Journal of Functional Analysis 180, 347403 (2001)

doi:10.1006jfan.2000.3698, available online at http:www.idealibrary.com on

The Dirichlet Problem for the Total Variation Flow

F. Andreu
Departamento de Analisis Matematico, Universitat de Valencia,
46100 Burjassot, Valencia, Spain

C. Ballester and V. Caselles

Department of Tecnologia, University of Pompeu-Fabra,
La Rambla 30-32, 08002 Barcelona, Spain

and

J. M. Mazon
Departamento de Analisis Matematico, Universitat de Valencia,
46100 Burjassot, Valencia, Spain
Communicated by H. Brezis
Received November 29, 1999; revised July 10, 2000; accepted October 6, 2000

We introduce a new concept of solution for the Dirichlet problem for the total
variational flow named entropy solution. Using Kruzhkov's method of doubling
variables both in space and in time we prove uniqueness and a comparison prin-
ciple in L 1 for entropy solutions. To prove the existence we use the nonlinear semi-
group theory and we show that when the initial and boundary data are nonnegative
the semigroup solutions are strong solutions.  2001 Academic Press

1. INTRODUCTION

Suppose that 0 is an open bounded domain with a Lipschitz boundary

and . # L  (0). Let %: 0  R N be a vector field (whose smoothness will
be made precise below) with |%| 1. Recently, in [7], a variational method
was proposed to extend the data . from 0 to a function u in 0 along the
integral curves of % =, the vector orthogonal to %, so that u is constant
along the integral curves of % =. Formally, we think of % as the vector field
made by the normals to the level sets of u, i.e., the sets [x # 0 : u(x)*],
* # R. In that case we would have that % } Du= |Du|. In the case that u is
a function of bounded variation, almost all levels sets are of finite perimeter
and, therefore, one can compute the normal along the boundary of the
level sets (modulo a set of H N&1 null measure). Moreover, to get . as a
trace of a function u in 0, the right function space is BV(0), the space of
347
0022-123601 35.00
Copyright  2001 by Academic Press
All rights of reproduction in any form reserved.
348 ANDREU ET AL.

functions of bounded variation in 0. Thus, to extend . from 0 to 0, it

was proposed in [7] to minimize the functional F(u)= 0 |{u| & 0 % } {u
defined in the set of functions of bounded variation BV(0) whose trace at
the boundary is given by .. Formally, if we integrate by parts in the second
term of F(u) we obtain

F(u)= |0
|{u| + | 0
div(%) } u& |
0
% } n u.

Since u, % are known at the boundary, minimizing F amounts to minimizing

E(u)= |
0
|{u| + |0
div(%) } u.

Let us comment on the class of admissible functions where E has to be

minimized. We assume that div(%) # L 1 (0) and . # L  (0). It seems
reasonable to impose that the solution u is a bounded function with an L 
bound given by &.&  . Then the second integral in the definition of E(u)
is well defined. The first integral requires the use of the space of bounded
variation functions. Thus our admissible class is A=[u # BV(0) : |u(x)| 
&.&  a.e. u| 0 =.]. The final model is [7]

Minimize
u#A
|0
|{u| + | 0
div(%) } u. (1.1)

As is well know [16, 19] the solution of this problem has to be understood
in a weak sense as the solution of the problem

Minimize
u # BV(0) |0
|{u| + | 0
div(%) } u+ |0
|u&.| dH 1. (1.2)
|u| &.&

Existence for this variational problem was proved in [19, Theorem 1.4]
when % # L 1loc(0) 2, div(%) # L 1 (R 2 ), . # L  (0).
This is one of our motivations to study the Dirichlet problem

u Du
t
=div
\ +
|Du|
+ f (t, x) in Q=(0, )_0

u(t, x)=.(x) on S=(0, )_0 (1.3)

u(0, x)=u 0 (x) in x # 0,

 TOTAL VARIATION FLOW 349

where u 0 # L 1 (0) and . # L  (0). This evolution equation is related to

the gradient descent method used to minimize the functional (1.2), if we
forget about the constraint |u| &.&  . The constraint would introduce a
further term in (1.3) but will not change the nature of the difficulties related
to the solution of the PDE. We shall even make a further simplification,
since we shall consider f (t, x)=0. Hence, our aim is to study existence and
uniqueness of solutions of the Dirichlet problem

u Du
t
=div
\ +
|Du|
in Q=(0, )_0

u(t, x)=.(x) on S=(0, )_0 (1.4)

u(0, x)=u 0 (x) in x # 0,

where 0 is an open bounded domain with a Lipschitz boundary, u 0 #

L 1 (0), and . # L  (0).
The other motivation for the study of (1.4) comes from [2], [3], and
[8]. The general purpose of the works [8] and [3] is the study of elliptic
and parabolic problems in divergence form with initial data in L 1. Exist-
ence and uniqueness results of entropy solutions when the associated varia-
tional energy has a growth at infinity of order p with p>1 are proved (see
also [4, 11]). In [2], the authors consider the equation

Du
u t =div
\ |Du| + (1.5)

in an open bounded Lipschitz domain with Neumann boundary condi-

tions, proving existence and uniqueness of entropy (or renormalized) solu-
tions (called weak solutions in [2]). Let us recall that this PDE appears
when one uses the steepest descent method to minimize the total variation,
a method introduced by Rudin and Osher [24, 25] in the context of image
denoising and reconstruction. The main point is that, in the case of
Neumann boundary conditions, this equation generates a nonlinear con-
traction semigroup in L 1 (0) which is homogeneous of degree 0, a fact
related to the regularity in time of the solutions on (1.5). Indeed, the
homogeneity of the operator permits one to conclude that u t (t) # L 1 (0) a.e.
for t>0. This was used to prove uniqueness of solutions of (1.5) in the case
of Neumann boundary conditions. This property is loosed when we con-
sider the case of Dirichlet boundary conditions. Thus, a different approach
is needed and we believe it to be helpful with a view to the general case of
energy functionals with linear growth in |Du|. A result about existence and
uniqueness of solutions (named pseudosolutions) for the Dirichlet problem
in the case of energy functionals with linear growth in |Du|, concretely for
350 ANDREU ET AL.

the Dirichlet problem for the time-dependent minimal surface equation, is

studied in [22].
The aim of this paper is to introduce a new concept of solution of the
problem (1.4), for which existence and uniqueness for initial data in L 1 (0)
are proved.
The paper is organized as follows: in Section 2 the results we need about
functions of bounded variation are summarized. In the next section we give
the definition of entropy solution and we state the main result. In Sections
4 and 5 we study the problem from the point of view of nonlinear semi-
group theory, showing that for initial data in L 2 (0), the semigroup solu-
tion is a strong solution. The next section is devoted to prove the existence
and uniqueness of entropy solutions. Finally, in the last section we obtain
that the time derivative of the entropy solution is an L 1loc function when the
initial data are nonnegative.

2. DEFINITIONS AND PRELIMINARY FACTS

To make precise our notion of solution let us recall several facts concerning
functions of bounded variation.
A function u # L 1 (0) whose partial derivatives in the sense of distribu-
tions are measures with finite total variation in 0 is called a function of
bounded variation. The class of such functions will be denoted by BV(0).
Thus u # BV(0) if and only if there are Radon measures + 1 , ..., + N defined
in 0 with finite total mass in 0 and

|0
uD i . dx=& | 0
. d+ i (2.1)

for all . # C 
0 (0), i=1, ..., N. Thus the gradient of u is a vector valued
measure with finite total variation

&Du&=sup
{|0
u div . dx : . # C  N
0 (0, R , |.(x)| 1 for x # 0.)
= (2.2)

The space BV(0) is endowed with the norm

&u& BV =&u& L1(0) +&Du&. (2.3)

For further information concerning functions of bounded variation we refer

to [1, 17, and 28].
We shall need several results from [5]. Following [5], let

X(0)=[z # L  (0, R N ) : div(z) # L 1 (0)]. (2.4)

 TOTAL VARIATION FLOW 351

If z # X(0) and w # BV(0) & L  (0) we define the functional (z, Dw) :
C0 (0)  R by the formula

( (z, Dw), .) =& | 0

w. div(z) dx& | 0
wz } {. dx. (2.5)

Then (z, Dw) is a Radon measure in 0,

|0
(z, Dw)= |0
z } {w dx (2.6)

for all w # W 1, 1 (0) & L  (0), and

}|B
(z, Dw) 
} | B
|(z, Dw)| &z&  | B
&Dw& (2.7)

for any Borel set B0. Moreover, (z, Dw) is absolutely continuous with
respect to &Dw& with the RadonNikodym derivative %(z, Dw, x) which is
a &Dw& measurable function from 0 to R such that

|B
(z, Dw)= |B
%(z, Dw, x) &Dw& (2.8)

for any Borel set B0. We also have that

&%(z, Dw, } )& L(0, &Dw&) &z& L(0, RN ) . (2.9)

In [5], a weak trace on 0 of the normal component of z # X(0) is

defined. Concretely, it is proved that there exists a linear operator
#: X(0)  L  (0) such that

&#(z)&  &z& 
#(z)(x)=z(x) } &(x) for all x # 0 if z # C 1 (0, R N ).

We shall denote #(z)(x) by [z, &](x). Moreover, the following Green's

formula relating the function [z, &] and the measure (z, Dw), for z # X(0)
and w # BV(0) & L  (0), is established:

|0
w div(z) dx+ |0
(z, Dw)= |0
[z, &] w dH N&1. (2.10)

We also need to introduce, as in [5], a weak trace on 0 of the normal

component of certain vector fields in 0. We define the space

Z(0) :=[(z, !) # L  (0, R N )_BV(0)* : div(z)=! in D$(0)].

352 ANDREU ET AL.

We denote R(0) :=W 1, 1 (0) & L  (0) & C(0). For (z, !) # Z(0) and
w # R(0) we define

( (z, !), w) 0 :=( !, w) BV(0)*, BV(0) + | 0

z } {w.

Then, working as in the proof of Theorem 1.1 of [5], we obtain that if w,

v # R(0) and w=v on 0 one has

( (z, !), w) 0 =( (z, !), v) 0 \(z, !) # Z(0). (2.11)

As a consequence of (2.11), we can give the following definition: Given

u # BV(0) & L  (0) and (z, !) # Z(0), we define ( (z, !), u) 0 by setting

( (z, !), u) 0 :=( (z, !), w) 0 ,

where w is any function in R(0) such that w=u on 0. Again, working
as in the proof of Theorem 1.1. of [5], we can prove that for every
(z, !) # Z(0) there exists M z, ! >0 such that

|( (z, !), u) 0 | M z, ! &u& L1(0) \u # BV(0) & L  (0). (2.12)

Now, taking a fixed (z, !) # Z(0), we consider the linear functional F :

L  (0)  R defined by

F(v) :=( (z, !), w) 0 ,

where v # L (0) and w # BV(0) & L  (0) is such that w |0 =v. By
estimate (2.12), there exists # z, ! # L  (0) such that

F(v)= | 0
# z, ! (x) v(x) dH N&1.

Consequently there exists a linear operator #: Z(0)  L  (0), with #(z, !)

:=# z, ! , satisfying

( (z, !), w) 0 = | 0
# z, ! (x) w(x) dH N&1 \w # BV(0) & L  (0).

In the case z # C 1 (0, R N ), we have # z (x)=z(x) } &(x) for all x # 0. Hence,
the function # z, ! (x) is the weak trace of the normal component of (z, !).
For simplicity of notation, we shall denote # z, ! (x) by [z, &](x).
 TOTAL VARIATION FLOW 353

We need to consider the space BV(0) 2 , defined as BV(0) & L 2 (0)

endowed with the norm

&w& BV(0)2 :=&w& L2(0) +&Dw&.

It easy to see that L 2 (0)/BV(0) 2* and

&w& BV(0)*2 &w& L2(0) \w # L 2 (0). (2.13)

Now, it is well known (see, for instance, [26]) that the dual space
(L 1 (0, T; BV(0) 2 ))* is isometric to the space L  (0, T; BV(0) 2* , BV(0) 2 )
of all weakly* measurable functions f : [0, T ]  BV(0) 2* , such that v( f ) #
L  ([0, T ]), where v( f ) denotes the supremum of the set [ |( w, f ) | :
&w& BV(0)2 1] in the vector lattice of measurable real functions. Moreover,
the dual paring of the isometric is defined by

T
( w, f ) = | 0
( w(t), f (t)) dt,

for w # L 1 (0, T; BV(0) 2 ) and f # L  (0, T; BV(0) *

2 , BV(0) 2 ).
By L 1w(0, T, BV(0)) we denote the space of weakly measurable functions
w: [0, T ]  BV(0) (i.e., t # [0, T ]  ( w(t), ,) is measurable for every
, # BV(0)*) such that  T0 &w(t)&<. Observe that, since BV(0) has a
separable predual (see [1]), it follows easily that the map t # [0, T ] 
&w(t)& is measurable.
To make precise our notion of solution we need the following definitions.

