is that generalized
solutions of these equations are strict solutions.
THEOREM 10.7. Under the above hypotheses, suppose u j has L,-derivatives in 9
of orders up to t i , j = 1, * , N, and ul, * * , u," satidy equations (1.1) in almost
everywhere. Then u, E Ctjfa, andfor every compact subdomain d of 9,
(10.14)

The constant depends on n, N, u, t' (Section 4), A (Section l ) , k , and on the minimum
distance between d and 6.
The facts stated in this theorem already were indicated briefly in a footnote in
ADN, p. 722. The theorem is proved by applying results of Sobolev, Calderon
and Zygmund, and E. Hopf to integral representations of u j and its derivatives in
a step by step procedure analogous to that of Appendix 5, ADN, pp. 7 19-722.
 CHAPTER V

APPLICATIONS OF ESTIMATES AND COMMENTS

Using the Schauder or L , estimates one may derive many properties of

solutions of our boundary value problem as in ADN, Chapter IV. For instance
one proves that the solutions are of class C" in the closure of the domain provided
this is true of the coefficients and the given data. Furthermore one sees that the
solutions are compact, that the space of solutions of the homogeneous problems are
finite dimensional, and that the differential operator has closed range. The
argument in Browder ([6] Theorem 4) may also be extended to show that the
range of the operator has finite codimension. Many of the results of Section 12 of
ADN can also be extended; we mention that a simpler proof of Theorem 12.8
there (in fact of a more general result) has been given by Agmon, Nirenberg (21,
Theorem 5.2. We shall merely present a few analogous results.

11. Necessity of Complementing Boundary Conditions and of

Supplementary Condition on L
In deriving the estimates of Chapters I11 and IV, we assumed the Comple-
menting Boundary Conditions (Section 2) and also the Supplementary Condition
on L (Section 1). These conditions are essential, at least in a sense we now describe.
Consider equations (1.1) in a hemisphere c, with boundary conditions (2.1)
relating to the flat face of X R , smoothness assumptions being made, say, as in
Theorem 9.2. When the Complementing Boundary Conditions and the Supple-
mentary Condition on L are valid, according to Theorem 9.2 a constant C exists
such that
2j I%lEo+tj+a 2 C E I2~isuilEo-s,+a+ ;Ih:I2j ~hiuj~t~--s,+a+ L:i IujIoI
i j

for any set ul, * * * , uN of infinitely differentiable, compactly supported functions

in C,. Our contention is that no such constant C exists when any one of the
Complementing Boundary Conditions, or the Supplementary Condition on L, is
violated. A similar statement can be made relative to the L , estimates of Section
10. These facts are easily proved by the method of ADN, p. 681, using the
functions (4.34)with u = I,.

12. Differentiability for Non-Linear Systems

. . , uh' satisfy a system of non-linear differential equations
Let ul, *

(12.1) Fi(P;u l , * * * , u A v ; 1,
.**;aSt+*lu aSi+*~u,~)=O, i = l , . * * , , V ,
3

in an (n + 1)-dimensional domain 9.We assume this system, for the given

83
84 S. AGMON, A. DOUGLIS AND L. NIRENBERG

solution, to be elliptic: the system of equations for first variations 6ui of the u j is
to be elliptic and, indeed, to fulfill all the requirements of Section 1. The latter
system of “variational” equations is, of course, linear with coefficients that
depend on the u j and their derivatives of orders up to si ti. +
Our first result is a n interior estimate.
THEOREM 12.1. Let FiE C1-Q with respect to all its arguments. If u i E Ctj in 9,
then, for any positive a’ < 1, u i E Ctj+n’ in 9.If FiE with 0 < u < 1 and
C1-si+z

with 1 an integer 2 I , then ui E C1+ti+“in 9.

