Calc. Var.

5, 391–407 (1997)

c Springer-Verlag 1997

Schauder estimates by scaling

Leon Simon?
Department of Mathematics, Stanford University, Stanford, CA 94305-2060, USA

Received June 6, 1996 / Accepted July 7, 1996

Introduction

There are a number of methods which may be used to derive Schauder estimates
for elliptic problems—see e.g. [Gia93], [Tru86] for some recent methods, and
[GT83], [ADN59], [ADN64], [Mor66] for classical derivations for second order
elliptic problems and general m th -order elliptic problems respectively.
Here we want to show that the Schauder estimates for elliptic and parabolic
operators can be proved by a very direct scaling argument. The method has the
nice feature that it works not just for elliptic and parabolic operators but also
for “scale elliptic” operators (operators which are hypoelliptic and which can
be written as κ-homogeneous plus lower order—the exact definitions are given
in Sect. 1 below).
Furthermore it has completely general applicability to boundary value prob-
lems, without special consideration of the nature of the boundary conditions and
boundary semi-norms beyond the natural assumptions that the corresponding
constant coefficient problem in a half-space is hypoelliptic and that the bound-
ary semi-norms are non-degenerate Hölder seminorms which when applied to
the boundary operators have the same κ-order as the Hölder seminorm being
estimated—again the precise terminology is described below.

1. Estimates for constant coefficient operators in R n

Let Pu be a partial differential operator on Rn of the form

X
Pu = aγδ D γ+δ u,
(γ,δ)∈I ×J

where aγδ are given constants and D = (D1 , . . . , Dn ), Dj = ∂/∂x j , D γ =

D1γ1 · · · Dnγn for multi-indices γ = (γ1 , . . . , γn ) ∈ Z+n (Z+ = the non-negative
integers); I , J are non-empty finite subcollections of Z+n (i.e. multi-indices)
given by
? Partially supported by NSF grants DMS-9504456 and DMS-9207704 at Stanford University
392 L. Simon

I = {γ ∈ Z+ : κ · γ = I }, J = {δ ∈ Z+ : κ · δ = J },

where I , J are given elements of Z+ with m = I + J > 0, and where κ =

(κ1 , . . . , κn ) is a given fixed n-tuple of positive integers.
Since κ · (γ + δ) = m ≡ I + J for each term of P , we say that P is “κ-
homogeneous” of “κ-order” m; notice that this gives

P (u ◦ Sy,σ ) ≡ σ m (Pu) ◦ (Sy,σ ), u ∈ C ∞ (Rn ), σ > 0,

where Sy,σ denotes the affine transformation

Sy,σ (x ) = Sy,σ (x 1 , . . . , x n ) = y + (σ κ1 x 1 , . . . , σ κn x n ), x ∈ Rn .

Bρκ (y) will denote the “κ-ball” of radius ρ and centre y defined by

X
n
Bρκ (y) = {x ∈ Rn : |x j − y j |2/κj < ρ2 }.
j =1

We also assume that P is hypoelliptic in the usual sense that if Ω ⊂ Rn is open

and connected, and if u ∈ L2loc (Ω), then

(1.1) Pu = 0 in Ω =⇒ u ∈ C ∞ in Ω,

where the equation P (u) = 0 is assumed to hold in the weak sense that
hu,PP ∗ ϕiL2 (Ω) = 0 for eachP ϕ ∈ Cc∞ (Ω), with P ∗ = the adjoint operator
|γ|+|δ| γ+δ
= (−1) aγδ D (or (−1)|γ|+|δ| a γδ D γ+δ in case the aγ are allowed
to be complex; in this case we of course must allow the solutions u to be com-
plex valued). Evidently the κ-homogeneity condition for P is equivalent to the
fact that p(ζ1κ1 , . . . , ζnκn ) is a homogeneous polynomial of degree m = I + J in
P ζ = (ζ1 , . . . , ζn γ+δ
the variable ) ∈ Rn , where p(ξ) √ is the (complex valued) “symbol”
p(ξ) = γ∈I , δ∈J aγδ (i ξ) , ξ ∈ Rn , i = −1. Also it is easily checked that
1.1 is equivalent (assuming this κ-homogeneity) to the requirement that

(1.10 ) / 0,
p(ξ) = ξ ∈ Rn \ {0}.

We shall not explicitly use this fact here, but the reader may check for example
that 1.1 0 =⇒ 1.1 for a κ-homogeneous operator
qP by noting that the κ-homogeneity
b
implies |p(ξ)| ≡ |ξ| |p(ξ)|, where |ξ|κ =
m
|ξi |2/κi and ξb = |ξ|−1 ξ ∈ ∂B κ (0);
κ i κ 1
then a standard regularity argument using Fourier transforms (see e.g. the dis-
cussion in [Sch77, Sect. 2.8]) establishes that 1.1 0 =⇒ 1.1. The converse is also
easy to check (e.g. by minor modification of the argument in [Sch77, Sect. 2.1]).
We now want to introduce some classes of Hölder seminorms. Let α =
(α1 , . . . , αn ) ∈ (0, 1)n be a set of “Hölder exponents” such that there is fixed
a > 0 with κj αj = a for each j = 1, . . . , n. Corresponding to such α and any
open connected Ω ⊂ Rn , we consider seminorms [u]I ,α;Ω , [f ]−J ,α;Ω , of the
form
Schauder estimates by scaling 393

X
[u]I ,α;Ω = [D γ u]α;Ω ,
γ∈I

and
X
[f ]−J ,α;Ω = inf [fδ ]α;Ω ,
δ∈J
P
where the inf is over all collections {fδ } in L1loc (Ω) such that f = δ∈J D δ fδ in
the sense of distributions on Ω and where
X
[g]α;Ω = ( sup |h|−αi |g(x + hei ) − g(x )|)
=0, x , x +hei ∈Ω
i ∈{1,...,n} h/

