Lecture notes

on the Stefan problem

Daniele Andreucci
Dipartimento di Metodi e Modelli Matematici
Universit`a di Roma La Sapienza
via Antonio Scarpa 16 00161 Roma, Italy
[email protected]
Introduction
These notes cover a portion of a course of the Doctorate Modelli e metodi mate-
matici per la tecnologia e la societ`a, given in 2001 in Rome. The title of the course
was Evolution equations and free boundary problems and its topics included,
essentially, an introduction to Stefan and Hele-Shaw problems.
Here only the material concerning the Stefan problem is partially reproduced.
The present notes assume the reader has some knowledge of the elementary theory
of L
p
and Sobolev spaces, as well as of the basic results of existence and regularity
of solutions to smooth parabolic equations.
The bibliography is minimal; only books and articles quoted in the text are
referenced. See [12], [18] and [20] for further references.
I thank the audience of the course for many stimulating comments and questions,
and prof. R. Ricci for interesting discussions on the subject of these notes.
Rome, January 2002
Revised version: Rome, January 2004
iii
Notation
The notation employed here is essentially standard.
Appendix D contains a list of the main symbols used in the text.
Constants denoted by , C, C

, . . . may change fom line to line.

v
Contents
Introduction iii
Notation v
Chapter 1. The classical formulation 1
1. The Stefan condition 1
2. The free boundary problem 4
3. Explicit examples of solutions 5
4. Basic estimates 7
5. Existence and uniqueness of the solution 8
6. Qualitative behaviour of the solution 12
7. Regularity of the free boundary 15
Chapter 2. Weak formulation of the Stefan problem 19
1. From the energy balance to the weak formulation 19
2. Comparing the weak and the classical formulations 20
3. Denition of weak solution 24
4. Uniqueness of the weak solution 25
5. Existence of weak solutions 29
6. A comparison result 37
Appendix A. Maximum principles for parabolic equations 39
1. The weak maximum principle 39
2. The strong maximum principle 41
3. Hopfs lemma (the boundary point principle) 45
4. Maximum principle for weak solutions 48
Appendix B. A theorem by Kruzhkov 51
1. Mollifying kernels 51
2. The main estimate 53
Appendix C. The spaces H
m+,
m+
2
(Q
T
) 55
1. Comments 55
Appendix D. Symbols used in text 57
Bibliography 59
vii
CHAPTER 1
The classical formulation
In this chapter we consider the formulation of the Stefan problem as a classical
initial boundary value problem for a parabolic partial dierential equation. A
portion of the boundary of the domain is a priori unknown (the free boundary),
and therefore two boundary conditions must be prescribed on it, instead than
only one, to obtain a well posed problem.
1. The Stefan condition
The Stefan problem ([17]) is probably the simplest mathematical model of a
phenomenon of change of phase. When a change of phase takes place, a latent
heat is either absorbed or released, while the temperature of the material changing
its phase remains constant. In the following we denote by L > 0 the latent heat
per unit of volume (p.u.v.), and neglect for the sake of simplicity any volume
change in the material undergoing the change of phase. We also assume the
critical temperature of change of phase to be a constant,
0
.
To be specic, consider at time t = t
0
a domain A divided by the plane x
1
= s
0
into two subdomains. At time t = t
0
the sub-domain A
1
= A x
1
< s
0
is
lled by water, while A
2
= A x
1
> s
0
is lled by ice. In the terminology of
problems of change of phase, A
1
is the liquid phase and A
2
is the solid phase.
The surface separating the two phases is referred to as the interface. Assume also
the setting is plane symmetric, that is the temperature is a function of x
1
only,
besides the time t, and the interface is a plane at all times. Denote by x
1
= s(t)
the position of the interface at time t. Note that, due to the natural assumption
that temperature is continuous,
(s(t)+, t) = (s(t), t) =
0
, for all t. (1.1)
Assume ice is changing its phase, that is the interface is advancing into the solid
phase. Due to the symmetry assumption we stipulate, we may conne ourselves
to consider any portion D, say a disk, of the interface at time t
0
. At a later
time t
1
> t
0
the interface occupies a position s(t
1
) > s(t
0
) = s
0
. The cylinder
D (s(t
1
), s(t
0
)) has been melted over the time interval (t
0
, t
1
) (see Figure 1).
The change of phase has therefore absorbed a quantity of heat
volume of the melted cylinder latent heat p.u.v. = area(D)(s(t
1
) s(t
0
))L.
(1.2)
The heat must be provided by diusion, as we assume that no heat source or sink
is present. We adopt for heat diusion Fouriers law
heat ux = k
i
D , (1.3)
where k
1
> 0 is the diusivity coecient in water, and k
2
> 0 is the diusivity
coecient in ice (in principle k
1
,= k
2
). Thus, the quantity of heat in (1.2) must
1
2 DANIELE ANDREUCCI
water
s(t
0
) s(t
1
)
ice
e
1
Figure 1. Melting ice
equal
t
1
_
t
0
_
D(t)
[k
1
D(s(t), t) e
1
k
2
D(s(t)+, t) (e
1
)] dx
2
dx
3
dt =
area(D)
t
1
_
t
0
[k
1

x
1
(s(t), t) +k
2

x
1
(s(t)+, t)] dt . (1.4)
Equating the two quantities, dividing the equation by t
1
t
0
and letting t
1
t
0
,
we nally nd
k
1

x
1
(s(t), t) +k
2

x
1
(s(t)+, t) = L s(t) , (1.5)
where we have substituted t
0
with the general time t, as the same procedure can
be obviously carried out at any time. This is called the Stefan condition on the
free boundary. We stress the fact that the Stefan condition is merely a law of
energetical balance.
Several remarks are in order.
Remark 1.1. Note that, although we did not assume anything on the values of
(x
1
, t) inside each one of the two phases, on physical grounds we should expect

0
, in water, i.e., in A
1
;
0
, in ice, i.e., in A
2
. (1.6)
The equality
0
in either one of the two phases (or in both) can not be ruled
out in the model. Rather, it corresponds to the case when a whole phase is at
critical temperature. Diusion of heat, according to (1.3), can not take place in
that phase, as
x
1
0 there. Thus (1.5) reduces to, e.g., if the solid phase is at
constant temperature,
k
1