Definition 1. Let 9 # L 1 (0, T, BV(0)). We say 9 admits a weak

derivative in L 1w(0, T, BV(0)) & L  (Q T ) if there is a function 3 # L 1w(0, T,
BV(0)) & L  (Q T ) such that 9(t)= t0 3(s) ds, the integral being taken as
a Pettis integral.

Definition 2. Let ! # (L 1 (0, T, BV(0) 2 ))*. We say that ! is the time

derivative in the space (L1 (0, T, BV(0) 2 ))* of a function u # L 1 ((0, T )_0) if

T T

|0
(!(t), 9(t)) dt=& | |
0 0
u(t, x) 3(t, x) dx dt

for all test functions 9 # L 1 (0, T, BV(0)) which admit a weak derivative
3 # L 1w(0, T, BV(0)) & L  (Q T ) and have compact support in time.
354 ANDREU ET AL.

Observe that if w # L 1 (0, T, BV(0)) & L  (Q T ) and z # L  (Q T , R N ) is

such that there exists ! # (L 1 (0, T, BV(0)))* with div(z)=! in D$(Q T ), we
can define, associated to the pair (z, !), the distribution (z, Dw) in Q T by
T
( (z, Dw), ,) := & | 0
( !(t), w(t) ,(t))

T
& | | 0 0
z(t, x) w(t, x) { x ,(t, x) (2.14)

for all , # D(Q T ).

Defintion 3. Let ! # (L 1 (0, T, BV(0) 2 ))*, z # L  (Q T , R N ). We say

that !=div(z) in (L 1 (0, T, BV(0) 2 ))* if (z, Dw) is a Radon measure in Q T
with normal boundary values [z, &] # L  ((0, T )_0), such that
T T

| QT
(z, Dw)+ | 0
( !(t), w(t)) dt= | |0 0
[z(t, x), &] w(t, x) dH N&1 dt,

for all w # L 1 (0, T, BV(0)) & L  (Q T ).

We shall denote by

1 if r>0
sign 0 (r) := 0
&1 { if r=0
if r<0

and by

1 if r>0

{
sign(r) := a # [&1, 1]
&1
if
if
r=0
r<0.

Let T k (r)=[k&(k& |r| ) + ] sign 0 (r), k0, r # R. We consider the set

T=[T k , T k+ , T k& : k>0]. We need to consider a more general set of
truncature functions, concretely, the set P of all nondecreasing continuous
functions p: R  R, such that there exists p$ except a finite set and supp( p$)
is compact. Obviously, T/P.

3. THE MAIN RESULT

In this section we give the concept of solution for the Dirichlet

problem (1.4) and we state the existence and uniqueness result for this type
of solution.
 TOTAL VARIATION FLOW 355

Definition 4. A measurable function u: (0, T )_0  R is an entropy

solution of (1.4) in Q T =(0, T )_0 if u # C([0, T ]; L 1 (0)), p(u( } )) # L 1w(0,
T, BV(0)) \p # T and there exist (z(t), !(t)) # Z(0) with &z(t)&  1, and
! # (L 1 (0, T, BV(0) 2 ))* such that ! is the time derivative of u in (L 1 (0, T,
BV(0) 2 ))*, !=div(z) in (L 1 (0, T, BV(0)))* and [z(t), &] # sign( p(.)&
p(u(t))) a.e. in t # [0, T ], satisfying

T T
& | |
0 0
j(u(t)&l ) ' t + | |
0 0
'(t) &Dp(u(t)&l )&

T
+z(t) } D'(t) p(u(t)&l ) | |
0 0
[z(t), &] '(t) p(u(t)&l ),

for all l # R, for all ' # C  (Q T ), with '0, '(t, x)=,(t) (x), being
, # D(]0, T [ ),  # C  (0 ), and p # T, where j(r)= r0 p(s) ds.
Our main result is:

Theorem 1. Let u 0 # L 1 (0), and . # L 1 (0). Then there exists a unique

entropy solution of (1.4) in (0, T )_0 for every T>0 such that u(0)=u 0 .
Moreover, if u(t), u^(t) are the entropy solutions corresponding to initial data
u 0 , u^ 0 , respectively, then

&(u(t)&u^(t)) +& 1 &(u 0 &u^ 0 ) +& 1 and &u(t)&u^(t)& 1 &u 0 &u^ 0 & (3.1)

for all t0.

4. THE SEMIGROUP SOLUTION

To prove Theorem 1 we shall use the techniques of completely accretive

operators and the CrandallLiggett semigroup generation theorem [14].
Let us recall the notion of completely accretive operators introduces in [9].
Let M(0) be the space of measurable functions in 0. Given u, v # M(0),
we shall write

u<
<v if and only if | 0
j(u) dx |
0
j(v) dx (4.1)

for all j # J 0 where

J 0 =[ j # R  [0, ], convex, l.s.c., j(0)=0] (4.2)

356 ANDREU ET AL.

(l.s.c. is an abbreviation for lower semicontinuous function). Let A be a

operator (possibly multivalued) in M(0), i.e., AM(0)_M(0). We shall
say that A is completely accretive if

u&u^ <
<u&u^ +*(v&v^ ) for all *>0 and all (u, v), (u^, v^ ) # A. (4.3)

Let

P 0 =[ p # C  (R) : 0p$1, supp( p$) is compact and 0  supp( p)].

If AL 1 (0)_L 1 (0), then A is completely accretive if and only if

|0
p(u&u^ )(v&v^ )0 for any p # P 0 , (u, v), (u^, v^ ) # A. (4.4)

A completely accretive operator in L 1 (0) is said to be m-completely

accretive if R(I+*A)=L 1 (0) for any *>0. In that case, by Crandall
Liggett's theorem, A generates a contraction semigroup in L 1 (0) given by
the exponential formula

&n
t
e &tAu 0 = lim
n \ I+ A
n + u0 for any u 0 # L 1 (0).

Let us write u(t)=e &tAu 0 . Then u # C([0, T], L 1 (0)), for any T>0, and
is a mild solution (a solution in the sense of semigroups [10]) of

du
+Au % 0 (4.5)
dt

such that u(0)=u 0 .

We shall use a stronger notion of solution of (4.5). We say that v #
C([0, T], L 1 (0)) is a strong solution of (4.5) on [0, T] if v # W 1, 1
loc ((0, T ),
1
L (0)) and v$(t)+Av(t) % 0 for almost all t # (0, T ). If u 0 # D(A)=
[u # L 1 (0): (u, v ) # A, for some v # L 1 (0)] (the domain of A) and A is
m-completely accretive, then u # W 1, 1 1
loc ((0, T ), L (0)) and u(t) is a strong
solution of (4.5) on (0, T ), for all T>0.
To prove Theorem 1 we shall associate a completely accretive operator
Du
A. to the formal differential expression &div( |Du| ) together with the
Dirichlet boundary condition.
Let us introduce the following operator A. in L 1 (0).

(u, v) # A. if and only if u, v # L 1 (0), p(u) # BV(0) for all p # P

 TOTAL VARIATION FLOW 357

and there exists z # X(0) with &z&  1, v=&div(z) in D$(0) such that

|0
(w& p(u)) v |0
z } {w&&Dp(u)&+ |0
|w& p(.)|

& |
0
| p(u)& p(.)|,

\w # W 1, 1 (0) & L  (0), and \p # P.

Theorem 2. Let . # L 1 (0). The operator A. is m-completely accretive

in L 1 (0) with dense domain.
To prove this theorem, we need first to consider the following operator,
which is related with the p-Laplacian operator with Dirichlet boundary
condition. For p>1, let . # W 1&1 p, p (0), and
p 1, p
W 1,
. (0) :=[u # W (0) : u | 0 =. H N&1-a.e. on 0].

We define the operator A ., p in L 1 (0) as

(u, v) # A ., p if and only if u # W 1,. p(0) & L  (0), v # L 1 (0) and

|
0
(w&u) v | 0
|{u| p&2 {u } {(w&u)

p 
for every w # W 1,
. (0) & L (0).

Proposition 1. Let . # L  (0) & W 1&1 p, p (0). The operator A ., p is

completely accretive and L  (0)R(I+A ., p ).
Proof. Let p # P 0 and (u, v), (u^, v^ ) # A ., p . Since (u, v) # A ., p , taking
w=u& p(u&u^ ) as a test function in the definition of the operator A ., p we
get

|0
p(u&u^ ) v |
0
|{u| p&2 {u } {p(u&u^ ).

Similarly, since (u^, v^ ) # A ., p , taking w=u^ + p(u&u^ ) as a test function in

the definition of the operator A ., p we get

|0
p(u&u^ ) v^  |
0
|{u^ | p&2 {u^ } {p(u&u^ ).
358 ANDREU ET AL.

Hence

|
0
(v&v^ ) p(u&u^ ) |0
(|{u| p&2 {u& |{u^ | p&2 {u^ ) } {p(u&u^ )0.

Therefore, A ., p is completely accretive.

Let us now see that L  (0)R(I+A ., p ). Let v # L  (0). We need to
prove that there exists u # W 1,. p(0) & L  (0) such that (u, v&u) # A ., p ; i.e.,

|
0
(w&u)(v&u) | 0
|{u| p&2 {u } {(w&u)

p 
\w # W 1,
. (0) & L (0). (4.6)

For n # N, let # n (s) :=T n (s)+ 1n |s| p&2 s, and consider the operators A n :
. (0)  (W
p 1, p
W 1, (0))*, defined by

( A n u, w) := |0
|{u| p&2 {u } {w+ |0
# n (u) w.

It is easy to see that A n is monotone, coercive, and continuous on finite

dimensional subspaces. Then, by classical results (see, for instance, [20]),
p
given v there exists u n # W 1,
. (0) such that

( A n u n , u n &w)  | 0
v(u n &w) \w # W 1, p
. (0).

That is

| 0
(w&u n )(v&# n (u n )) | 0
|{u n | p&2 {u n } {(w&u n )

p
\w # W 1,
. (0). (4.7)

Let k>0 be such that &.&  k. If we take w=T k (u n ) in (4.7), we get

| (T (u )&u )(v&# (u ))|

0
k n n n n
0
|{u n | p&2 {u n } {(T k (u n )&u n ).

Hence, if

A n (k) :=[x # 0 : |u n (x)| >k],

 TOTAL VARIATION FLOW 359

we have that

|0
|{(u n &T k (u n ))| p = | An(k)
|{u n | p&2 {u n } {u n

= | 0
|{u n | p&2 {u n } {(u n &T k (u n ))

 | 0
v(u n &T k (u n ))& | 0
# n (u n )(u n &T k (u n ))

 | 0
v(u n &T k (u n )).

Now, by Young's inequality

|0
p$
v(u n &T k (u n ))C = &v&  * N (A n (k))+=C | 0
|u n &T k (u n )| p.

p
From here, since u n &T k (u n ) # W 1,
0 (0), using Poincare's inequality, we
obtain that

&u n &T k (u n )& 1, p R* N (A n (k)) 1 p,

from which it follows, applying the classical Stampacchia methods (see

for instance, Appendix B in [20]), that there exists a constant M 1 =
M 1 (&v&  , &.&  ) such that

&u n &  M 1 \n # N. (4.8)

On the other hand, taking w 0 as a test function in (4.7) and applying

Young's inequality, we obtain

|0
|{u n | p  |
0
|{u n | p&2 {u n } {w 0 + | 0
v(u n &w 0 )

+ |
0
# n (u n )(w 0 &u n )

=C | 0
|{u n | p +C = | 0
|{w 0 | p + |0
vu n

+ |
0
w 0 (# n (u n )&v).
360 ANDREU ET AL.

From this it follows that there exists a constant M 2 =M 2 (* N (0), &v&  ,

&.&  , &w 0 & 1, p ) such that

|0
|{u n | p M 2 \n # N. (4.9)

As a consequence of (4.8) and (4.9), [u n ] n # N is bounded in W 1, p (0).

Hence there exists a subsequence, still denoted u n , such that u n 
u # W 1, p (0) weakly in W 1, p (0). Moreover, by the RellichKondrachov
theorem, u n  u in L p (0), and by Theorem 3.4.5 in [23], u n  u in L p (0).
After passing to a suitable subsequence, we can assume that u n  u a.e. in
p 
0. So, by (4.8), &u&  M 1 .Therefore we have that u # W 1,. (0) & L (0).
Proceeding as in the proof of step 3 of Theorem 2.1 in [3], we obtain
that

|{u n | p&2 {u n  |{u| p&2 {u in measure, and a.e.

