194 7 The Hille–Yosida Theorem

A is maximal monotone ⇐⇒ A is maximal monotone

⇐⇒ A is closed, D(A) is dense, A and A are monotone.

A more general version of this result appears in Problem 16.

• Theorem 7.7. Let A be a self-adjoint maximal monotone operator. Then for every
u0 ∈ H there exists a unique function5

u ∈ C([0, +∞); H ) ∩ C 1 ((0, +∞); H ) ∩ C((0, +∞); D(A))

such that ⎧
⎨ du
+ Au = 0 on (0, +∞),
dt
⎩u(0) = u .
0

Moreover, we have
 
 du  1
|u(t)| ≤ |u0 | and  (t) = |Au(t)| ≤ |u0 | ∀t > 0,

dt t
(26) u ∈ C k ((0, +∞); D(A )) ∀k,  integers.

Proof.
Uniqueness. Let u and u be two solutions. By the monotonicity of A we see
that ϕ(t) = |u(t) − u(t)|2 is nonincreasing on (0, +∞). On the other hand, ϕ is
continuous on [0, +∞) and ϕ(0) = 0. Thus ϕ ≡ 0.
Existence. The proof is divided into two steps:
Step 1. Assume first that u0 ∈ D(A2 ) and let u be the solution of (6) given by
Theorem 7.4. We claim that
 
 du  1
(27)  (t) ≤ |u0 | ∀t > 0.
 dt  t

As in the proof of Proposition 7.6 we have

Jλ = Jλ and Aλ = Aλ ∀λ > 0.

We go back to the approximate problem introduced in the proof of Theorem 7.4:

duλ
(28) + Aλ uλ = 0 on [0, +∞), uλ (0) = u0 .
dt
Taking the scalar product of (28) with uλ and integrating on [0, T ], we obtain

5 Let us emphasize the difference between Theorems 7.4 and 7.7. Here u0 ∈ H (instead of u0 ∈
D(A)); the conclusion is that there is a solution of (6), which is smooth away from t = 0. However,
dt (t)| may possibly “blow up” as t → 0.
| du
7.4 The Self-Adjoint Case 195
 T
1 1
(29) |uλ (T )|2 + (Aλ uλ , uλ )dt = |u0 |2 .
2 0 2

Taking the scalar product of (28) with t du

dt and integrating over [0, T ], we obtain
λ

    T 
T  duλ 2 duλ
(30)  (t) t dt + A u (t), (t) t dt = 0.
 dt  λ λ
dt
0 0

But
     
d duλ duλ duλ
(Aλ uλ , uλ ) = Aλ , uλ + Aλ uλ , = 2 Aλ uλ , ,
dt dt dt dt

since Aλ = Aλ . Integrating the second integral in (30) by parts, we are led to
 T  
duλ 1 Td
Aλ uλ (t), (t) t dt = [(Aλ uλ , uλ )]t dt
0 dt 2 0 dt
(31)  T
1 1
= (Aλ uλ (T ), uλ (T )) T − (Aλ uλ , uλ ) dt.
2 2 0

On the other hand, since the function t → | du λ

dt (t)| is nonincreasing (by Lemma 7.1),
we have
 T   2 2
 duλ 2  
(32)  (t) t dt ≥  duλ (T ) T .
 dt   dt  2
0

Combining (29), (30), (31), and (32), we obtain

 2
1  duλ  1
|uλ (T )|2 + T (Aλ uλ (T ), uλ (T )) + T 2  (T ) ≤ |u0 |2 ;
2 dt 2

it follows, in particular, that

 
 duλ  1
 
(33)  dt (T ) ≤ T |u0 | ∀T > 0.

Finally, we pass to the limit in (33) as λ → 0. This completes the proof of (27), since
duλ
dt → dt (see Step 5 in the proof of Theorem 7.4).
du

Step 2. Assume now that u0 ∈ H . Let (u0n ) be a sequence in D(A2 ) such that
u0n → u0 (recall that D(A2 ) is dense in D(A) and that D(A) is dense in H ; thus
D(A2 ) is dense in H ). Let un be the solution of
⎧
⎨ dun
+ Aun = 0 on [0, +∞),
dt
⎩u (0) = u .
n 0n

We know (by Theorem 7.4) that

276 9 Sobolev Spaces and the Variational Formulation of Elliptic BVPs in N Dimensions

Then fix any function ψ ∈ Cc1 () ˜ such that ψ = 1 on . Denote by P u the

function ψ ũ extended to R by 0 outside .
2 ˜ It is easily shown that the operator
P : W 1,p () → W 1,p (R2 ) satisfies (i), (ii), and (iii).

