ON ELLIPTIO PARTIAL DIPPERBNTIAL EQUATIONS

by L. NIRENBERG (New Pork) (4

Outline.

This series of lectures will to~icllon a 11111nberof topics in the theory

of elliptii: tlifferential eqiiatioris. 111 Lecture I we tliscuss the fundamental
solution for equations with collsta~~t coefficients. Lecture 2 is concerned
with Calculux ir~eqna~lities i~lcludi~lgtlre well kuomli ones of Sobolev. I n le-
o t ~ ~ r e3s and 4 we preserlt tlle Hilbert space approach to the Dirichlet pro-
blern for strongly elliptic systems, alrd describe various il~eqoalities.Lectures
5 and 6 comprise a self contained proof of the well k11ow11fact that < weaku
solutions of elliptic equations with sufficiently t smooth >> coefficients are
classical solut'ions.
In Lectures 7 and 8 we describe some work of Agmon, Douglis,
Nirenberg [14] collcerni~~gestimates ]]ear the bou~ldsry for solutions of
elliptic eqiietions satisfying boulldary conditions. This work is based 011
explicit forlnulas, given by Poisson kernels, for solutions of homogeneoils
elliptic equation with consbdut coefficiellts in a half space.
Tlrronghout, for sin~plicity we treat one equation in one unknown.
The material will on the whole be self contained, though of course not
a11 proofs can be included. However, we shall attempt to indicate those
of the main results.

r) Qnesto oiolo di conferenze B stato tennto a Pisa dtll loa1 10 sottembre 1958, e
ha fatto parte del oorso del C. I. W. E. olie ha tlvnto per tema : t I l prinoipio di minimo
e sue appliaazioni alle eqnazioni fnneiouali )). Tale oorso si B svolto in oollaborazione con
l a Sonola Narmale Snperiore e lJIstitnto Mate~natico dell7Universitil di Pisa. I n qnesti
Annali saranno snooessivamente pnbblioati i oorsi d i oonferenze tennti dai professori
C, B. Morrey e L. Bere,
 L. NIUENBERB
: On elliptic partial

Lecture I. !llhe Funda~nental8olution.

I would like to start wit11 a few geueral and somewbat unrelated
colnments. I n studying differential equations one is usually interested in
obtiil~ii~guvbique solutions by imposing suitable boundary or ir~itialcol~di-
tions, the kind depending on the so - called <type)> of the equation - elliptic,
i~yperbolic,etc. However, the type classification for general equations has
not been carried out, and ill ttlrt~~yciises it is not known what bo~iudary
conditions to impose. Indeed for equations that change type - a ~ we d
are a11 familliar with the initial work in this field due to Professor Tricomi -
the nature of the boundary conditions is far from obvious.
Thus if one considers an arbitrary equiltion witboilt regard to type it
is s natural question to ask whether there exist solutio~lsa t all. In fast
there are occl~sions when one simply wants some solutions. Such occur
often in differentjal geometry. Take a well know~l case: to introduce
isothermal coordinates with respect to a given Riemanuean lrretric OII i\
two diluensiollal manifold. This reduces t o . a local probleln of finding
nontrivial solutions of a differential equatiou i l l s8neighb[~l.l~oodof a point.
Auothel question is: are there solutions in the large of a given
equation. For the preceding this is answered by nniformization theory for
Riemann surfaces.
I n this t d k we will consider for some special cases the question: For
a given differential operator L are there solt~tionsof Ltc =f for << well
behaved B fiunctions f . Of course equations with analytic coefficients nlwi~ys
have local solutions, obtained for instance by power series expitnsions
(Cauchy-Kowilewski).
Recelltly Hans Lewy [1] exhibited an equatiou with Cw coefficieuts
having no solutions ewen locally. Since it is easy to describe, we present i t :
In 3-space with coordinates x, y , t , set x = x + i y , write the Cauchy-
Rieml~nnoperator as 7= -
d
dx
l d
2dm
(- + ) , and consider the differential

equation

where the right hand side is a continnous real fuuction of t alone which,
for couvel~ier~ce, is writbell as 21. derivirti~eof a real fi~llctiol~ y.
THEOREM:If there is a co~rti~auotcslydjfe~enticcble solutiolc $1 of tile
eqcifctio~bis tr vteiylrbovliood 07' 1 1 oviyij~,
~ / h , e ~y ( t ) i s real analytic.
 differential equations

Thus for ;illy ion-a~lalytic y there is no solutio~lnear the origin. (The

proof lnay be erlsily modified to show tllat there are also no a generalized
solrltions P).
dl4
I'roof: If we integrate --d 0 over a circle 1 z 12 = s 2 0 , z = 8112 egg,
dz

Now set l = s + i t and I

U ( l ) = x u (10
over the circle we fiud that U satisfies
. Integrating the equation for zl

+
It follows tlriit V(l) = U 2 n i y is a l i o l o ~ ~ ~ o r lfunction
~hic of C = s i t+
in H donlt~i~l > .
ilettr the origin with re 5 = s 0 But 011 s = 0 the function
U , i. e. the real part of V , vairishes, and therefore V can be co~ltinued
analj tically across s = 0 . Hence y is analytic.
1;) [I] Lewy also constructs t~ furlatinu F s~lcll that tlre equation
LU= F h t ~ s 110 a smooth u solutioo in the ileigllborhood of airy p o i ~ ~ t .
Lewy also coujectures t h i ~ tthere are l ~ o ~ ~ ~ o g v equatio~ls
~ ~ e o a e wit11 CCQcoeffi-
cients Iraviug IIO solutiolls in the ileighborhood of any ~)oiut.
The simplest cle,ss of differential ope~.atorsL of arbitrary type, for
which one tnigbt expect s o l u t i o ~ ~us of

to exist, for all well behaved fi~nctions f , are operators with constaut
coefficientn. 111 the last few years a convideri~blestudy has been made of
general differential operators with conskaut coefficie~its,(See Ebrenpreis [2],
Hijrniairder [3], Milgrailge [4]. Solntious of (1.1) cliu be found, at least
locqlly, if olle lrnows that i~ f~~udameukalsolutiori E of L 1= d (the Dirac
8 ftu~ction)exists. This is a (possibly generalized) function E such that
for all C" functions u with compact sapport. We shall denote the class
of such functions by Cr. Here x denotes convolution. Then i f f is in C ?
the function u = E x f is a solution of (1.1).
Malgrange [4] and Ehrenpreis [2] proved the existence of a fundamental
solution for any differential operator with constant coefficients. However it
is not difficult to construct one explicitily, as Hormander, and also Trbves [5],
have shown, i~iidwe shall now describe such a colistruction.
First we fix our
NOTATION : We consider functions u (a) of n variables x = (xi x,),... ,
and denote the differentiation vector by D = (D, ,...
,D,) Di= d/dsi. ,
, , ,
The letters j y ,u v will denote vectors j = (fi, ,... ,
j,) with lion-negative
integral coefficients Pi, and we set I ,8 I = 2 gj . Otherwise for ally vector
5 = (ti,... , ,
tN)1 5 1 will represent its Euclidean length 1 5 l2 = 2 I ti 12, and
E.)7=2&qi. We write

. B
, B
D=D: ... D , , ;
fit,

for conveliience me shiill also, on occasion, express a gel~erwlmth order

partial derivative of B fi~nctiollu by Dn'u. C r will deuote the class of Om
functions with compact support.
We consider now a, differential operator L of order k with consti~nt
coefficients, which me may write as a polynomial ill D of order k .

I n constructing the fundamental solution let us first argue in a heuristic

manner. Introduce the Fourier transform of the function u (x)

J
u (E) = e - & ' h (x)d x ,
-
integration heillg over the eutiro n-space. Tl~eli

Soifu=ErrLu=
J E ( d - y ) L u ( y ) d y then
 diferential equations

Problem : give formula (1.2) a msanilzg.

In attempting to do this (and there are many ways) there a~retwo
difficulties that occur. The first is the son-integrability at iufilrity, duo to
the fact that we are integrating over the full %-space. The s e c o ~ ~difficulty
d
is caused by tlre real roots 5 of the polynomiel L (i5).
The first difficulty is easily overcome. It essentitllly expresses the filct
that is general E is a distxibution, i. e. a finite derivative of a coatiaoous
fiurctio~r. Instead of cor~structiagE directly we shall constrt~ctthe funda-
mental solution EN of the operator (I. - A)N L = (1- 2 D?)NL( L ) ) We
i
.
solution EN having continuous derivatives
shall const~~ucta funda~rte~~hal
up to any given order, by &king N sufficiently large. We m:~y the11 trke,
ill the distribution sense,

i. e. for f in Cr the function

is a solution of Lu =f.
Thus we consider, for p (t)= 1 + 2

Taking N large elirnirrates the first difficulty, i. e. the trouble a t ilrfil~it~y.

Xow to handle the second difficnlly. We ma.y assume, after a possible
rotation of coordinates, that the coefficiel~t of D: in L (D) is $. 0 , say
unity. Cousider L ( i t ) as a poly~romialin 5,. We shall first illbegrate i ~ t
(1.4) with respect to the variable E,, keeping E f = (El, ...,
E,,-l) fixed,
however we slr~ll 1nove the line of integration from the real line to a
pnrallel liue lying irt the colnplex t,, plane.
For fixed real 5' there are k roots 5, of L (i 5). 111 the strip
1
) 3 n b tflI < - ill the complex 5, plane there is therefore a line parallel to
2
the real axis whose distance from any root is a t least (2 k +
3)-1, as one
easily sees. Let us a choose one such line 31%5, = c(Y) whose distance to
RIIY +
root is a t least (4 k 4)-l . The clloice of c ( [ ' ) depends on but i t r,
is easy to see that c = c (5') may be chosen so as to be c o ~ ~ t i ~ l u oexcept
iis
on a set of 5' of (n. - 1)-di~nensioll~.lIneastlre zero.
Settiug tj = q (5') = (0 , ... ,
c (f')) me IIOW take a.s detieition

wl~ereintegration is first mit,b respect to [,, .

Since

me see that EN has derivatives up to ally give11 order, if N is 1~1,ge

enougl~.
W e have 611ally to verify that for u E 0;

,
Setting ( 1 - A)N L ( D )= LN( I ) ) the right: hand side equals

ei(=-~).(ESis)
LN(I(E iq)) + a E Ln ( 1 ) ) u (y) d Y .