The second part of the theorem follows from the first part and Theorem 5 of
[lo], p. 532. To prove the first part of the theorem, consider subdomains d and
33 o f 9 such that 2 c 33, 3 c 9,and let 5 denote a non-negative C“ function
= 1 in d and = 0 outside 33. We begin by deriving, from (12.1), equations for
difference quotients

then obtaining from these analogous relations for [ du, , functions with compact
support. If 33 is of sufficiently small diameter, Theorem 10.3 will furnish a bound
for II[du, llf,I,, with any p > 1, and, in particular, with p = ( n + 1)/(1 - a’).
This bound will apply to The bound being independent of A x k
(assumed sufficientlysmall) and ofk (andj), we easily conclude that u, E Hi,+
with ~ \ u ~ I I $ + ~ , ~ also
~ subject to the aforementioned bound. By a theorem of
Sobolev, it now follows that u j E C‘J’~’in d.Since the position o f d is arbitrary,
the same statement holds in 9, as contended.
The remaining results are concerned with differentiability of the u, near a
portion l’of the boundary of 9 on which m (possibly non-linear) conditions
(12.2) G , ( ~ ; ul, . . . , u,; . .. . arhitl ul, ... a r h i t.vu
*v) = Q
are satisfied. Not losing generality, we take to include a hemisphere C, and
take I’ as the hemisphere’s flat face. As in Section 4, coordinates x l , . * * , x , , t
r
are assumed chosen in which has the equation t = 0; then an arbitrary point
of E R is denoted by P = ( x , t ) , x representing the vector (xl, * * , x n ) . -
On I‘, the du, are subject to linear boundary conditions (the “variational”
boundary conditions) consistent with (12.2). These will be assumed to satisfy the
Complementing Boundary Conditions and other restrictions of Section 2. Two
statements then are made which, combined, yield a general theorem concerning
differentiability at the boundary in non-linear problems. The first statement we
formulate as
THEOREM 12.2. Let 0 < a < 1 and I , = max (0, rh)>and assume u , E C’ ‘1

in C,. Let 1 2 1, be a fixed integer, and assume F , E C1--si+xand G,, E C1-rh+’ with
respect to all arguments. Then u i E in C,.
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 85

The proof of this result is parallel to that of Theorem 1 1.1, ADN; only the case
1 > I,
is of interest. Theorem 9.2 first is applied to difference quotients tangent
to I?, these difference quotients satisfying a linear elliptic system in C,, and
Complementing Boundary Conditions on r,
that fulfill all the required assump-
tions. By a limiting procedure, the derivatives of the form azalo+t5uj are found to
belong to C". From the equations 8,' -'i+'0Fi = 0, as in an analogous argument in
the proof of Theorem 8.2, the pure t-derivatives a,t+tit 'uiare expressible, how-
ever, in terms of mixed and lower order derivatives known already to belong to C".
Hence, all the derivatives of u, of order 1, ti + +
1 are in C", and the theorem is
seen to hold for 1 = 1, + 1. Repeating the above argument proves it for any
given 1.
Our second statement as to differentiability near the boundary we formulate
in
THEOREM 12.3. Let 1, = max (0, rh +
I), and assume ui E C1l+tJ in 2 R'
Assume Fi E C1l-'i+l and Gh E: CZ1-*h+l with respect io all arguments. Then for every
positive a' < 1, u E Cll+tj+r'in C,.
To prove this, we first consider any set of difference quotients du,, * . ,duN -
tangent to the flat face r of C ,. For any sufficiently small, positive Y < R and
anyp > 1, reasoning as in the proofofTheorem 12.1 shows the norms ~~dui@+tj,Lu
to be bounded; therefore, a,uj E Hl,tti,L,in C,. Choosing! = (n 1)/(1 - a'), +
we thus have from Sobolev's lemma that a,alt+tj-lu, E C"' in &, and all that
remains is to show
a p uj E c"
in 2,. This is done, as at the end of the proof of Theorem 12.2, by expressing the
pure t-derivatives in question in terms of derivatives of the u, , and other functions,
now known to be in C".
We see, in particular, that if Fi and G , are of class C" in 5,then so is the
solution.
13. Perturbation of Non-Linear Problems
It is convenient to symbolize, say, f ( x ; yll, * * * ,yls; * - * ; yT1,* * ,xS)
by
f(x,yj,&=,, ... ,r,. B = l , . .., or, more briefly, just by f ( x , y i , J . Analogous notation
is used when b, for instance, denotes a multi-index.
L e t j be an index, and p and (T multi-indices, as in previous usage. For a
parameter r in a neighborhood of zero, let
~ i ~ ~ , ~ , ~ j , p ~ , p , = o , . . .j,=8l i; .+. f. X
,>; i= I,***,N,
and
G~(~,P,~j,,)/,/=~;..,r~+t~;j=l;..,~, h = l)*.*,rn,
be functions with continuous first and second derivatives with respect to r
+ a Czo-rht ", respectively, with
belonging, with these derivatives, to C z ~ - s ~ and
respect to the other arguments. (Again, m = +(Xsi +
C tj) is assumed to be an
86 S. AGMON, A. DOUGLIS AND L. NIRENBERG