x
1
(s(t), t) = L s(t) . (1.7)
Note that if >
0
in water, (1.7) predicts that s(t) > 0. In other words,
melting of ice is predicted by the model, instead of solidication of water. This
is consistent with obvious physical considerations.
Problems where one of the two phases is everywhere at the critical temperature
1. THE STEFAN CONDITION 3
are usually referred to (somehow misleadingly) as one phase problems, while the
general case where (1.5) is prescribed is the two phases problem. In the latter
case, the sign of s(t), and therefore the physical behaviour of the system water/ice
predicted by the mathematical model, depends on the relative magnitude of the
two heat uxes at the interface.
Remark 1.2. On the interface, which is also known as the free boundary two
conditions are therefore prescribed: (1.1) and (1.5).
In the case of a one phase problem, this fact has the following meaningful inter-
pretation in terms of the general theory of parabolic PDE: The boundary of the
domain where the heat equation (a parabolic equation based on (1.3)) is posed,
contains an a priori unknown portion, corresponding to the interface separat-
ing the liquid phase from the solid one. Clearly, if only the Dirichlet boundary
condition (1.1) was imposed on it, we could choose arbitrarily this part of the
boundary, and solve the corresponding initial value boundary problem. Appar-
ently the solution would not, in general, satisfy (1.7). A similar remark applies
to solutions found imposing just (1.7) (where now the functional form of the ar-
bitrarily given boundary x
1
= s(t) is explicitly taken into account).
It is therefore evident that on the free boundary both conditions (1.1) and (1.7)
should be prescribed in order to have a well posed problem. (Or, anyway, in more
general free boundary problems, two dierent boundary conditions are required.)
Incidentally, this circle of ideas provides the basic ingredient of a possible proof
of the existence of solutions: we assign arbitrarily a candidate free boundary s

and consider the solution to the problem, say, corresponding to the data (1.1).
Then we dene a transformed boundary s

exploiting (1.7), i.e.,

k
1

x
1
(s

(t), t) = L s

(t) .
A xed point of this transform corresponds to a solution of the complete problem.
Remark 1.3. The meaning of conditions (1.1) and (1.5) in the context of two
phases problems is probably better understood in terms of the weak formulation
of the Stefan problem, which is discussed below in Chapter 2. We remark here
that, actually, one phase problems are just two phases problems with one phase
at constant temperature, so that the discussion in Chapter 2 applies to them too.
Remark 1.4. Problems where the temperature restriction (1.6) is not fullled,
are sometimes called undercooled Stefan problems. We do not treat them here,
albeit their mathematical and physical interest (see however Subsection 2.4 of
Chapter 2); let us only recall that they are, in some sense, ill posed.
1.1. Exercises.
1.1. Write the analogs of Stefan condition (1.5) in the cases of cylindrical and
spherical symmetry in R
3
.
1.2. Note that if we assume, in (1.5), <
0
in A
1
and
0
in A
2
, we nd
s(t) < 0. This appears to be inconsistent with physical intuition: a phase of
ice at sub critical temperature should grow into a phase of water at identically
critical temperature. Indeed, (1.5) is not a suitable model for the physical setting
considered here. In other words, the Stefan condition in the form given above
keeps memory of which side of the interface is occupied by which phase. Find
out where we implicitly took into account this piece of information and write the
Stefan condition when ice and water switch places.
4 DANIELE ANDREUCCI
1.3. Prove that Stefan condition (1.5) does not change its form if bounded volu-
metric heat sources are present (i.e., if the heat equation is not homogeneous).
2. The free boundary problem
Keeping the plane symmetry setting considered above, we may of course assume
the problem is one dimensional. Denoting by x the space variable, the complete
two phases problem can be written as
c
1

t
k
1

xx
= 0 , in Q
1
, (2.1)
c
2

t
k
2

xx
= 0 , in Q
2
, (2.2)
k
1

x
(0, t) = h
1
(t) , 0 < t < T , (2.3)
k
2

x
(d, t) = h
2
(t) , 0 < t < T , (2.4)
(x, 0) = (x) , 0 < x < d , (2.5)
k
1

x
(s(t), t) +k
2

x
(s(t)+, t) = L s(t) , 0 < t < T , (2.6)
(s(t), t) = (s(t)+, t) =
0
, 0 < t < T , (2.7)
s(0) = b . (2.8)
Here 0 < b < d, T and c
1
, c
2
, k
1
, k
2
are given positive numbers. The c
i
represent
the thermal capacities in the two phases. The liquid phase occupies at the initial
time t = 0 the interval (0, b), while the solid phase occupies (b, d). The problem
is posed in the time interval (0, T). Moreover we have set
Q
1
= (x, t) [ 0 < x < s(t) , 0 < t < T ,
Q
2
= (x, t) [ s(t) < x < d , 0 < t < T .
We are assuming that 0 < s(t) < d for all 0 < t < T. If the free boundary
hits one of the two xed boundaries x = 0 and x = d, say at time t

, of course
the formulation above should be changed. In practice, one of the two phases
disappears at t = t

. We leave to the reader the simple task of writing the

mathematical model for t > t

.
One could impose other types of boundary data, instead of (2.3), (2.4), e.g.,
Dirichlet data.
If we are to attach the physical meaning of a change of phase model to the problem
above, the data must satisfy suitable conditions. At any rate
(x)
0
, 0 < x < b ; (x)
0
, b < x < d .
Essentially, we need
0
in Q
1
and
0
in Q
2
.
Actually, we will deal mainly with the one phase version of (2.1)(2.8) where
the solid phase is at constant temperature. Namely, after adimensionalization,
we look at
u
t
u
xx
= 0 , in Q
s,T
, (2.9)
u
x
(0, t) = h(t) > 0 , 0 < t < T , (2.10)
u(x, 0) = u
0
(x) 0 , 0 < x < b , (2.11)
u
x
(s(t), t) = s(t) , 0 < t < T , (2.12)
u(s(t), t) = 0 , 0 < t < T , (2.13)
s(0) = b . (2.14)
3. EXPLICIT EXAMPLES OF SOLUTIONS 5
(We have kept the old names for all variables excepting the unknown u.) Here
we denote for each positive function s C([0, T]), such s(0) = b,
Q
s,T
= (x, t) [ 0 < x < s(t) , 0 < t < T .
We regard the rescaled temperature u as a function dened in Q
s,T
. The solid
phase therefore does not appear explicitly in the problem. As a matter of fact
we assume it to be unbounded in the positive x direction (i.e., d = +), so that
no upper limit has to be imposed on the growth of the free boundary s. The
sign restrictions in (2.10) and in (2.11) are imposed so that u > 0 in Q
s,T
, see
Proposition 4.1 below.
Definition 2.1. A solution to problem (2.9)(2.14) is a pair (u, s) with
s C
1
((0, T]) C([0, T]) , s(0) = b , s(t) > 0 , 0 t T ;
u C(Q
s,T
) C
2,1
(Q
s,T
) , u
x
C(Q
s,T
t = 0) ,
and such that (2.9)(2.14) are satised in a classical pointwise sense.
We prove a theorem of existence and uniqueness of solutions to (2.9)(2.14),
under the assumptions
h C([0, T]) , h(t) > 0 , 0 t T ; (2.15)
u
0
C([0, b]) , 0 u
0
(x) H(b x) , 0 x b . (2.16)
We also study some qualitative behaviour of the solution. The adimensionaliza-
tion of the problem does not play a substantial role in the mathematical theory
we develop here.
For further reading on the one-phase Stefan problem, we refer the reader to [5],
[3]; we employ in this chapter the techniques found there, with some changes.
Remark 2.1. The free boundary problem (2.9)(2.14) is strongly non linear, in
spite of the linearity of the PDE and of the boundary conditions there. Indeed,
recall that the free boundary s itself is an unknown of the problem; its dependence
on the data is not linear (as, e.g., the explicit examples of Section 3 show).
2.1. Exercises.
2.1. Prove that the change of variables
u +
0
, x , t ,
allows one to write the one phase problem in the adimensionalized form (2.9)
(2.14), for a suitable choice of the constants , , .
Also note that adimensionalizing the complete two phases problem similarly is in
general impossible.
3. Explicit examples of solutions
Example 1. An explicit solution of the heat equation (2.9) is given by
v(x, t) = erf
_
x
2