The next lemma is very useful.

Lemma 9.3 (partition of unity). Let  be a compact subset of RN and let U1 , U2 ,

. . . , Uk be an open covering of , i.e.,  ⊂ ki=1 Ui . Then there exist functions θ0 ,
θ1 , θ2 , . . . , θk ∈ C ∞ (RN ) such that


k
(i) 0 ≤ θi ≤ 1 ∀i = 0, 1, 2, . . . , k and θi = 1 on RN ,
i=0

supp θi is compact and supp θi ⊂ Ui ∀i = 1, 2, . . . ,
(ii)
supp θ0 ⊂ RN \.

If  is an open bounded set and  = ∂, then θ0| ∈ Cc∞ ().

Proof. This lemma is classical; similar statements can be found, for example, in
S. Agmon [1], R. Adams [1], G. Folland [1], P. Malliavin [1].

Proof of Theorem 9.7. We “rectify”  = ∂ by local charts and use a partition of

unity.7 More precisely, since  is
compact and of class C 1 , there exist open sets
k
(Ui )1≤i≤k in R such that  ⊂ i=1 Ui and bijective maps Hi = Q → Ui such
N

that

Hi ∈ C 1 (Q), Hi−1 ∈ C 1 (U i ), Hi (Q+ ) = Ui ∩ , and Hi (Q0 ) = Ui ∩ .

Consider the functions θ0 , θ1 , θ2 , . . . , θk introduced in Lemma 9.3. Given u ∈

W 1,p (), write


k 
k
u= θi u = ui , where ui = θi u.
i=0 i=0

Now we extend each of the functions ui to RN , distinguishing u0 and (ui )1≤i≤k .

(a) Extension of u0 . We define the extension of u0 to RN by


u0 (x) if x ∈ ,
ū0 (x) =
0 if x ∈ RN \.

7 In the following we shall often use this technique to transfer a result proved on RN
+ (or Q+ ) to the
same conclusion on a smooth open set .
9.2 Extension Operators 277

Recall that θ0 ∈ C 1 (RN ) ∩ L∞ (RN ), ∇θ0 ∈ L∞ (RN ), since ∇θ0 = − ki=1 ∇θi
has compact support, and that supp θ0 ⊂ RN \. It follows (by Remark 4(ii)) that

∂ ∂u ∂θ0
ū0 ∈ W 1,p (RN ) and ū0 = θ0 + ū.
∂xi ∂xi ∂xi
Thus

ū0
W 1,p (RN ) ≤ C
u
W 1,p () .
(b) Extension of ui , 1 ≤ i ≤ k.
Consider the restriction of u to Ui ∩  and “transfer” this function to Q+ with the
help of Hi . More precisely, set vi (y) = u(Hi (y)) for y ∈ Q+ . We know (Proposition
9.6) that vi ∈ W 1,p (Q+ ). Then define the extension on Q by reflection of vi (Lemma
9.2); call it vi . We know that vi ∈ W 1,p (Q). “Retransfer” vi to Ui using Hi−1 and
call it wi :
wi (x) = vi [Hi−1 (x)] for x ∈ Ui .
Then wi ∈ W 1,p (Ui ), wi = u on Ui ∩ , and

wi
W 1,p (Ui ) ≤ C
u
W 1,p (Ui ∩) .

Finally, set for x ∈ RN ,


θi (x)wi (x) if x ∈ Ui ,
ûi (x) =
0 if x ∈ RN \Ui ,

so that ûi ∈ W 1,p (RN ) (see Remark 4(ii)), ûi = ui on , and

ûi
W 1,p (RN ) ≤ C
u
W 1,p (Ui ∩) .
N
(c) Conclusion. The operator P u = ū0 + i=1 ûi possesses all the desired properties.

• Corollary 9.8 (density). Assume that  is of class C 1 , and let u ∈ W 1,p () with
1 ≤ p < ∞. Then there exists a sequence (un ) from Cc∞ (RN ) such that un| → u
in W 1,p (). In other words, the restrictions to  of Cc∞ (RN ) functions form a dense
subspace of W 1,p ().

Proof. Suppose first that  is bounded. Then there exists an extension operator P (by
Theorem 9.7). The sequence8 ζn (ρn  P u) converges to P u in W 1,p (RN ) and thus it
answers the problem. When  is not bounded we start by considering the sequence
ζn u. Given ε > 0, fix n0 such that
ζn0 u − u
W 1,p < ε. One may then construct an
extension v ∈ W 1,p (RN ) of ζn0 u (since this only involves the intersection of  with
a large ball). We finally pick any w ∈ Cc∞ (RN ) such that
w − v
W 1,p (RN ) < ε.