C I ,α (Ω) will denote the linear space of functions u ∈ C 0 (Ω) such that D γ u ∈
C 0 (Ω) for each γ ∈ If ≡ {multi-indices δ : ∃ β ∈ I with δj ≤ βj ∀ j =
1, . . . , n} and [u]I ,α;Ω < ∞. C I (Ω) denotes the linear space of functions
u ∈ C 0 (Ω) such that D γ u ∈ C 0 (Ω) for each γ ∈ If; C I is equipped with the
locally convex topology such that uk → u if and only if D γ uk → D γ u in C 0 (Ω)
(i.e. local uniform convergence) for each γ ∈ If.
We assume that the index set I is “κ-proper” in the sense that for an arbitrary
C I (Ω) function on a convex domain Ω there is a “κ-Taylor polynomial type”
expansion up to κ-order I :
X xγ X xγ
(1.2) u(x ) = D γ u(y) + D γ u(xγ ), x , y ∈ Ω,
γ! γ!
γ∈Ie\I γ∈I

for suitable xγ ∈ xy; this can then be written

X xγ
(1.20 ) u(x ) − u (I ,y) (x ) = (D γ u(xγ ) − D γ u(y)),
γ!
γ∈I

P γ
where u (I ,y) = γ∈Ie (x −y) γ
γ! D u(y). Of course in the case when all κj = 1, 1.2
is simply the usual order I Taylor approximation for u at y, and hence holds
automatically. In the “parabolic case”, when we have n + 1 in place of n with
variables (x , t) = (x 1 , . . . , x n , t), κ1 = κ2 = · · · = κn = 1, κn+1 = I , then we can
write u(x , t) = (u(x , t) − u(x , s)) + u(x , s), and using the mean-value theorem
with respect to the t variable in the first term and the usual order I Taylor
approximation at y with respect to the x -variables for u(x , s), we see that in this
case also we get 1.2 with (y, s) in place of y and (x , t) in place of x .
Of course 1.2 0 gives

(1.3) sup |u(x ) − u (I ,y) | ≤ ρI +a [u]I ,α;Bρκ (y)

Bρκ (y)

over any ball Bρκ (y) on which u is of class C I ,α ; in view of the arbitrariness of
y and ρ this evidently implies that for any ball Bρ (x0 ) ⊂ Ω we have
394 L. Simon

sup |u (I ,y) − u (I ,z ) | ≤ 2σ I +a [u]I ,α;Bρκ (x0 ) , y, z ∈ Bσκ (x0 ), σ ≤ ρ/4

Bσκ (x0 )

and hence families C of u ∈ C I ,α (Ω) with [u]I ,α bounded on compact sets are
such that D γ u, γ ∈ If, u ∈ C form an equicontinuous family on each compact
set, and hence by Arzela-Ascoli

sup(kuk kC 0 (K ) + [uk ]I ,α;K ) < ∞∀ compact K ⊂ Ω

k

(1.4) =⇒ uk 0 → u in C I (Ω)

for some subsequence {uk 0 } of {uk }. In addition we have the interpolation in-
equalities

(1.5) ρκ·γ sup |D γ u| ≤ ρI +a [u]I ,α;Bρκ (x0 ) + C sup |u|, γ ∈ If,
Bρκ (x0 ) Bρκ (x0 )

valid for any  > 0, where C depends only on I , , α; this also follows
from 1.2 0 , because 1.2 0 implies |u (I ,y) | ≤ |u| + ρI +a [u]I ,α;Bρ (x0 ) on any ball
Bρκ (x0 ) for any y ∈ Bρκ (x0 ).
The following is the basic estimate for scale elliptic constant coefficient op-
erators.

Theorem 1. Subject to the assumptions 1.1, 1.2 and the assumption that P is a
κ-homogeneous operator, there is a constant C such that

[u]I ,α;Rn ≤ C [Pu]−J ,α;Rn

for every u ∈ C I ,α (Rn ), where C = C (P , IP, J , α). Actually C can be taken

fixed over any family F of operators P = aγδ D γ+δ satisfying 1.1 and cor-
responding to compact sets of coefficients {(aγδ )(γ,δ)∈I ×J } in RM (M = the
number of elements in I × J ).

Remark. Of course, in view of 1.1 0 , the uniformity of the estimate for compact
classes of F of P is equivalent to the fact that the constant C depends only
on n, α, M , and any positive constants λ, Λ such that λ ≤ infξ∈∂B1κ (0) |p(ξ)| and
|aγδ | ≤ Λ ∀ (γ, δ) ∈ I × J .

We give the proof in Sect. 3, but first we want to illustrate the result with
some examples.

2. Examples
Pn
(1) Pu = 4u = j =1 Dj2 u, I = {γ : |γ| = 2}, κ = (1, . . . , 1), α = (µ, . . . , µ)
with µ ∈ (0, 1), J = {0}, (so [f ]−J ,α;Ω = [f ]µ;Ω , where [f ]µ;Ω denotes the
usual Hölder seminorm of f on Ω). Evidently 1.1 is satisfied in this case and 1.2
holds trivially because all κj = 1, so Theorem 1 gives
Schauder estimates by scaling 395

[D 2 u]µ;Rn ≤ C [4u]µ;Rn ,

which is the standard second order Hölder estimate for 4 on Rn .

(2) Pu = 4u again, I = J = {γ : |γ| = 1}, κ = (1, . . .P , 1), α = (µ, . . . , µ) with
µ ∈ (0, 1), (so [f ]−J ,α;Ω is equivalent to [f ]∗;Ω ≡ inf j [fj ]µ;Ω , where the inf
Pn
is over all fj such that f = j =1 Dj fj in the distribution sense, and where [fj ]µ;Ω
denotes the usual Hölder seminorm on Ω). Evidently 1.1 is again satisfied in this
case, so Theorem 1 gives

[Du]µ;Rn ≤ C [4u]∗;Rn ,

which is the standard C 1,µ Hölder estimate for 4 on Rn .