integer.) We shall consider the problem of solving the system of differential

equations
(13.1) P, a P u j )
Fi(7, = 0, i = l,*..,N,
in an ( n + 1)-dimensional region 9 under the boundary conditions
(13.2) Gh(7, P, =0 on 9, h =l,..*,m.
For P = 0 a solution up' E will be assumed known; it will be taken as
ujO) = 0.The question as to whether a solution exists for any P in a neighborhood
of 0 is made to depend on the properties of the linear problem

(13.3)

THEOREM 13.1. For P = 0, let u j 0 solve (13.1), (13.2). Assume the linear
problem (1 3.3) to satigy all the conditions of Theorem 9.3 and to possess a unique solution
(wl, * * * , wN),w j belonging to C'o+tj+'*,for arbitrary fi E C z ~ - s i + aand #h E Cl~-~h+".
-
Then for 171 suficientb small, problem (13.1), (13.2) has a unique solution ( u l , * * , uN)
with u j E C'O+tj+PL.
The proof is based on Schauder estimates and is similar to that of Theorem
12.6, ADN.
Remark. It should be noted that one can use L , estimates in place of the
Schauder estimates. This is because the product of any two functions belonging
to Hj,L, is again in Hj,Lpprovided p j > dimension. Thus one may prove a n
analogue of Theorem 13.1 for equations (1 3. l), ( 1 3.2), obtaining a solution u with
u j E Hl+t,,Lpprovided max (1, 1 - rh) > (n +
l)/p.

14. Schauder Estimates for Systems of Semi-Linear Equations

Let lij and Bhj denote linear differential operators defined, respectively, in,
and on the boundary of, a finite domain 9 and satisfying the conditions of
Theorem 9.3 with &' = 9, = 9, r 1 = 1,. Here we shall consider semi-linear
problems
2
lijuj = Fi in 9,
(14.1)
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 87

are permitted to depend non-linearly on the unknown functions and their deriv-
atives of the indicated orders. When the non-linearities are not severe, the
solutions of these problems are subject to estimates of Schauder type generalizing
those for a single semi-linear equation of order 2m (ADN, Section 13). These
estimates are to be obtained again by the method of Nagumo, who studied
such matters in connection with second order equations.
j = 1,
We suppose that, when any set of functions u j ( P ) E C‘O+~~+“, , N, - -
are inserted into Fiand $ h , the resulting functions of P, namely

satisfy conditions of the form

where (a) I is a finite sum of terms each of which is a finite product

with ajk < I, + t i + t ~ ,pjk 2 0, and X ( l , + t , + a)-I 2k a,kjrk < 1, and also
j
(b) J is a finite sum of similar products with ajk < I, + t j + a, j j k2 0, and with
2 ( I , + ti + cc)-lzajnpJK= 1 ; K is a constant, and p(S) is a fixed function de-
3 k
fined for positive S which is o(S) as S + a.
Under these hypotheses, we state
THEOREM 14.1. r f bounds are known for Iull,, - , luNIO, it ispossible to estimate
*