t
_
=
2

x/2

t
_
0
e
z
2
dz , x, t > 0 ,
v(x, 0) = 1 , x > 0 .
6 DANIELE ANDREUCCI
Here erf denotes the well known error function. Fix C > 0 arbitrarily, and set,
for a > 0 to be chosen presently,
u(x, t) = C
_
erf erf
_
x
2

t
__
.
Dene also s(t) = 2

t; note that s(0) = 0. Thus u > 0 in Q

s,T
, and (2.9) as
well as (2.13) are satised. By direct calculation
u
x
(x, t) =
C

t
e

x
2
4t
.
Hence,
u
x
(s(t), t) = u
x
(2

t, t) =
C

t
e

2
= s(t) =

t
,
if and only if
C =

2
.
Note that on the xed boundary x = 0 we may select either one of the conditions
u
x
(0, t) =
C

t
, or u(0, t) = C erf .
We have proven that, when is chosen as above,
u(x, t) = 2e

_
x/2

t
e
z
2
dz (3.1)
solves the problem sketched in Figure 2. Note however that u is not continuous
at (0, 0); the notion of solution in this connection should be suitably redened.
u
x
= C/

t
or
u = C erf
t
x
s(t) = 2

t
u = 0
u
x
= s(t) = /

t
u
t
u
xx
= 0
Figure 2. The Stefan problem solved by u in (3.1)
Example 2. It is obvious by direct inspection that the function
u(x, t) = e
tx
1 , (3.2)
solves the Stefan problem in Figure 3, corresponding to the free boundary s(t) = t.
Note that we are forced to prescribe an exponentially increasing ux on x = 0 in
order to obtain a linear growth for s(t).
4. BASIC ESTIMATES 7
u
x
= e
t
or
u = e
t
1
t
x
s(t) = t
u = 0
u
x
= s(t) = 1
u
t
u
xx
= 0
Figure 3. The Stefan problem solved by u in (3.2)
3.1. Exercises.
3.1. Convince yourself that the solutions corresponding to u
x
(0, t) = 2C/

t,
in the case of Example 1, and to u
x
(0, t) = 2e
t
in the case of Example 2, can
not be obtained by linearity from the ones given above.
4. Basic estimates
Proposition 4.1. If (u, s) is a solution to (2.9)(2.14), then
u(x, t) > 0 , in Q
s,T
; (4.1)
s(t) > 0 , for all t > 0. (4.2)
Proof. By virtue of the weak maximum principle, u must attain its maximum
on the parabolic boundary of Q
s,T
, i.e., on
Q
s,T
(x, t) [ t = T , 0 < x < s(T) .
The data h being positive, the maximum is attained on t = 0 or on x = s(t).
Therefore u 0 (remember that u
0
0). If we had u( x,

t) = 0 in some ( x,

t)
Q
s,T
, invoking the strong maximum principle we would obtain u 0 in Q
s,T
t <

t. This is again inconsistent with h > 0. Thus (4.1) is proven.

Then, the value u = 0 attained on the free boundary is a minimum for u. Re-
calling the parabolic version of Hopfs lemma, we infer
s(t) = u
x
(s(t), t) > 0 , for all t > 0.

Remark 4.1. The proof of estimate (4.1) does not make use of the Stefan condi-
tion (2.12). In the same spirit, we consider in the following results solutions to the
initial value boundary problem obtained removing Stefan condition from the for-
mulation. The rationale for this approach is that we want to apply those results
to approximating solutions constructed according to the ideas of Remark 1.2.
Lemma 4.1. Let u be a solution of (2.9), (2.10), (2.11), (2.13), where s is assumed
to be a positive non decreasing Lipschitz continuous function in [0, T], such that
s(0) = b. Let (2.15) and (2.16) be in force. Then u
x
is continuous up to all the
points of the boundary of Q
s,T
of the form (0, t), (s(t), t), with T t > 0.
8 DANIELE ANDREUCCI
The regularity of u
x
up to x = 0 is classical; the proof of this result will be
completed in Section 7, see Lemma 7.1.
Proposition 4.2. Let u, s be as in Lemma 4.1. Then
0 < u(x, t) M(s(t) x) , in Q
s,T
, (4.3)
where M = max(|h|

, H).
Proof. Dene v(x, t) = M(s(t) x). It follows immediately
v
t
v
xx
= M s(t) 0 , in Q
s,T
;
v(s(t), t) = u(s(t), t) = 0 , 0 t T ,
v
x
(0, t) = M h = u
x
(0, t) , 0 < t < T ,
v(x, 0) = M(b x) u
0
(x) = u(x, 0) , 0 x b .
Therefore, taking into account the results of Appendix A,
v(x, t) u(x, t) , in Q
s,T
.

Corollary 4.1. If u, s are as in Proposition 4.2, then

0 > u
x
(s(t), t) M , 0 < t < T , (4.4)
where M is the constant dened in Proposition 4.2.
Proof. A trivial consequence of Proposition 4.1 and of Proposition 4.2, as well
as of Lemma 4.1. We also keep in mind Remark 4.1, and of course make use of
Hopfs lemma for the strict inequality in (4.4).
Remark 4.2. We stress the fact that the barrier construction of Proposition 4.2
is made possible by the fact that s 0. In turn, this is for solutions of the Stefan
problem a consequence of the positivity of u. Thus, such a barrier function, and in
general any similar barrier function, does not exist in the case of the undercooled
Stefan problem.
4.1. Exercises.
4.1. Note that the function v dened in the proof of Proposition 4.2 is just
Lipschitz continuous in t. In spite of this fact, one may apply to v u the weak
maximum principle in the form given in Section 4 of Appendix A. Carry out the
proof in detail.
5. Existence and uniqueness of the solution
We prove here
Theorem 5.1. Assume (2.15), (2.16). Then there exists a unique solution to
(2.9)(2.14).
Let u, s be as in Proposition 4.2. Moreover assume that
s := Lip([0, T]) [ 0 M , (0) = b .
Here M is the constant dened in Proposition 4.2. The set is a convex compact
subset of the Banach space C([0, T]), equipped with the max norm. A useful
property of all s is
b +Mt s(t) b , 0 t T .
5. EXISTENCE AND UNIQUENESS OF THE SOLUTION 9
Dene the transform T (s) by
T (s)(t) = b
t
_
0
u
x
(s(), ) d , T t 0 .
Note that, as a consequence of Lemma 4.1 and of Corollary 4.1,
T (s) Lip([0, T]) C
1
((0, T]) ,
and
M
d
dt
T (s)(t) = u
x
(s(t), t) 0 , t > 0 .
Then T : . Also note that a xed point of T corresponds to a solution of
our Stefan problem.
We use the divergence theorem to transform the boundary ux integral dening
T (s). Namely we write
0 =
t
_
0
s()
_
0
(u