Now, by (4.9), we have that [ |{u n | p&2 {u n ] n # N is bounded in (L p$ (0)) N.

Hence

|{u n | p&2 {u n  |{u| p&2 {u weakly in (L p$ (0)) N. (4.10)

p 
Given w # W 1,
. (0) & L (0), by (4.10), we get

|0
|{u n | p&2 {u n } {w  |
0
|{u| p&2 {u } {w, (4.11)

and by Fatou's lemma, we have

|0
|{u| p&2 {u } {ulim inf
n
| | 0 0
|{u n | p&2 {u n } {u n . (4.12)

On the other hand, since u n  u in L p (0) we have

lim
n
|0
(w&u n )(v&# n (u n ))= | 0
(w&u)(v&u). (4.13)

From (4.11), (4.12), and (4.13), passing to the limit in (4.7) we get (4.6),
and the proof concludes. K

To prove Theorem 2, we need to give the following characterization of

the operator A. .
 TOTAL VARIATION FLOW 361

Proposition 2. The following assertions are equivalent:

(a) (u, v) # A.
(b) u, v # L 1 (0), p(u) # BV(0) for all p # P, and there exists z # X(0),
with &z&  1, v=&div(z) in D$(0) such that

|0
(w& p(u)) v | 0
(z, Dw)=&Dp(u)&+ | 0
|w& p(.)|

& | 0
| p(u)& p(.)| (4.14)

for every w # BV(0) & L  (0) and p # P.

(c) u, v # L 1 (0), p(u) # BV(0) for all p # P, and there exists z # X(0),
with &z&  1, v=&div(z) in D$(0) such that

|0
(w& p(u)) v | 0
(z, Dw)&&Dp(u)&& |
0
[z, &](w& p(.))

& |
0
| p(u)& p(.)| (4.15)

for every w # BV(0) & L  (0) and p # P.

(d) u, v # L 1 (0), p(u) # BV(0) for all p # P, and there exists z # X(0),
with &z&  1, v=&div(z) in D$(0) such that

| 0
(z, Dp(u))=&Dp(u)& \p # P (4.16)

[z, &] # sign( p(.)& p(u)) H N&1-a.e. on 0, \p # P. (4.17)

Proof. Let (u, v) # A. . Then, there exists z # X(0) with &z&  1, v=
&div(z) in D$(0), such that

| 0
(w& p(u)) v |
0
z } Dw&&Dp(u)&+ | 0
|w& p(.)|

& | 0
| p(u)& p(.)| (4.18)
362 ANDREU ET AL.

for every w # W 1, 1 (0) & L  (0) and every p # P. Let w # BV(0) & L  (0),
p # P. Using Lemmas 5.2 and 1.8 in [5] we know that there exists a
sequence w n # W 1, 1 (0) & L  (0) such that

wn  w in L 1 (0),

|0
|{w n |  &Dw&, (4.19)

|
0
z } {w n = |0
(z, Dw n )  | 0
(z, Dw).

and w n | 0 =w | 0 , &w n &  &w&  , \n # N. Then taking w n as a test func-

tion in (4.18) and letting n   we get that (4.18) holds for all
w # BV(0) & L  (0) and all p # P. Thus (a) and (b) are equivalent.
Since

& |
0
[z, &](w& p(.)) | 0
|w& p(.)|,

to prove the equivalence between (b) and (c), it is enough to show that if
(u, v) # A. , then (4.15) is satisfied. In fact, since (u, v) # A. , there exists
z # X(0) with &z&  1, v=&div(z) in D$(0), such that

|
0
(w& p(u)) v | 0
(z, Dw)&&Dp(u)&+ | 0
|w& p(.)|

& | 0
| p(u)& p(.)| (4.20)

for every w # BV(0) & L  (0) and every p # P. Now, given w # BV(0) &
L  (0) and p # P, by Lemmas 5.2 and 5.5 of [5], there exists w n #
W 1, 1 (0) & L  (0) such that w n  w in L 1 (0), w n | 0 = p(.), and &w n & 
&w&  +&p(.)&  , \n # N. Then taking w n as a test function in (4.18) and
using Green's formula, we get

| 0
(w n & p(u)) v |0
(z, Dw n )&&Dp(u)&& | 0
| p(u)& p(.)|

=& | 0
div(z) w n + | 0
[z, &] p(.)&&Dp(u)&

& | 0
| p(u)& p(.)|.
 TOTAL VARIATION FLOW 363

Letting n  , it follows that

| 0
(w& p(u)) v & | 0
div(z) w+ | 0
[z, &] p(.)&&Dp(u)&

& |
0
| p(u)& p(.)|.

Therefore, applying Green's formula again, we obtain (4.15).

Suppose now that (b) or, equivalent, (c) is satisfied. Taking w= p(u) in
(4.18) we obtain

0 | 0
(z, Dp(u))&&Dp(u)&.

Thus,

|
0
(z, Dp(u))&z&  &Dp(u)&&Dp(u)& |0
(z, Dp(u)),

and (4.16) holds. Let us prove (4.17). Since p(.) # L  (0), by Lemma 5.5
in [5], there exist w n # W 1, 1 (0) & L  (0) satisfying:

w n | 0 = p(.) \n # N,

1
|0
|{w n |  |
0
| p(.)| +
n
\n # N,

1
&w n & 1  , &w n &  &p(.)&  \n # N.
n

Taking w=w n in (4.18) and using Green's formula (2.10), we get

| 0
(w n & p(u)) v & | 0
div(z) w n + |
0
[z, &] p(.)&&Dp(u)&

& |
0
| p(u)& p(.)|. (4.21)
364 ANDREU ET AL.

Then, letting n   in (4.21), we obtain

& | 0
p(u) v | 0
[z, &] p(.)&&Dp(u)&& | 0
| p(u)& p(.)|.

Now, by (4.16), and applying Green's formula, we have that

&Dp(u)&= |0
(z, Dp(u))= | 0
vp(u)+ |
0
[z, &] p(u).

Hence,

0 | 0
([z, &]( p(.)& p(u))& | p(u)& p(.)|).

Since

[z, &]( p(.)& p(u))& | p(u)& p(.)| 0,

we have that

[z, &]( p(.)& p(u))= | p(u)& p(.)| H N&1-a.e. on 0,

and we obtain (4.17). Finally, to prove that (d) implies (c), we only need
to apply Green's formula. K

Remark 1. (1) As a consequence of the proof of the above proposi-

tion we can put equality in the definition of the operator; that is, the
following characterization of the operator A. holds.

(u, v) # A. if and only if u, v # L 1 (0), p(u) # BV(0) for all p # P

and there exists z # X(0) with &z&  1, v=&div(z) in D$(0) such that

|0
(w& p(u)) v= | 0
(z, Dw)&&Dp(u)&+ | 0
|w& p(.)|

& | 0
| p(u)& p(.)|,

\w # BV(0) & L  (0) and \p # P.

 TOTAL VARIATION FLOW 365

(2) As a consequence of the above proposition, if (u, v) # A. , we

have that %(z, DT k (u), x)=1 a.e. with respect to the measure &DT k (u)&. In
the case that z # C(0, R N ), this implies that

DT k (u)
z(x) } =1, &DT k (u)&-a.e.,
&DT k (u)&

where DT k (u)&DT k (u)& denotes the density of DT k (u) with respect to

Du
&DT k (u)&. Heuristically, this amounts to saying that z= &Du& . When z is
not continuous we have that

DT k (u)
z(x) } =1, &{T k (u)&-a.e.,
&DT k (u)&

where &{T k (u)& denotes the absolutely continuous part of &DT k (u)& with
respect to the Lebesgue measure in R N [5]. In particular, if u # W 1, 1 (0) &
L  (0) we have that

{u
z(x) } =1, &{u&-a.e.
&{u&

(3) Observe that by (d) in the above proposition, if u # L  (0), then

the truncatures are redundant in the definition of A. .
To prove the following result, we need to introduce the function
8: L 1 (0)  (&, +] defined by

8(u)=
{ &Du&+

+
| 0
|u&.| if

if
u # BV(0)
1
u # L (0)"BV(0).
(4.22)

The functional 8 is convex and lower semicontinuous in L 1 (0) (see [6] or

[27]).

Proposition 3. Let . # L 1 (0). Then L  (0)/R(I+A. ) and D(A. ) is

dense in L 1 (0).
Proof. Suppose first that . # W 12, 2 (0) & L  (0). Let v # L  (0). We
shall find u # BV(0) & L  (0) such that (u, v&u) # A. , i.e., there is
z # X(0) with &z&  1 such that v&u=&div(z) and

|0
(w&u)(v&u) |
0
z } {w dx&&Du&+ |
0
|w&.| & |
0
|u&.| (4.23)

for every w # W 1, 1 (0) & L  (0).

366 ANDREU ET AL.

Since . # W 1&1 p, p (0) for all p>1, by Proposition 1, we know that for
p 
any 1<p2 there is u p # W 1, . (0) & L (0) such that (u p , v&u p ) # A ., p .
Hence

|0
(w&u p )(v&u p ) | 0
|{u p | p&2 {u p } {(w&u p ), (4.24)

p 
for every w # W 1,
. (0) & L (0).
Let M :=sup [&.&  , &v&  ]. Then, taking w=u p &(u p &M) + as a test
function in (4.24), we obtain

|
0
(u p &M) + (u p &v)0.

Hence,

| [up >M]
(u p &M) 2  | [up >M]
(u p &M)(u p &v)

= | 0
(u p &M) + (u p &v)0.

Consequently, u p M a.e. in 0. Analogously, taking w=u p +(u p +M) & as

a test function, we get &Mu p a.e. in 0. Therefore,

&u p &  M for all 1<p2. (4.25)

p 
Taking w = w 0 # W 1,. (0) & L (0) in (4.24) and applying Young's
inequality we obtain

|0
|{u p | p  | 0
|{u p | p&2 {u p } {w 0 & |0
(w 0 &u p )(v&u p )

= | 0
|{u p | p +C = | 0
|{w 0 | p +C(&v&  , &w 0 &  ).

Thus

| 0
|{u p | p M 1 \1<p2, (4.26)
 TOTAL VARIATION FLOW 367

where M 1 depends on &v&  , &w 0 &  , and &w 0 & 1, 2 . Using Holder's
inequality we also have that

|
0
|{u p | M 2 \1<p2, (4.27)

where M 2 does not depend on p. Thus, [u p ] p>1 is bounded in W 1, 1 (0)

and we may extract a subsequence such that u p converges in L 1 (0) and
almost everywhere to some u # L 1 (0) as p  1+. Now, by (4.25) and
(4.27), we have that u # BV(0) & L  (0).
Let us prove that [|{u p | p&2 {u p ] p>1 is weakly relatively compact in
L (0, R N ). For that, using (4.26), we observe that
1

( p&1) p

|0
|{u p | p&1 
\| 0
|{u p | p
+ * N (0) 1 p M 3 ,

where M 3 does not depend on p. On the other hand, for any measurable
subset E0 such that * N (E)<1, we have

}|E
|{u p | p&2 {u p 
} | E
|{u p | p&1 M (1p&1) p * N (E) 1 p M 4 * N (E) 12

Thus, [ |{u p | p&2 {u p ] p>1 , being bounded and equiintegrable in L 1 (0, R N ),

is weakly relatively compact in L 1 (0, R N ). We may assume that

|{u p | p&2 {u p ( z as p  1 +, weakly in L 1 (0, R N ). (4.28)

0 (0), taking w=u p \ in (4.24) and letting p  1 , we

Given  # C  +

obtain

| 0
(v&u) = |0
z } {,

that is, v&u=&div(z) in D$(0). Moreover, the same technique that we

use in the proof of Lemma 1 in [2] shows that &z&  1.

For every w # W 1, 2
. (0) & L (0), by (4.24) and Young's inequality, we
get

p&1 1
| 0
|{u p | + |
0
|u p &.| 
p
* N (0)&
p | 0
(w&u p )(v&u p )

1
+
p |
0
|{u p | p&2 {u p } {w.
368 ANDREU ET AL.

Then, using the lower semicontinuity of the functional 8 defined by (4.22),

letting p  1 +, we obtain

&Du&+ |0
|u&.|  & | 0
(w&u)(v&u)+ | 0
z } {w, (4.29)


for every w # W 1, 2
. (0) & L (0).
Now, to prove (4.23), we assume first that there exists w 0 # W 1, 2 (0) &

L (0) such that .=w 0 | 0 (i.e., . is the trace of w 0 ). Let w # W 1, 1 (0) &
. (0) & L (0) be such that w n  w in L (0) as
L  (0) and let w n # W 1, 2  1

n   and &w n &  &w&  . Using w n as a test function in (4.29) and apply-
ing Green's formula (2.10), we may write

| 0
(w n &u)(v&u) |
0
z } {w n &&Du&& |
0
|u&.|

=& | 0
div(z) w n + | 0
[z, &] .&&Du&& | 0
|u&.|.