Since zc has compact support its Fourier transform ;(E) cau be extended
to colnplex vectors E a8s an entire analytic fuuctiou, and since u Cm the
derivatives of die do,wll faster that any yon7er of I El as we go to infinity
I<
in a strip I 3 %5 ~ constant. Thus, interchangiag the order of iiltegratio~i
in the above, we find that it equals

Because of the behaviour of ; of infinity we may shift the line of inte-

gratiou of the E, parallel to itself aud dud that this expressiou

J
-
a
,

= ( 2.)-#I 6k.E u (5)d 5 = tc (a)

 differential equations

Thus. the fu~letion EN defined by (1.4)' is a fii~~cIarnent;~l

soliitiol~for the
operator L N . The desired fit~~d:lmeutalsolution of Lu then is given by (1.3).
One sees easily that the fni~tlamental solation EN given by (1.4)' has
exponential growth iu the x,, variable.
For fiirtl~er important work on fundament~l solutioi~sfor equatious
wit11 oo~rsttmtcoefticie~~ts
we refer to Hortrla~ider(61.
Co~~siderrow elliptic difleere~~tialoperators with eoastal~tcoefficieuts.
These are open~tors 1, wlrose leading part L' - consistiug of the terms
of l~igl~estorder - satisfy

L'(f)+ 0 for real (+ 0 .

We shall have need later of the fu~ldameiitalsolution for a I~olnoge-
neous elliptic operator wit11 cor~slautcoeffifioients, i. e. L' = 1, . For snelr,
of course, the f u ~ ~ d a n ~ e i solntio~~
~tal first ~coastructedby Herglotz is well
belraved at i~~finity.We shall use the following ford of it, given iu P.
John's book [7].

where i~~tegl.ation is over the full unit sphere with d o t as the elenlent of
area, q is ic non-negative ir~tegerof the same parity t t , i. e. q +n is even,
a11d the principal bnl.11~11of tlre logarithm is taker1 with the plane slit
along the uegative real axis.
From (1.5) me obtain as n special case, for L = A power, the followi~rg
idelltiby which is due to P. John and used extensively in [71, represen-
g 8 function in t e r m of plaue waves: For u iu 17;
t i ~ ~the

I n [7] John derives (1.6) from the kuomii expressioli for the fundamen-
tal solution for a power of the Laplacean, and then derives (1.5) from (1.6).
This Itlay be done as folloms. Suppose E ( x . . () satisfies

L K (0 - 6)= (x . 5)Qlog 3-
.E
, Z

then a fundamental solution of the operator L is given by

8 L. NIKLENBEL~G
: On elliptic partial

But sucl~n K is easily found. If we set rx: [ = then K(o) satisfies

(,"J
L (€) - K (o) = on log o/i ,

a solution of which is

with ck,, an appropriate constant. If we insert this into the above expres-
sion for the fundament,al solution of L we obtain the expression

which differs from (l.S).only by the term i~rvolvingc ~ , But~ . this term is
a polyno~sialof degree k - 21 wllic11 is therefore tl solution of L v = 0 ,
and so may be ignored.
It slro111d also be possible to derive (1.5) from the heuristic: formula
(1.2). (1.5) aserts that

n+q
is a fundamental solution for .the operator A ~ LLet. us attempt, to de-
rive this expression from the corresponding exl)ression of (1.2):

Arguiug heurisitically a,gain let us modify the expression by introduci~rg

polar coordinates in the 5 space

Then (1.8) becomes

(1.8)'
 diferential equations

Let us now write the heuristic expression

as a well defined contour integral

where the contour (? is a curve which goes from +

oo ill t,he complex Q
plaee, ellcircles t)he origin counterclockwise ant1 returns to +
co ;11o11g
the real axis, the breach for the logarithm is the sa~neas above, w,nd the
constaut o is chosen so thict

The expressio~~s(1.9)' may be eval~~ated explicity, aud on insertion

into (1.8)', yields the expression (1.7). We leave the calculation to t,he
reader.

Lecture 11. Calculus Inequalities.

A priori estimates play a central role in the theory of partial diffe-

rential equatior~s. They are of v~,riouskinds - pointwise estimates for
derivatives of solutions end their modulus of coatinuity, and estiwates of,
say,,l,, norms of solutio~ls a,nd their derivatives - alld it is 11atura1Iy
imporla'nt to ni~derstandthe relatiouships betweeu t,hese various estimates.
For instauce, the well ki~ownre'sults of Sobolev assert that if the m'tl~
order derivatives 1)"' u of a fullctiou 21 (x, , ...,
x,,) (with C0111~)il~t s~tpport)
are in L,, 1 r< < oo the11 lower order derivatives U i 26, j < wz belong to
,
L,,for some p or, if r is s ~ ~ f f i c i e ~ ~high,
t l y the Dj u are bounded and
satisfy a Holder condition with n certaiu exponent a .
Since we shall often make use of it, let 11s recall here the ~lotioliof
H ~ L D ECONTINUI'I'Y.
R A ful~ctionf ( x ) defined 011 iL set fi ill a Euclitlean
< < ,
space satisfies a, Holder conditiou there with expollent a , 0 a 1 if
 L. N ~ ~ ~ N N B:EOn
R Gelliptic partial

is finite. I t is Holder eo~ltinuoue( e x p o ~ ~ a) e ~in

~ t a d o n ~ a i lif~ i t satisfies a
Holder eolldition with expollent a in every eempact subset of the domain.
This lecture is co~laenled wil)l~c i ~ l c ~ ~ lilleqnelities
us relating integral
and l)oi~ltwise estimates of f i i ~ ~ e t i o ~al s~ ~their d clerivatives. The recent
ituportaut reanlt of de Giorgi [11] 011 the differelltiability of solutions
of regulnr vib~*iatio~lal l~roblelltsseems ill fitct to be basetl OII e caloulus ine-
quality assertil~gthat cellail1 ir~tegra~lestim;ltes imply Holder conti~ruity.
We shall eonsider f i i ~ ~ c t i o ~ l(s x defitled
) ill ~ c ~ t l i ~ ~ ~ e a aEuclidean
i o ~ r a l space
g Lp , and whose (1erivabives of order nt belong to L , ,
aud b e l o ~ ~ g i n to
, .
1< q r < CG W e shall present i~~terpolative i~lequalitiesfor the L p and
Holder norms [ 1, of derivatives Dj u 0 (j , < 911, for the maximal range
of p alld a . Our inequalities are a c o m b i ~ l a t i o ~of,~ a.rld include, those
usua.lly called of Sobolev type (wllich lloltl also for fi~actionalderivatives,
a r d of wl~icllnlily be found in [8]),illld fami-
and rather s t ~ ' : ~ i g b t , f o ~ . ~proofs
liar inter1)olative illequalities s u c l ~a s

~ , <2 constant M , . M~

. .
where illliis e u . b of the L, norms of the derivatives of order i of a
,
functiou u , i = 0 , 1 2 . The proofs use o~llyfirst principles and are ellti-
rely elemeataty. (No attempt will be made here to obtain best constants).
The inequalities is this sectio~t were pl.esented at the 111t'e Congress in
Edinbnrgl~August 1958, where we learned Lhat almost equivalent results
had also been proved by E. Ga,gliardo.
111 this Ieeture we shall iise tlre followillg
1
NOTATION : For - w < <
- w we defins the norms and se~l~inorms
P
121 1, for f n n c t ~ i o ~11~(x)
s defiued in a domain cD in 11-ctimensional spaces :
For p > O
[ u 1, = the L , norm of u in cD .

For p < 0 set s = [- n/p] , - a = s + tc/p and define

 . .
where e u . b is bake11 wit11 respect to all partial det~ivirtivesDS of order
,
s and over points in 0 .
W e define I D J u 1, a s the mnximatn of the I ,1 ilortrls of all j - t h order
derivatives of u .
W e ahall express o w resnlt for fii~~ctionsu detiued it1 the entire
a-space En. ~ x t e t l s i o n to otlrer dotnail~s will be described briefly in the
remtirks after the theoreel.
THEOREY:Let u belong to L, it,. ond its tlelivcctives qf order w ,
Dm , t ~ ,belong to L,. , 1 < q r , .
co For the devicutivez D j u 0 <j , 91t< ,
the following inequalities hold

(2.2) I ~j u 1, constant I Dn' u ; 1 / I( IFa,

where

for all a in the interval

, ,
(the constccnt depeltding only ow n 9 1 2 , j , q , r n ) , with the follo~aing
exceptional ccrses
< ,
1. If j = 0 , r ~8 n q = co ,then we mnke the additional nsstmirption
that either u tends to zero crt ittflrrity or u E 1;9 $IY some .fi,tite [> 0 .
< ,
2. If 1 r < w trnd 111 -j - vl/r is a won negative intege~thrw (2.2)
holds only for cs strti.sfying j/nh (a< 1 .
W e shall not give i~ cotr~plete proof of the lheore~n here but shall
indicate the meill steps. First some corntr~e~tts.
1. The value of is determined simply by d i m e l ~ s i o n ~a~lalysis.
l
2. For a = 1 the fact that u is coutained it1 Cq does not enter in
the estimate (2.9), and the estimate is equivaletlt to the results of Sobolev
(note tlrat 'we permit r to be unity).
3. That j/nh is the sma,llest possible value for n may be seen by
taking 16 = sin a xi 5' (8) where is in C; : For large 1 me have I u 14. =0(1),
I D j u l p = 0 (lj),I Dm u 1, = 0 (An$) where no 0 can be replaced by o .
4. I t will be clear from the proof that the result l~oldsalso for u
defined in a product domaiu

and hence for any domain that clin be mapped in a one-to-one wa,y onto
such a donlain by a sufficiently < nice >> mapping.
 : 0% elliptic partial
L. NIRICNB~RG