(3) Pu = ut − 4u (the heat operator) on Rn+1 , where t is the (n + 1)st coordinate
of the general point (x , t) ∈ Rn+1 , κ = (1, . . . , 1, 2), I = 2, J = 0, I = {(γ, 0) =
(γ1 , . . . , γn , 0) : |γ| = 2} ∪ {(0, . . . , 0, 1)}, α = (µ, . . . , µ, µ2 ), J = {0}, (so
[f ]−J ,α;Ω = [f ](µ,...,µ, µ2 ) ). 1.1 is again satisfied in this case, and 1.2 is also
satisfied by virtue of the discussion following 1.2 0 , so Theorem 1 gives

[ut ](µ,...,µ, µ2 );Rn + [Dx2 u](µ,...,µ, µ2 );Rn ≤ C [ut − 4u](µ,...,µ, µ2 );Rn ,

which is the standard order 2 Hölder estimate for the heat equation.

3. Proof of Theorem 1

First we make a point concerning the action of P on polynomials; for any multi-
γ
indices γ, δ we have by definition that D γ x δ = Πjn=1 Dx jj (x j )δj , and hence, as one
readily checks,

γ δ δ! if γ = δ
(3.1) D x =
0 if κ · γ ≥ κ · δ and γ=
/ δ,

the latter being true because κ · γ ≥ κ · δ implies that at least one γi is strictly
greater that the corresponding δi , unless γ = δ. Hence we see that for any v ∈ C I

P v(0) if J = {0}
(3.2) Pv (I )
≡
0 if J =
/ {0},
P γ
where v (I ) denotes the “κ-Taylor polynomial” γ∈Ie D γv(0) γ
! x .
We also need a preliminary lemma, as follows:

Lemma 1. If P is κ-homogeneous of κ-order m, if 1.1 holds, if Pu = 0 in Rn and

if there are constants C , q > 0 such that supBR (0) |u| ≤ CR q , ∀ R ≥ 1, then u is
a polynomial.
396 L. Simon

Proof. According to an argument of Hörmander [Hör55] based on the closed

graph theorem, 1.1 implies that if Ω e ⊂⊂ Ω (with Ω open in Rn ) and P (u) = 0
in Ω, then u satisfies the bounds
Z
(1) e
sup |D k u|2 ≤ C (k , P , Ω, Ω) |u|2 , k ≥ 1.
Ωe Ω

Indeed Hörmander’s argument goes as follows: Let δ be any multi-index with

|δ| = k ≥ 1, let W be the (closed, by hypoellipticity 1.1) subspace of L2 (Ω)
consisting of the L2 (Ω) ∩ C ∞ (Ω) functions w with P w = 0 in Ω, and B :
W → C 0 (Ω)e be defined by B w = D δ w| . Evidently B has a closed graph in
e
Ω
e whence (1) holds by virtue of the closed graph theorem.
W × C 0 (Ω),
By virtue of the κ-homogeneity of P , (1) evidently implies
Z
(2) sup |D γ u|2 ≤ C (κ, P , n, γ)R −2κ·γ |BRκ (y)|−1 |u|2
κ
BR/2 BRκ (y)

for any R > 0 and any multi-index γ, provided u is a solution of P (u) = 0 on

BRκ (0). (Notice that if u is a solution on BRκ (0), then u ◦S0,R is a solution on B1κ (0),
and we can therefore apply (1) with u ◦S0,R in place of u and with B1κ (0), B1/2 κ
(0)
e
in place of Ω and Ω respectively; by making the change of variable y = S0,R (x )
in the integral we then get (2) as claimed.) By letting R ↑ ∞ in (2) and using
the given upper bound on |u|, we then see that D k u = 0 for k sufficiently large,
which is the required result.

Proof of Theorem
P 1. If this (kis) false we get a sequence of κ-homogeneous op-
γ+δ
erators Pk = γ∈I , δ∈J aγδ D
(k )
u with aγδ → aγδ such that Pk and P ≡
P
aγδ D γ+δ satisfies 1.1, and a sequence {uk } of C I ,α functions with

[Pk uk ]α < k −1 [uk ]I ,α , k = 1, 2, . . .

By replacing uk by λk uk , with λk = [uk ]−1

I ,α , we assume without loss of generality
that [uk ]I ,α = 1 for each k , and hence

(1) [Pk uk ]−J ,α < k −1 , [uk ]I ,α = 1

for each k . Since

X XX
n
|D β u(x + hei ) − D β u(x )|
[u]I ,α = [D β u]α ≡ sup ,
x ∈Rn , h>0 h αi
β∈I β∈I i =1

we can select β ∈ I and i ∈ {1, . . . , n} and, for infinitely many k , there are
xk ∈ Rn and hk > 0 such that
1
hk−αi |D β uk (xk + hk ei ) − D β uk (xk )| ≥ ,
2Mn
Schauder estimates by scaling 397

where M is the number of elements in the set I . Defining uk0 = sk−I −a uk ◦ Sxk ,sk ,
1/κ
sk = hk i , and keeping in mind that κ · β + αi κi = I + a for each i = 1, . . . , n,
we then see that

(2) [uk0 ]I ,α = 1, |D β uk0 (ei ) − D β uk0 (0)| ≥ C , [Puk0 ]−J ,α ≤ k −1

for infinitely many k , where C > 0 is independent of k . Now let ũk = uk0 −(uk0 )(I ) ,
where the notation is as in 3.2. By 3.2 we have P ((uk0 )(I ) ) = const. if J = {0}
and P ((uk0 )(I ) ) = 0 if J =
/ {0}, and hence

(3) [P (ũk )]−J ,α;Rn = [P (uk )]−J ,α;Rn ≤ k −1 .