From here, letting n   and applying again the Green's formula, we get

| 0
(w&u)(v&u)

& |
0
div(z) w+ |0
[z, &] .&&Du&& |0
|u&.|

= | 0
z } {w& | 0
[z, &] w+ | 0
[z, &] .&&Du&& |0
|u&.|

 | 0
z } {w&&Du&+ | 0
|w&.| & | 0
|u&.|,

and the proof of (4.23), in this particular case, concludes.

Suppose now we are in the general case, that is, . # L 1 (0). Take
v n # W 1, 2 (0) & L  (0) such that . n :=v n | 0  . in L 1 (0). From the
above, there exists u n # BV(0) & L  (0) and z # X(0) with &z n &  1 such
that v&u n =&div(z n ) and

|0
(w&u n )(v&u n ) |
0
z n } {w dx&&Du n &+ |0
|w&. n |

& |0
|u n &. n | (4.30)
 TOTAL VARIATION FLOW 369

for every w # W 1, 1 (0) & L  (0). Moreover, by (4.25), we have &u n &  
max[&v&  , &. n &  ]. We can assume that z n  z weakly* in L  (0). Now,
taking w=0 in (4.30), we get

& |
0
u n v+ | 0
(u n ) 2 +&Du n &+ |0
|u n &. n |  | 0
|. n |.

Hence,

&u n & 22 +&Du n & | 0

u n v+ | 0
|. n |  12 &u n & 22 + 12 &v& 22 + |0
|. n |.

Thus, [u n ] is a bounded sequence in BV(0) & L 2 (0). Then, since BV(0)

is compactly embedded in L 1 (0) (see [28] or [17]), there is a sub-
sequence, still denoted by [u n ] such that u n  u in L 1 (0). Finally, taking
limits in (4.30), we obtain that (u, v&u) # A. .
To prove the density of D(A. ) in L 1 (0), we prove that C  0 (0)
1
D(A. ) L (0). Let v # C 
0 (0). By the above, v # R(I+ 1
A
n . ) for all n # N.
Thus, for each n # N there exists u n # D(A. ) such that (u n , n(v&u n )) # A.
and, therefore, there exists some z n # X(0) with &z n &  1, n(v&u n )=
&div(z n ) in D$(0) such that

| 0
(w&T k (u n )) n(v&u n ) | 0
z n } {w&&DT k (u n )&+ | 0
|w&T k (.)|

& |0
|T k (u n )&T k (.)|

for every w # W 1, 1 (0) & L  (0). Taking w=T k (v) and applying Fatou's
lemma we have that

1
|0
(v&u n ) 2 
n \| 0
|{u| + |
0 +
|.| .

Letting n  , it follows that u n  v in L 2 (0). Therefore v #

1
D(A. ) L (0). K

Proof of Theorem 2. Let (u, v), (u^, v^ ) # A. , p # P 0 . We have to prove

that

|0
p(u&u^ )(v&u^ )0. (4.31)
370 ANDREU ET AL.

Let z, z^ # X(0), &z&  1, &z^&  1, be such that v=&div(z), v^ = &div(z^ )
and

|
0
(w&T k (u)) v | 0
(z, Dw)&&DT k (u)&& | 0
[z, &](w&T k (.))

& | 0
|T k (u)&T k (.)|, (4.32)

|
0
(w&T k (u^ )) v^  | 0
(z^, Dw)&&DT k (u^ )&& | 0
[z^, &](w&T k (.))

& | 0
|T k (u^ )&T k (.)|, (4.33)

for any w # BV(0) & L  (0) and any k>0. As observed in the previous
remark, %(z, DT k (u), x)=1 &DT k (u)&-a.e., and, using Corollary 1.6 in [5],
we obtain that

| B
(z, DT k (u))= | B
%(z, DT k (u), x) &DT k (u)&= | B
&DT k (u)&,

}| B }
(z^, DT k (u))  | B
&DT k (u)&

for any Borel set B0. Similarly,

| B
(z^, DT k (u^ ))= |B
&DT k (u^ )&,

} | (z, DT (u^))} | &DT (u^)&

B
k
B
k

for any Borel set B0. It follows that

| B
(z&z^, D(T k (u)&T k (u^ )))0

for any Borel set B0. This implies that

%(z&z^, D(T k (u)&T k (u^ )), x)0, &D(T k (u)&T k (u^ ))&-a.e.

Since, according to Proposition 2.8 in [5], we have that

%(z&z^, Dp(T k (u)&T k (u^ )), x)=%(z&z^, D(T k (u)&T k (u^ )), x)
 TOTAL VARIATION FLOW 371

a.e. with respect to the measures &D(T k (u)&T k (u^ ))& and &Dp(T k (u)&
T k (u^ ))&. We conclude that

%(z&z^, Dp(T k (u)&T k (u^ )), x)0, &Dp(T k (u)&T k (u^ ))&-a.e. (4.34)

Taking w=T k (u)& p(T k (u)&T k (u^ )) in (4.32) and w=T k (u^ )+p(T k (u)&
T k (u^ )) in (4.33), adding both terms, and using (4.17) and (4.34), we obtain

|0
p(T k (u)&T k (u^ ))(v^ &v)

 |
0
(z^ &z, Dp(T k (u)&T k (u^ )))

+ | 0
([z, &]&[z^, &]) p(T k (u)&T k (u^ ))

=& | 0
%(z&z^, Dp(T k (u)&T k (u^ )), x) &Dp(T k (u)&T k (u^ ))&

+ | 0
([z, &]&[z^, &]) p(T k (u)&T k (u^ ))0.

The inequality (4.31) follows by letting k  . Therefore A. is completely

accretive.
In view of Proposition 3, to prove that A. satisfies the range condition,
it is enough to prove that A. is closed. Let (u n , v n ) # A. , such that
(u n , v n )  (u, v) in L 1 (0)_L 1 (0). Let us see that (u, v) # A. . Since
(u n , v n ) # A. , there exists z n # X(0), &z n &  1 with v n =&div(z n ) in D$(0)
such that

|
0
(w& p(u n )) v n  | 0
(z n , Dw)&&Dp(u n )&+ |
0
|w& p(.)|

& | 0
| p(u n )& p(.)| (4.35)

for every w # BV(0) & L  (0) and all p # P. Since &z n &  1 we may
assume that z n ( z in the weak* topology of L  (0, R N ) with &z&  1.
Moreover, since v n  v in L 1 (0), we have v=&div(z) in D$(0), and

lim
n
| 0
(z n , Dw)= |0
(z, Dw).
372 ANDREU ET AL.

Now, letting n   in (4.35), and having in mind the lower semicontinuity

of the function 8, defined in (4.22), we obtain that

|0
(w& p(u)) v | 0
(z, Dw)&&Dp(u)&+ | 0
|w& p(.)|

& | 0
| p(u)& p(.)|.

Consequently, (u, v) # A. . K

5. STRONG SOLUTIONS FOR DATA IN L 2 (0)

In this section we are going to see that when the initial datum is in
L 2 (0), then the semigroup solution is a strong solution.
Let [S(t)] t0 be the contraction semigroup in L 1 (0) generated by
the operator A. via CrandallLiggett's exponential formula. Since A. is an
m-completely accretive operator, S(t)(L 2 (0))/L 2 (0). Let 9 . : L 2 (0) 
]&, +], the restriction to L 2 (0) of the functional 8 defined by
(4.22), i.e.,

&Du&+ | |u&.| u # BV(0) & L 2 (0)

9 . (u)=
{ +
0
if

if u # L 2 (0)"BV(0) & L 2 (0).

(5.1)

Since the function 9 . is convex and lower semicontinuous in L 2 (0), we

have that 9 . is a maximal monotone operator in L 2 (0), and conse-
quently (see [12]), if [T(t)] t0 is the semigroup in L 2 (0) generated by
9 . , for every u 0 # L 2 (0), u(t) :=T(t) u 0 is a strong solution of the
problem

du
+9 . u(t) % 0
dt (5.2)
u(0)=u 0 .

Recall that the operator 9 . is defined by

(u, v) # 9 . if and only if u, v # L 2 (0), and

9 . (w)9 . (u)+ (w&u) v, | 0

\w # L 2 (0).
 TOTAL VARIATION FLOW 373

Lemma 1. Let B . :=A. & (L 2 (0)_L 2 (0)). Then B . =9 . .

Proof. Let (u, v) # B . . Then, u, v # L 2 (0), p(u) # BV(0) for all p # P,

and there exists z # X(0) with &z&  1, v=&div(z) in D$(0) such that

|
0
(w& p(u)) v | 0
(z, Dw)&&Dp(u)&+ | 0
|w& p(.)|

& | 0
| p(u)& p(.)|,

\w # BV(0) & L  (0) and \p # P. Letting p=T k and k   we obtain that

| 0
(w&u) v | 0
(z, Dw)&&Du&+ |0
|w&.| & |0
|u&.|,

\w # BV(0) & L  (0). To prove that (u, v) # 9 . , we have to prove that

| 0
(w&u) v&Dw&&&Du&+ | 0
|w&.| & |
0
|u&.| (5.3)

for every w # L 2 (0) & BV(0). Now, given w # L 2 (0) & BV(0), since
(u, v) # B . , by the first observation of the lemma, there exists z # X(0), with
&z&  1, v=&div(z) in D$(0) such that

| 0
(T k (w)&u) v |0
(z, DT k (w))&&Du&+ | 0
|T k (w)&.|

& | 0
|u&.|,

for every k>0. From this, it follows that

|
0
(T k (w)&u) v&DT k (w)&&&Du&+ |
0
|T k (w)&.| & |
0
|u&.|. (5.4)

Now, since lim k   T k (w)=w in L 2 (0),

&Dw&lim inf &DT k (w)&.

k
374 ANDREU ET AL.

Moreover, since &DT k (w)&&w&, we also have that

lim sup &DT k (w)&&Dw&.

k

Thus

lim &DT k (w)&=&Dw&.

k

Therefore, letting k   in (5.4), we obtain (5.3). We have proved that

B . /9 . .
By Proposition 3, we have that L  (0)/R(I+B . ). Hence, 9 . =
L2(0)
B . . It follows that 9 . =A. & (L 2 (0)_L 2 (0)). K
Using this lemma and having in mind Proposition 2, we have the follow-
ing result.

Theorem 3. Let . # L 1 (0). Given u 0 # L 2 (0), u(t)=S(t) u 0 is a strong

solution of (5.2). Moreover, u$(t) # L 2 (0), p(u(t)) # BV(0) for all p # P, and
there exists z(t) # X(0), &z(t)&  1, and u$(t)=div(z(t)) in D$(0) a.e.
t # [0, +[, satisfying

| 0
(w& p(u(t))) u$(t)

 |
0
(z(t), Dw)&&Dp(u(t))&

& |
0
[z(t), &](w& p(.))& |
0
| p(u(t))& p( .)| (5.5)

for every w # BV(0) & L  (0) and p # P.

Moreover, u(t) is also characterized as follows: there exists z(t) # X(0),
&z(t)&  1, and u$(t)=div(z(t)) in D$(0) a.e. t # [0, +[, satisfying

|
0
(z(t), Dp(u(t)))=&Dp(u(t))& \p # P (5.6)

[z(t), &] # sign( p(,)& p(u(t))) H N&1-a.e. on 0, \p # P. (5.7)

Remark 2. Note that under the assumptions of Theorem 3, since u(t) #

BV(0), applying the lower semicontinuity of 9 . , if we take p=T k and
take limits when k  , we obtain that (5.5), (5.6), and (5.7) are true when
p is the identity map.
 TOTAL VARIATION FLOW 375

We have the following weak form of the maximum principle.

Theorem 4. Let u 1 and u 2 be two strong solutions of

du i
+9 .i u i (t) % 0
dt (5.8)
u i (0)=u i, 0 , i=1, 2,

where u i, 0 # L 2 (0) and . i # L 1 (0). Suppose that u 1, 0 u 2, 0 and . 1 . 2 .