5. For a bounded domain (with t smooth % boundary) the result

holds if we add to the right slde of (2.1) the term

constant I tt .1;
for any >0. The col~~tiiuts then depelld also on the domain.
6. Similar estiu~ittes hold for tlie L, uorms of D j u on linear subs-
paces of lower dimension, for suitable p .
7. Si~r~ilar interpolatioa iaequalities also hold for fractional derivtl.
tives, but their proof is not so elementary.
The theoreni, ill its full geuerirlity should be useful in treating nonli-
uear problems. We mentioil in parf,iciilar tltat from (2.2) for a =jlnt ,
q = w it follows that the set of fur~ctious $4 mhich are bounded and have
derivatives of order nl, belotlgiug to L,. forlris a Banach Algebra. For r = 2
this is a ~ l l e dtlie Schauder ring.
The proof of ttie theorem is elemelrtary and contains in particular an
elementary proof for the Sobolev case a = 1. Iu order to prove (2.2) for
any given j one has o111y to prove i t for the extreme values of a ,j/w and
unity. (If Case 2 ltolds eome additioual remark has to be made.) For in
general there is a simple
<
Interpolation Lemma : if - oo 1<p <r <oo the@

where G i8 indepettdeut of u .
Tlie lemma is easily proved; for R > 0 it is merely the usual ii~terpo-
lati011 iueqnality for L, norms.
Let us tarn now to the proof of the theore~n, or a t lea.st to tile main
poir~ts. Consider first the Sobolev ease, a = 1. It saffices to coiisider the
case j = 0 , 111 = 1, frorri mhich the ge~teralresult ntay then be derived.
>
If r n (2.2) asserts that u satisfies :I oertaiu Bolder coi~dition,A I I ~a11 ele-
mentary proof due to Morrey has loug been k~toan.We shitll sketch it here
for functiorls defined ill a genet,al dorl~airi (i3.
De.fittition : A domaill (2) is said to Itave the s t r o ~ gcoue property if there
exist positive constauts I / , i;r~ida closed solid right sghel.ica1 colle V of fixed
a
opeuing aud height sue11 that ally points P , Q i n (the closure of q)with
are vertices of coues Vp, VQ lying in (i3 which tire courguent to V and
have the followi~~g property : the volume of the intersection of the sets:
Vp, VQ, alld the two spheres with contor8 P , Q w,nd radius P- Q 1 , is I
I
not less than R P - Q In.
We 11ow prove the assertion
If u hne first tlerivntiver in L,., v > a, is n dowtailc Q hanoing the strong
cone property, then for points P,Q i ~ r9 with L' - Q 5 a we h.ave
I I ,

wltere the constant depends only on a , 1 , V n nird

(From this follows easily an estimate for [a]
,
I-;
.,
1..
depending on the
domain).
I I
Proof: Set 8 = P - Q and let Sp(SQ) be the intersection of Vp ( VQ)
with the sphere about, P ( Q ) radius 8 . Set BP n .SQ= S If R is a poi~lt .
in IS we have, OII i~ltegratingwith respectt to R over S,

Volume of 8 . I z c ( P ) - w ( Q ) I 5
S
I Izc(P)-a(R)ltlR+

+I/ S
- u (Q) l d B -
r (4

Because of the strong cone property the left l~andside is not less than

The first term on the right may be estilrlated as follows. Introducing polar
coordinates Q , 7 , about P , where 9 is A unit vector, we fiud easily that
the first term in the right is bouuded by

(where d o is the element of area on the unit sphere, and d o is the ele-
ment of volume)
 L. NIRENBKRG
: 012 elliptic partial

by Holder7s inequality,
< constant . s
-
? (SP/ 1 lvd $)+*

A silnilar estilnate holds for the t e r ~ u

result follows.
.'
i
I
I
a (R)
- 16 Id
(0) 12 , and the

We retrlrr~i ~ o mto f i ~ a c t i o ~defined

~s in the fill1 11-space.
<
Soppose r vc. We shall prove a strouger for~nulatiollof (2.2), namely

IU,,,.
-
<---nlel?;.
r s-1
2 n-r i 8 5 ,
11-r

< <
For 1 r vc (2.4) follows from the special case r = 1 as o11e readily ,
verifies, by sir~lply agplyil~g the illequality for v = 1 to the fuuction
n-1
-
V = l u le-rr aald u~illgHiilderls illequality in a sniixable way. T h i ~ sit silf-
fices to prove (2.4) for the ease r = 1 .

We shall prove (2.4)' here for 28 = 3 . One sees easily that

where
I,
i
denotes illtegratiou i110~gtile ftdl lil~e tllroiigl~ x parallel to

the xi axis. Thus

Integrating with respect to xi the11 x z , and the11 x3 me find with the aid
of Scl~marz~sinequalitmy
and finally

that is, (2.4)'.

For general n the inequality i~ provetl in the same way wit11 the aid
of Hiilder's ineqru~lity.
Suppose fiually, for j = 0 , r t = 1, tllat = 98 ; tllis is the exceptional
case 2. We claim that

.
where the const,ant tlepends o~llyn , q i~nd p I t sllffices to show this for
large p and tltis is e;~sily done by rrppliyng (2.4)' to the fulrctioli
v = 1 u l p ( l - l i a ) , and using Hiildel.'s illeqltality i t 1 a jutlic'l ~ o ~1n11
i s tl~ler.
~ e ' us t now consider the otlter extreme case a = j / m . I t suffices to
,
cousider the cAse j = 1 w = 2 , the general case mtly then be proved by
iuduction on m . We clrrinl tlli~tthe followiug lrolds

1 1
2 1 1
(2.5) I D u l < e l Z l 2 u l 7 / u l : for
P-
-=-+-;
p 9 . q
l < q , v ~ m ,
-

with c nth absolute cottxtant. Incide~ltally, :IS. Utlgar poillted ont, we ala,y
pennit q to be any positive nurnl)er, 1)ut I sllall cotrtitle rnyself to the case
< < .
cited, in fact to the case q finite, 1 r 00 The gelleral case may be
obtsiued by a slightly difft$re~lti ~ r g ~ ~ ~ t lor
e l ljnst
t , by lethitrg q tel~dto a,
and 1. teud to 1 or m ill (2.6).
Inequality (2.5) follows fro111 the corrisl~ot~dinginequality in one
dimension

which ltolds for the full, or Italf-infinite line (with c an absolute constant),
by iutegrating with respect to the other vt~riablesand appliyng Hiilder's
inequality.
Our proof of (?.6), though elementary, is slightly tricky. Peter Ungar
has found another slightly longer proof wliich furnishes a better value for c.
The proof is bssecl on }L sinrpla lemma wllicll we leave as a,n
exercise.
 L. NIRENB~RG
: On elliptie partial

Oa an intevval 1, whose lenglbt we also derrote by 1, we have

LEMMA:

wit8 o
an absolute co~tstnnt.
We s b l l prove tlrat for any interval L : 0 < x L the following
inequality holds

(2.6) follows easily from (2.8).

I n proving (2.8) we may suppose that 1 ttx, ,1 = I . We shall cover the
... ,
interval L by a finite nurnber of successive intervals A,, 12, each one
having as initial poilrt the ead poiirt of tlre preceditrg. For k a fixed yosi-
L
tive integer, clroose first the i~rtervel1:0 (XI-,
aild consider (2.7) for
k
this ieterval. If the first term OII the right of (2.7) is greaber thiun the
second set 1, = 1; we then have

since I urn ,1 = 1. If however tlre seco~rdt e r a of (2.7) is the greater extend

the interval 1 (keeping its left endpoint fixed) 1111til t l ~ etwo terms of the
right of (2.7) becotne equal. Since 1 + 8, - Er > 0 eqnillity of these two
terms nils st occnr for a finite value of 1. Let 1, be the resulting interval.
We thei hive

Starting a t the end poii~t of 1, repeat this process, keeping k fixed,

choosing A,, 1,, ...,until L is covered. There are cletbrly a t most lc such
r
iutervals 4.If we now sum our estimates for
J
hi
IP d r we fitrd, with
 differential equations

the aitl of Holder's illequality (recall that P -+

2r
P = 1) that
-
2p

If we uow let 16 - w tlte first term 011 the right of the preceding tends
to zero, bec:ruse > 1 , at111we obtain (2.8), completi~rgthe proof of (2.6).

Lecture 111. The Dirichlet Problem.

We co~~sider now elliptic tlifferential operators, confi~ringourselves for

simplicit,y to a single equw,tiolr for one unkr~own. Let L ( z , D) be a partial
differel~tiill ol~erator with co~tlplex valued coefficients, w.nd let L' be the
part of highest order. L is elliptic if there are no real clraracteristics, i. e;,

L'(x,fl+O, real E+O.

I t is ettsily see11 tlrat: for more tlrat~two vari~tbles,a > 2 , ellipticity implies
t11at tire order k of L is evetl. In treaitiug tile Dirichlet problem we shall
assume that k = 2 918 is eve]) and hirat the operi~toris strongly elliptic,
i. e. that (efter possibly mu1tiplying by a stlitable coinplex function)

The Diriolrlet problem consist,^ of finding B solution in a domain 9of

where a/& represeuts differentiation l~orlnalto the boundary. Here f and

Q,, are given functions in 9and 4 respectively.
We shall describe here the Hilbert space approach to the Dirichlet
problem, which is based on sotne form of the projectioll theorenl, and is
related to the classical method of minimizing the Dirichlet integral, In its
 L. N I ~ ~ N B ~ :C On
R Gelliptic partiat
0
preseut form the existellee theory is ~naiulydue to Garding, Vislrik, Browdela
and others ; we refer t l ~ ereirder to [9] aud [a]for expositions and refere~rces.
This and the followi~~g lecture comprise a brief description of [9]. The
0
theory is based on a sitrgle L2 inequality. Gardi~rg's inequality, expressi~~g
the positive defiuiteness of the Dirichlet integral associated with the diffe-
rential operator.
Since this approach to the Dirichlet problem requires cousiderable
differentiability assumptions on the coefficients we shall ~ssilmefor sim-
plicity that they are of class Om in cZ) and that the boundary d is
sufficiently smooth. We shall also assume Q to be bounded. Parthermore
if the Qj are sufficiently smooth we may subtract from u a furlction having
the same Dirichlet data as u, so we shall consider the case where the
Qj vanish
(3.3) Lu=f

The Hilbert space approach yields at first <generalized aolutior~so of

(3.3) which we ~nulrt define. A f u ~ ~ c t i oul ~which belongs, say, to 5, in
every compact subdomain of 0 is a u meako solutiou of Ltc =f if

for every wlrich belongs to C r (Q), i. e. is of the class Cm and has

,
compact support ill 0. Here ( ) denotes the L, scalar product, and LX
is the formal adjoint of L . I n addition to the L, norm we also introduce
0
the Hilbert spaces Hj(Hj), j a non-negative inleger. These are the closures
in the norm (using the notation of Lecture I)

of the spaces C m ( 9 ) (Q). The associated norm end spaces relative to a

subdom~ina will be denoted by 1 1 11. ,
a a ou
, .
H, Hi Clearly H,, =g=
L2.
We remark that for j i Hj c Hi and the set 11 u 1
> 1s
,constant is
compact in Hi.
Pollowiug Sobolev and Friedrichs we say that a fimction u in 9has
strong derivatives in 4 up to order j ia 00)if u belongs to H~(H:) for
every compact subdomain a of 9. With the aid of the results of the
 differential equations

preceding lecture we see tllat e fiulction ill Hj is conti~luousif 2 j a . >

Functions h 2,satisfy the bou~tdary conditions of (3.3) in a geaelxlized
sense.
We now formulate the
GENERALIZED DIRICHLETPROBLEM : #ioen j' in H,, fwd a weak so/ution
0
u in Hm of h2c= f .
Using the notation of Lecture 1 we may write the operator L in the
form
L = 2 Dfl ap,,Dy
lei, lulsm
.
0
If u is weak solution in H,,, we lnay then carry out some partial iute.
gration in equation (3.4) and write it as

,
B [ u , v] is linear in u antilinear in v and satisfies, by Schwarz~iuequality

We sl~all assulne the strong ellipticity (3.2) to hold uniformly, i. e.

for some positive constant c,

8.6 (- 1Irn 2 ab,, (8)tflP .>c, 1 8 l Z r n , 8 real,

IBI lul-m

for all 3 in (D. Our main result is

THEOREM:For ssl~,flcientlylarge the generalized Dirichlet proble?,, for
+
the equation (L )': 14 =f admit8 a unique solutio?z. For the eqztntion Lzc =f
we have the Wedholm alternative.
The .L, estimate on which t l ~ etheorem is based is
G ~ B D I N G ~INEQUALITY
S : There exist constants c >
0 and C such that

holds for every pl in C: ( 0 ) .