8 As usual, (ρ
n ) is a sequence of mollifiers and (ζn ) is a sequence of cut-off functions as in the proof
of Theorem 9.2.
278 9 Sobolev Spaces and the Variational Formulation of Elliptic BVPs in N Dimensions

9.3 Sobolev Inequalities

In Chapter 8 we saw that if  has dimension 1, then W 1,p () ⊂ L∞ () with
continuous injection, for all 1 ≤ p ≤ ∞. In dimension N ≥ 2 this inclusion is true
only for p > N; when p ≤ N one may construct functions in W 1,p that do not
belong to L∞ (see Remark 16). Nevertheless, an important result, essentially due

to Sobolev, asserts that if 1 ≤ p < N then W 1,p () ⊂ Lp () with continuous
injection, for some p ∈ (p, +∞). This result is often called the Sobolev embedding
theorem. We begin by considering the following case:
A. The case  = R N .

• Theorem 9.9 (Sobolev, Gagliardo, Nirenberg). Let 1 ≤ p < N . Then

 1 1 1
W 1,p (RN ) ⊂ Lp (RN ), where p is given by 
= − ,
p p N

and there exists a constant9 C = C(p, N) such that

(17)
u
p ≤ C
∇u
p ∀u ∈ W 1,p (RN ).

Remark 10. The value p can be obtained by a very simple scaling argument (scaling
arguments, dear to the physicists, sometimes give useful information with a minimum
of effort). Indeed, assume that there exist constants C and q (1 ≤ q ≤ ∞) such that

(18)
u
q ≤ C
∇u
p ∀u ∈ Cc∞ (RN ).

Then necessarily q = p  . To see this, fix any function u ∈ Cc∞ (RN ), and plug into
(18) uλ (x) = u(λx). We obtain
(1+ Nq − N
p)

u
q ≤ Cλ
∇u
p ∀λ > 0,

which implies 1+ Nq − Np = 0, i.e., q = p (provided u does not vanish identically).

The proof of Theorem 9.9 relies on the following lemma:

Lemma 9.4. Let N ≥ 2 and let f1 , f2 , . . . , fN ∈ LN−1 (RN−1 ). For x ∈ RN and

1 ≤ i ≤ N set

x̃i = (x1 , x2 , . . . , xi−1 , xi+1 , . . . , xN ) ∈ RN−1 ,

i.e., xi is omitted from the list. Then the function

f (x) = f1 (x̃1 )f2 (x̃2 ) · · · fN (x̃N ), x ∈ RN ,

belongs to L1 (RN ) and

9 We can take C(p, N ) = (N − 1)p/(N − p), but this constant is not optimal. The best constant
is known (but it is not simple!), see Th. Aubin [1], G. Talenti [1], and E. Lieb [1].
9.3 Sobolev Inequalities 279

&
N

f
L1 (RN ) ≤
fi
LN−1 (RN−1 ) .
i=1

Proof. The case N = 2 is trivial (why?). Let us consider the case N = 3.

We have
 
|f (x)|dx3 = |f3 (x1 , x2 )| |f1 (x2 , x3 )||f2 (x1 , x3 )|dx3
R R
 1/2  1/2
≤ |f3 (x1 , x2 )| |f1 (x2 , x3 )|2 dx3 |f2 (x1 , x3 )|2 dx3
R R

(by Cauchy–Schwarz). Applying Cauchy–Schwarz once more gives


|f (x)|dx ≤
f3
L2 (R2 )
f1
L2 (R2 )
f2
L2 (R2 ) .
R3

The general case is obtained by induction—assuming the result for N and then
deducing it for N + 1. Fix xN+1 ∈ R; because of Hölder’s inequality,

|f (x)|dx1 dx2 · · · dxN
RN
- .1/N 
N
≤
fN+1
LN (RN ) |f1 f2 · · · fN | dx1 dx2 · · · dxN

(with N  = N/(N − 1)). Applying the induction assumption to the functions

  
|f1 |N , |f2 |N , . . . , |fN |N , we obtain
 &
N
N N 
|f1 | · · · |fN | dx1 · · · dxN ≤
fi
N
LN (RN−1 )
,
RN i=1

from which it follows that

 &
N
|f (x)|dx1 · · · dxN ≤
fN+1
LN (RN )
fi
LN (RN−1 ) .
RN i=1

Now vary xN+1 . Each of the functions xN+1 →

fi
LN (RN−1 ) belongs to LN (R),
1
1 ≤ i ≤ N. As a consequence, their product N i=1
fi
LN (RN−1 ) belongs to L (R)
1