Of course we also have ũk ∈ C I ,α (Rn ) and [ũk ]I ,α;Rn = [uk0 ]I ,α;Rn = 1. Then
by (2) we have a subsequence {ũk 0 } ⊂ {ũk } such that

[ũk 0 ]I ,α = 1, |D β ũk 0 (ei )| ≥ C , (ũk )(I ) ≡ 0, [P ũk 0 ]−J ,α;Rn ≤ k −1 .

By this and the compactness 1.4 we have ũk 0 → v in the space C I (Rn ), with

v ∈ C I ,α , D β v(ei ) =
/ 0, D β v(0) = 0, [P v]−J ,α;Rn = 0, [v]I ,α;Rn ≤ 1.

But [P v]−J ,α = 0 implies P v = 0 if J = / {0} while P v = const. if J = {0};

on the other hand, in the latter case we have by 3.2 that P ũk (0)(≡ P (uk0 )(0) −
P ((uk0 )I )) = 0 for each k , and so P v(0) = lim P ũk (0) = 0 in case J = {0}. Thus
P v ≡ 0 and [v]I ,α;Rn ≤ 1 in either case, and by Lemma 1 we have then that
v is a polynomial. Thus D β v(tei ) is a non-constant polynomial in t, and hence,
since αi ∈ (0, 1),

|D β v(tei ) − D β v(0)|
sup = ∞,
t∈R\{0} |t|αi

contradicting the fact that [v]I ,α < ∞.

4. Local estimates

Here we show that only a slight modification of the scaling argument of The-
orem 1 establishes an analogous estimate locally on balls. (This can of course
also be proved by applying Theorem 1 directly to u ζ, where ζ is a C ∞ cut-off
function.)
We continue to use the notation that Bρκ (y) denotes the “κ-ball” of radius ρ
and centre y defined by

Bρκ (y) = {x ∈ Rn : |x − y|κ < ρ},

qP
n
where |x − y|κ = j =1 |x j − y j |2/κj .
398 L. Simon

Theorem 2. Suppose P is κ-homogeneous and 1.1, 1.2 hold. Then ∀ θ ∈ (0, 1),
−I −a
κ (0) ≤ C ([Pu]−J ,α;B κ (0) + ρ
[u]I ,α;Bθρ ρ
kukL∞ (Bρκ (0)) ), u ∈ C I ,α (Bρκ (0)),

where C depends only on n, θ and P ; as in Theorem 1, C can be chosen fixed for

families F of P determined by compact families of coefficients (aγδ )(γ,δ)∈I ×J .

Proof. We first establish that for any P ∈ F and any σ,  > 0, y ∈ Rn there
are δ = δ(, F , I ) ∈ (0, 1) and C = C (, F , I ) > 0 such that

(1) κ (y) ≤ [u]I ,α;B κ (y) + C [Pu]−J ,α;B κ (y) ,

[u]I ,α;Bδσ σ σ

provided u ∈ C I ,α (Bσ (y)). This is proved by very minor modification of the

argument of Sect. 3. In fact by translation and scaling x 7→ (δσ)−1 (x − y), we
see that if the result is false for some  > 0 and with δ = k −1 and β = k ,
k = 1, 2, . . . , then there is a sequence uk ∈ C I ,α (Bkκ (0)) and Pk ∈ F with

 [uk ]I ,α;Bkκ (0) + k [Pk uk ]−J ,α;Bkκ (0) < [uk ]I ,α;B1κ (0) .

In particular normalizing so that [uk ]I ,α;B1κ (y) = 1 we have that [uk ]I ,α;Bkκ (0) is
bounded independent of k . Now we can easily modify the rest of the argument
from Sect. 3 (taking xk , yk ∈ B1κ ) to give a contradiction. So (1) is established.
Now we note that (1) implies

κ (y) ≤ [u]I ,α;B κ (y) + C ([Pu]−J ,α;B κ (y) + σ

−I −a
(2) [u]I ,α;Bθσ σ σ
kukL∞ (Bσκ (y)) ),

for every θ ∈ (0, 1) and every ball Bσκ (y) ⊂ Bρκ (0), where C = C (, θ, F ). In fact
this trivially follows from (1) by considering the quotient |h|−αi |D γ u(x + hei ) −
D γ u(x )|, where x ∈ Bθσκ
(y), subject to the 2 alternatives 0 < |h| < δ(1 − θ)σ and
κ
|h| ≥ δ(1 − θ)σ. In the first instance we can apply (1) with the ball B(1−θ)σ (x )
κ
in place of Bσ (y), thus giving

|h|−αi |D γ u(x + hei ) − D γ u(x )| ≤ [u]I ,α;Bδ(1−θ)σ (x )

≤ [u]I ,α;Bσκ (y) + C [Pu]−J ,α;Bσκ (y)

in this case, while in case of the second alternative we have trivially |h|−αi |D γ u(x +
hei ) − D γ u(x )| ≤ 2(1 − θ)−1 δ −αi supBσκ (y) |D γ u|, and then we can use the inter-
polation inequality 1.5 to again deduce (2).
Now the rest of the proof easily follows if we note that by (2) we can use the
following general lemma with S (A) = [u]I ,α;A and k = I +a (= the constant value
of κ · γ + κi αi for γ ∈ I and with E = C (ρI +a [Pu]−J ,α;Bρκ (0) + kukL∞ (Bρκ (0)) ).

Lemma. If S is a monotone subadditivePN function on the convex subsets of a ball

B = Bρκ0 (y0 ) (so that S (A) ≤ j =1 S (A j ) whenever A, A1 , . . . , AN are convex
subsets of B with A ⊂ ∪Nj=1 Aj ) and if θ0 ∈ (0, 12 ], k > 0 are given constants, then
there is  = (θ0 , k , κ) ∈ (0, 1) such that if E ≥ 0 is a constant and

σ k S (Bθκ0 σ (y)) ≤ σ k S (Bσκ (y)) + E

Schauder estimates by scaling 399

for all balls Bσκ (y) ⊂ B , then for any ball Bρκ (y) ⊂ B we have

κ
ρk S (Bθρ (y) ≤ CE

for each θ ∈ (0, 1), where C = C (n, θ0 , θ, k , κ).