u
xx
) dxd =
b
_
0
u
0
(x) dx
t
_
0
u(s(), ) s() d
+
s(t)
_
0
u(x, t) dx
t
_
0
u
x
(s(), ) d
t
_
0
h() d .
Therefore
T (s)(t) = b
t
_
0
u
x
(s(), ) d =
b +
b
_
0
u
0
(x) dx +
t
_
0
h() d
s(t)
_
0
u(x, t) dx =: F(t)
s(t)
_
0
u(x, t) dx. (5.1)
Note that this equality allows us to express T (s) in terms of more regular func-
tions than the ux u
x
(s(t), t), which appeared in its original denition. We are
now in a position to prove that T is continuous in the max norm.
Let s
1
, s
2
. Let us dene
(t) = min(s
1
(t), s
2
(t)) , (t) = max(s
1
(t), s
2
(t)) ,
i = 1 , if (t) = s
1
(t), i = 2 , otherwise.
(The number i is a function of time; this will not have any specic relevance.)
Let us also dene
v(x, t) = u
1
(x, t) u
2
(x, t) .
Then v satises
v
t
v
xx
= 0 , in Q
,T
, (5.2)
v
x
(0, t) = 0 , 0 < t < T , (5.3)
v(x, 0) = 0 , 0 < x < b , (5.4)
[v((t), t)[ = [u
i
((t), t)[ M((t) (t)) , 0 < t < T . (5.5)
10 DANIELE ANDREUCCI
Therefore, we may invoke the maximum principle to obtain
|v|
,t
:= max
Q
,t
[v[ M|s
1
s
2
|
,t
, (5.6)
where we also denote
|s
1
s
2
|
,t
= max
0t
[s
1
() s
2
()[ .
On the other hand, we have
T (s
1
)(t) T (s
2
)(t) =
s
2
(t)
_
0
u
2
(x, t) dx
s
1
(t)
_
0
u
1
(x, t) dx
=
(t)
_
0
v(x, t) dx + (1)
i
(t)
_
(t)
u
i
(x, t) dx.
Therefore
[T (s
1
)(t) T (s
2
)(t)[ [(t)[|v|
,t
+M((t) (t))
2
(b +MT)M|s
1
s
2
|
,t
+M|s
1
s
2
|
2
,t
, (5.7)
and the continuity of T : is proven. By Schauders theorem, it fol-
lows that a xed point of T exists, and thus existence of a solution. Uniqueness
might be proven invoking the monotone dependence result given below (see The-
orem 6.1).
However, mainly with the purpose of elucidating the role played in the theory of
free boundary problems by local integral estimates, we proceed to give a direct
proof of uniqueness of solutions. More explicitly, we prove the contractive char-
acter (for small t) of T , thereby obtaining existence and uniqueness of a xed
point.
5.1. Local estimates vs the maximum principle. Estimates like (5.6), ob-
tained through the maximum principle, have the advantage of providing an im-
mediate sup estimate of the solution in the whole domain of denition. However,
the bound they give may be too rough, at least in some regions of the domain.
Consider for example, in the setting above, a point (b/2, ), with 1. The
maximum principle predicts for v(b/2, ) a bound of order M|s
1
s
2
|
,
, that
is, obviously, the same bound satised by the boundary data for v. On the other
hand, taking into account (5.3), (5.4) one might expect v(b/2, ) to be much
smaller than v((), ).
This is indeed the case, as we show below. In order to do so, we exploit local
integral estimates of the solution, that is, estimates involving only values of v in
the region of interest, in this case, away from the boundary. We aim at proving
that T is a contraction, that is,
|T (s
1
) T (s
2
)|
,t
d|s
1
s
2
|
,t
, (5.8)
with d < 1 for small enough t. A quick glance at (5.7) shows that (5.8) does not,
indeed, follow from there (unless bM < 1). This failure is due only to the term
5. EXISTENCE AND UNIQUENESS OF THE SOLUTION 11
originating from the estimate of
(t)
_
0
v(x, t) dx.
Thus, a better estimate of this integral is needed.
A key step in any local estimation is a good choice of cut o functions. These
are, typically, non negative smooth functions equal to 1 in the region we want to
single out, and identically vanishing away from it.
Let b/2 > > 0, and dene the cut o function (x), such that
(x) = 1 , 0 x b 2 , (x) = 0 , b x,
2

x
(x) 0 .
Multiply (5.2) by v
2
and integrate by parts in Q
b,t
= [0, b] [0, t]. We get
0 =
__
Q
b,t
(v

v
2
v
xx
v
2
) dxd
=
1
2
b
_
0
v(x, t)
2
(x)
2
dx +
__
Q
b,t
v
2
x

2
dxd + 2
__
Q
b,t
vv
x

x
dxd , (5.9)
whence (using the inequality 2ab a
2
+b
2
/, > 0)
1
2
b
_
0
v(x, t)
2
(x)
2
dx +
__
Q
b,t
v
2
x

2
dxd 2
__
Q
b,t
[v
x
[ [v[ [
x
[ dxd
2
__
Q
b,t
v
2

2
x
dxd +
1
2
__
Q
b,t
v
2
x

2
dxd .
By absorbing the last integral into the left hand side, we nd
b
_
0
v(x, t)
2
(x)
2
dx +
__
Q
b,t
v
2
x

2
dxd 4
__
Q
b,t
v
2

2
x
dxd
= 4
t
_
0
b
_
b2
v
2

2
x
dxd
16

2
t|v|
2
,t
=
16

t|v|
2
,t
. (5.10)
Let us now go back to
T (s
1
)(t) T (s
2
)(t) =
(t)
_
0
v(x, t) dx + (1)
i
(t)
_
(t)
u
i
(x, t) dx
=
b2
_
0
v(x, t) dx +
(t)
_
b2
v(x, t) dx + (1)
i
(t)
_
(t)
u
i
(x, t) dx.
12 DANIELE ANDREUCCI
Use Holders inequality and (5.10) to bound the integral over (0, b 2) (recall
that 1 there), and nd
[T (s
1
)(t) T (s
2
)(t)[
_
b2
_
0
v(x, t)
2
dx
_
1/2