(see Remark 2 following Hölder’s inequality in Chapter 4) and

 &
N+1
|f (x)|dx1 dx2 · · · dxN dxN+1 ≤
fi
LN (RN ) .
R N+1 i=1

Proof of Theorem 9.9. We begin with the case p = 1 and u ∈ Cc1 (RN ). We
have
280 9 Sobolev Spaces and the Variational Formulation of Elliptic BVPs in N Dimensions
  x1 
 ∂u 

|u(x1 , x2 , . . . , xN )| =  (t, x2 , . . . , xN )dt 
−∞ ∂x1
 +∞  
 ∂u 
≤  
 ∂x (t, x2 , . . . , xN )dt,
−∞ 1

and similarly, for each 1 ≤ i ≤ N ,

 +∞  
 ∂u  def
|u(x1 , x2 , . . . , xN )| ≤  
 ∂x (x1 , x2 , . . . , xi−1 , t, xi+1 , . . . xN )dt ≡ fi (x̃i ).
−∞ i

Thus
&
N
|u(x)| ≤
N
fi (x̃i ).
i=1
We deduce from Lemma 9.4 that
 & N 
&

N
 ∂u 1/(N−1)
|u(x)| N/(N−1)
dx ≤
1/(N−1)

fi
L1 (RN−1 ) =  
 ∂x  1 N .
RN i=1 i L (R )
i=1

As a consequence, we have
N 
&

 ∂u 1/N
(19)
u
LN/(N−1) (RN ) ≤   .
 ∂x  1
i=1 i L (RN )

This completes the proof of (17) when p = 1 and u ∈ Cc1 (RN ). We turn now to the
case 1 < p < N , still with u ∈ Cc1 (RN ). Let m ≥ 1; apply (19) to |u|m−1 u instead
of u. We obtain

&N   N 
&

 m−1 ∂u 1/N  ∂u 1/N
(20)
u
m ≤ m  |u|  ≤ m
u
m−1  
mN/(N −1)  ∂x  p (m−1)  ∂x  .
i=1 i 1 i=1 i p

Then choose m such that mN/(N − 1) = p (m − 1), which gives m =

(N − 1)p  /N (m ≥ 1 since 1 < p < N). We obtain
N 
&

 ∂u 1/N

u
p ≤ m  
 ∂x  ,
i p
i=1

and thus

u
p ≤ C
∇u
p ∀u ∈ C 1 (RN ).
To complete the proof let u ∈ W 1,p (RN ), and let (un ) be a sequence from Cc1 (RN )
such that un → u in W 1,p (RN ). One can also suppose, by extracting a subsequence
if necesary, that un → u a.e. We have

un
p ≤ C
∇un
p .
9.5 Variational Formulation of Some Boundary Value Problems 297

Step B: Existence and uniqueness of the weak solution.

Proposition 9.24. For every f ∈ L2 (), there exists a unique weak solution u∈H1 ()
of (44). Furthermore, u is obtained by
   
1
min (|∇v|2 + v 2 ) − fv .
v∈H 1 () 2  

Proof. Apply Lax–Milgram in H = H 1 ().

Step C: Regularity of the weak solution.

This will be discussed in Section 9.6.
Step D: Recovery of a classical solution.
If u ∈ C 2 () is a weak solution of (44), we have from (46)
  
∂u
(47) (−u + u)v + vdσ = f v ∀v ∈ C 1 ().
  ∂n 

In (47) first choose v ∈ Cc1 () to deduce

−u + u = f in .

Then return to (47) with v ∈ C 1 (); one obtains


∂u
vdσ = 0 ∀v ∈ C 1 ()
 ∂n

and therefore ∂u
∂n = 0 on .