Remarks. Special cases of the above lemma are commonly used in many different
contexts—it is therefore somewhat surprising that it seems not to be very widely
known that a general lemma of the above type is available. A crucial point of the
above lemma is that  can be chosen independent of E ; this is important because
in many applications (and in the present application) E depends on .
(2) The same holds if B is replaced by the half-ball B1κ,+ (0) ≡ B1κ (0) ∩ {x :
x n > 0} and if all balls Bρκ (y) in the hypotheses and conclusion are replaced by
Bρκ,+ (y) ≡ Bρκ (y) ∩ {x : x n > 0}, where y is always required to lie in the upper
half space Rn+ .

Proof of Lemma. First of all note that it suffices to prove the lemma in the special
case when θ = 12 , because for any θ ∈ ( 12 , 1) and for any ball Bρκ (y) ⊂ B we can
κ
find a collection B(1−θ)ρ (yj ) ⊂ Bρκ (y), j = 1, . . . , N , with N = N (n, θ, κ) and
κ κ
with Bθρ (y) ⊂ ∪j =1 B(1−θ)ρ/2 (yj ). So assume θ = 12 , and let
N

κ
Q= sup σ k S (Bσ/2 (y)) < ∞.
Bσκ (y)⊂B

Then in view of the given inequality we have

(1) 2−k σ k S (Bθκ0 σ/2 (y)) ≤ Q + E

for each ball Bσκ (y) ⊂ B . Take any ball Bσκ (y) with Bσκ (y) ⊂ B and select points
κ κ
yj ∈ Bσ/2 (y) such that Bσ/2 (y) ⊂ ∪Nj=1 Bθκ0 σ/4 (yj ) and with N ≤ C , where C is
κ
a constant depending only on θ0 , n, κ. Since each Bσ/2 (yj ) ⊂ B , we can then
κ κ
apply (1) with Bσ/2 (yj ) in place of Bσ (y) and use the given subadditivity of S ;
this gives
κ
σ k S (Bσ/2 (y)) ≤ C (Q + E ), C = C (n, k , θ0 , κ).

In view of the arbitrariness of Bσκ (y), this gives

Q ≤ C Q + CE , C = C (n, k , θ0 , κ),

so if  ≤ 12 C −1 we get Q ≤ 2CE , and in particular

κ
σ k S (Bσ/2 (y)) ≤ 2CE ,

which is the required result (with θ = 12 ).

400 L. Simon

5. Variable coefficients and lower order terms

Of course the usual Schauder estimates for variable coefficients in the presence
of lower order terms follow easily from the local results for constant coefficients
proved in the previous section.
Indeed one can handle the case when in place of P we have a variable
coefficient operator P of the form
X X
Pu= D δ (Aγδ D γ u) + D δ (Aγδ D γ u),
(γ,δ)∈I ×J (γ,δ)∈Ie×Je, κ·(γ+δ)<m

The coefficients Aγδ appearing in the definition of P are given functions on a

κ
closed ball B ρ (0) with
X X
(5.1) ρa [Aγδ ]α + ρm−κ·(γ+δ) |Aγδ | ≤ Λ,
(γ,δ)∈I ×J (γ,δ)∈Ie×Je

where a continues to denote the common value of κi αi , i = 1, . . . , n.

κ
We also assume that, at any given point y ∈ B ρ (0), the top order part Py
P
given by Py = γ∈I , δ∈J Dxδ (Aγδ (y)Dxγ ) of the operator P satisfies the same
condition as P of Sect. 1:
κ
(5.2) ∀ y ∈ B ρ (0), Py is κ-homogeneous and 1.1 holds with P = Py .

Although we don’t explicit use it here, we point out that this P

is equivalent to the
assumption 1.1 0 , with py (ξ), the symbol defined by py (ξ) = γ∈I , δ∈J Aγδ (y)
κ
(i ξ)γ+δ , in place of p(ξ), for each y ∈ B ρ (0); that is,

(5.20 ) / 0 ∀y ∈ Bρκ (0), ξ ∈ ∂B1κ (0).

py (ξ) =

Theorem 3. Subject to the conditions of Theorem 2 and the conditions 5.1, 5.2
we have
−I −a
κ (0) ≤ C ([P u]−J ,α;B κ (0) + ρ
[u]I ,α;Bθρ ρ
kukL∞ (Bρκ (0)) ), u ∈ C I ,α (Bρκ (0))

for each θ ∈ (0, 1), where C depends only on n, β, θ, Λ, σ and P ; the dependence
on P is uniform for any class satisfying 5.1, 5.2 such that the set {Py }y∈Bρκ (0)
lies in a compact family F as in Theorem 1.

Remark. Of course, in view of 5.2 0 , the uniformity of the estimate for compact
classes F of P is equivalent to the fact that the constant C depends only on
n, β, θ, Λ, and λ, where 0 < λ ≤ infξ∈∂B1κ (0), y∈Bρκ (0) |py (ξ)| and |aγδ | ≤ Λ for
(γ, δ) ∈ I × J .
Schauder estimates by scaling 401

Proof. First note that for each y ∈ Bρκ (0) and each σ ∈ (0, ρ) such that Bσκ (y) ⊂
Bρκ (0) we can write

P = Py + Ry
P P
where Py = γ∈I , δ∈J Dxδ (Aγδ (y)Dxγ ) and where Ry = γ∈I ,δ∈J Dxδ ((Aγδ −
P
Aγδ (y))D γ ) + (γ,δ)∈S D δ (Aγδ D γ ), where S denotes If× J
f\ I × J . Notice
that then Py satisfies the same restriction as P in Sect. 4 and, for each η ∈ (0, 1),
Ry satisfies the condition that
X
(1) [Ry u]−J ,α,Bσκ (y) ≤ C Λη a [u]I ,α;Bσκ (y) + C Λ σ κ·δ−I −a sup |D δ u|
Bσκ (y)
η∈Ie

for u ∈ C I ,α (Bσκ (y)), σ ≤ ηρ. According to Theorem 2 we have

κ (y) ≤ C ([Py u]−J ,α;B κ (y) + σ
−I −a
(2) [u]I ,α;Bσ/2 σ
kukL∞ (Bσκ (y)) ).