b + ((t) b + 2)|v|
,t
+M((t) (t))
2

b
4

t|v|
,t
+ (Mt + 2)|v|
,t
+M((t) (t))
2
.
Take now =

t (t is xed in this argument), and apply again the sup estimate

(5.6), nally obtaining
|T (s
1
) T (s
2
)|
,t

_
4M

b
4

t + M(Mt + 2

t) + M
2
t
_
|s
1
s
2
|
,t
.
Clearly, for t = t
0
(M, b), T is a contractive mapping.
5.2. Exercises.
5.1. In Subsection 5.1 we proved existence and uniqueness of a xed point of
T in a small time interval (locally in time). Show how the argument can be
completed to give existence and uniqueness of a xed point in any time interval
(global existence).
5.2. Why uniqueness of solutions (u, s) with s is equivalent to uniqueness of
solutions in the class of Denition 2.1, without further restrictions?
5.3. Give an interpretation of (5.1) as an energetical balance.
5.4. Prove that T has the property (see also Figure 4)
s
1
(t) < s
2
(t) , 0 < t < T = T (s
1
)(t) > T (s
2
)(t) , 0 < t < T .
5.5. Extend the proof of existence and uniqueness of solutions to the case when
h 0. What happens if h 0, u
0
0?
5.6. To carry out rigorously the calculations in (5.9) actually we need an ap-
proximation procedure: i.e., we need rst perform integration in a smaller 2-
dimensional domain, bounded away from the boundaries x = 0, t = 0. Recognize
the need of this approach, and go over the (easy) details.
6. Qualitative behaviour of the solution
Theorem 6.1. (Monotone dependence) Let (u
i
, s
i
) be solutions of (2.9)
(2.14), i = 1, 2, respectively corresponding to data h = h
i
, b = b
i
, u
0
= u
0i
.
Assume both sets of data satisfy (2.15), (2.16). If
h
1
(t) h
2
(t) , 0 < t < T ; b
1
b
2
; u
01
(x) u
02
(x) , 0 < x < b
1
; (6.1)
then
s
1
(t) s
2
(t) , 0 < t < T . (6.2)
Proof. 1) Let us assume rst b
1
< b
2
. Reasoning by contradiction, assume

t = inft [ s
1
(t) = s
2
(t) (0, T) .
Then the function v = u
2
u
1
is strictly positive in
0 < t <

t , 0 < x < s
1
(t) ,
6. QUALITATIVE BEHAVIOUR OF THE SOLUTION 13
t
x
b
T (s) = s
x = (t) x = T ()(t)
Figure 4. Behaviour of the transform T
by virtue of the strong maximum principle and Hopfs lemma. Indeed,
v(s
1
(t), t) > 0 , 0 < t <

t .
Then v attains a minimum at (s
1
(

t),

t) = (s
2
(

t),

t), where
v(s
1
(

t),

t) = 0 .
Thus, due to Hopfs lemma,
v
x
(s
1
(

t),

t) < 0 .
But we compute
v
x
(s
1
(

t),

t) = u
2x
(s
2
(

t),

t) u
1x
(s
1
(

t),

t) = s
2
(

t) + s
1
(

t) .
Hence s
2
(

t) > s
1
(

t), which is not consistent with the denition of

t.
2) Assume now b
1
= b
2
. Let us extend the data u
02
to zero over (b, b + ),
where 0 < < 1 is arbitrary. Let (u

, s

) be the solution of problem (2.9)(2.14)

corresponding to the data h = h
2
, b = b
2
+ , u
0
= u
02
. Then, by the rst part
of the proof, s
2
< s

, s
1
< s

, and for all 0 < t < T

s

(t) s
2
(t) =
s
2
(t)
_
0
[u

u
2
](x, t) dx
s

(t)
_
s
2
(t)
u

(x, t) dx .
Therefore s

s
2
+ , so that s
1
< s

s
2
+ . On letting 0 we recover
s
1
s
2
.
Let us investigate the behaviour of the solution of the Stefan problem for large
times. In doing so, we of course assume that T = . Note that the result
of existence and uniqueness applies over each nite time interval; a standard
extension technique allows us to prove existence and uniqueness of a solution
dened for all positive times.
Owing to Proposition 4.2, we have
s(t) S , 0 < t < = u(x, t) MS , in Q
s,
. (6.3)
14 DANIELE ANDREUCCI
Moreover, s being monotonic, certainly there exists
s

= lim
t
s(t) . (6.4)
Theorem 6.2. Let (u, s) be the solution of Theorem 5.1. Then
s

= lim
t
s(t) = b +
b
_
0
u
0
(x) dx +

_
0
h(t) dt . (6.5)
Proof. 1) Assume rst

_
0
h(t) dt = +. (6.6)
We only need show s is unbounded. Let us recall that, for all t > 0,
s(t) = b +
b
_
0
u
0
(x) dx +
t
_
0
h() d
s(t)
_
0
u(x, t) dx. (6.7)
From (6.3), it follows that, if s is bounded over (0, ), then u is also bounded
over (0, ). This is clearly inconsistent with (6.7), when we keep in mind (6.6).
2) Assume

_
0
h(t) dt < +. (6.8)
The balance law (6.7), together with u > 0, immediately yields
s(t) < b +
b
_
0
u
0
(x) dx +

_
0
h(t) dt < +. (6.9)
Then we have s(t) s

< ; it is only left to identify s

as the quantity
indicated above. Owing to (6.7) again, we only need show
lim
t
s(t)
_
0
u(x, t) dx = 0 . (6.10)
On multiplying (2.9) by u and integrating by parts in Q
s,t
, we get
1
2
s(t)
_
0
u(x, t)
2
dx +
__
Q
s,t
u
2
x
dxd =
1
2
b
_
0
u
0
(x)
2
dx +
t
_
0
u(0, )h() d . (6.11)
Recalling (6.3) and (6.9), the last integral in (6.11) is majorised by

_
0
Ms

h() d < .
Thus a consequence of (6.11) is
__
Q
s,
u
x
(x, t)
2
dxdt < +. (6.12)
7. REGULARITY OF THE FREE BOUNDARY 15
Elementary calculus then shows that
__
Q
s,
u(x, t)
2
dxdt =
__
Q
s,
_
s(t)
_
x
u

(, t) d
_
2
dxdt
s

__
Q
s,
s(t)
_
x
u

(, t)
2
d dxdt s
2

__
Q
s,
u

(, t)
2
d dt < .
Then there exists a sequence t
n
, t
n
, such that
s(t
n
)
_
0
u(x, t
n
)
2
dx 0 .
But the function
t
s(t)
_
0
u(x, t)
2
dx,
when we take into account (6.11), is easily seen to have limit as t , so that
this limit is 0. By Holders inequality,
s(t)
_
0
u(x, t) dx

s

_
s(t)
_
0
u(x, t)
2
dx
_
1/2
0 , t ,
completing the proof of (6.10).
6.1. Exercises.
6.1. Find conditions ensuring that the inequality in (6.2) is strict.
6.2. Discuss the necessity of assumptions (2.15), (2.16) in Theorem 6.1.
7. Regularity of the free boundary
The approach in this Section is taken from [16], and provides an example of
bootstrap argument, i.e., of an inductive proof where any given smoothness of
the solution allows us to prove even more regularity for it. Our rst result will
become the rst step in the induction procedure, and is however required to prove
Lemma 4.1.
Lemma 7.1. Let u C
2,1
(Q
s,T
) C(Q
s,T
), where s Lip([0, T]), and s(t) > 0
for 0 t T. Assume u fulls
u
t
u
xx
= 0 , in Q
s,T
, (7.1)
u(s(t), t) = 0 , 0 t T . (7.2)
Then for each small enough > 0, u
x
is continuous in P