Iujllo+t,+a f o y j = 1, * * * N .
9

15. On the Representation of Solutions by Singular Integrals

The Poisson kernels of Section 4, which give integral representations in terms
of boundary data for the solutions of equations with constant coefficients in a
half-space, have been of obvious utility; this paper has been based upon them.
Analogous Poisson kernels for equations with variable coefficients also would be
useful, for instance in obtaining expressions for derivatives of a solution at the
boundary as linear combinations of derivatives of singular integrals involving the
boundary data. Such purposes are served almost as well, however, by “approxi-
mate” Poisson kernels that give, as singular integrals involving the prescribed
boundary data, not the actual solution u, but a function u which approximates
u in the sense that u - u is more regular than u or u individually might be. We
now indicate briefly how such “approximate” Poisson kernels might be obtained
in the case of equations with sufficiently differentiable leading coefficients and a
sufficiently smooth boundary. For brevity, we confine our discussion to a local
problem and to homogeneous equations.
Consider a solution of (1.1) and (2.1) with Fi = 0, and suppose that near a
point on the boundary, which we take to be the origin, the boundary has been
88 S. AGMON, A. DOUGLIS AND L. NIRENBERG

flattened and represented by the equation x , + ~ = t = 0, the domain lying on the

side t > 0. As before, we use the coordinates ( x , t ) , where x represents coordinates
in the plane t = 0. In some neighborhood C,: / x i 2 t 2 < R2, t 2 0, our system +
has the form (the summation convention is used)
f&, t ; a, a J u j = 0, J
i = l,-*.,N,
(15.1)
B d x ; a, Y at)u,l,=o = 4&)> h = l;.*,m.
For (xo, t o ) E C,, let Kih(x0,t o ; x, t ) be the Poisson kernels, as constructed in
Section 4, for problems with constant coefficients of the associated type
I ~ ~ ( x O t, o ; 8, for t > 0,
, a,) vi = o
(15.2)
Bij(xo;a,, i3,)Vj = y h ( x ) for t = 0,
with sufficiently differentiable yh(x) having compact support. These Poisson
kernels and their derivatives 8 K h j with respect to x , t are governed by inequalities
of the form (4.28)'. Moreover, if the coefficients in the operators & ( x , t ; a,, a,)
and BLj(x; a,, a,)
have derivatives of a sufficiently high order, the derivatives
of a corresponding order of 8 K h i ( x 0 ,t o ; x, t ) with respect to xoJ to exist and also
are subject to inequalities of this form. In what follows, we shall suppose the
number of such derivatives to be sufficient for our needs.
The solution of (15.2) is expressed in terms of the Poisson kernels as

(15.2)' V f ( X O , to; x, t ) =
s Kj*(XO,t o ; x --y, t)yJ&)

We apply this formula with yI,(x) = { ( x , O ) & ( x ) , where { ( x , t ) is a non-negative,

4.

infinitely differentiable function defined for t 0, vanishing outside the hemi-

sphere C, , and equalling one in a concentric hemisphere & , r < R. Then we
define
(15.2)" v j ( x J l) = v $ ( x J x, '), j = 1 , * . * ,N .
These ui are approximations to the u i of the desired kind; for, the differences
w i = v j - u j are subject, for instance, to inequalities of the form

( 1 5.3)
with constant C and with l/c = ( 1 +
l/n)/p, 1 being a n integer not less than zero
or 1 + max rh. (For the u j and the v j individually, derivatives of $h of orders up
to 1 - rh would be expected to appear.)
We prove (15.3) from Theorem 10.6, or rather from its consequence,
(wt'-l
11 aZ+liwjii L,,(xr,2) <
=C
2-8,

2 II
2i h!=o a"ij,Wj,II L,(x,)
(1 5.4) I

+ c, 2 Id
x
lIBhj'~j'11/7,-~,p,t'+l, + c, 2Y I l ~ i 4 L 1 ( X r ~ ~
in which a superscript attached to a norm indicates the region to which the norm
refers, and C, represents a constant expressible as a polynomial in r. C,, C,, and
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 89