Then we have u 1 u 2 .
Proof. By Theorem 3 and the above remark, we have that u i (t), u$i (t) #
L 2 (0), and there exist z i (t) # X(0), &z i (t)&  1, and u$i (t)=div(z i (t)) in
D$(0), satisfying:

| 0
(z i (t), D(u i (t)))=&D(u i (t))& (5.9)

[z i (t), &] # sign(. i &u i (t)) H N&1-a.e. on 0. (5.10)

Since dtd (u 2 (t)&u 1 (t))=div(z 2 (t)&z 1 (t)) in L 2 (0), multiplying by

(u 2 (t)&u 1 (t)) +, integrating, and using Green's formula, we get

1 d
2 | 0 dt
[(u 2 (t)&u 1 (t)) + ] 2

= |
0
div(z 2 (t)&z 1 (t))(u 2 (t)&u 1 (t)) +

=& | 0
(z 2 (t)&z 1 (t), D((u 2 (t)&u 1 (t)) + ))

+ | 0
[z 2 (t)&z 1 (t), &](u 2 (t)&u 1 (t)) +. (5.11)

Now, by (5.9) it follows that

%(z 2 (t)&z 1 (t), D(u 2 (t)&u 1 (t)), x)0 &D(u 2 (t)&u 1 (t))&-a.e.

According to Proposition 2.8 in [5], we have

%(z 2 (t)&z 1 (t), D(u 2 (t)&u 1 (t)), x)

=%(z 2 (t)&z 1 (t), D(u 2 (t)&u 1 (t)) +, x)
376 ANDREU ET AL.

a.e. with respect to &D(u 2 (t)&u 1 (t))& and &D(u 2 (t)&u 1 (t)) +&. Hence we
can conclude that

%(z 2 (t)&z 1 (t), D(u 2 (t)&u 1 (t)) +, x)0, &D(u 2 (t)&u 1 (t)) +&-a.e.

Consequently, we have

|
0
(z 2 (t)&z 1 (t), D((u 2 (t)&u 1 (t)) + ))

= |
0
%(z 2 (t)&z 1 (t), D(u 2 (t)&u 1 (t)) +, x) &D(u 2 (t)&u 1 (t)) +&

0. (5.12)

On the other hand, since . 1 . 2 , from (5.10), it is easy to see that

|0
[z 2 (t)&z 1 (t), &](u 2 (t)&u 1 (t)) + 0. (5.13)

From (5.11), (5.12), and (5.13), we obtain that

1 d
2 |0 dt
[(u 2 (t)&u 1 (t)) + ] 2 0.

Hence the initial condition u 1, 0  u 2, 0 gives u 1  u 2 , and the proof

concludes. K

Proposition 4. Let 0u 0 # L 2 (0) and 0. # L 1 (0). Then, if u is the

strong solution of the problem (5.1), we have

u(t)
u$(t) for t>0.
t

The opposite inequality holds if u 0 , .0.

Proof. We shall prove the proposition only when u 0 , .0, the other
case being similar. First, let us see that for *>0, we have

* &1u(*t)=e &tA* &1. (* &1u 0 ). (5.14)

 TOTAL VARIATION FLOW 377

By CrandallLiggett's exponential formula, it is enough to prove that for

all +>0,

(I++A* &1. ) &1 (* &1u 0 )=* &1 (I+*+A. ) &1 (u 0 ). (5.15)

In fact, v + :=(I++A* &1. ) &1 (* &1u 0 ) if and only if (v + , (* &1u 0 &v * )+) #
A* &1. , which is equivalent to the existence of z + # X(0), such that

* &1u 0 &v *
&div(z + )= ,
+

| 0
(z + , Dv + ) =&Dv + &,

[z + , &] # sign(* &1.&v + ).

Then, we have

u 0 &*v *
&div(z + )= ,
*+

| 0
(z + , D*v + ) =&D*v + &,

[z + , &] # sign(.&*v + ),

which is equivalent to saying that (*v + , (u 0 &*v + )*+) # A. , that is, v + =

* &1 (I+*+A. ) &1 (* &1u 0 ), and (5.15) holds.
Fix t>0. For h>0, if *t=t+h, applying (5.14), we obtain

u(t+h)&u(t)=u(*t)&u(t)=(1&* &1 ) u(*t)+* &1u(*t)&u(t)

h
= u(t+h)+e &tA* &1. (* &1u 0 )&u(t).
t+h

Now, since * &1u 0 u 0 and * &1.., by Theorem 4, we get

e &tA* &1. (* &1u 0 )u(t).

Consequently,

h
u(t+h)&u(t) u(t+h),
t+h

and the result follows. K

378 ANDREU ET AL.

6. EXISTENCE AND UNIQUENESS FOR DATA IN L 1 (0)

In this section we are going to prove Theorem 1.

Proof of Theorem 1 (Existence).

Let u 0 # L 1 (0) and [S(t)] t0 the contraction semigroup in L 1 (0)
generated by A. . We shall prove that u(t) :=S(t) u 0 is an entropy solution
of problem (1.4). We divide the proof in different steps.
Step 1. Since D(A. ) & L  (0) is dense in L 1 (0), given u 0 # L 1 (0)
there exists a sequence u 0, n # D(A. ) & L  (0) such that u 0, n  u 0 in L 1 (0).
Then, if u n (t) :=S(t) u 0, n , we have that u n  u in C([0, T]; L 1 (0)) for
every T>0. As a consequence of Theorem 3, u n (t), u$n (t) # L 2 (0), p(u n (t))
# BV(0) for all p # P and there exists z n (t) # X(0), &z n (t)&  1, and u$n (t)
=div(z n (t)) in D$(0) a.e. t # [0, +[, satisfying

& |0
(w& p(u n (t))) u$n (t)

 |0
(z n (t), Dw)&&Dp(u n (t))&

& |
0
[z n (t), &](w& p(.))& | 0
| p(u n (t))& p(.)| (6.1)

for every w # BV(0) & L  (0) and p # P. Moreover,

|0
(z n (t), Dp(u n (t)))=&Dp(u n (t))& \p # P (6.2)

and

[z n (t), &] # sign( p(.)& p(u n (t))) H N&1-a.e. on 0, \p # P. (6.3)

Since &[z n (t), &]&  &z n (t)&  1, we can suppose (up to extraction of
a subsequence, if necessary) that

[z n ( } ), &]  \ _(L  (S T ), L 1 (S t )).

Step 2. Convergence of the derivatives and identification of the limit.

Since the map t [ u$n (t) is strongly measurable from [0, T] into L 2 (0),
and by (2.13),

&u$n (t)& BV(0)*2 &u$n (t)& L2(0) ,

 TOTAL VARIATION FLOW 379

it follows that this map is strongly measurable from [0, T] into BV(0) 2* .
Moreover, for every w # BV(0) 2 , by Green's formula we have

|
0
u$n (t) w= |
0
div(z n (t)) w=& |0
(z n (t), Dw)+ | 0
[z n (t), &] w.

Hence

}|0 }
u$n (t) w &Dw&+ |
0
|w| M &w& BV(0)2 \n # N.

Thus,

&u$n (t)& BV(0)*2 M \n # N and t # [0, T].

Consequently, [u$n ] n # N is a bounded sequence in L  (0, T; BV(0) *).

2 Since
L  (0, T; BV(0) 2*) is a vector subspace of the dual space (L 1 (0, T;
BV(0) 2 ))*, we can find a net [u$: ] such that

u$:  ! # (L 1 (0, T; BV(0) 2 ))* weakly*. (6.4)

Since &z n (t)&  1 for all n # N and a.e. t # [0, T], we can suppose that

z n  z # L  (Q T , R N ) weakly*. (6.5)

Given ' # D(Q T ), since ' # L 1 (0, T; BV(0) 2 ), we have

T
( !, ') =lim ( u$: , ') =lim
: :
| 0
( u$: (t), '(t)) dt

T
=lim
:
| | 0 0
u$: (t) '(t) dx dt

T
=lim
:
| | 0 0
div(z : (t)) '(t) dx dt

T
= &lim
:
| |
0 0
z : (t) } {'(t) dx dt

=& |QT
z } {'=( div x (z), ').

Hence,

!=div x (z) in D$(Q T ). (6.6)

380 ANDREU ET AL.

On the other hand, if we take '(t, x)=,(t) (x) with , # D( ]0, T [ ) and
 # D(0), the same calculation as above shows that

!(t)=div x (z(t)) in D$(0) a.e. t # [0, T]. (6.7)

Consequently, (z(t), !(t)) # Z(0) for almost all t # [0, T]; therefore we can
consider [z(t), &] defined as in Section 2.

Lemma 2. ! is the time derivative of u in the sense of Definition 2.

Proof. Let 9 # L 1 (0, T, BV(0)) be the weak derivative of 3 # L 1w(0, T,
BV(0)) & L  (Q T ), i.e., 9(t)= t0 3(s) ds, the integral being taken as a Pettis
integral. By (6.4) we have that

T T

|0
( !(t), 9(t)) dt=lim
:
| 0
( u$: (t), 9(t)).

Now,

T T u : (t+h)&u(t)
| 0
( u$: (t), 9(t)) =lim
h
| | 0 0
9(t)
h
dx dt

T 9(t&h)&9(t)
=lim
h | | 0 0 h
u : (t) dx dt

T 1 t
= &lim
h
| |
0 0 h | t&h
3(s) ds u : (t) dx dt

T
=& | |
0 0
3(t, x) u : (t, x) dx dt.

Passing to the limit in : in the above expression, we obtain

T T

|
0
( !(t), 9(t)) dt=& | |
0 0
3(t, x) u(t, x) dx ds. K (6.8)

Step 3. Convergence of the energy. In this step we shall prove that for
any p # P, we have

T T
lim
n | 0
&Dp(u n (t))&+ | |
0 0
| p(u n (t))& p(.)|

T T
= | 0
&Dp(u(t))&+ | | 0 0
| p(u(t))& p(.)|. (6.9)
 TOTAL VARIATION FLOW 381

Taking w=0 in (6.1) we get

&Dp(u n (t))&+ | 0
| p(u n (t))& p(.)|

& |
0
p(u n (t)) u$n (t)+ | 0
[z n (t), &] p(.).

If we denote J p (r) := r0 p(s) ds, it follows that

t T

| 0
&Dp(u n (t))&+ | |
0 0
| p(u n (t))& p(.)|

T d T
& | 0 dt | 0
J p (u n (t))+ | |
0 0
| p(.)|

T
= | 0
(J p (u 0, n )&J p (u n (T )))+ | | 0 0
| p(.)| M p .

Since the functional 8 p : L 1 (0)  ]&, +], defined by

|
8 p (w)=
{ &Du&+

+
0
|u& p(.)| if w # BV(0)

if u # L 1 (0)"BV(0),
(6.10)

is lower semicontinuous in L 1 (0), we have

8 p ( p(u(t)))lim inf 8 p ( p(u n (t)))

n

n \
=lim inf &Dp(u n (t))&+ | 0
| p(u n (t))& p(.)| .
+ (6.11)

On the other hand, by the Fatou's lemma, it follows that

|0 \
lim inf &Dp(u n (t))&+
n
|
0
| p(u n (t))& p(.)|
+
T
lim inf
n
| 0 \&Dp(u (t))&+| n
0
| p(u n (t))& p(.)| M p .
+ (6.12)

As a consequence of (6.11) and (6.12), we obtain that p(u(t)) # BV(0) for

almost all t # [0, T].
382 ANDREU ET AL.

Lemma 3. The map t [ p(u(t)) from [0, T] into BV(0) is weakly

measurable.
Proof. Let E :=C c (0) N+1 and S: BV(0)  E* be the map defined by

w w
\
S(w) := w dx,
x 1
, ...,
x N
.
+
Then, &w& BV(0) &S(w)& E* N &w& BV(0) . If we denote by F the closure in
E of the set

[(, 0 , , 1 , ..., , N ): , i # D(0), and , 0 =div(, 1 , ..., , N )],

in [1] it is shown that S(BV(0)) is isomorphic to ( EF )*; that is, G := EF is

the predual of the space BV(0). Now, if ,=(, 0 , , 1 , ..., , N ) # D(0) N+1,
N
,
|
( S( p(u(t))), ,) = p(u(t)) , 0 & :
i=1
| 0
p(u(t))
x i
.

Hence, the map t [ ( S( p(u(t))), ,) is measurable. From here, approxi-

mating the functions of C c (0) N+1 by functions of D(0) N+1, we get that
for every , # G, the function t [ ( S( p(u(t))), ,) is measurable. Thus, since
G is separable, it follows that the map

t [ &p(u(t))& BV(0) = sup ( S( p(u(t))), ,)

, # G, &,&1

is measurable.
Given w # BV(0)*, let g(t) :=( p(u(t)), w). To see that g is measurable,
consider w : # G such that w :  w with respect to _(G**, G*)=
_(BV(0)*, BV(0)). From the above, we know that if g : (t) :=( S( p(u(t))),
w : ), g : is measurable, and g : (t)  g(t). Now, since

| g : (t)| &p(u(t))& BV(0) &w : & BV(0)*

R &p(u(t))& BV(0) =F(t) # L 1 (0, T ),

and the order interval [&F, F] in L 1 (0, T ) is _(L 1 (0, T ), L  (0, T ))-
relatively compact, there exists a sequence g :n such that

g :n  g in _(L 2 (0, T ), L  (0, T )).