, where P

= (x, t) [
< x < s(t) , < t < T.
16 DANIELE ANDREUCCI
Proof. The proof is based on standard local regularity estimates for solutions
of parabolic equations. Introduce the following change of variables
_
y =
x
s(t)
,
= t ;
v(y, ) = u(ys(), ) .
The set Q
s,T
is mapped onto R = (0, 1) (0, T), where v solves
v

1
s()
2
v
yy

s()
s()
yv
y
= 0 , in R, (7.3)
v(1, ) = 0 , 0 T . (7.4)
More explicitly, (7.3) is solved a.e. in R, as v is locally a Sobolev function in R.
Classical results, see [11] Chapter IV, Section 10, imply that for any xed > 0,
v

, v
y
, v
yy
L
q
((, 1) (, T)) ,
for all q > 1. Then we use the embedding Lemma 3.3 of [11] Chapter II, to infer
that, for q > 3,
v
y
H
,

2
([, 1] [, T]) , (7.5)
where = 13/q (see also Remark 11.2 of [11], p. 218; the space H
,

2
is dened
in Appendix C). Since
u
x
(x, t) =
1
s(t)
v
y
_
x
s(t)
, t
_
,
the result follows.
Our next result implies that the free boundary in the Stefan problem (2.9)(2.14)
is of class C

(0, T).
Theorem 7.1. Assume u and s are as in Lemma 7.1, and moreover
u
x
(s(t), t) = c s(t) , 0 < t < T , (7.6)
where c ,= 0 is a given constant. Then s C

(0, T).
Proof. For v dened as in the proof of Lemma 7.1, we rewrite (7.6) as
s() =
1
cs()
v
y
(1, ) , 0 < < T . (7.7)
Choose (0, 1). Then (7.5) and (7.7) yield at once
s H
,

2
([, 1] [, T]) , for each xed > 0. (7.8)
Next we make use of the following classical result:
If the coecients in (7.3) (i.e., s), are of class
H
m+,
m+
2
([, 1] [, T]), then v is of class
H
2+m+,
2+m+
2
([2, 1] [2, T]).
(7.9)
Here m 0 is any integer. We prove by induction that for all m 0
s H
m+,
m+
2
([, 1] [, T]) , for each xed > 0. (7.10)
We already know this is the case when m = 0, from (7.8). Assume then (7.10) is
in force for a given m 0. Then, owing to (7.9),
v H
2+m+,
2+m+
2
([, 1] [, T]) , for each xed > 0. (7.11)
7. REGULARITY OF THE FREE BOUNDARY 17
Therefore, by the denition of the spaces H
,

2
(see Appendix C), we have that
v
y
H
1+m+,
1+m+
2
([, 1] [, T]) , for each xed > 0. (7.12)
Thus, taking (7.7) into account,
s H
1+m+,
1+m+
2
([, 1] [, T]) , for each xed > 0 (7.13)
(we have used also (1.1) of Appendix C). The induction step, and the proof, are
completed.
CHAPTER 2
Weak formulation of the Stefan problem
In this chapter we consider the weak formulation of the Stefan problem. As in
other PDE problems, the weak formulation actually takes the form of an integral
equality. It is to be noted that any explicit reference to the free boundary is
dropped from the weak formulation.
Well comment on the modeling dierences between the classical and the weak
formulations, and give the basic mathematical results for the latter.
We work in the multi-dimensional case of a spatial domain R
N
. We still
denote the temperature by u in this chapter.
1. From the energy balance to the weak formulation
The heat equation
u
t
= div(Du) +f , (1.1)
amounts to an energy balance equating the local change in time of energy (ex-
pressed by u
t
; various physical constants are normalized to 1 here), to the diver-
gence of the energy ux, plus the contribution of volumetric sources, represented
by f.
The weak formulation is based on the extension of this idea to the case where the
energy exhibits a jump at the critical temperature, due to the change of phase,
as shown in Figure 1. Then we write (formally)

t
v = div(Du) +f , (1.2)
where the energy (more exactly, the enthalpy) v jumps at the change of phase.
v
u
E(u)
u
v
(v)
Figure 1. Enthalpy v as a graph of temperature u, and viceversa.
Specically, solid at the critical temperature u = 0 corresponds to v = 0, while
liquid at temperature u = 0 corresponds to v = 1 (we assume the latent heat
19
20 DANIELE ANDREUCCI
is normalized to unity). Where 0 < v < 1, therefore, change of phase is taking
place, and the corresponding region is lled with a material whose state is neither
pure solid nor pure liquid. Such regions are usually called mushy regions.
The standard heat equation is assumed to hold in the pure phases (i.e., where
v > 1 or v < 0). This essentially amounts to
v =
_
u, u < 0 ,
u + 1 , u > 0 .
(1.3)
Where 0 < v < 1, u must equal the critical temperature u = 0. It is therefore
convenient to express the relation between v and u as follows
u(x, t) =
_

_
v(x, t) , v(x, t) 0 ,
0 , 0 < v(x, t) < 1 ,
v(x, t) 1 , v(x, t) 1 .
(1.4)
Note that v (not u) carries all the information on the state of the material. We
can rephrase (1.4) in the language of graphs:
v E(u) , (1.5)
where E is the graph dened by
E(s) =
_

_
s , s < 0 ,
[0, 1] , s = 0 ,
s + 1 , s > 0 .
(1.6)
When v, u satisfy (1.5), we say that v is an admissible enthalpy for u, or that u
is an admissible temperature for v.
Obviously, (1.2) can not be given a classical pointwise interpretation, since v
is general not continuous (see (1.5)). Following an usual procedure, we obtain the
weak formulation of (1.2) on multiplying both sides of it by a testing function
C

0
(Q
T
), and integrating (formally) by parts. In this way some of the
derivatives appearing in (1.2) are unloaded on the smooth testing function. We
obtain
__
Q
T
v
t
+Du D dxdt =
__
Q
T
fdxdt . (1.7)
Note that this formulation requires only we give a meaning to the rst spatial
derivatives of u (for example, u may be a Sobolev function). The complete
formulation of the Stefan problem will be given below (see Section 3).
The notion of weak solutions to the Stefan problem was introduced in [14], [9].
2. Comparing the weak and the classical formulations
2.1. The spatial normal. Let S be a smooth surface of R
N+1
, which we may
assume for our purposes to be locally represented in the form
(x, t) = 0 , (2.1)
with C
1
(R
N+1
), and D ,= 0 everywhere. Here we denote by D the
gradient of with respect to x, and by = (D,
t
) the complete gradient of
with respect to (x, t).
2. COMPARING THE WEAK AND THE CLASSICAL FORMULATIONS 21
We may think of S as of a moving surface in R
N
. More exactly, at each xed
instant t the surface takes the position
S(t) = x R
N
[ (x, t) = 0 .
A moving point x(t) belongs to S(t) for all t if and only if
(x(t), t) = 0 , for all t,
which is equivalent, up to the choice of suitable initial data, to
D(x(t), t) x(t) +
t
(x(t), t) = 0 , for all t.
Dene the spatial normal on S by
n =
D(x(t), t)
[D(x(t), t)[
. (2.2)
The spatial normal, of course, is dened up to a change in sign. We have for all
motions t x(t) as above
x(t) n =