C, below will denote constants of a similar description. Since Bhjwj= 0 for

1x1 5 r, t = 0, the second summation on the right-hand side of (15.4) is zero. To
estimate the terms in the first summation, we note that li,wl = lipj is expressible
in terms of the integral operators on the $h with kernels that, by (15.2), are less
singular than, say, the kernels occurring in the expressions for derivatives of v j
of order si + t j . For any possible differentiation ak,
aklijwjis similarly repre-
sentable in terms of singular integrals, and, moreover, the strength of the singular-
ities can be controlled to some extent by integrating by parts. From these remarks,
and in view of (5.28)’, we readily have, for instance, that in Xr

the inner summation extending over all derivatives of the indicated orders unless
k + si - rh - I < 0, in which case only the term for k’ = 0 appears. To (15.5)
applies the general inequality

the L, norm on the left, with 1 < 1< CO, extending over the half of the x, t-space
for which t > 0. Hence, for I 2 0, I 2 max rh 1, +

l-rh-1
5 c* k2
=O
II ak$h II LJZ,).
These inequalities together with the preceding considerations imply inequality
(15.3).
Inequalities comparable to (1 5.3) appraising lower order derivatives of the
w j in terms of lower order derivatives of the $h can be obtained by applying L ,
estimates for equations written in variational form such as have been discussed
in ADN.
When restricted to the boundary 1 = 0 the formulas (15.2)‘, (15.2)“ yield
expressions for the derivatives of the solution in terms of derivatives of the $h
via singular integral operators in the sense of Calderon and Zygmund [8].

Bibliography
[l] Agmon, S., Douglis, A., and Nirenberg, L., Estimates near the boundagv for solutions of elliptic
partial diffkrmtial equations satisfuing general boundary conditions, I , Comm. Pure Appl. Math.,
V O ~12,
. 1959, pp. 623-727.
[2] Agrnon, S., and Nirenberg, L., Properties of solutions of ordinary differential equations i n Banach
spaces, Comm. Pure Appl. Math., Vol. 16, 1963, pp. 121-239.
[3] Agranovich, M. S., and Dynin, A. S., General boundary problems for elliptic systems in multi-
dimensional domains, Doklady Akad. Nauk SSSR, Vol. 146, 1962, pp. 51 1-514. Soviet Math.,
Vol. 3, 1962, pp. 1323-1327.
90 S. AGMON, A. DOUGLIS AND L. NIRENBERG