Hence, g is measurable. K
From the above, if 0' # D( ]0, T [), the map t [ p(u(t)) '(t) from
[0, T] into BV(0) is weakly measurable.
 TOTAL VARIATION FLOW 383

Lemma 4. For any {>0, we define the function  { as the Dunford

integral (see [15])

1 t
 { :=
{ | t&{
'(s) p(u(s)) ds # BV(0)**,

that is,

1 t
( { (t), w) =
{ | t&{
( '(s) p(u(s)), w) ds,

for any w # BV(0)*. Then  { # C([0, T]; BV(0)). Moreover,  { (t) # L 2 (0),
and, thus,  { (t) # BV(0) 2 .

Proof. Given , # D(0),

1 t
|(  { (t), ,) | 
{ | t&{
|'(s)| |( p(u(s)), ,) | ds

1 t
=
{ | t&{
|'(s)|
\| 0 +
| p(u(s))| |,| dx dsC &,&  .

Consequently,  { (t) is a finite Radon measure in 0. Moreover, a similar

{(t)
calculation shows that for every i=1, 2, ..., N, xi is also a finite Radon
measure in 0. Hence, we have  { (t) # BV(0) (see Exercise 3.2 in [1]), and
the Dunford integral of the definition of  { (t) is a Pettis integral.
Moreover, if a n  0 (for simplicity suppose that a n >0), given w # BV(0)*
with &w&1, we have

|( { (t+a n )& { (t), w) |

1 t+an 1 t
=
} { | t+an &{
'(s)( p(u(s)), w) ds&
{ | t&{
'(s)( p(u(s)), w) ds
}
1 t+an 1 t&{+an

}{
 |
t
'(s)( p(u(s)), w) ds&
{ | t&{
'(s)( p(u(s)), w) ds
}
1 t+an

{ |t
|'(s)| &p(u(s))& BV(0) ds

1 t&{+an
+
{ | t&{
|'(s)| &p(u(s))& BV(0) ds.
384 ANDREU ET AL.

Since the function s [ |'(s)| &p(u(s))& BV(0) is in L 1 ([0, T]),

lim & { (t+a n )& { (t)& BV(0) =0.

n

Thus,  { # C([0, T]; BV(0)).

Moreover,  { (t) # L 2 (0). In fact, given g # L  (0), with &g& 2 1, since
g # BV(0)*, we have

1 t

}
|(  { (t), g) | =
{ | t&{
'(s)( p(u(s)), g) ds
}
1 t

}{
= |
t&{
'(s)
\|0
p(u(s)) g dx ds
+ }
1 t

{ | t&{
|'(s)| &p(u(s))& 2 &g& 2 M.

From the density of L  (0) in L 2 (0), we obtain that  { (t) # L 2 (0). K

Lemma 5. For {>0 small enough, we have

T T '(t&{)&'(t)
| 0
(  { (t), !(t)) dt & | |
0 0 &{
J p (u(t)). (6.13)

Proof. Since  { # C([0, T], BV(0)) admits a weak derivative in L 1w(0,

T, BV(0)) & L  (Q T ), using (6.8) we have for {>0 small enough that

T T u(t+{)&u(t)
|0
(  { (t), !(t)) dt= | | 0 0 {
'(t) p(u(t)).

Now, since p is nondecreasing, we have

J p (u(t))&J p (u(t+{))(u(t)&u(t+{)) p(u(t))

and consequently, for {>0 small enough, we obtain

T u(t+{)&u(t) T J p (u(t+{))&J p (u(t))

| |
0 0 {
'(t) p(u(t)) | |
0 0 {
'(t)

T '(s&{)&'(s)
= | |
0 0 {
J p (u(s)),

and we finish the proof of (6.13). K

 TOTAL VARIATION FLOW 385

Now, we can conclude the proof of Step 3. As a consequence of (6.13),

using Green's formula, we have

T '(t&{)&'(t)
| |
0 0 &{
J p (u(t))

T T
& | | 0 0
(  { (t), !(t)) dt

T
=&lim
:
| 0
(  { (t), u$: (t)) dt

T 1 t
=&lim
:
| 0 \ { | t&{
'(s)( p(u(s)), u$: (t)) ds dt
+
T 1 t
=&lim
:
| 0 \ { | t&{
'(s)
\| 0
p(u(s)) div z : (t) ds dt
+ +
T 1 t
=lim
: _| \ 0 { | t&{
'(s)
\| 0
(z : (t), Dp(u(s))) ds dt
+ +
T 1 t
& | 0 \ { | t&{
'(s)
\| 0
[z : (t), &] p(u(s)) ds dt
+ + &
T 1 t
 |0 \ { | t&{
'(s) &Dp(u(s))& ds dt
+
T 1 t
& | 0 \ { | t&{
'(s)
\| 0
\(t) p(u(s)) ds dt.
+ +
Then, taking the limit as {  0 +, we get

T T T

| |
0 0
'$(t) J p (u(t)) | 0
'(t) &Dp(u(t))&& | 0
'(t) |
0
\(t) p(u(t)).

Now, since this is true for all 0' # D(]0, T [), it follows that

d
&
dt | 0
J p (u(t))&Dp(u(t))&& |0
\(t) p(u(t)),
386 ANDREU ET AL.

and consequently
T T

| 0
(J p (u 0 )&J p (u(T ))) | 0
&Dp(u(t))&& | |
0 0
\(t) p(u(t)). (6.14)

Finally, using (6.14), we obtain

T T

| 0
&Dp(u(t))&+ | | 0 0
| p(u(t))& p(.)|

T T
lim inf
n
| 0
&Dp(u n (t))&+ | | 0 0
| p(u n (t))& p(.)|

T T
lim inf
n \| 0
| 0
p(u n (t)) u$n (t)+ | | 0 0
[z n (t), &] p(.)
+
T
= |0
J p (u 0 )&J p (u(T ))+ | |
0 0
\(t) p(.)

T T
 |0
&Dp(u(t))&+ | | 0 0
\(t)( p(.)& p(u(t)))

T T
 |0
&Dp(u(t))&+ | | 0 0
| p(u(t))& p(.)|,

which concludes the proof of (6.9). Moreover, we get that

\(t) # sign( p(.)& p(u(t))) H N&1-a.e. on 0, a.e. t # [0, T]. (6.15)

Step 4. The boundary condition. Let us now see that

\(t)=[z(t), &] H N&1-a.e. on 0, a.e. t # [0, T]. (6.16)

In fact, if w # BV(0) & L  (0) and v # R(0) such that v | 0 =w | 0 , we have

that
t t t

|0
( z : (s), w) 0 = | 0
( div(z : (s)), v) + ||
0 0
z : (s) } {v.

Hence
t t t
lim
:
| 0
( z : (s), w) 0 = | 0
( !(s), v) + || 0 0
z(s) } {v

t t
= | 0
( z(s), w) 0 = ||
0 0
[z(s), &] dH N&1. (6.17)
 TOTAL VARIATION FLOW 387

On the other hand, since z : (s) # X(0), if we apply Green's formula we have
that

t t t

|
0
(div(z : (s)), v) =& ||
0 0
z : (s) } {v+ ||
0 0
[z : (s), &] w.

Consequently,

t t

| 0
(z : (s), w) 0 = ||
0 0
[z : (s), &] w.

From here, taking limits in :, we get

t t

||
0 0
\(s) w= ||
0 0
[z(s), &] w

\w # BV(0) & L  (0), and t # [0, T]. (6.18)

Now, if w # L 1 (0), we take w k # BV(0) & L  (0) such that w k| 0 =T k (w).

By (6.18), we have

||
0 0
\(s) w k = |0
[z(s), &] w k .

Letting k  , it follows that

t t

|| 0 0
\(s) w= || 0 0
[z(s), &] w \w # L 1 (0), and t # [0, T],

and consequently (6.16) holds.

Step 5. Next, we prove that !=div(z) in (L 1 (0, T, BV(0) 2 ))* in the

sense of Definition 3. To do that let us first observe that (z, Dw), defined
by (2.14), is a Radon measure in Q T for all w # L 1w(0, T, BV(0)) & L  (Q T ).
Let , # D(Q T ), then

T
( (z, Dw), ,) = & | 0
( !(t)&u$: (t), w(t) ,(t)) & | QT
w(z&z : ) } { x ,

T
+ | 0
( (z : (t), Dw(t)) ,(t)).
388 ANDREU ET AL.

Then by (6.4), taking limits in :, we get

T
( (z, Dw), ,) =lim
:
|
0
( (z : (t), Dw(t)), ,(t)). (6.19)

Therefore

T
|( (z, Dw), ,) | &,&  | 0
&Dw(t)& dt,

from which it follows that (z, Dw) is a Radon measure in Q T . Moreover,

from (6.19), applying Green's formula we obtain that

|
QT
(z, Dw)=lim
:
| 0
(z : (t), Dw(t))

T T
=lim &
: \ | | 0 0
div(z : (t)) w(t)+ | |
0 0
[z : (t), &] w(t)
+
T T
=& | 0
( !(t), w(t)) + | | 0 0
[z(t), &] w(t).

Consequently

T T

|QT
(z, Dw)+ | 0
( !(t), w(t)) = | | 0 0
[z(t), &] w(t). (6.20)

Step 6. Conclusion. Finally, we are going to prove that u verifies:

T T
& | |
0 0
j(u(t)&l ) ' t + | |
0 0
'(t) &Dp(u(t)&l )&
T
+ | |
0 0
z(t) } D'(t) p(u(t)&l )

T
 | | 0 0
[z(t), &] '(t) p(u(t)&l ), (6.21)

for all ' # C  (Q T ), with '0, '(t, x)=,(t) (x), being , # D( ]0, T [ ),
 # C  (0 ), and p # P, where j(r)= r0 p(s) ds.
Let ' # C  (Q T ), with '0, '(t, x)=,(t) (x), being , # D( ]0, T [ ),
 # C  (0 ), and p # P, a # R. Let H p (r) :=  ra p(s) ds. Since u$n (t) =
div(z n (t)), multiplying by p(u n (t)) '(t) and integrating, we obtain that
 TOTAL VARIATION FLOW 389

T d
| |
0 0 dt
H p (u n (t)) '(t)

T
= | |
0 0
p(u n (t)) u$n (t) '(t)

T
= | |
0 0
div(z n (t)) p(u n (t)) '(t)

T t
=& | |0 0
(z n (t), D( p(u n (t)) '(t)))+ || 0 0
[z n (t), &] p(u n (t)) '(t)

T T
=& | |0 0
'(t) &Dp(u n (t))&& | |
0 0
z n (t) } {'(t) p(u n (t))

T
+ | |0 0
[z n (t), &] p(u n (t)) '(t)

T T
=& | |0 0
'(t) &Dp(u n (t))&& | |
0 0
z n (t) } {'(t) p(u n (t))

T T
& | |0 0
| p(u n (t))& p(.)| '(t)+ | |
0 0
[z n (t), &] p(.) '(t).

Hence, having in mind that '(0)='(T )=0, we get

T T

| |
0 0
'(t) &Dp(u n (t))&+ | |
0 0
| p(u n (t))& p(.)| '(t)

T T
=& | |
0 0
z n (t) } {'(t) p(u n (t))+ | | 0 0
[z n (t), &] p(.) '(t)

T d
& | |
0 0
H (u (t)) '(t)
dt p n
T T
=& | |
0 0
z n (t) } {'(t) p(u n (t))+ | | 0 0
[z n (t), &] p(.) '(t)

T d T
& | |
0 0 dt
(H p (u n (t)) '(t))+ | |
0 0
H p (u n (t)) ' t

T T
=& | |
0 0
z n (t) } {'(t) p(u n (t))+ | | 0 0
[z n (t), &] p(.) '(t)

T
+ | |
0 0
H p (u n (t)) ' t .
390 ANDREU ET AL.

Letting n  , it follows that

T T

| | 0 0
'(t) &Dp(u(t))&+ | |
0 0
| p(u(t))& p(.)| '(t)

T T
lim inf
n _| | 0 0
'(t) &Dp(u n (t))&+ | |
0 0
| p(u n (t))& p(.)| '(t)
&
T
=lim inf &
n _ | |
0 0
z n (t) } {'(t) p(u n (t))

T T
+ | |
0 0
[z n (t), &] p(.) '(t)+ | |
0 0
H p (u n (t)) ' t
&
T T
=& | |
0 0
z(t) } {'(t) p(u(t))+ | |
0 0
[z(t), &] p(.) '(t)

T
+ | |
0 0
H p (u(t)) ' t .