This ail1 be proved in the next lecture. I t is clear from (3.5) that tbe
0
inequality extends also to functions in Hm, and it follows from (3.6) that
0
+
the only solution in H, of (L 0)u = 0 is = 0 .
Let us now prove the theorem. Suppose first that the operator is
symmetric, i. e. B [pl , p] is real. and that the coastant 0 in (3.6) vanishes
+
- which we may achieve by consideriug L C in place of L . It follows
 : On elliptic partiat
L. NIRICNBERG

,
from (3.5), (3.6) (with C = 0) that B [u v] serves as an alternative scalar
product in the Hilbert space $,; the norms B [u u] and 1 u 1 , are eqai- ,
valeut. We see that the antilinear functional ( f , v ) defined for all y in in,
satisfies

and is therefore a bounded functional. By the well known represel~tation

0
theorem there exists therefore a function u in the Hilbert space H,, such that

'U i~ then the solution of the Dirichlet problem, and we have proved the
Brat part of the theorem with C! = C . To prove the secoud part we write
the equation Lu =f in the form (L C) = Cu f or + +

Since ( L C)-1 maps H, boundedly into k,, it is completely coutinuous

+
,
in H,, by a previous remark, aud from the Riesz theory for completely
continuous operators we derive the second part of the theorem.
,
Suppose now that B [ y y] is not symmetric. If we add C ( y y ) to B ,
so that it satisfies

then we may still rely on a generalized rep~~esentation

theorem due to Lax
alld Milgram. We conclude the lecture with this
REPRESENTATION ,
THEOREM : Let B ( x y) be tc jovm de$ned l o r pairs
of vector x 7 y in n Hilbert space H(uorn111 I/),
which is littear in a ,
,
antilinear in y and satisjies

Suppose that for some positive cottstant c the inequality

(3.8) I B (x 7 $1 l 2 c 11 x 112
holds for every x i s H. Then every bounded antilinear functional P ( x )
admits tile representation

F ( x )= B ( v , x) = H ( x , w ) .
For $xed e1ei)tents 11, ,w to11iclr art: u~iiqne.
 diferential equations

Proof: For any Bxed element v , B (v, x) is a bout~dedantilinear fun-

ctional of x and therefore admits the representation

, ,
for some element y where ( )H denotes the scalar prodect in H. This
defines a ~ n a p p i ~y~=
g A v which is clearly liuear. L e t t i ~ ~xg = v and
applying (3.8)we find that

It follows that the operator A has a bounded inverse and t l ~ a tits range
is closed. Fort,hermore the v corresponding to any y is unique. To see that
the rarge of A is the whole space H suppose that z is orthogo~~sl to it.
,
Then we hr~ve B (v z)= 0 for a11 u . From (3.8) it follows, by setting
v = z , that z = 0 . Thus A maps onto the elltire space, and therefore
every a~ltili~learfunctional ' P ( x ) being of the form (y, x ) admits
~ the
,
representation P(x)= B (V $2.). The other representation is proved in a
similar way.

Lecture IV. A Priori Estimates.

Before provil~gGardiugls inequality let us make some general remarks

about a priori estimates. Consider a differential equation h =f of order
k and assume that the solution has been made unique by some auxilirry
conditions. One wauts to study the inverse operator - to see, for
instance, to what class of functions the solution belongs, i f f belonge to rt
give11 class. For this problem, and also for the existence theory, a priori
ineq~u~litiesplay a basic role. Let us suppose that the auxiliary couditions
are homogeneous, then a typical a priori estimate would assert that for
some uorm 11 11
11 Dfl u 11 zz constant 11 Lzc 11
For instance, if we kr~owthat the equatiol~has a solution of class OK for
all f uf class 0 then indeed, by a simple application of the closed graph
theorem, we would have

II Dfi 21 1) < constant )ILzc 11 , 1 B I s; K -j ,

 L. NIEENBEBG
: 0s elliptic partial

with 11 11 the usual uorm.in Cj. 111 general if Lu has fit~ite (1 11 norm we
will not obtaili such an inequality for K = k , rather K < k ; that is we
c a ~ ~ ~ esti~riate
lot iudividu;~lly all derivatives enterir~g ill L . However I
believe that elliptic equatiolis oil11 be cltari~cterizedas tllose for which oue
can e s t i ~ ~ t a all
t e derivatives, i. e.

for a wide class of n o r m (this is skated its it co~iviction~ i o ti~ tlieorem).

Cousider ttow an elliptic equatiott Lu =f with saitwble lto~nogel~eoas
bouriditry coaditions. Most i h priori estimates are just of tlie type (4.1) or,
if oue does not assume uniql~el~ess, of the for111

Iudeed ~ n u c hof the theory of elliptic e q ~ ~ a t i ois~ concerned

~s witlt proving
s u c l ~estilnt~tesfor various liorms 11 ,
11 a11(1proviug attalogo~lsestimates
for fuuctious with no boundary restrietioos :

Here a is ally compact srtbdomai~tof 9, and the norm 11 1 " is cousidered

only for fu~rctio~is defined in a . I
A word qf oautiow: The estilrlate do not hold for the most obious
norm tliat orhe would try, ~iamelytlie 111axi1n111n (or Go) norlrl lior i l l getteral
for Cj nonns, however they do hold for Cj+"iorn~s, 0 a < < l , and for
many illtegral norms.
W e quote some immediate colisequence of (4.2), (4.2)'.
1. If f aud the coeffieiet~ts of L are in Cm t l ~ e na solutio~t of
Lu =f is also in COO. This follows fairly easily from (4.2)'.
2. 80lutiotis of Lu = 0 with bounded llorlrr 11 11 form a compact
family. This follows from (4.2)' itutl the
Calculus Levma: The set 11 # (1 +11 Du 11 coltstant is conlpuct in the
q a o e with (1 11 as norm.
This lemma holds for a wide class of norms.
3. The set of sol~itionsof L.u = 0 sitisfyiug the bouudary conditions
(so that (4.2) holds) is fiuite dimensional. This follows with the aid of the
Oalculus Lemma.
I would like to describe briefly a general recipe for proving such
estimaten. This cousists of several steps :
1. I11 case of (4.4)' prove it fil~stfor equations with constant coeffi-
oiente and only highest order terms, and for functions of compact support.
 diferential equations

111 case (4.2), prove it also for such eqnations and for functions defioed in
a half spitce, v;cnishing ueir infi~lity,and satisfiyng (on the p1au:lr bouudary)
the boundary conditions. These are also assrimed to have c o ~ ~ s t a ncoeffi-
t
cients (i. e. to be transliltio~~
invibriant).
2. Now elitnit~atethe hypothesis of compact support.
3. Extellti the estirn~teto variable coefficients as follows : with the
aid of a partitions of unity write the funchion u as a sum of fiinctions % with
s~nallsupport, in each of which the leadiug coefficients are close to coastiints,
ant1 treat the variation from aoustant as an error term, using the results
of Step 2 end the followiug lemma which lnay also be used ill the proof
of Step 2.
Caloz~lusLumnza : For appropriute constants o, o, ,

where for ,/unctiotu of compact support we ?nay take c, = 0 a d c, indepejtdetzt

o f the support of u .
This holds for a wide class of norms.
In case the support of ui touclreu the boundary, make a local change
of variable to flatten out the bol~ndaryso t h ~ ,Steps t 1 ; ~ n d2 can be i~pplied.
Tlre main step h'ere is Step 1. We retnark that in Step 3 we rely on
( t i t , least) the c o ~ ~ t i n ~ i iof ,
t y tlre leading coefficients of L or on the fact
that tl~eydiffer little from constunts in small domains. Because of this olle
does not obtain in this way the Inore refined estimates required for tt.eating
nouliue~r probletns, such as those in Bers, Nireuberg [lo], de Giorgi [li],
or Nash [12].
T l ~ ellor~rls for which such estimates are easiest to derive are the L,
norms for functions and their derivatives, and we shall illustrate the recipe
0
for these by proving Garding's inequality in its general form.
Consider a quadratic integral form d e f i ~ ~ efor d Om functions with
compact support in a bouuded domain cZ)

and suppose that the (complex valued) coefficients c b , are cotrtiuuol~sin (i3.
A necessary and sufloient condition jor the existence of positive constants
c ,0 so that the ifiequality
 : On elliptic partial
L. NIKINBEKG

holds for all u E C r (9i)

s that for some positive cotbsttrtlt co

3e 2 O P , ~5P 57 2 c0 I 5 12m for a11 real 5.

IS1 > lyl-ns

Here the notatio~lsof Lecture 3 is used.

I'roof: We prove first the sufficiency, following our recipe. The Calculus
Lelr~rne(4.3) will be used in the forlrr: For every E > 0 there is :I aol~stant
C(E)such that for every Gm function u with compact support

Tl~isis contained in our inequalities of Lecture 2, but is most easily proved

with the aid of Fourier transforms.
We consider now t l ~ edifferent steps in provir~g(4.5), the Step 2 of the
recipe doe8 not occur here since our fnrlctio~~s have comlract support.
~ l t vr~irishunless IbI =l y l = m .
1. Suppose that the oP,, are c o ~ ~ s t t tat111
We ilrtroduce the Fourier t,ral~sformof u

By Parseval's theorem we have

proving (4.5) for this special case.