Example 5 (unbounded domains). In the case that  is an unbounded open set in RN

one imposes—in addition to the usual boundary conditions on  = ∂—a boundary
condition at infinity, for example u(x) → 0 as |x| → ∞. This “translates,” at the
level of a weak solution, by the condition u ∈ H 1 . Of course, one must first prove
that if u is a classical solution such that u(x) → 0 as |x| → ∞, then u must belong to
H 1 (see the discussion in Example 8 of Chapter 8). Here are a few typical examples:

(a)  = R N ; given f ∈ L2 (RN ) the equation

−u + u = f in RN

has a unique weak solution in the following sense:

  
u ∈ H (R ) and
1 N
∇u · ∇v + uv = fv ∀v ∈ H 1 (RN ).
RN RN RN

+ ; given f ∈ L (R+ ) the problem

(b)  = R N 2 N
298 9 Sobolev Spaces and the Variational Formulation of Elliptic BVPs in N Dimensions

−u + u = f in RN+,
u(x  , 0) = 0 for x  ∈ RN−1 ,

has a unique weak solution in the following sense:

  
u ∈ H0 () and
1
∇u · ∇v + uv = fv ∀v ∈ H01 ().
  

+ ; given f ∈ L (R+ ) the problem

(c)  = R N 2 N

⎧
⎨ −u + u = f in RN
+,
∂u 
⎩ (x , 0) = 0 for x  ∈ RN−1 ,
∂xN
has a unique weak solution in the following sense:
  
u ∈ H () and
1
∇u · ∇v + uv = fv ∀v ∈ H 1 ().
  

9.6 Regularity of Weak Solutions

Definition. We say that an open set  is of class C m , m ≥ 1 an integer, if for every

x ∈  there exist a neighborhood U of x in RN and a bijective mapping H : Q → U
such that

H ∈ C m (Q), H −1 ∈ C m (U ), H (Q+ ) = U ∩ , H (Q0 ) = U ∩ .

We say that  is of class C ∞ if it is of class C m for all m.

The main regularity results are the following.

• Theorem 9.25 (regularity for the Dirichlet problem). Let  be an open set of
class C 2 with  bounded (or else  = RN+ ). Let f ∈ L () and let u ∈ H0 ()
2 1

satisfy
  
(48) ∇u∇ϕ + uϕ = f ϕ ∀ϕ ∈ H01 ().
  

Then u ∈ H 2 () and

u
H 2 ≤ C
f
L2 , where C is a constant depending only on
. Furthermore, if  is of class C m+2 and f ∈ H m (), then

u ∈ H m+2 () and

u
H m+2 ≤ C
f
H m .

In particular, if f ∈ H m () with m > N/2, then u ∈ C 2 (). Finally, if  is of class

C ∞ and if f ∈ C ∞ (), then u ∈ C ∞ ().
9.6 Regularity of Weak Solutions 299

Theorem 9.26 (regularity for the Neumann problem). With the same assumptions
as in Theorem 9.25 one obtains the same conclusions for the solution of the Neumann
problem, i.e., for u ∈ H 1 () such that
  
(49) ∇u · ∇ϕ + uϕ = f ϕ ∀ϕ ∈ H 1 ().
  

Remark 24. One would obtain the same conclusions for the solution of the Dirichlet
(or Neumann) problem associated to a general second-order elliptic operator, i.e., if
u ∈ H01 () is such that
     
∂u ∂ϕ ∂u
aij + ai ϕ+ a0 uϕ = f ϕ ∀ϕ ∈ H01 ();
 ∂xi ∂xj  ∂xi  
i,j i

then26

[f ∈ L2 (), aij ∈ C 1 () and ai ∈ C()] ⇒ u ∈ H 2 (),

and for m ≥ 1,

[f ∈ H m (), aij ∈ C m+1 () and ai ∈ C m ()] ⇒ u ∈ H m+2 ().

We shall prove only Theorem 9.25; the proof of Theorem 9.26 is entirely analo-
gous. The main idea of the proof is the following. We consider first the case  = RN ,
then the case  = RN+ . For a general domain  we proceed in two steps:
1. Interior regularity, i.e., u is regular on every domain ω ⊂⊂ . Here, the proof
follows the same pattern as  = RN .
2. Boundary regularity, i.e., u is regular on some neighborhood of the boundary.
Here, the proof resembles, in local charts, the case  = RN+.

We recommend that the reader study well the cases  = RN and  = RN+ before
tackling the general case. The plan of this section is the following:
A. The case  = RN .
B. The case  = RN +.
C. The general case:
C1 . Interior estimates.
C2 . Estimates near the boundary.
The essential ingredient of the proof is the method of translations27 due to
L. Nirenberg.
A. The case  = R N .

Notation. Given h ∈ RN , h = 0, set

26 If  is not bounded we must also assume that D α aij ∈ L∞ () ∀α, |α| ≤ m + 1 and D α ai ∈
L∞ () ∀α, |α| ≤ m.
27 Also called the technique of difference quotients.