Now, using that | p(u(t))& p(.)| =[z(t), &]( p(.)& p(u(t))), we have

T T
& | |
0 0
H p (u(t)) ' t + | |
0 0
'(t) &Dp(u(t))&

T T
+ | |
0 0
z(t) } {'(t) p(u(t)) | |
0 0
[z(t), &] p(u(t)) '(t). (6.22)

Finally, given l # R and p # P, since q(r) :=p(r&l ) is an element of P, and

taking a=l, we obtain (6.21) as a consequence of (6.22) and the proof of
the existence is finished.

Proof of Theorem 1 (Uniqueness)

To prove uniqueness we shall show that the entropy solutions and semi-
group solutions coincide. As a consequence of the semigroup theory (3.1)
is satisfied. Our technique is inspired by a method introduced by Kruzhkov
[21] to prove L 1-contraction for entropy solutions for scalar conservation
laws: the doubling of variables. Carrillo [13] probably was the first to
apply Kruzhkov's method to second order equations (see also [18]).
 TOTAL VARIATION FLOW 391

Let u(t) be an entropy solution with initial datum u 0 # L 1 (0) and

u(t)=S(t) u 0 the semigroup solution with initial datum u 0 # L  (0). Then,
there exist z(t), z(t) # Z(0) with &z(t)&  1, &z(t)&  1, and

[z(t), &] # sign(T k+ (.)&T k+ (u(t))) a.e. in t # [0, T], (6.23)

[z(t), &] # sign(T k+ (.)&T k+ (u(t))) a.e. in t # [0, T], (6.24)

and such that, if r, r # R N, with &r&1, &r &1, and l 1 , l 2 # R, then

T T
& | |
0 0
j k+ (u(t)&l 1 ) ' t + | |
0 0
'(t) &DT k+ (u(t)&l 1 )&

T
+ | |
0 0
(z(t)&r) } D'(t) T k+ (u(t)&l 1 )

T
+ | |
0 0
r } D'(t) T k+ (u(t)&l 1 )

T
 | |
0 0
[z(t), &] '(t) T k+ (u(t)&l 1 ), (6.25)

and

T T
& | |
0 0
j k& (u(t)&l 2 ) ' t + | |
0 0
'(t) &DT k& (u(t)&l 2 )&

T
+ | |
0 0
(z(t)&r ) } D'(t) T k& (u(t)&l 2 )

T
+ | |
0 0
r } D'(t) T k& (u(t)&l 2 )

T
 | |
0 0
[z(t), &] '(t) T k& (u(t)&l 2 ), (6.26)

for all ' # C  (Q T ), with '0, '(t, x)=,(t) (x), being , # D( ]0, T [ ),
 # C  (0 ), and j k+ (r)= r0 T k+ (s) ds, j k& (r)= r0 T k& (s) ds.
We choose two different pairs of variables (t, x), (s, y) and consider u,
z as functions in (t, x), u, z in (s, y). Let 0, # D(]0, T [ ), 0 # D(0),
392 ANDREU ET AL.

\ n a classical sequence of mollifiers in R N and \~ n a sequence of mollifiers

in R. Define

t+s x+ y
' n (t, x, s, y) :=\ n (x& y) \~ n (t&s) ,
\ + \ +
2

2
.

Note that for n sufficiently large,

(t, x) [ ' n (t, x, s, y) # D( ]0, T [_0) \(s, y) # Q T ,

(s, y) [ ' n (t, x, s, y) # D( ]0, T [_0) \(t, x) # Q T .

Hence, for (s, y) fixed, if we take in (6.25) l 1 =u(s, y) and r=z(s, y), we
get
T T
& | |
0 0
j k+ (u(t, x)&u(s, y))(' n ) t + | |
0 0
' n &D x T k+ (u(t, x)&u(s, y))&

T
+ | |
0 0
(z(t, x)&z(s, y)) } { x ' n T k+ (u(t, x)&u(s, y))

T
+ | |
0 0
z(s, y) } { x ' n T k+ (u(t, x)&u(s, y))0. (6.27)

Similarly, for (t, x) fixed, if we take in (6.26) l 2 =u(t, x) and r =z(t, x), we
get
T T
& | |
0 0
j k& (u(s, y)&u(t, x))(' n ) s + | |
0 0
' n &D y T &
k (u(s, y)&u(t, x))&

T
+ | |
0 0
(z(s, y)&z(t, x)) } { y ' n T k& (u(s, y)&u(t, x))

T
+ | |
0 0
z(t, x) } { y ' n T k& (u(s, y)&u(t, x))0. (6.28)

Now, since T k& (r)=&T k+ (&r) and j k& (r)= j k+ (&r), we can rewrite (6.28)
as
T T
& | |
0 0
j k+ (u(t, x)&u(s, y))(' n ) s + | |
0 0
' n &D y T k+ (u(t, x)&u(s, y))&

T
+ | |
0 0
(z(t, x)&z(s, y)) } { y ' n T k+ (u(t, x)&u(s, y))

T
& | |
0 0
z(s, y) } { y ' n T k+ (u(t, x)&u(s, y))0. (6.29)
 TOTAL VARIATION FLOW 393

Integrating (6.27) in (s, y) and (6.29) in (t, x) and taking their sum yields

& | QT_QT
j k+ (u(t, x)&u(s, y))((' n ) t +(' n ) s )

+ |
QT_QT
' n &D x T k+ (u(t, x)&u(s, y))&

+ |
QT_QT
' n &D y T k+ (u(t, x)&u(s, y))&

+ |
QT_QT
(z(t, x)&z(s, y)) } ({ x ' n +{ y ' n ) T k+ (u(t, x)&u(s, y))

+ |
QT_QT
z(s, y) } { x 'T k+ (u(t, x)&u(s, y))

& |
QT_QT
z(t, x) } { y ' n T k+ (u(t, x)&u(s, y))0. (6.30)

Now, by Green's formula we have

| QT_QT
z(s, y) } { x ' n T k+ (u(t, x)&u(s, y))

+ |
QT_QT
' n &D x T k+ (u(t, x)&u(s, y))&

=& |
QT_QT
' n z(s, y) } D x T k+ (u(t, x)&u(s, y))

+ |
QT_QT
' n &D x T k+ (u(t, x)&u(s, y))&0,

and

& |QT_QT
z(t, x) } { y ' n T k+ (u(t, x)&u(s, y))

+ | QT_QT
' n &D y T k+ (u(t, x)&u(s, y))&

= |
QT_QT
' n z(t, x) } D y T k+ (u(t, x)&u(x, y))

+ | QT_QT
' n &D y T k+ (u(t, x)&u(s, y))&0.
394 ANDREU ET AL.

Hence, from (6.30), it follows that

& |
QT_QT
j k+ (u(t, x)&u(s, y))((' n ) t +(' n ) s )

+ | QT_QT
(z(t, x)&z(s, y))

_({ x ' n +{ y ' n ) T k+ (u(t, x)&u(s, y))0. (6.31)

Since,

t+s x+ y
(' n ) t +(' n ) s =\ n (x& y) \~ n (t&s) ,$
\ + \ +
2

2

and

t+s x+ y
{ x ' n +{ y ' n =\ n (x& y) \~ n (t&s) ,
\ 2 + { \ 2 + ,
passing to the limit in (6.31), it yields

& | QT
j k+ (u(t, x)&u(t, x)) .$(t) (x)

+ | QT
(z(t, x)&z(t, x)) } {(x) ,(t) T k+ (u(t, x)&u(t, x))0. (6.32)

We have to prove that

lim
n
| QT
(z(t, x)&z(t, x)) } { n (x) ,(t) T k+ (u(t, x)&u(t, x))0

for any sequence  n A 1 0 . Since !=div(z), ! =div(z ) in (L 1 (0, T,

BV(0) 2 ))*, the following integration by parts formula holds

|QT
(z&z, Dw)+ | 0
( !(t)&!(t), w(t)) dt

T
= | |
0 0
[z(t, x)&z(t, x), &] w(t, x) dH N&1 dt,
 TOTAL VARIATION FLOW 395

for all w # L 1 (0, T, BV(0)) & L  (Q T ). Set

w=((&1) ,T k+ (u&u )) { (t, x)=((x)&1)(,T k+ (u&u )) { (t, x),

where

1 t+{
(,T k+ (u&u )) { (t, x)=
{ |
t
,(s) T k+ (u(s, x)&u(s, x)) dt,

in the above formula to obtain

|QT
(z(t, x)&z(t, x)) } {((x)&1)(,T k+ (u&u )) { (t, x) dx dt

=& | QT
(z(t, x)&z(t, x)) } ((x)&1) D(,T k+ (u&u )) { (t, x) dx dt

& | QT
(!(t)&!(t))((x)&1)(,T k+ (u&u )) { (t, x)

T
+ | | 0 0
[z(t, x)&z(t, x), &]((x)&1)

_(,T k+ (u&u )) { (t, x) dH N&1 dt.

Since

| QT
(z&z ) } {(&1) ,T k+ (u&u ) dx dt

= lim
{  0+
|
QT
(z&z ) } {(&1)(,T k+ (u&u )) { dx dt,

and, using that | 0 =0, also

T
& | |
0 0
[z&z, &] ,T k+ (u&u ) dH N&1 dt

T
= lim
{  0+
| |
0 0
[z&z, &] (&1)(,T k+ (u&u )) { dH N&1 dt
396 ANDREU ET AL.

we may write

|QT
(z&z ) {,T k+ (u&u )

= | QT
(z&z ) {(&1) ,T k+ (u&u )

=lim
{
| QT
(z&z ) } (1&) D(,T k+ (u&u )) { dx dt

+ |QT
(!&! )(1&)(,T k+ (u&u )) {

T
& | | 0 0
[z&z, &] ,T k+ (u&u ) dH N&1 dt.

Now, since !, ! are the time derivatives of u, resp. u, in (L 1 (0, T, BV(0) 2 ))*,
we have that

| |
0 0
(!&! )(1&)(,T k+ (u&u )) {

T
= | | 0 0
(!&! )((1&) ,T k+ (u&u )) {

T 1
= | | 0 0
(1&) ,T k+ (u&u ) 2 &
{ {
(u&u ),

where 2 &
{ (u&u )=(u&u )(t)&(u&u )(t&{). Let v=u&u. Since

T k+ (v(t))(v(t)&v(t&{))J T+ (v(t))&J T+ (v(t&{))

k k

(J T + being the primitive of T k+ ), and ,, (1&)0, we have for { small

k
enough that

| |
0 0
(!&! )(1&)(,T k+ (u&u )) {

T J T k+ (v(t))&J T + (v(t&{))
 | | 0 0
(1&) ,
{
k

T ,(t+{)&,(t)
=& | | 0 0 {
(1&) J T+ (u&u ).
k
 TOTAL VARIATION FLOW 397

Thus, we have

|QT
(z&z ) {,T k+ (u&u )

lim
{ \| QT
(z&z ) } (1&) D(,T k+ (u&u )) { dx dt

T ,(t+{)&,(t)
& | |
0 0 {
(1&) J T + (u&u )
k +
T
& | |
0 0
[z&z, &] ,T k+ (u&u ) dH N&1 dt.

Finally, we observe that

lim
{ }| QT
(z&z )(1&) D(,T k+ (u&u )) { dx dt
}
2 | QT
(1&) , &DT k+ (u&u )& dx dt,

which enables us to write that

|QT
(z&z ) {,T k+ (u&u ) &2 |QT
(1&) , &DT k+ (u&u )& dx dt

T
& | |
0 0
,$(t)(1&) J T + (u&u )
k

T
& | |
0 0
[z&z, &] ,T k+ (u&u ) dH N&1 dt.

Let = n where  n A 1 0 in the last expression. Using that &DT k+ (u(t)&
u(t))& is a Radon measure a.e. in t with &DT k+ (u(t)&u(t))& # L 1 (0, T ),
letting n  , we obtain

lim
n | QT
(z&z ) { n ,T k+ (u&u )

T
& | |
0 0
[z&z, &] ,T k+ (u&u ) dH N&1 dt.
398 ANDREU ET AL.

Thus

| QT
j k+ (u(t, x)&u(t, x)) ,$(t)

T
& | |
0 0
[z&z, &] ,T k+ (u&u ) dH N&1 dt0. (6.33)

Since this is true for all 0, # D( ]0, T [ ), we get

d
dt | 0
j k+ (u(t, x)&u(t, x))0.

Hence

|0
j k+ (u(t, x)&u(t, x)) | 0
j k+ (u 0 &u 0 ).

Then, letting k  , we obtain

|0
(u(t, x)&u(t, x)) +  |0
(u 0 &u 0 ) +.