Since Py u = P u − Ry u, this, together with (1), gives

−κ·γ
κ (y) ≤ [u]I ,α;B κ (y) + C ([P u]−J ,α;B κ (y) + σ
(3) [u]I ,α;Bσ/2 ησ ρ
sup |D γ u|)
Bρκ (y)

provided η = η(, Λ, κ) is sufficiently small, where C = C (Λ, , I ). By virtue of

the interpolation inequality 1.5 this gives
κ (y) ≤ [u]I ,α;B κ (y) + C ([P u]−J ,α;B κ (y) + σ
−I −a
(4) [u]I ,α;Bσ/2 σ σ
sup |u|)
Bσκ (y)

for any given  > 0, where C = C (, Λ, I ). But now (4) is an inequality exactly
of the form of inequality (2) in the proof of Theorem 2, hence we can repeat
exactly the same argument as in the last part of the proof of Theorem 2 to give
the required result.

6. Bounds in a half-space

Here

Rn+ = {(x 1 , . . . , x n ) : x n > 0},

N ≥ 1 is an integer, and m1 , . . . , mN are non-negative integers in {0, . . . , I }.

We consider boundary conditions

Qu = 0 on ∂Rn+

where Q has the form Q = (Q1 , . . . , QN ) with N ≥ 1 and with

X
Qj (u) = bj γ D γ , j = 1, . . . , N ,
γ∈Bj

for some constants bj γ , where Bj = { multi-indices γ in Rn : κ · γ = mj }.

402 L. Simon

Kj will denote { multi-indices δ in Rn−1 : κ · (δ, 0) = I − mj }. We assume

that the corresponding Hölder seminorms
X
[f ]Kj ,α 0 ;Rn−1 ×{0} ≡ [f ]α 0 ; Rn−1 ,
δ∈Kj

with α 0 = (α1 , . . . , αn−1 ), are are non-degenerate in the sense that

X xγ
(6.1) [f ]Kj ,α 0 = 0 =⇒ f (x ) ≡ cγ
γ!
fj
γ∈K

for suitable constants cγ , j = 1, . . . , N , where (Cf. the notation for If) K fj

denotes the set of multi-indices {γ = (γ1 , . . . , γn−1 ) : ∃ δ ∈ Kj with γj ≤
δj for each j = 1, . . . , n − 1}. Notice of course that 6.1 is automatic if all the κj
are equal to one.
Analogous to 1.1 we assume the regularity condition
κ,+
u ∈ C I (B ρ (y)) with Pu = 0 on Bρκ,+ (y) and Qu = 0 on Bρκ (y) ∩ (Rn−1 × {0})
κ,+
(6.2) =⇒ u ∈ C ∞ (B σ (y)) ∀ σ ∈ (0, ρ).

Then we have the following analogue of Theorem 1:

Theorem 4. Subject to 1.1, 1.2 and 6.1, 6.2 we have

X
N
[u]I ,α;Rn+ ≤ C ([Pu]−J ,α;Rn+ + [Qj (u)]Kj ,α0 ;Rn−1 ×{0} )
j =1

n P
for each u ∈ C I ,α (R+ ) with [u]I ,α;Rn+ < ∞; here [f ]Kj ,α0 = δ∈Kj [D δ f ]α 0 ,
where C depends on P , Q; this dependence is uniform with respect to P, Q cor-
responding to compact subsets of coefficients {aγδ } and {bj γ }.

6.3 Examples

(1) Everything as in Example

P 1 of Sect. 2, N = 1, m1 = 0, and Qu = u. Then
[Qu]K1 ,α 0 ;Rn−1 ×{0} ≡ |δ|=2 [D δ u]α 0 ;Rn−1 ×{0} and Theorem 4 gives
Pn−1
[D 2 u]µ;Rn+ ≤ C ([4u]µ;Rn+ + i ,j =1 [Di Dj u]µ;Ω∩Rn−1 ×{0} )

n
for all u ∈ C 2,µ (R+ ) which is the usual second order Hölder estimate for Pois-
son’s equation with Dirichlet boundary conditions on a half-space.
(2) Similar to (1) but with Hölder coefficients of first derivatives rather than
second. In this case the estimate is

[Du]µ;Rn+ ≤ C ([4u]∗;Rn+ + [D 0 u]µ;Ω∩Rn−1 ×{0} )

Schauder estimates by scaling 403

n
for all u ∈ C 2,µ (R+ ), where [4u]∗ is the seminorm of Example (2) of Sect. 2,
and D 0 is the gradient on Rn−1 × {0}.
(3) Similar to (1) but with m1 = 1 and Neumann boundary condition Qu = Dn u.
In this case the estimate is

[D 2 u]µ;Rn+ ≤ C ([4u]µ;Rn+ + [D 0 Dn u]µ;Rn−1 ×{0} )

(4) Similar to (2) but with Neumann boundary condition Qu = Dn u. In this case
the estimate is

[Du]µ;Rn+ ≤ C ([4u]∗;Rn+ + [Dn u]µ;Ω∩Rn−1 ×{0} ).

(5) Heat operator ut −4u with the initial data Qu = u and with other parameters
as in Example (3) of Sect. 2, with N = 1 and m1 = 0; in this case Theorem 4
gives the usual second-order Schauder parabolic estimate

[ut ](µ,...,µ, µ2 );Rn+1

+
+ [Dx2 u](µ,...,µ, µ2 );Rn+1
+

≤ C ([ut − 4u](µ,...,µ, µ2 );Rn+ + [Dx2 u](µ,...,µ, µ2 );Rn ×{0} ).