[4] Atiyah, M. F., and Singer, I. M., The index of elliptic operators on compact manifolds, Bull. Amer.
Math. SOC.,Vol. 69, 1963, pp. 422-432.
[5] Bade, W. G., and Freeman, R. S., Closed extensions of the Laplace operator determined by a general
class of boundary conditions, Pacific J . Math., Vol. 12, 1962, pp. 355-410.
[6] Browder, F., Estimates and existence theoremsfor elliptic boundary value problems, Proc. Sat. Acad.
Sci. U.S.A., Vol. 45, 1959, pp. 365-372.
[7] Browder, F., A priori estimates for solutions f o r elliptic boundary nalue problems, Nederl. Akad.
Wetensch. Indag. Math., Vol. 22, 1960, pp. 145-159 and pp. 160-169. Ibid, Vol. 23, 1961,
pp. 404-410.
[8] Calderon, A, P., and Zygmund, A,, Singular integral operators and diyerential equations, Amer.
J. Math., Vol. 79, 1957, pp. 901-921.
[9] Cordes, H. O., The basic a priori estimates for elliptic systems of arbitrary order, Lecture Notes,
Univ. of California, Berkeley, 1962.
[ 101 Doughs, A., and Nirenberg, L., Interior estimates for elliptic systems of partial diyerential equations,
Comm. Pure Appl. Math., Vol. 8, 1955, pp. 503-538.
[ l 13 Fife, P., Schauder estimates under incomplete Holder continuity asrumptions, Pacific J . Math., to ap-
pear.
[ 121 Hormander, L., Linear Partial Differential Operators, Springer, Berlin, 1963.
[13] Ince, E. L., Ordinary Dzyerential Equations, Dover Publications, New York, 1944.
[14] John, F., Plane Waves and Spherical A4eans Applied to Partial Dgerential Equations, Interscience
Publishers, Sew York, 1955.
[15] Lions, J. L., and Magenes, E., Problemi ai limiti non omogenei. Problemes a m limites non homogenes,
I , Ann. Scuola Norm. Sup. Pisa, Vol. 14, 1960, pp. 269-308. I l , Ann. Inst. Fourier, Vol. 11,
1961, pp. 137-178. III, I V , Ann. Scuola Norm. Sup. Pisa, Vol. 15, 1961, pp. 39-101 and
311-326. V, ibid, Vol. 16, 1962, pp. 1-44.
[161 Lopatinskii, Ya. B., On a method of reducing boundary problems f o r a system of dz&ential equations
of elliptic trpe to regular equations, Ukrain. Mat. Z., Vol. 5, 1953, pp. 123-151.
[17] Miranda, C., Su un problema di Dirichlet per le eguazione fortamente ellittiche a coe@cienti con-
stanti in un dominio a frontiera degenere, Bull. SOC.Roy. Sci. Liege, Vol. 31, 1962, pp. 614-
636.
[ 181 Miranda, C., Teorema di unicita in domini non limitati e teoremi di Liouville per le soluzioni de prob-
lemi a1 contorno relativi alle equarioniellittiche, Ann. Mat. Pur. Appl., Val. 59, 1962, pp. 189-212.
[19] Nirenberg, L., Remarks on strongly elliptic partial diyerential equations, Comm. Pure Appl. Math.,
V O ~8,. 1955, pp. 648-674.
[20] Nirenberg, L., On elliptic partial dij'erential equations, Ann. Scuola Norm. Sup. Pisa, Vol. 13,
1959. pp. 115-162.
[21] Peetre, J., Elliptic partial diffirential equations of higher order, Univ. of Maryland, Inst. Fluid
Dynamics Appl. Math., Lecture Series So. 40, 1962.
[22] Schechter, M., On L, estimates and regularity, Amer. J. Math., to appear.
[23] Taibleson, M., Smoothness and differentiability conditions f o r functions and distributions in En, Thesis,
Univ. of Chicago, 1962.
[24] Volevich, L. R., On the theory of boundary valueproblems for general ellijticsystems, Doklady Akad.
Nauk SSSR, Vol. 148, 1963, pp. 489-492. English translation, Soviet Math., Vol. 4, 1963,
pp. 97-100.
[25] Volevich, L. R., A problem of linear programming arising in daxerential equations, Uspekhi Mat.
Sauk, Vol. 18, No. 3, 1963, pp. 155-162.

Supplementary Bibliography
From the Russian translation of ADN by L. R. Volevich, Moscow, 1962.
1 * Agmon, S., The L , approach to the Dirirhlet problem, Ann. Scuola Norm. Sup. Pisa, Vol. 13,
1959, pp. 405-448.
 BOUNDARY ESTIMATES FOR ELLIPTIC EQUATIONS 91