We uow consider' the verii~blecoefficie~rtcase and breitk Step 3 into
two parts.
2. Suppose that the support of u is sufficieatly small, contained,
say, in a small sphere about the o r i g i ~ .Then accorfing to the preceding
inequality we have
 diff6rsntial equations

If now the support of u is so srr~allthat cp,, hi18 smi~lloscillation there

we see that the second term on tlre riglrt may be bounded by

11 u ( I ,
The third term is trivially bolrnded by oo~lsta~it (1 u
Thus we find that

from which follows the inequality

(4.5) now follows with the aid of (4.7).

3. Consider )low the general case. Uonstruct e partition of unity in a,

with the support of each wj as small as desired. Then

by the preceding Case 2,

2 constant (1 u 1 ; + 0 (11 u I)m . )Iu (In,-1)

and the desired result now follows easily with the aid of (4.7).
 L. NIRENBERG
': On, elliptic partial

,
We see that the constants c 0 iu (4.5) depend OII c,, a11 upper
I
bonr~d for the cg,, 1 , i111d 011 the modulus of co~~tiunity of the leadi~rg
cBly with I = y I I
I = nb , a11d finally on tlle iion~ail~
0 .
Now for the proof of the nacessity of (4.6). Suppose that (4.5) holds
aud that the left Itand side of (4.5) va~~islles for solne real I[ I , [=r, I/=
and some poiut ill (i3, say the origin. Followiug the ergument iu Step 2
in the proof of sufficiency we see that tlie inequl~lity

llolds for all Cw u with support in sonle fixed neighborhood U about the
,
origin and in (2). Set u = e""ra. c (a) for real I where 5 (x) is a fixed real
Cm f u a c t i o ~with
~ s~rpportin U and in 9 , One sees reitdily that as I w -
the left hand side of (4.5)' is 0 ( P I ) and not o (IZm)while the right hand
side is 0 (1%"'-I), so that (4.5)' does not hold.
C
Garding's inequality (4.5) is a t one end of a whole spectrim of inter-
esting and useful ineql~alities making tliffere~ltrequirelr~entson u at the
0
bou~ldary,Gardiugls inequality niaking the maximill reqaire~nent- that
all derivatives of u of order less than nt va~~islr a t the boundary. A t the
other end of the spectrunl is the ir~equalityof Arouszi~ju[13] involving no
boulrdary conditions whatsoever.
Aro~lszaju co~lsiders a 11ulnber of dillerential operators Lj (8,D),
... ,
j = 1 , N of order ~ 8 with , c:oefficients eoutinuons ill the closure of a
bou~~detldomain (a, and solves the following problern : Under what con-
ditions can one assert that for 1111 C" functio~lsu in 9 the inequality

holds, with ttle ao~lstant iudepeudeut of , u ? He gives necessary and

sufficieut conditions :
(a) the operibtor 2 Lj L; is elliptic, here Lf is the formal adjoint of L j .
(b) A t any boundary point - 8 of 9 , if is the unit normal to
ci) m d 5 . 10 is any real vector taugent to 6 then the polynomials in z ,
+ -+
Lj (x, 4 z n) have no common complex root z . Here Lj is the leadiug
part of Lj .
An exs~nple of Aronzsaju7s illequality is the following; for functions
u (x, y) in a bounded domain in the plane

J1 uW 12 ilx dy sz constant
I(1 urn l2 + 1 u, l2 + I u le) dx dy .
 diffevential equations

Even this simple extcmple is not trivial to prove.

Since the report of Aronszajn a 11nmber of people Ileve coottsidered
the problem *of proving (4.5) for v n r i o ~ ~quatlrat,ia
s forms (4.4) and under
various differential boundary oo~lditions. For one operator L j Agmon,
Douglis, Nireuberg [14], (in a fortheotning paper which will be discussed
later) have characterized these dilTerential boundary coilditions whicb are
m/2 in unmber aud for which (4.8) holds. Scheclrter [l5] has treated N
operators a l ~ dgeneral boundary conclitions. Aronszaju, in i ~ n p ~ ~ b l i swork,
l~ed
has treated the geaeral problrtn (4.5). Also Horl~landerand Agmon [l6]
have solved the general problem for (4.5) a,nd genelnal ciifferel~tiiclboundary
conditions. The proofs follow the recipe outlitled above, the trraii~step being
the first, for functions in a llalf spibce.
We conclude the lect~ire with a result that will be used in proving
the differe~rtiabilit~y r ~ t the boundary of solutiolls of elliptic equi~tions. Iu
< ,
the followil~g ,ZR denotes the henlisphere I x I R x,, 2 0 We shall .
,
denote the varia,ble x,, by t (xi, ... , , ,
x,,-,) by x and (xi ... x,) by (x, t) .
Lemma: Let u be a weak solutio~t of a diflr.etttia1 equatios (of order k)
with, for simnplioity, CW coeflcienfs,

i n the interior of a hemisphere Z R , where .b are giver&f~cnctions,mid assultre

that the plane t = 0 i s ~iowhevecharacteristic, ill jirct that the coeflcient a of
>
I$ i n L does not vtrlzish. I f fol* every 6 0 the funelions .fP,IIb lc fbr
I /?1 <j , D, Dj w belong to Lz i r ~2R-sthen trlro the ftrt~ctioftDft' 11 hcts
this property.
For j 2 k - I there is notl~ing to prove, as we may solve for the
functioll ~ ( " z c froin the differential equatiou (4.9) operated on by D::+'-~.
Thus we suppose j < k -1 .
The proof ltlakes use of a well known formula giving explicitly a
smooth extension of a function v defined ill a half spitce t > O to a fu~~ction
defined in the full space:

with the ;lj chosen so that

28 : On elliptic partial
L. NIKENBLEHG

We observe that,

Here the norm on the left is over the full space while on the right it is
over the half space t 0 > .
> ,
A o o f of the Leircma : Choose a fixed 6 0 , let [ (s t ) be a fixed Cm
fuaction with sapport in I x l2 + <
t2 R2 and wirich equals one in , Z R 4 ,
and set [ au = v . If we can prove that D/+' v belongs to L2 then, since
a+0 it follows easily that D/+'u is ill L2 ill 2R-S. FPOII~our assumptions
we see that v is a weak solution of a differential oquation of the form

where the us,, b e l o ~ ~ so

g L,, autl that derivatives D, 1)jv and v itself
belorrg to L, .,
For N sufficierltly large we now exterrd tire filrrctiolls v , v,,, to nega-
,
tive t defining UN by (4.10)a l ~ dv , , , , ~by

One may then verify that the equation

holds in the entire 8pace iu the weak sense, R I I ~that tire v , , , , ~ the
, deri-
vatives D, DjvN and V N itself belong to L 2 .
Let us now take Folirier tr:rr~sforr~~s with respect to 8 and t , and
write ( E l , ..., .
E,,-,) = 5, En = t Denotirrg the trsasfonn of a function f
7
by we find that

with
w

VS,,,~, GN a r ~ d1 5 1 (1 E I j f I t I j )
belougirrg to L, in the (E t ) spice. ,
To conclnde the proof we have to show that belougs to L 2 .
To this end write
 differential equations

We shall show tJhat ertch term on the right belongs to L,. Pron~(431)
we
find that the first tern1 on the right is bout~tled by

Since s $ I y I < k -j - 1 it follows t h ~ the

t fact,or of v , , ,is~ uniformly
bounded, and hence t h i ~ tthis term belongs to I,,, since the V,,~,N do.
The second term on the right of (4.12) is bounded by

with o an absolafe constant, and hence belongs also to L , , by an earlier

remark.
This completes the proof of the Le~n~na.

Lecture V. The 1)ilferentiabilitj- of Weak Solutions of

Elliptic Equations

In this and the next lecture we slrall preseut a self contained proof of
the well known result thiht solutions of elliptic eqllations with C m coeffi-
cients are of class Cm.
Many proofs exist in the literatoi~eincluding proofs for Inore general
classes of equations, see Hiirn~;u~~ler 1171, Malgrange [18]. The proof here
seems rather straigtforward; it i~ based essent.ially OII a proof given by
Lax [19] and is closely related to proofs give11 in lectures by Bers [20]
a ~ l dSchwartz [21] (see also [9]). We confine oiirselves as before t,o a single
equation (not l~ecessarily strongly elliptic) although the argument extends
also to systems.
Dieretitiability !L'heorenz : If u is a locally epucrre ittteq~ableweak soh-
tion of the elliptic epuatiofo L u =f, alzd f E Cm tlbelz u E Om.
Remark : If u is a distribution solution theu u = Ak v for some conti-
nuous v (here A is the Laplace operator), and v is theu a weak solution
of L A h =f . The Theorem holds therefore for this case also.
The proof consists in showing that u has L, derivatives of all orders
in every compact subdomain. That u E GM" tl~en follows from the Sobolev
estimates proved in Lecture 2. However since we ollly need a very simple
case of the Sobolev lemmas me present a separate proof of it here.
 Lemzcin (Sobolev): In n u ~vnoothu dot~aaitt9 i f u hrrs I;, derivtctives up
to order s in a >
for s 9812, thes u is continttous i n 0 .
In fact

Proqf: The first assertion follows easily from tile iaequality. To prove
the i~lequalily let x, be an inner point in 0 (for 8irlrplicif.y take x, = 0)
alld suppose tlrere is 'a sphere about x,, in 0 with radius R . Let
<
furthermorl 5 ( r ) be a function iir Cm, e q ~ l ~tol 1 for 0 T L R I 2 , and
.
vanishing for r 2 R By ii~tegratiou alolrg ally rl~tliusfrom zo= 0 , aud
by repeated partial integration we see that

iutegrating over the unit sphere (with area Q) of radial directions one
finds

>
using ScLwarz i~leqnality.For. s n/2 the last iutegral is finite.
If the boundary of 0 is such that a t any point in there exists a
cone with a fixed ope~iiogalrd length contailled in then tire same proof
holds; instead of illtegrating over the full sphere of radial directions, we
merely integrate over the directions lying in the cone.
The proof of the Differentiabilih Theorem co~lsistsmainly of a series
of simple lemmas of c a l c u l ~ ~coucer~~ed
s with a special situatioa) that of
periodic functious, and tlris lecture will confirled to these oalculus statements.
We consider peviodic fulrctio~rsu E C* with period 2 n in each z,. For
such f u ~ ~ c t i o utire
s Fourier series

,u=XuEei..f, E = (E, .... .5,,)

5

( f j = in tegor) converges uuifornll y.