Before we begin the proof of Theorem 4, we need to establish an analogue

of Lemma 1 of Sect. 3:
n
Lemma 2. If u ∈ C I (R+ ) with Pu = 0 on Rn+ and each Qj u = 0 on Rn−1 × {0},
and if there are constants C , q such that kukC I (B κ,+ (0)) ≤ CR q for all R ≥ 1, then
R
n
P to R+ of a polynomial
u is the restriction on Rn . Here we use the notation that
kukC I (Ω∩Rn+ ) = γ∈Ie supΩ∩Rn+ |D γ u|.

Proof. We first establish (Cf. inequality (1) in the proof of Lemma 1) that if
0∈Ω e ⊂⊂ Ω ⊂⊂ Rn , then

e
sup |D k u| ≤ C (k , Ω, Ω)kukC I (Ω∩Rn+ )
e n+
Ω∩R

n
for any u ∈ C I (Ω ∩ R+ ) with Pu = 0 in Ω ∩ Rn+ and Qu = 0 on Ω ∩ Rn−1 × {0}.
This is proved by a minor variant of the argument used in the proof of Lemma 1
in Sect. 3. Namely, let C be the Banach space of functions v ∈ C 0 (Ω e ∩ Rn )
+
with kvkC ≡ supΩ∩Re n |v| < ∞, equipped with the norm kvkC and let W be
+
n n
the subspace of C I (Ω ∩ R+ ) consisting of all functions w ∈ C I (Ω ∩ R+ ) with
P w = 0 in Ω ∩ Rn+ and Qw = 0 on Ω ∩ Rn−1 × {0} and with kwkC I (Ω∩Rn+ ) < ∞,
equipped with the norm kwkC I (Ω∩Rn+ ) . Using the hypoellipticity assumptions 1.1
and 6.1 we can check that W is a Banach space, that all w ∈ W are C ∞ (Ω e ∩ Rn )
+
and that, for any multi-index γ with |γ| = k ≥ 1, the operator B : W → C
defined by B w = D γ w|Ω∩R
e n+ is well-defined and has closed graph in W × C .
But then by the closed graph theorem we obtain (1) as claimed.
In view of the κ-homogeneity of P we deduce immediately from (1) that
404 L. Simon

X
sup |D γ u| ≤ C (γ)R −κ·γ R κ·δ sup |D δ u|,
κ,+
BRκ,+ (0)
δ∈Ie
BR/2 (0)

and in view of the given bound on kukC I (B κ,+ ) this shows that D k u = 0 on Rn+
R
for all sufficiently large k , and the lemma is proved.
Proof of Theorem 4. The proof initially proceeds exactly as for Theorem 1: if
false there are sequences of operators Pk , Q (k ) and xk ∈ Rn+ , hk > 0 with
[uk ]I ,α = 1, [Pk uk ]−J ,α < k −1 , [Qj(k ) (u)]Kj ,α;Rn−1 ×{0} < k −1 , j = 1, . . . , N ,
and
1
hk−αi |D β uk (xk + hk ei ) − D β uk (xk )| ≥
2Mn
for some β ∈ I and i ∈ {1, . . . , n}, where M is the number of elements of I
as in the proof of Theorem 1. The rescaling in the present half-space setting is
just slightly different: we consider the alternatives
−κn /κi −κn /κi
(a) sup hk xkn = ∞, (b) sup hk xkn < ∞.
k k

−κ /κ
In case (a), we can assume without loss of generality that limk →∞ hk n i xkn = ∞
and we can define uk0 and uek exactly as in the proof of Theorem 1 (noting that
−κ /κ
in the present setting these are defined on the half-space x n > −hk n i xkn ),
giving a sequence uek → u locally on Rn with Pu = 0 on Rn , [u]I ,α;Rn ≤ 1 and
D β u not constant, which gives a contradiction by exactly the same argument as
before. (Notice in particular that for case (a) we did not need Lemma 2.)
In case (b) the argument is just slightly different. In this case we define
uk0 = sk−I −a uk ◦ Sxk0 ,sk where xk0 = (xk1 , . . . , xkn−1 , 0) (with sk = hk i as before), so
1/κ

that now [uk0 ]I ,α;Rn+ = 1, [Pk uk0 ]−J ,α;Rn+ < k −1 , [Qj (uk0 )]Kj ,α;Rn−1 ×{0} < k −1 ,
j = 1, . . . , N , and
1
|D β uk0 (0, tk ) − D β uk0 ((0, tk ) + ei )| ≥ ,
2Mn
−κ /κ
where tk = hk n i xkn ≥ 0 is a bounded sequence (because we are in case (b)).
Notice that we still have the boundary condition Quk0 = 0 on Rn−1 × {0}. Now
we
P can x γas before define uek (x ) = uk0 (x ) − Tk (x ), where Tk is the polynomial
γ 0
γ∈Ie γ!
D uk (0) and we note that we now have the boundary conditions

(1) [Qj(k ) uek ]Kj ,α = [Qj(k ) uk0 ]Kj ,α ≤ k −1 on Rn−1× {0}, γ ∈ Kj ,

P
(because δ∈Kj D δ Qj(k ) Tk ≡ const. for each k , j by virtue of the fact that each
D δ Qj(k ) is an operator of homogeneous κ-order I and Tk is a polynomial of
κ-degree I ). Also, we have

(2) fj ,
Dxγ0 Qj(k ) uek (0) = Dxγ0 Qj(k ) uk0 (0) − Dxγ0 Qj(k ) Tk (0) = 0, γ ∈ K

fj (⊂ If) is as in 6.1.
where x 0 = (x 1 , . . . , x n−1 ) and K
Schauder estimates by scaling 405