From here we deduce that

&u(t)&u(t)& 1 &u 0 &u 0 & 1 , \t0.

Hence, taking u n (t)=S(t) u 0, n , u 0, n # L  (0), and u 0, n  u 0 in L 1 (0), we

have

&u(t)&u n (t)& 1 &u 0 &u 0, n & 1 , \t0.

Consequently, letting n  , u(t)=S(t) u 0 , and the proof of the unique-

ness concludes. K

7. REGULARITY FOR POSITIVE INITIAL DATA

In this section we shall see that when the initial data are nonnegative, the
semigroup solutions are strong solution.
We need to consider truncatures T a, b , a<b, defined by

a if r<a
T a, b (r) := r
{ b
if arb
if r>b.
 TOTAL VARIATION FLOW 399

Proposition 5. Let u 0 # L 1 (0), . # L 1 (0). Let (S(t)) t0 be the semi-

group generated by A. . Then, if u(t)=S(t) u 0 ,

(i)

t t

| 0
j(u(t))+ | 0
8( p(u(s))) || 0 0
| p(.)| + |
0
j(u 0 ),

where p is a truncature ( p=T a, b ), j is the primitive of p, and 8 is the

functional defined by (4.22).
(ii) p(u) t # L 2loc(0, {, L 2 (0)), for every truncature p as above. Moreover,
we have the estimate

t
8( p(u(t)))+ || s 0
| p(u) t | 2 C,

where C>0 depends on s, &u 0 & L1 , &.& L1 , and p.

Proof. (i) First, assume that u 0 # L 2 (0)

d
dt |
0
j(u)= | 0
p(u) u t = | 0
p(u) div(z)

=& |
0
z } Dp(u)+ |
0
[z, &] p(u)

=& |
0
|Dp(u)| + |0
[z, &]( p(u)& p(.)+ p(.))

=& |
0
|Dp(u)| & |0
| p(u)& p(.)| + |
0
[z, &] p(.).

Integrating this expression, we obtain

T T

|0
j(u(t))+ |0
8( p(u(t))) | |0 0
| p(.)| + | 0
j(u 0 ). (7.1)

Since j has linear growth at infinity, if u 0 # L 1 (0), the estimate in (i)

follows by approximating u 0 by functions u 0n # L 2 (0) and passing to the
limit.
400 ANDREU ET AL.

(ii) Assume first that u 0 # L 2 (0). Let $>0 and t, s$ such that
(u(t), &u t (t)), (u(s), &u t (s)) # A. . We know that

| 0
( p(u(t))&w) u t (t) | 0
(z(t), Dw)&&Dp(u(t))&+ |
0
|w& p(.)|

& |
0
| p(u(t))& p(.)|

for all w # BV(0) & L  (0). Setting w= p(u(s)) in the above expression we
have

8( p(u(t)))&8( p(u(s))) |
0
u t (t)( p(u(s))& p(u(t))).

Using the estimate for semigroups generated by subdifferentials in L 2 (see

for instance [12, Theorem 3.2]) we have

8( p(u(t)))&8( p(u(s)))C($) &u 0 & 2 &p(u(s))& p(u(t))& 2 .

Since a similar estimate holds with s and t interchanged, we have

|8( p(u(t)))&8( p(u(s)))| C($) &u 0 & 2 &p(u(s))& p(u(t))& 2 . (7.2)

Since u # W 1, 1 2
loc ((0, {), L (0)), i.e., is a locally absolutely continuous func-
tion of time, then p(u) is also, and, from (7.2), we deduce that 8( p(u)) is
absolutely continuous in [0, {] for all {>0. Let t # [0, ) be such that u,
p(u), 8( p(u)) are differentiable at t and (u(t), &u t (t)) # A. . Set w=
p(u(t+=)), w= p(u(t&=)) in the above expression to obtain

|0
( p(u(t))& p(u(t\=))) u t (t)8( p(u(t\=)))&8( p(u(t))).

Letting =  0+ we have

d
dt
8( p(u(t)))+ | 0
p$(u(t)) u t (t) 2 =0.

In particular, since p$ is either 0 or 1, we have

d
dt
8( p(u(t)))+ | 0
| p(u) t (t)| 2 0.
 TOTAL VARIATION FLOW 401

In particular 8( p(u(t)) is a decreasing function of t. If u 0 # BV(0) & L 2 (0),

integrating from 0 to t we get

t
8( p(u(t)))+ ||
0 0
| p(u) t | 2 8(u 0 ).

Observe that by the estimate in (i), if u 0 # L 2 (0), then u(s) # BV(0) &
L 2 (0) for almost all s>0 and we have

t
8( p(u)(t))+ || s 0
| p(u) t | 2 8( p(u)(s)),

for almost all 0<s<t. Now, let u 0 # L 1 (0) and u 0n # L 2 (0) be such that
u 0n  u 0 in L 1 (0). Then, if u n (t, x) denotes the solution corresponding to
initial datum u 0n we have

t
8( p(u n )(t))+ ||
s 0
| p(u n ) t | 2 8( p(u n )(s)), (7.3)

for almost all 0<s<t and all n. Now, observe that by the estimate in (i),

|0
8( p(u n )({))C

for some constant C>0. Let $>0. Then

t $

|0
8( p(u n )({)) |
0
8( p(u n )({))8( p(u n )($)) $.

Consequently, 8( p(u n )(s)) is a bounded sequence for almost all s>0.

Thus, for a.e. s>0, the left hand side of (7.3) is bounded. Hence, we may
assume that p(u n (t))  p(u(t)) in L 1 (0) for a.e. t>0. Now, we may pass
to the limit in (7.3) and use the lower semicontinuity of the left hand side
to obtain that

t
8( p(u)(t))+ ||
s 0
| p(u) t | 2 C, (7.4)

where C depends on s, &u 0 & L1 , &.& L1 , p. K

Theorem 5. Let u 0 # L 1 (0), . # L 1 (0). Suppose that u 0 + M  0,

.+M0 (or u 0 &M0, .&M0) for some M0. Let (S(t)) t0 be the
semigroup generated by A. . Then, if u(t)=S(t) u 0 , u t # L 1loc(0, T, L 1 (0)).
402 ANDREU ET AL.

Proof. It is easy to check via the resolvents that the semigroup solution
corresponding to the data u 0 \M, .\M coincides with the semigroup
solution corresponding to the data u 0 , . plus the constant \M; i.e.,
S(t)(u 0 \M, .\M)=S(t)(u 0 , .)\M. Thus, without loss of generality we
may assume that M=0. Let us prove the theorem in case u 0 , .0, the
other case being analogous. We know, by the homogeneity estimate,
Proposition 4, that u t is a Randon measure in (s, t)_0, for all 0<s<t.
Thus, its mass is bounded; i.e.,
t

||
s 0
|u t | <.

Now, taking p=T a, b , the estimate in (ii) of the previous proposition says
that u t is a function in L 2 (Q a, b ), for all a<b, where Q a, b =
[(t, x) # Q: a<u(t, x)<b]. Thus, this, with the last integral bound, proves
that u t # L 1loc(0, {, L 1 (0)). K
Remark 3. Under the assumption of the above theorem, since u t #
L 1loc(0, T, L 1 (0)), working as in [2] we can prove that u is a strong solu-
tion. Consequently, existence and uniqueness can be obtained in an easier
way than in the general case using the same technique as in [2].

ACKNOWLEDGMENTS

The first and fourth authors have been partially supported by the Spanish DGICYT,
Project PB98-1442. The second and third authors acknowledge partial support by the TMR
European Project ``Viscosity Solutions and their Applications'' reference FMRX-CT98-0234.
This paper was written while the third author was visiting the Centre de Recerca Matematica
de la Universitat Autonoma de Barcelona. He expresses his gratitude to this center for its kind
hospitality.

REFERENCES

1. L. Ambrosio, N. Fusco, and D. Pallara, ``Functions of Bounded Variation and Free

Discontinuity Problems,'' Oxford Mathematical Monographs, Oxford Univ. Press, Oxford,
2000.
2. F. Andreu, C. Ballester, V. Caselles, and J. M. Mazon, Minimizing total variation flow,
Differential Integral Equations, in press.
3. F. Andreu, J. M. Mazon, S. Segura, and J. Toledo, Quasilinear elliptic and parabolic
equations in L 1 with nonlinear boundary conditions, Adv. Math. Sci. Appl. 7 (1997),
183213.
4. F. Andreu, J. M. Mazon, S. Segura, and J. Toledo, Existence and uniqueness for a
degenerate parabolic equation with L 1-data, Trans. Amer. Math. Soc. 315 (1999), 285306.
 TOTAL VARIATION FLOW 403

5. G. Anzellotti, Pairings between measures and bounded functions and compensated com-
pactness, Ann. Mat. Pura Appl. (4) 135 (1983), 293318.
6. G. Anzellotti, The Euler equation for functionals with linear growth, Trans. Amer. Math.
Soc. 290 (1985), 483501.
7. C. Ballester, M. Bertalmio, V. Caselles, and G. Sapiro, Interpolation of Plane Vector
Fields and Applications, preprint.
8. Ph. Benilan, L. Boccardo, T. Gallouet, R. Gariepy, M. Pierre, and J. L. Vazquez, An
L 1-theory of existence and uniqueness of solutions of nonlinear elliptic equations, Ann.
Scuola Norm. Sup. Pisa (4) 22 (1995), 241273.
9. Ph. Benilan and M. G. Crandall, Completely accretive operators, in ``Semigroups Theory
and Evolution Equations'' (Ph. Clement et al., Eds.), pp. 4176, Dekker, New York, 1991.
10. Ph. Benilan, M. G. Crandall, and A. Pazy, ``Evolution Equations Governed by Accretive
Operators,'' in press.
11. D. Blanchard and F. Murat, Renormalized solutions of nonlinear parabolic problems with
L 1-data: existence and uniqueness, Proc. Roy. Soc. Edinburgh A 127 (1997), 11371152.
12. H. Brezis, ``Operateurs Maximaux Monotones,'' North-Holland, Amsterdam, 1973.
13. J. Carrillo, On the uniqueness of solution of the evolution dam problem, Nonlinear Anal.
22 (1994), 573607.
14. M. G. Crandall and T. M. Liggett, Generation of semigroups of nonlinear transformations
on general Banach spaces, Amer. J. Math. 93 (1971), 265298.
15. J. Diestel and J. J. Uhl, Jr., ``Vector Measures,'' Math. Surveys Monogr., Vol. 15, Amer.
Math. Soc., Providence, RI, 1977.
16. I. Ekeland and R. Temam, ``Convex Analysis and Variational Problems,'' North-Holland,
Amsterdam, 1976.
17. L. C. Evans and R. F. Gariepy, ``Measure Theory and Fine Properties of Functions,''
Studies in Advanced Math., CRC Press, Boca Raton, FL, 1992.
18. G. Gagneux and M. Madaune-Tort, Unicite des solutions faibles d'equations de diffusion-
convection, C. R. Acad. Sci. Paris Ser. I Math. 318 (1994), 919924.
19. M. Giaquinta, G. Modica, and J. Soucek, Functionals with linear growth in the calculus
of variations I, Comment. Math. Univ. Carolinae 20 (1979), 143156.
20. D. Kinderlehrer and G. Stampacchia, ``An Introduction to Variational Inequalities and
Their Applications,'' Academic Press, San Diego, 1980.
21. S. N. Kruzhkov, First order quasilinear equations in several independent variables, Mat.
Sb. 10 (1970), 217243.
22. A. Lichnewski and R. Temam, Pseudosolutions of the time dependent minimal surface
problem, J. Differential Equations 30 (1978), 340364.
23. C. B. Jr. Morrey, ``Multiple Integrals in the Calculus of Variations,'' Springer-Verlag,
BerlinNew York, 1966.
24. L. Rudin and S. Osher, Total variation based image restoration with free local constraints,
in ``Proc. of the IEEE ICIP-94 Austin, TX,'' Vol. 1, pp. 3135, 1994.
25. L. Rudin, S. Osher, and E. Fatemi, Nonlinear total variation based noise removal
algorithms, Physica D 60 (1992), 259268.
26. L. Schwartz, ``Fonctions mesurables et *-scalairement mesurables, mesures banachiques
majores, martingales banachiques, et propriete de RadonNikodym,'' Sem. Maurey
Schwartz 19674-75, Ecole Polytech., Centre de Math.
27. R. Temam, On the continuity of the trace of vector functions with bounded deformation,
Appl. Anal. 11 (1981), 291302.
28. W. P. Ziemer, ``Weakly Differentiable Functions,'' Graduate Texts in Mathematics,
Vol. 120, Springer-Verlag, BerlinNew York, 1989.