 By Parseval's equality me have the following estimate for each non-
negative 8
(5.1) , canstant 2 (1
5
+1 E
- constant 2 (1
I
C
+ I iI? )S( U E I'
where the i ~ ~ t e g r aisl taken over :L period cube.
For any integer 8 we introdl~cethe followiag scalar prodnct and norm,
differing slightsly from our previous notation,

, ,
We write (u u), = (u u ) and proceed with the
Calculus :
.
1. 11 u Il is inocretisillg in 8 Flirlhern~ore for t, < < t,
8 and any
E >0 there is a constant C (6) such that

< +
Proof: For any a 2 0 , as E ot9 C ( 8 ) ot1.
2. Set p , = ( l - A ) t u , y = ( l - A ) t v , sethatp,=2ue(L+I&12)teb.~.
From this we bud e

As a consequence we have
Lemma : If o E Cm, then

(5.5) ,v)t = (a W v)t 4- 0 (I1 llt I1 v + I/

(W 11 1 91 ]It-1 Ilt-I I1 v (It).
Proof: Assunie t < 0 . Using (5.4), (5.3), (5.1), and partial integration, we
find
( w u , ~ ) ~ = ((1
w- A ) - t p , ! ~ ) = ( ( l - A)-tp, zy)

Iu the case t / 0 the proof is similar.

 L. NIRENBERG
: On elliptic partial

3. Sckwar(t)t'a inequality :

(Clear I)

Proof: According to (5.6) tile left side of (5.7) is 11ot smaller than the
right side. If however we set v = (1- A ) t t z , then, by (5.4)

proving (5.7).
We can now form Hilbert espace Hs by colnyletil~g Cm fullotions in
the norms 11 11,. For a > 0 these agree wit11 our previous definitious.
> .
Obviously H,c Ht f6r a t All the previous results llold for fuuctions
with the appropriate i~orlnsfinite, for insti~~lce
(5.7). We lnay regard H, as
give11 by a forlnal Fourier series with fiuite 11 11, noun.
We remark that the scalar product

, ,
is defined, by extension, for any fiu~ctions11 E Hg v E H-, and that ally
bounded linear fu~~ctional
f (u)defined on Hs lnay be represented in the form

with v E H-,; this follows imlnediately from the Fourier series representa-
tion, so that we ]nay regclrd H-, as dual to H s .
Though me shell not use this, we reulark thirt the closed ulrit ball
< >.
11 u 11, 1 in H, is compact, i l l Ht for s t
We continue with the calculus.
4. Consider any differential operator L of order k with Cm coefficients.
Claim :

More precisely

, ,
where c = G (k u) K is a bound for the leading coefficieuts, and K' is a
bound for all coefficients and their derivatives u p to order 1s 1.
 P ~ . o o f :S i ~ ~ cobviously
e 11 Di 24,II 5 C O I I S ~11 14 it soffices, in order to
prove (5.9), to show that i f a E Ca theu

1 I
where k' a11t1k" are bounds for a 1 aud 1 Dj (1 (j 1 s 1 ) respectively.
P ~ o o fof (5.10) : Co~~siderfirst the cave s 0 . Set 97 = (1- A)' < t~
-
ly' = (1 A ) S a11 theu we I~iive,by (5.4), and partial integratiou,

lntegratiug t l ~ el ~ , s by
t pt~rts(- 8) times we tind it is not greater than

So dividiug by 11 y we have, with the aid of (5.3),

=o k 11 [ I s
21 + c k' 11 16 Ils-l by (5.3).
I u case s 2 0 we have
IIauII:=(a u , (1. - A)'au)

and may integrate by parts as above.

So L car1 be extendetl to all of Hs and maps it boundedly into Hs-k.
This operation of L agrees with that of L actiug on u , regarded as a
distribution.
Teckdcal Lemma: Suppose w is a Ca real function, then

To conclude this leoture we consider

Diferenoe Quotients: For given vector h let
 L. NIRIENBIERG
: Om elliptic partial

be the difference quotielit. One verifies easily : 11 u (8 + h) 11, = 11 u (x)111,,

, <
Furthermore : If tc 6 Ha uh E H, and 11 uh 1 , k for each h, then )I u llsfl 5 k .
< ,
Corolltrry :If t c E H,,I1 u: 11 k for each 8 then a E H,+, and 11 u I ,+, 5 k.
Proof; Let u = 8 t c c &.a7 and let u~ = 8 uc e b . 5 . One filids
€ IblSi

Lecture ?I. Proof of the Differelltiability Theorelti.

Let now L be an elliptic operator of order k . I n the pe~,iodiccase we

prove the Differelltiability Theorem in the form
D~FFERENTIAB~L~TY , ,
M :11 E H, L u E H s - k + ~ tlteu IL E Ha+,
T I I B ~ ~ ~ E Ij' .
So it. follows that iJ",rE H, atid L u E Ht-k, thelt. t i E Ht .
The non-periodic case is easily reduced to this as follows; We prove
successively that u has L2 first order derivatives, thee second order deri-
vatives, then second order derivatives, and so on.
To cwry out this reduction let [ be a C m fnnetiou defined ill a
neighborhood of s poiut and with colnpact support. Let v = [ 14 and extend
v and the coefficients of L s s periodic functions. SO

where f = [ L t i , g = L (( u)- ( L TL ; g contai~~s only derivatives of t c up

to order k - 1 , and so, as is easily seen with aid of (5.8) llas finite
11 111-k norm.
So L v E H l-k, therefore u E Hi and so u has L, derivatives in a nei-
ghborhood of the point. Using this one repeats the argn~nentfor a sn~aller
neigborl~ood,and sees that 5 v E ,
so v E H 2 , itnd so on.
The proof of tlre Differentiability Theoreni i l l tlie periodic c u e follows
0
easily, in turn, from the followi~igestin~iitewhich is tlie analogue of Garding's
inequality.
Basic E~timata: For any s a,, ,
11 u lls+k < constant 11 L u [Is + constant 11 18 ,1 .
Postponing the proof of (6.1) let us prove the Differentiability Theorem.
Consider a difference quotient uh. If Lh represents the operator obtained by
replacing each coefficient in L by its differeuce quotieut we see that h h =
 diferenlial equations

-
= ( L w ) ~ Lh.u (8 + h) Thus we have, from (6.1)
11 uh ] I s <constant 11 L uh + constant 11 zlF.
IJ,-k

- 11
C constant ( L 2 Jls-k ~ ) ~ + constant 11 Lh u (x + 1 ~ )
f constant, 11 u 11,

< constant 11 1,u lIs-k+l

- + co~istant11 u [Is by (5.1 2): (5.8)

The desired result follows from the corollary after (6.12).

Proof of the Rnsio Estimate : Clearly only so s $ k is of interest. <
The proof cotrsists of several steps, following our recipe, ant1 the proof of
0
G~~rdiug'sinequality in Lecture 4.
1. L has constant coefficients with leading terms only. Then

> constant 2 ua
-
I
I l 2 1 l jZk (1 + I l JZ)s
while

Hence
11 L u 1 ; + 11 tc )>:I constant 2 (
I
l 2 ( 1 + I 612)S+k =
2 1 ~

= constant 11 u .
2. L has variable coeffieier~tswit11 leading coefficients differing from
constant value by less than 8 , E sufficiently small : Let L, be the operi~tor
with these constant coefficients. By case 1. we have

11 u I l s + k I coustant 11 Lo u 11, + aonataut 11 21 I,

< constant (11 L 21 1 , + 11 (Lo- Lu) 118) + constant 11 21 1 ,
-< constant 11 Tr u 11% + constant E 11 u Ils+k + constant 11 21 Ils+k-l, by (5.9),

(constant 11 L u (1, f co~iatrute 11 u ll.+k + 11 u + constent 11 tc 1 ,

1

by (5.2),
from which (6.1) follows.
Note: If tc has its support in a small set then the leading coefficients,
being continuous, difrer little from constant values. So (6.1) holds iu
that case.
 3. General case: Intrnduce a partition of unity over the closed
period cube

with each w , having its suppokt in a small region.

< constal~t2 (L wj u, L a j us)+ constant 2 (1 @j u

- + 0 11
( u l18+k: 11 u ((s+k-~)
i j

(since wj u has its support in a small set)

<
- constant + 3
(1 5 21 112 1 (I + II
0 - , by (5.2)

from which (6.1) follows.

We co~~clude this lecture with some remarks concerning the differen-
tiability uear the boundary of the solution of the Dirichlet problem obtai-
ned in Lecture 3. Using the notation of that lecture we recall that the
solutiol~of L u = f belonged to the space H,. Since the divcussion is
local we may assume that the bou~ldaryis given by zN= 0 and thitt we
have a solution of the equation in the hemisphere Z R (see Lecture 4). I t
suffices, by Sobolev, to show that the sol~ition u has derivatives of all
orders in L2 in Z R 4 , for f in Om. We shall merely indicate the first
step - +
the proof that u has derivatives of order (st 1) in L2 (see [9]) -
and sllsll use the notation of the l e ~ n ~ nata the end of Lecture 4. By that
lemlna it suffices to prove that derivatives of the form D, Dmu belong to
L2 in .ZR4 for any 6 0 .>
This we do with the aid of difference quotients as above. For fixed
>
6 0 let 5 (2,t) be a Cm function with support in I x "1 tt< R2, and which
equals one in ,ZR-d. Since the function u satisties
for all g, ill it fo11ows that the function u = 6 u satisfies
H, (ZR)

for such g,. If we IIOV form difference quotients v h as above with A parallel
to the boundary t = 0 we find easily that

for cp in iWr .
(Zx)Setting g, = v h and applying ~ i r d i n ~inequality
~s for
strongly elliptic operators of Lecture 3 we obtain a bound for

,
which is independent of h and it follows easily that the derivatives 1lm v
,
are in Lz hence that D, Dm u E L2 in ZEl--g.

Lecture VII. A l'riori Estimates Near the Boundary.

111 the relrii~iuigtime we shall discuss briefly the derivation of estimates
Ileiir the boundary for solntio~rs of elliptic eq~~ations iu, for simplicity, it
bounded domain (3'. Thihl material is taken from a paper by Agmon,
I)onglis, Nirenberg [14] which is coucerlrecl wit11 both Sohauder a ~ l dL,
estimates near the boundary for solutions satisfying general boundary
coutlition. As remibrked in Lectnre 4 the estimates for p = 2 are special
cases of more general results. We wish also to draw attention to a pitper
[52] by Honnander concerned with eqi~ationsL u = O with constant coeffi-
cients in a lralf space, aud solntious satisfyil~ga number of bot~ndary
coliditiolis Kju = O described by differeutial operators Bj with consta~lt
coefficierlts. Horma~lder characterizes i~ll'such systems for which the solu.
tioris belong to Om on the boundary, and also those for which the solutions
are a~lalytica t the boundary.
Since the time is li~i~ited we shall restrict ourselves here to the Di-
richlet problem for a single elliptic equations of order 2 m

where represents the uuit normal td the boundary.