2* Agmon, S., The coerciveness problem for integro-diyerential forms, J . Analyse Math., Vol. 6, 1958,
pp. 183-223.
3* Agronovich, M. S., On questions concerning the index of elliptic operators, Doklady Akad. Nauk
SSSR, Vol. 142, No. 5, 1962, pp. 983-985. English translation, Soviet Math., Vol. 3, 1962,
pp. 194-197.
4* Boyarskii, B. V., On thefitst boundary problem for a system of elliptic equations of second order in the
plane, Bull. Acad. Polon. Sci., Ser. Sci. Math. Astr. Phys., Vol. 7, 1959, pp. 565-570.
5* Boyarskii, B. V., On Dirichlet's problem for elliptic systems of equations in space, Bull. Acad. Polon.
Sci., Ser. Sci. Math. Astr. Phys., Vol. 8, 1960, pp. 19-23.
6* Calderon, A. P., and Zygmund, A., Local properties of solutions of ellipticpartial diyerential equations,
Studia Math., Vol. 20, 1961, pp. 171-225.
7* Dynin, A. S., Singular operators of arbitrary order on manifolds, Doklady Akad. Nauk SSSR,
Vol. 141, No. 1, 1961, pp. 21-23. English translation, Soviet Math., Vol. 2, 1961, pp. 1375-
1377.
8* Dynin, A. S., Multidimensional elliptic boundary problem with one unknown function, Doklady Akad.
Sauk SSSR, Vol. 141, No. 2, 1961, pp. 285-287. English translation, Soviet Math., Vol. 2,
1961, pp. 1431-1433.
9' Dynin, A. S., Fredholm elliptic operators on manifolds, Uspekhi Mat. Nauk, Vol. 17, No. 2, 1962,
pp. 194-195.
10' Gagliardo, E., Caratterizrnzioni delle tracce sulla frontiera relative ad alcune classi di funzioni in N
variabili, Rend. Sem. Mat. Univ. Padova, Vol. 27, 1957, pp. 284-305.
l l * Gel'fand, I. M., Some questions of anabsis and dzyerential equations, Uspekhi Mat. Nauk, Vol. 14,
90.3, 1959, PP. 3-20.
12' ' Gel'fand, I. M-.,On elliptic equations, Uspekhi Mat. Nauk, Vol. 15, No. 3, 1960, pp. 121-132.
13* Gel'fand, I. M., and Shilov, G. E., Generalized Functions and Operations on Them, (30s. Izd.
Fiz.-Mat. Lit., Moscow, 1958.
14* Gel'fand, I. M., Petrovskii, I. G., and Shilov, G. E., Theory of systems of partial d@erential
equations, Trudy Vsesoyuznogo Mat. S'ezda, Vol. 111, 1956, pp. 65-72, Moscow, 1958.
15* Mikhlin, S. G., On the theory of multi-dimensional singular integral equations, Vestnik Leningrad
Univ., No. 1, 1956, pp. 3-24.
16* Mikhlin, S. G., Singular integrals in L,, Doklady Akad. Nauk SSSR, Vol. 117, No. 1, 1957,
pp. 28-3 1.
17* Mikhlin, S. G., Remarks on thesolution of integral equations, Doklady Akad. Nauk SSSR, Vol. 131,
1960, pp. 1019-1021.
18* Nikolskii, S. M., Concerning theorems on the imbedding, continuation, and approximation of dzjerentiable
functions of several variables, Uspekhi Mat. Nauk, Vol. 16, No. 5, 1961, pp. 63-114.
19* Peetre, J., Another ap&oach to elliptic boundary problems, Comm. Pure Appl. Math., Vol. 14,
1961, pp. 711-731.
20' Schechter, M., Remarks on elliptic boundary ualue problems, Comm. Pure Appl. Math., Vol. 12,
1959, pp. 561-578.
2 I * Schechter, M., Mixed boundary problems for general elliptic equations, Comm. Pure Appl. Math.,
Vol. 13, 1960, pp. 183-201.
22* Schechter, M., Elliptic problems in which the boundary conditions do not form a normal set, Bull.
Amer. Math. Soc., Vol. 66, 1960, pp. 84-85.
23* Schechter, M., V&ous types of boundary conditionsfor elliptic equations, Comm. Pure Appl. Math.,
Vol. 13, 1960, pp. 407-425.
24* Schechter, M., Negative norms and boundary problems, Ann. of Math., Vol. 72, 1960, pp. 581-593.
25* Schechter, M., Observations concerning a paper of Peetre, Comm. Pure Appl. Math., Vol. 14,
1961, pp. 733-736.
26* Seeley, R. T., Singular integrals on compact manifolds, Amer. J. Math., Vol. 81, 1959, pp. 658-690.
27* Seeley, R. T., Regularization of singular integral operators on compact manifoldss,Amer. J. Math.,
Vol. 83, 1961, pp. 265-275.
92 S. AGMON, A. DOUGLIS AND L. NIRENBERG