t
(x(t), t)
[D(x(t), t)[
.
This shows that, at a given position on S, the component of the velocity x along
the spatial normal is independent of the motion x. This quantity is referred to
as the normal velocity V of S. Therefore, we have by denition
V (x, t) =

t
(x(t), t)
[D(x(t), t)[
.
Again, note that V is dened up to a change in sign.
The complete normal to S at (x, t) is clearly = (
x
,
t
), where

x
=
D(x, t)
[(x, t)[
= n
[D(x, t)[
[(x, t)[
,

t
=

t
(x, t)
[(x, t)[
.
2.2. Smooth weak solutions, with smooth interfaces, are classical so-
lutions. Let us assume that a function u satises
u > 0 , in A,
u < 0 , in B,
and that u = 0 on the common portion of the boundaries of the open sets A,
B R
N+1
. We assume this portion to be a smooth surface S, with complete
normal = (
x
,
t
) and spatial normal n, according to the notation above. Let
be the outer normal to B. Moreover assume
u, f C(A B) , u
|A
C
2,1
(A) , u
|B
C
2,1
(B) .
Finally, assume v (dened as in (1.3)) is a weak solution of (1.7), for any smooth
whose support is contained in the interior of A B S. It will be apparent
from our calculations that the denition of v at u = 0 is not relevant, in this case,
essentially because the N +1-dimensional measure of the free boundary u = 0 is
zero.
22 DANIELE ANDREUCCI
By direct calculation we have, owing to the regularity of u and of v,
__
A
v
t
dxdt =
_
S
E(0+)
t
d
__
A
v
t
dxdt ,
__
B
v
t
dxdt =
_
S
E(0)
t
d
__
B
v
t
dxdt .
On adding these two equalities we nd, recalling the denition of E,
__
AB
v
t
dxdt =
_
S

t
d
__
AB
u
t
dxdt , (2.3)
since v
t
= u
t
both in A and in B. The space part of the dierential operator in
(1.7) is treated similarly
__
A
Du Ddxdt =
_
S
Du
x
d
__
A
udxdt ,
__
B
Du Ddxdt =
_
S
Du
x
d
__
B
udxdt .
Again, on adding these two equalities we nd
__
AB
Du Ddxdt =
_
S
[Du
B
Du
A
]
x
d
__
AB
udxdt , (2.4)
where we denote by Du
A
[Du
B
] the trace on S of the spatial gradient of the
restriction of u to A [B]. Combining (2.3) with (2.4) we arrive at
__
AB
fdxdt =
__
AB
v
t
+Du D dxdt =
_
S
[
t
+Du
B

x
Du
A

x
] d
+
__
AB
u
t
u dxdt . (2.5)
Taking an arbitrary smooth supported in A, we immediately nd that in A
u
t
u = f . (2.6)
Of course the same PDE holds in B, by the same token.
Hence, we may drop the last integral in (2.5). Then take =

, where for all

> 0

|S
= C
1
0
(S) ; [

[ 1 ; [supp

[
N+1
0 , as 0.
Then, taking 0 in (2.5) we get
_
S
[
t
+Du
B

x
Du
A

x
] d = 0 .
As is reasonably arbitrary, it follows that on S
Du
B

x
Du
A

x
=
t
.
This condition can be rewritten as
V = Du
B
n Du
A
n, (2.7)
2. COMPARING THE WEAK AND THE CLASSICAL FORMULATIONS 23
and is the multi-dimensional equivalent of the Stefan condition (1.5) of Chapter 1.
In fact, it could be directly derived from an energetical balance argument, as we
did for (1.5) of Chapter 1. In this last approach, the weak formulation of the
Stefan problem follows from (2.7) and from the heat equation which we assume
to hold in A and in B separately: we only need go over our previous calculations
in reverse order.
The case where either v 1 in A or v 0 in B can be treated similarly.
2.3. Some smooth weak solutions are not classical solutions. Let us con-
sider the following problem
v
t
u
xx
= 1 , in Q
T
= (0, 1) (0, +), (2.8)
u(x, 0) = 1 , 0 < x < 1 , (2.9)
u
x
(0, t) = 0 , 0 < t , (2.10)
u
x
(1, t) = 0 , 0 < t (2.11)
(see [15]). We perform only a local analyis of the problem in the interior of the
domain Q
T
, giving for granted the solutions below actually take the boundary
data (in a suitable sense).
If, instead of (2.8), the standard heat equation
u
t
u
xx
= 1
was prescribed, clearly the solution to the initial value boundary problem would
be
u(x, t) = 1 +t , 0 < x < 1 , 0 < t .
Let us check that this function can not be a solution to the Stefan problem above.
Otherwise, we would have
v(x, t) = 1 +t , 0 < t < 1 ; v(x, t) = t , 1 < t .
Thus for every C

0
(Q
T
),
__
Q
T
v
t
+u
x

x
dxdt =
__
Q
T
v
t
dxdt =
__
Q
T
dxdt +
1
_
0
(x, 1) dx.
The last integral in this equality is evidently spurious, on comparison with the
weak formulation (1.7).
Let us instead check that a solution (actually the unique solution, see Section 4)
to (2.8)(2.11) is given by
v(x, t) = 1 +t , 0 < t ; u(x, t) =
_

_
1 +t , 0 < t < 1 ,
0 , 1 < t < 2 ,
2 +t , 2 < t .
It is immediately checked that (1.7) is fullled. We still have to check that (1.5)
holds, that is that v is an admissible enthalpy for u. Again, this follows immedi-
ately from the denitions.
The above can be interpreted as follows: the enthalpy v grows in time accordingly
to the prescribed volumetric source; the change of phase takes place over the time
interval 1 < t < 2, because this is the time interval where v (0, 1); over this
time interval, therefore, the temperature equals the critical temperature u = 0;
24 DANIELE ANDREUCCI
for all other times, (2.8) coincides with the standard heat equation, and thus u
is simply the solution to a suitable problem for the heat equation.
Note that the set u = 0 has in this example positive measure, in contrast with the
classical formulation of Chapter 1. See also [8], [1] for a discussion of existence
and non existence of mushy regions in weak solutions to change of phase problems.
2.4. Without sign restrictions, classical solutions may not be weak
solutions. Let us go back to problem (2.9)(2.14) of Chapter 1, where we now
assume u
0
C
1
([0, b]), u
0
(b) = 0, u
0
(x) < 0 for 0 x < b, and, e.g., h 0. It
can be shown (see [6]) that this problem has a classical solution, in the sense of
Denition 2.1 of Chapter 1, at least for a small enough T > 0. Note that the
Stefan condition (2.12) of Chapter 1 has now the wrong sign (cf. Exercise 1.2
of Chapter 1). Therefore the classical solution at hand is not a solution of the
weak formulation. Indeed, otherwise it would be a smooth weak solution with
a smooth free boundary, and we would be able to infer the Stefan condition as
above. However, as shown above, this condition would be the one correctly
corresponding to the actual sign of the solution, and therefore would be dierent
from the one we prescribed.
More generally, no undercooling is possible in the weak formulation introduced
here. In fact, the liquid and the solid phases are identied solely by the value of
v. Thus, whenever u changes its sign a change of phase must take place. This
is not the case in the classical formulation, where the liquid and solid phases are
essentially identied by a topological argument, as the two connected components
of the domain, separated by the special level surface which is dened as the free
boundary. Other level surfaces corresponding to the value u = 0 may exist inside
both phases.
2.5. Exercises.
2.1. Assume that the interior M of the region u = 0 is non empty, where u is
given by (1.4), and v satises (1.7). Show that, in a suitable weak sense, v
t
= f
in M.
3. Denition of weak solution
Let us dene the inverse of the graph E in (1.6). This is the function given
by
(r) =
_