 : On elliptic partial
L. NILLIENBIHG

The operator. L will be required to satisfy a certain condition.

Condition O I L L : lf L ' ( x , D) is the leading ptrvt of L , we reqrive that
for every pair of independent real vectors ti,t2the yolysoniiccl i n r

huve ezactly m roots oft either side of the real z axis.

In three or more di~rlensiol~sthis conditiol~ is automatically satisfied.
We see however that the operator (&.+i &y in two dinlensions violates
tlle condition. (This operator and others in two di~nellsiollscome under the
theory when treated as a system).
111 describing the estimates we make use of t l ~ efollowing norms and
selr~iuorms.In lecture 2 we already met the Holder norm, for 0 < a 1 <

For functions of classe Ck in (i3 we also introduce (differing from the no-
tation in Lecture 2)

where the 1. u. b. is taken over all derivdtives of order k , and a11 points
in (i3. 111 addition, for functiol~sin C k with Holder oonti~ruous(exponent a)
derivatives of order k , we introduce

where the max, is taken over all derivntives of order k . The space of
fiumtions with finite I Ik+, llorlrl is denoted by Ck+" ((i3). (We also uae the
notation of Lecture 3 and 4).
For functions p, defined on the (smooth) boundary @ of (i3 we also
have analogous ~ ~ o r m sdetilled
, ill a tather obvious way en terms of local
coordinates, and which we deuote in the same way.
 diferertial eyuations

We now summarize the resnlts without specifying the exact smooth-

ness oor~ditionsqn the boundary; k will denote a non-negative integer.
The integral estimates will be stated only for p = 2 .
L, Estima.tes : If uE H2, and sntis$es, for simplicity, homoye~ceous
Dirichlet dtrta, i. e. y j = 0 , aiid if L u E Hk, aicd tile coef3oients of L belong
to Ck, then u belongs to and

Similar reszilts hold jbr equations in integral, or variational, form.

Sohauder Estimates :I j y o r some positive a 1 ,uE C 2m+"((i3),LU E 0 k+a(q),
<
yj E C2~$+~+l-j+~, awd the coefjcie~tsof L belong to C ~ + Rthen
, u E C2m+Na and

Si9)tilar result^ hold for vqrctctions in varicctional form. fiom these oae
,
rcay deriue, tior instarcce, the following result for solutions of L u = 0 under
srtitable srtoothqaess &ssustptions on the coafjcieuts :
) y j E Cm-J+k+d,j = 1
I f u E C ~ - l + ~ ( gand ,.,. ,
m , the% u E Cm-l+k+a @I
and
(7.4) I u Im-I +k+a 5 constant (2I Im-j+k+a +I 10) *

.
Tlre coustants in the above are indeper~dentof w I n case of uuiqueness
of the solntion in the class considered bhe terms (1 u [Io or / u ,1 may be
dropped. Mirauda [23], using results of Agmon [25], has recently proved
au extended maximum principle for solutions of strougly elliptic equatioas
(7.1) in two dimensions, which asserts that (7.4) holds for k = a = 0 I .
believe that this holds true in general for operators satisfying the con-
dition on L .
Tlle Schauder estimates have a number of useful consequences. With
their aid oue may prove the existence of solutions of strougly ellilbtic
equations having merely Holder colltinuous coefficieats. I n particular, with
the aid of (7.4) one may solve such equations with the given Bj in class
Qm-j+a.

In addition one can also solve the Dirichlet problem for a wide
class of equations which are not strol~gly elliptic. Puthermore, and
this is perh;tps the most useful feature of the estin~ates, with their
aid one may prove looal ~ ~ a r t u r b a t i theorellis
o~~ for nonlinear elliptic
squations. For example if FA(s u , , ...,
DZmzc) = 0 is 21, nonlinear equa-
tion depending on a parameter 1 and <c smooC11Iyu 011 all variables,
such Chat for 1= 0 the fuuctiou uo is a solution with, say, zero Dirichlet
data, and if the a first variation * of P at u, is a linear elliptic operator
L whicll is invertible (i. e. for which the Dirichlet problen~(7.1) has one
aud ouly one solution) the11 for I R I sr~fficie~ltly small there exists a uuique
solutiol~u~ of the n o ~ ~ l i ~ iequation
ear wit11 zero Dirichlet data. The estimates
also yield differentirtbilihy theorems a t tlre boundary for solutions of non-
liuetrr elliptic equations.
The estimates are derived followilrg the <<recipe$of Lecture 4, the
main step being the first one. That is, one col~siderseqnatiol~s(7.1) in a
>
half space x,, 0 , for operators L with constant coefficients and only
higlrest order terms, and Cm functious u ill t 2 0 vi~uishir~g outside some
splrere. This syste~r~ is the11 trei~tetlwit11 tile aid of explicitly constructed
Poisson kernels, wlriclr will be described in the next lecture, with whicl~
one solves the system (7.1) with f = 0 . Witlr the aid of the explicit
represeetations for u and its derivatives so obtibined, tlre desired estimates
for this cor~stant coefficieat case are then obtitiued with the aid of certain
potential theoretic results.
I mould like tb describe these results, whic11 I believe should prove
useful for other problems. Si~lccwe are operating in a 11alf space x,,> 0
, ,
it is convenient to reaame the coordinates, set (xl ... xn-;) = x , x, = t ,
...
(2,, ,xn) = ( 8 , t) = P.
~ s t) ,
We consider integral tmusforrns of functions f (2) iuto f u ~ l c t i o ~u(x,
t>O. Let K(x,t) be a kernel defiued ie the half space t>O aud l ~ o ~ ~ ~ o g e u e o t ~ s
of degree 1- n .

I (I
here PI = x l2 +
t2)i/2; Asstinre that Q is colltiuuous on the half sphere
IPI = 1 , t 2 0 and assiltne also (this condition can be weakened conside-
rably) that has continuot~s first derivatives OII tlre half splrere whioh,
together with B itself, are bou~rded in absolute value by x . Ia additiou
we make the basic sssumptiou

I
181-1
L'(x,O)d~,=o.

Here integration is over the unib sphere ( x = 1 I , with d w, as element

of area.
Consider the transformation

u(x,t)=
IK(x-y,t)f(y)dy
 diferential equations

integratiol~ beilig over the elltire y space. Deuote tlre L, norm o f f by

If Ig, aud that of u ( x ,t) iu x , for ally fixed t , by I u l L p, t .
THEOREX:1. For 0 < a < 1

acheye c dqends only ow a rrtrd n . Here the novels vefev to the half space
> ,
t 0 jbr u and the plane t = 0 for f.
>
2. For every t 0 and 1< p oo < ,

where o depends only os P and n .

where c is an absolute oonstnst. Here (f)-112is defined in tavms of the

Fourier transforvn T(E)
o f f by

There is at1 L, analog~ie of 3, which is however more complicated

to state.
We call Part 1 of the theorem a result of Privaloff type. I t is a
simple extension of classictal results of Holder, Giraud and others, to wllicl~
it reduces if we set t = 0 . Part 2, a result of Riesz type, is a stnlighfor-
wlttd extension of recent results of Ualtlero~~ a1111 Zygmund [24], to which
it reduces if we set t = 0 . For the special case of the HiIbert transfonn
for ,n= 2 it is due to Riesz, a ~ ill ~ dfact it is proved by reduction to the
Riesz result witlr the aid of a, device of [24]. Part 3, is proved with the
A

aid of Fourier transfor~ns- one sl~owsthat the Fourier tri~llsfornl K ( 5 ,t)

of K ( x ,t) with respect to the a variables is bou~~ded i l l absolnte value by
+1
constant (1 t I)-' , from which the result follows easily. Part 3 plays
an essential role in the derivation of the L, estimates.

Lecture VIII. The Boundary Value Problem in a Half Space;

The P o i s s o ~Kernels.
~
I n this lecture we shell show how to solve explicity the elliptic
system (7.1) with coasbant coefficiei~tsfor the special case of a half space.
Making n slight change of llotation we sb@ll co~18iderthe space to be
 L. Nrnlc~slcaa: On elliptic parNal

+
n 1 itimeusional, mitli tlie first n coordinates denoted by a = (3, x,~), ...,
>
and tlle last coordinate by t . 111 the half spikce t 0 me cousider-for sim-
a
plicity the homoge~ieousequation, with D

where L is an elliptic operator of order 2 m with only highest order terms,

satisfying the e coutlition 011 L B of the ~reviouslecture, i. e. for fixed real
5 =(l, ,... , +
tn) 0 the polyuolnial L ( E l z) lias exactly nb roots z on each
side of the real axis.
011t = 0 we prescribe the derivatives

I$-1,'= Gj(x)

with the Qii in G r , for simplicity.

,
Tlie solution w'ill be given i l l terms of kerliels Kj ( a , t ) j = l ,...,m,
the Poisson kernels,

(8.3) Kj(~-y,t)Qij(y)dy=2Kj*Qij,

where (: deuotes co~ivolution. Our constrrlction of the Ki is an exteusion

of the coustructio~lgiveu by Agnion [25] in two dilneasions, n = 1 , but it
is based OIL the Fritz Jehu ideutity (1.6) of Lecture 1 : For QJ (a) in Go"

where q is a non-negative integer of the stluie parity as s, A is the

Laplacean, ant1 the priucipitl bnruch of the l~garithm is taken with the
plane slit along the negative real exis.
+
First some preliininaries. For fixed real t 0 denote by z t = z$ (6) ,
k =1 ,...,
m , the roots z with positive ilnaginary parts of L ( E l z) = 0 ,
and set
."
The coefficieuts at +
are an~lyt,icill 6 for real 8 0 , a l ~ dIio~~iogeneous
of
degree p . With M+ we associate the polynomials (in z)
 diflerential equations

The following relations are easily verified.

where y is a rectifiable Jordan coutour in the complex a plane enclosing

all the roots r+(5) ill its interior; 8; is the Kroi~eckerdelta.
We call uow writhe dowll the
Poisson Kernels : For j - 1 2 a

for j-1<%

Here

and y is a Jordau c o ~ ~ t o uinr 3tit.z> 0 eiiclosiug ell the roots z of M + ( [ , a )

for all 1 61 = 1, l real.
Before proving thiit these for~riulwsrepresent Poisso~~ kernels we ob-
serve, with the aid of the identities

(- 1y+r
(p+A)!(-l-p)!(2Y(
2"~
3
log -r = z p , p<O, I+p>O

for 1,zc integers, and ~ 1 , ~sott1e

' apl)rnl)riate co~~stants,
that sr may rel)rese~rt
the fal~ctioi~sK, iir the forn~- with q a I I O I I - I I ~ ~ ~ integer
L ~ ~ V ~ 11avi11g
the
same parity as n. -
44 : On elliptic partial
L. NIBENBERG

where, for j - 1 2 n ,

and for j - l < n

I t is easily seen that Kj,, and all its derivatives up to order j q +

are contiuuous in the closed half space t 2 0 .
We uom prove that the kel.aels Kj give11 by (8.7), (8.7)' are indeed
Poissorr kernels. By iuspectior~ me see t l ~ a tthe Kj are analytic solutious
>
of Lu = 0 for t 0 . H e l m a defined by (8.3) is ib solution. Setting

we shall show that uj beloltgs to Cw iu t 2 0 alrd that

for t=O, k=1, ...,m.