28* Seeley, R. T., On boundary value problems f o r Laplace's equation, Amer. J. Math., Vol. 83, 1961,
pp. 698-708.
29* Shapiro, Z. Ya., On genzral boundary problems f o r elliptic eguations, Izv. Akad. Nauk SSSR, Ser.
Mat., Vol. 17, 1953, pp. 539-562.
30* Slobodetskii, L. N., Estimates in L , f o r the solution of elliptic systems, Doklady Akad. Nauk. SSSR,
Vol. 123, 1958, pp. 616-619.
31 * Slobodetskii, L. N., Estimates in L , for the solutions of linear ellipticandparabolicsystems, I , Vestnik
Leningrad Univ., Ser. Mat. Mech. Astr., No. 7, 1960, pp. 28-47.
32* Slobodetskii, L. N., SoboleL' spaces of fractional order and their application to boundary problems f o r
partial diJerential equations, Doklady Akad. Nauk SSSR, Vol. 118, No. 2, 1958, pp. 243-246.
33* Slobodetskii, L. N., Generalized spaces of Sobolev and their application to boundary problems f o r partial
differential equations, Scientific Notes of the Gertsen Leningrad State Pedagogical Institute, 197,
1958, pp. 54-1 12.
34* Sobolev, S. L., Some Applications of Functional Analysis to Mathematical Physics, Leningrad, 1950.
35* Stein, E. M., The characterisation of functions arising aspotentials, Rull. Amer. Math. SOC.,Vol. 67,
1961, pp. 102-104.
36* Titchmarsh, E., Introduction to the Theory of Fourier Integrals, The Clarendon Press, Oxford, 1937.
37* Vishik, M. I., On general boundary problems f o r elliptic diJerentia1 equations, Trudy Moskov. Mat.
Obshch., Vol. 1, 1952, pp. 187-246.
38* Volpert, A. I., On the index and normal solutions of boundary problems f o r an elliptic system of dty-
ferential equations in theplane, Trudy Moskov. Mat. Obshch., Vol. 10, 1961, pp. 41-87.
39* Volpert, A. I., On the index in boundary problems f o r systems of harmonic functions in three independent
variables. Doklady Akad. Nauk SSSR, Vol. 133. No. 1, pp. 13-15.

Bibliography Added in Proof

I * * Agranovich, M. S., Volevich, L. R., and Dvnin, A. S., Solvability of zeneral boundary value
problems f o r elliptic systems in domains of arbitrary dimension, Joint Soviet-American Symposium
on Partial Differential Equations, Novosibirsk, August, 1963, pp. 3-9.
2** Agranovich, M. S., and Vishek, M. I., Elliptic boundap value problems depending on a parameter.
Doklady Akad. Nauk SSSR, Vol. 149, No. 2, 1963, pp. 223-226. English translation, Soviet
Math., Vol. 4, 1963, pp. 325-329.
3** Bermansky, Yu. M., Krein, S. G., and Roitberg, Ya. A., A theorem on homeomorphisms and
local increase in smoothness up to the boundary of solutions to elliptic equations, Doklady Akad. Nauk
SSSR, Vol. 148, 1963, pp. 745-748. English translation, Soviet Math., Vol. 4, 1963, pp. 152-
155.
4* * Roitberg, Ja. A., and Sheftel, 2. G., Energy inequalities f o r elliptic operators with discontinuous
roejicients and f o r general boundary conditions and conjugation conditions, Doklady Akad. Nauk SSSR,
Vol. 148, 1963, pp. 531-533. English translation, Soviet Math., Vol. 4, 1963, pp. 141-143.,
5** Roitberg, Ja. X., and Sheftel, 2. G., General boundary value problems f o r elliptic equations with
discontinuous coefiients, Doklady Akad. Nauk SSSR, Vol. 148, 1963, pp. 1034-1037. English
translation, Soviet Math., Vol. 4, 1963, pp. 231-234.
6** Sheftel, Z.G., Estimates in L, of solutions of elliptic equations with discontinuous coescients and satis-
Lying general boundary conditions and conjugary conditions, Doklady Akad. Nauk SSSR., Vol. 149,
1963, pp. 48-51. English translation, Soviet Math., Vol. 4, 1963, pp. 321-324.
7** Solonnikov, V. A., Estimates for solutions o f general boundaiy ralue problems f o r elliptic systems
Doklady Akad. Nauk. SSSR, Vol. 151, 1963. pp. 783-785. English translation, Soviet Math.,
Vol. 4, 1963, pp. 1089-1091.

Received May, 1963.