Now analogous to the interior case, we have (after passing to a subsequence)

that the coefficients of Pk and Qj(k ) converge to give limiting operators P , Qj , and
n
at the same time uek → u locally in R+ with [u]I ,α;Rn+ ≤ 1, D β u((0, t0 ) + ei ) = /
D u(0, t0 ) (t0 = lim tk ), Pu = 0 on Rn+ and Dxγ0 Qj u = 0 on Rn−1 × {0} for
β

γ∈K gj and j = 1, . . . , N , and [Qj u]K ,α = 0. We now complete the argument

j
essentially as for the interior case, except that we use Lemma 2 in place of
Lemma 1. Notice that the hypothesis kukC I (B κ,+ (0)) ≤ CR q needed for Lemma 2
R
is a trivial consequence of the fact that [u]I ,α;Rn+ < ∞, and the fact that Qj u = 0
on Rn−1 × {0} follows from the facts that [Qj u]Kj ,α = 0 (hence, by 6.1, Qj u
is a polynomial with exponents given by multi-indices in K fj ) and, by (2),
γ
D 0 Qj u(0) = 0 for each γ ∈ K fj . Thus we can apply Lemma 2, and we conclude
x
that u is a polynomial, and hence D β u((0, t0 ) + tei ) is a non-constant polynomial
contradicting the fact that [u]I ,α;Rn < ∞.

7. Boundary value problems

It is of course routine to go from the estimates for constant coefficient opera-

tors in a half-space to general boundary-value problems for variable-coefficient
operators.
Here we continue to use the notation

Bρκ,+ (y) = Bρκ (y) ∩ Rn+ ,

and we write

∂Bρκ,+ (y) = Sρ (y) ∪ Tρ (y),

where Sρ = ∂Rn+ ∩ ∂Bρκ (y) and Tρ = ∂Bρκ,+ (y) \ Sρ (so Tρ (y) is the “flat” part of
the boundary of ∂Bρκ,+ (y)).
We assume that Bj , Kj are as in Sect. 6 and that Q = (Q1 , . . . , QN ) is an
operator on C I ,α (Bρκ,+ (y)) of the form
X
Qj u = Qj γ D γ u
γ∈Bj ∪Lj

where

Li = {γ : κ · γ < mi }, i = 1, . . . , N ,

and
X X
(7.1) ρa [Qj γ ]Kj ,α 0 ;Tρ (0) + ρI −κ·(γ+δ) sup |Dxδ 0 Qj γ | ≤ Λ,
Tρ (0)
γ∈Bj γ∈Bj ∪Lj , δ∈Kj

where a continues to denote the common value of the κi αi .

Then we have the following theorem:
406 L. Simon

κ,+
Theorem 5. P Suppose 1.1, 1.2, 5.1, 5.2 hold, and, for each y ∈ B ρ (0), 6.1 holds
with Qj ,y ≡ γ∈Bj Qj γ (y)D γ in place of Qj , and suppose that 6.2 and 7.1 hold.
Then

[u]I ,α,B κ,+ (0) ≤ C ([P u]−J ,α,Bρκ,+ (0)

θρ

X
N
+ [Qj u]Kj ,α 0 ;Tρ (0) + ρ−I −a kukL∞ (Bρκ,+ (0)) )
j =1

for each u ∈ C I ,α (Bρ+ (0) ∪ Tρ (0)) with [u]I ,α;Bρκ,+ (0) < ∞.

The proof follows direct from Theorem 4 in a way that is completely analo-
gous to the manner in which Theorem 3 was derived from Theorem 2; the details
are left to the reader.

8. Higher order estimates

For these we just use the J = {0} estimates applied to the derivatives of u in
order to inductively prove the appropriate estimates. To get higher order boundary
estimates we also have to keep in mind that by using the equation directly
(regardless of boundary conditions) all high-order (of κ-order > I ) can be written
0
pointwise in terms of mixed derivatives of the form Dxj n Dxγ0 u for j = 0, . . . , m −1
(m = order of the equation, x 0 = (x 1 , . . . , x n−1 ), γ 0 = (γ1 , . . . , γn−1 )), and of
course these are obtained by applying the J = {0} estimates to v = Dxγ0 u, which
satisfies boundary conditions Qj v = Dxγ0 fj , where fj = Qj u (which is known).

9. Systems of equations

It is very easy to check that all the foregoing carries over, with only notational
changes, to the case of systems of equations (i.e. to the
P case when u is vector-
valued u = (u1 , . . . , uM ) for some M ) and when P = aγδ D γ+δ with now aγδ
matrices of size L × M for some L.

References

[ADN59] S. Agmon, A. Douglis, L. Nirenberg. Estimates near the boundary for solutions of elliptic
partial differential equations satisfying general boundary conditions I. Comm. Pure Appl.
Math. 12 (1959), 623–727.
[ADN64] S. Agmon, A. Douglis, L. Nirenberg. Estimates near the boundary for solutions of elliptic
partial differential equations satisfying general boundary conditions II. Comm. Pure Appl.
Math. 17 (1964), 35–92.
[Gia93] M. Giaquinta. Introduction to regularity theory of non-linear elliptic systems. Lectures in
Mathematics, ETH Zürich. Birkhäuser, Basel, 1993.
[GT83] D. Gilbarg, N. Trudinger. Elliptic Partial Differential Equations of Second Order. Springer-
Verlag, 1983.
Schauder estimates by scaling 407

[Hör55] L. Hörmander. On the theory of partial differential operators. Acta Math. 94 (1955),
161–248.
[Mor66] C. B. Morrey. Multiple integrals in the calculus of variations. Springer Verlag, Berlin-
Heidelberg-New York, 1966.
[Sch77] M Schechter. Modern methods in partial differential equations. McGraw-Hill, 1977.
[Tru86] N Trudinger. A new approach to the Schauder estimates for linear elliptic equations.
volume 14 of Proceedings of the CMA, pages 52–59. Australian National University,
Canberra, 1986.