Col~sitleriury pa,rt,ial derivative of ortler s of tcj . Clloosing au iuteger
q of the same parity as n , altd such that q 2 s -j + 2 we have, for t>O

after partial integration, recalli~igthat !Dj€ G.Since, as remarked iibove,

D-j,, is continnous in the closed half space t 2 0 it follows that DSuj c m
be exte~rdedas ;G contilrtious fiil~ctionill the entire closed l~alfspace t > O .
Since s is arbitra,ry we have proved that uj E Cm iri t > 0 .
 To verify (8.1 1) choose q snfficiently large so t l ~ a tq >j - 1c +1 .
j =1 ,..., .
m Usillg (8.12) we have, for t = 0 ,

after a change of variable.

Assume first that k * j . Usirlg (8.10)', (8.10)" we find, for t = 0 , and
appropriate constants c', c"

by (8.6). Thus (8.11) is proved for k j +.

. >
Now suppose k =j If j - 1 n we have, usil~g(8.9) (8.10)' arid (8.6))
for some constant C'

n i q-
- pj (ja- ! I(y .E)q (log 7+ c/)/%
eSdW6
ICI-1 Y
M+ rj-l

where yq (y) is a hornogeueous polynot~tislof degree q .

<
Similarly if j - 1 n we firlil, usirlg (8.10)" ilncl (8.6)

(8.14)" D{-' ,
(y 0) = (- l)*lB'/
(n - j)!q!
(y . 6)q log Y . E d q f yq (y)
l€l=l

where again y, dellotes a bomogeaeous poly~lo~uial of degree q :

Prom (8.13), (8.14)', (8.14)" we fiud, after iliserting the value of fij from
(8.8), and rechanging variables, that
Here we have used the fact that

sil~ce ly, is a ~)olynombl of degree p and is therefore annihilated by

~ t + q ) ' ~ By
. Joh1~7sideutity (8.4) the right side of (8.15) equtrls (Pi(%) , and
the proof that the K , are Poisson kernels is co~nplete.
We remark that the fiuletions K,,, are actually analytic ill t 2 0 except
at the origin, and that for s >j +
q , D",,, is ho~noge~~eous o f degree
+
j - 1 q - s . I t follows from onr proof above that if s = a 4 j- k > 0 ++
then
t
I Di , I
K,,, ( x t ) < constant . x 12 t2)(1,+1)/2 l
(I +
Purlier~nore we see that becanse of tbe reprodnciug properties (8.11) of
the K j we mHy ssqert that

for x +0 ,
t
~ / - lK j ,
( x t ) g constant
(I m +
12 t2)(nS1)/2'

With the aid of (8.16), (8.16)' it is not diffioult to establish the following
Extended Yaximuln Pi.ittciple : The so1ut;or~(8.3) of the Divichlet problem
(8.1), (8.2) satisfies

where the least upper bounds are token with respect to all derivatives of ordev
,
m - 1 and, 012 the left, with respect to all jx t) in the ha(f space, on the
right with vespect to nll x .
This is an a~ialogueof a ~pecialcase of Mirauda7e'extended nltixi~num
prilrciple of [23].
 BIBLIOGRAPHY

[ I ] H. LEWY, A n ezample of a smooth liaear partial difetwtlial eqnation tuithout solution.

Annals o f Math. 66 (1957) p. 155-158.
121 L. E H I ~ K N P R ESolutions
IS, of some problems of division I, II. American Jourl~alo f Math
76 (1964) p. 883-903, 77 (19.55) p. 286-292. 7'he division problsni for dietributio*e, P~.oo.
Nat. Aoad. Soi. 41-10 (1955) p. 756-758.
[3] L. H ~ H M A N D EOn R , the theory of general partial differential operators. Aota Math. 94
(1965) p. 160-348.
[4] B. MALGRANGE,Exietence et approximation des solvtions des dqaatiolls aux ddrirdes partiellee
et den Cquatioas de convolntiors. Annales de L'lt~at. Fourier 6 (1955-6) p. 271-355.
[ 5 ] F. T I ~ $ V E SSolutions
, dlimei~laire d'lquations ~ I U Xdhivdea partiellea dPpead~ntd'un para-
mdtre, C. R. Aoad. Soi. P;rris 242 (1956) p. 1250-1252.
[6] 1,. H ~ R M A N D ELocalI I , and glubal properties of futtdaete~ttal solutions. Math. Soandinavioa
6 (1957) p. 27-39. On the divieion oj distributions by polyaomiala. Arlciv. fiir Mnt. 3 Nu.
53 (1968) p. 555-568.
[7] F. J O H N ,Plane waves and qherical meaas applied to partial differential uquations. Inter-
eoionoe, New Pork, 1955.
[8] N . DU I'~.rasrs, Some theorama about the Hieez fractional integral. Trans. Amer. Math.
Soo. 80 (1955) p. 124-134.
[9] L. N I I ~ R N B E I Il&marks
G, on strowgly elliptic partial dvferetilial eqitations. Colnm. Pure
Appl. Math. 8 (1955) p. 649-675.
[ l o ] L. Bmns, L. h'IltENBERG, ( a ) On a repreee*tafion theorem for linear elliptic systems toitli
diuoontinuous coefficients and its applications. (b) Oti linear and ronliaear elliptic boundary
value problema in the plane. Aiti de Ct~nvegnoInter. sol. Eql~azionialle derivate Par-
ziali, Trieste, 1954 (pnblished 1955).
[ I l l DE GIOKGI,S~llla difJresziabilitlt e l'aaaliticitb delle estremali degli integrali mztllipli
wgolari. Mem. della Aooad. delle Soienze di Torino. Ser. 3, V o l . 3 (1957) p. 25-43.
[ l a ] J. NASH, Cantinuity of solution8 of parabolic and elliptic equations. A n ~ e r .Jonrn. Ma.th.
SO (1958) p. 931-954
[I31 N. AI~ONSZAJN,On coercive integro differential qsadratic fornie. Conference on Partial
Differential Equations, Univereily o f Kansas, 1954, Teohnioal Report No. 14, p. 94-106.
[I41 8. AGMON, A. DOUGLIB,I.. N I B I C N B I C H IE~tinaaiee
G aear the boundary for eotutiotts of
elliptic partial differential eqtcntione eatisfying general bonadaiy conditions I. T o appear i n
Comm. Pnre Appl. Mrth.
[15] M. S C H R C H T I EIntegral
I~, inequalities for partial differential operators and functions sati-
afying general boundary conditions. T o appear in C o n ~ mPnre . Appl. Math. V o l . 12, No. 1
(1969).
[I61 8. AGMON, The coerciveness problem for integro diferential forms, Journal dlAnalyse
Math. 6 (1968) p. 184-223.
[17] L. H ~ ~ X A N D EOnR ,the interior regularity of the 8olutioh.8 of partial difereittial eqnationa.
Comm. Pure Appl. ~ a t h 11 . (1958) p. 197-218.
El81 B. MALGRANGE,Snr nne classe d'oplratetrre diffdrentiele hypoelliptiques. Bull. Soo. Math.
Franoe. 85, 3 (1957) p. 283-306.
 L. NIRPNBPRC: On elliptic

[I91 ?. D. L A X , On Cauch,y1s problenb for l~yperbolic eq~rntiona and the differertiabilily of

solutio~te of elliptic equations. Comlu. 1'11re Appl. Math. 8 (1955) p. 615 633.
[20] L. BICI<S,Elliptic partial differential eqnatioss. Lecture notes o f t h e Senlinar it1
Applied Mathematics, University o f Colorado, June 1967.
[21] L. S c a w a e ~ z ,Ecuaciones difetfel.enciales parcialee elipticas, L e c t ~ ~ r ests Bogota Colombia,
1956.
[%4] L. H ~ ~ I ~ M A NOn D Pthe
K ,regularity of tile soltilions of boundary problems. Aota Math 99
(1958) p. 225-264.
1231 C. M I R A N D A Teorema
, del massimo modulo e teorema di eeisten~ae di unicitir per il pro-
blema di Dirichlet relativo alle equaziotii ellitliclie in due variabili. Alli~ali di Mitt. Pura
ed Appl. Ser. 4, Vol. 46 (1958) p. 265-312.
[a41 A. P. CAI.I)IERON, A. ZYGMUND,On singular integrals, Amer. Journal Math. 79 (1956)
p. 289-309.
[!5] S. AGMON,Mulli~de layer potetttials and the Diricl~let pt'oblem for higher order elliptic
equations iqi the plane I. Comm. I'nre Appl. Math. 10 (1957) p. 179-239.

ADDITIONAL GENERAL BIBLIOGRAPHY

[ A ] C. M I I ~ A N DEquazioni
A, alle derivate parziali di tip0 ellittico, Springer, Berlin 1955.
[ B ] l'vaaeactions of the Synbposit6m ot~ Partial Differealial Equqsntiotts. Berkeley California,
1955, published in Con~nl.Pnre Appl. Math., Vol. 9, No. 3 11966).
[C] E. MAGICNEB, G. S T A M P A C C H I Aproblemi
, a1 contorno pet- le equazioni dijfevenziali lineari
di tip0 ellittico. Annitli dells Souola Norm. Snp. di Pisa Ser. 3, Vol. 12, Fasc. 3
(1958) p. 297-358.