_
r , r 0 ,
0 , 0 < r < 1 ,
r 1 , 1 r .
(3.1)
Let Q
T
= (0, T), where R
N
is a bounded open set with boundary of
class C

.
Let us consider the Stefan problem
v
t
(v) = f(v) , in Q
T
, (3.2)
v(x, 0) = v
0
(x) , x , (3.3)
(v)
n
= 0 , on (0, T), (3.4)
4. UNIQUENESS OF THE WEAK SOLUTION 25
where v
0
L

() and f L

(R) C

(R) are given functions. We assume

that for a xed > 0
[f(v
1
) f(v
2
)[ [v
1
v
2
[ , for all v
1
, v
2
R. (3.5)
Definition 3.1. A function v L

(Q
T
) is a weak solution to (3.2)(3.4) if
u := (v) L
2
(0, T; W
2
1
()) , (3.6)
and for all W
2
1
(Q
T
) such that (x, T) = 0 we have
__
Q
T
v
t
+Du D dxdt =
_

v
0
(x)(x, 0) dx +
__
Q
T
f(v)dxdt . (3.7)
Note that equation (3.7) is obtained integrating formally by parts equation (3.2),
after multiplying it by a as above.
Of course the structure of the problem could be generalized; for example the
smoothness required of can be reduced by approximating with more regular
domains (but (3.5) is going to play an essential role). However we aim here at
showing some basic techniques in the simple setting above, which is suitable for
our purposes.
Remark 3.1. Actually, if v is a weak solution to the Stefan problem (in a sense
similar to ours), the corresponding temperature u is continuous in Q
T
(see [4]),
a fact which however we wont use here.
3.1. Exercises.
3.1. Assume u C(Q
T
) (see also Remark 3.1). Show that u is of class C

where
it is not zero.
4. Uniqueness of the weak solution
4.1. A dierent notion of weak solution. If we assume in (3.7) that
i) W
2
2,1
(Q
T
) ; ii) (x, T) = 0 ; iii)

n
= 0 , on (0, T);
(4.1)
we immediately obtain, on integrating once by parts,

__
Q
T
v
t
+u dxdt =
_

v
0
(x)(x, 0) dx +
__
Q
T
f(v)dxdt . (4.2)
We need drop the requirement (x, T) = 0, for technical reasons. This can be
done as follows. Choose a satisfying i), iii) of (4.1), but not necessarily ii). Fix
t (0, T), and dene for 0 < < t (see Figure 2)

() = min
_
1,
1

(t )
+
_
.
The function

satises requirement (4.1) in full, so that it can be taken as a

testing function in (4.2). Let us rewrite it as

__
Q
T
v[

dxd =
_

v
0
(x)(x, 0) dx +
__
Q
T
u +f(v)

dxd .
26 DANIELE ANDREUCCI

()

1
t t
Figure 2. The auxiliary function

.
The behaviour of the right hand side as 0 is obvious. The left hand side
equals

__
Q
T
v

dxd +
1

t
_
t
_

v(x, )(x, ) dx d .
On letting 0 in this quantity we get, for almost all t (0, T)

__
Q
t
v

dxd +
_

v(x, t)(x, t) dx.

Thus the denition of weak solution given above actually implies the new (and
weaker) one
Definition 4.1. A function v L

(Q
T
) is a weak solution of class L

to
(3.2)(3.4) if for all W
2
2,1
(Q
T
) such that

n
= 0 on (0, T) we have
_

v(x, t)(x, t) dx
__
Q
t
v

+u dxd
=
_

v
0
(x)(x, 0) dx +
__
Q
t
f(v)dxd , (4.3)
for almost all t (0, T). Here u = (v).
Note that we dropped in Denition 4.1 any regularity requirement for v (excepting
boundedness).
Remark 4.1. It follows from (4.3) that the function
t
_

v(x, t)(x, t) dx
is actually continuous over [0, T], up to modication of v over sets of zero measure.
4. UNIQUENESS OF THE WEAK SOLUTION 27
4.2. Continuous dependence on the initial data. We are now in a position
to prove
Theorem 4.1. (Continuous dependence on the data) Let v
1
, v
2
be two
weak solutions of class L

to (3.2)(3.4), in the sense of Denition 4.1 (or two

weak solutions in the sense of Denition 3.1), corresponding to bounded initial
data v
01
, v
02
respectively. Then, for almost all 0 < t < T
_

[v
1
(x, t) v
2
(x, t)[ dx e
t
_

[v
01
(x) v
02
(x)[ dx. (4.4)
Corollary 4.1. (Uniqueness) Let v
1
, v
2
be two weak solutions of class L

to
(3.2)(3.4), in the sense of Denition 4.1 (or two weak solutions in the sense of
Denition 3.1), corresponding to the same bounded initial data. Then v
1
v
2
in
Q
T
.
Remark 4.2. (Solutions of class L
1
) Both Theorem 4.1 and its immediate
Corollary 4.1 actually hold for a more general class of weak solutions, obtained
replacing the requirement v L

(Q
T
) in Denition 4.1 with v L
1
(Q
T
). Also
the initial data may be selected out of L
1
(). In this connection, in order to
keep the integrals in (4.3) meaningful, we have to assume that is a Lipschitz
continuous function in Q
T
, with
x
i
x
j
L

(Q
T
), i, j = 1 , . . . , N, and

n
= 0
on (0, T).
Once existence of solutions in the sense of Denition 4.1 has been obtained,
existence of solutions of class L
1
can be proven as follows. Assume v
0
L
1
(),
and v
i
0
v
0
in L
1
(), v
i
0
L

(). Note that the solutions v

i
of class L

corresponding to the approximating initial data v

i
0
satisfy (4.4). Therefore v
i

is a Cauchy sequence in L
1
(Q
T
), and we may assume it converges to a v L
1
(Q
T
)
both in the sense of L
1
(Q
T
), and a.e. in Q
T
. It is now a trivial task to take the
limit in the weak formulation (4.3) satised by v
i
and obtain the corresponding
formulation for v.
Standard references for the material in this Section are [11], Chapter V, Section 9,
and [12], whose approach we follow, with some modications.
4.3. Proof of Theorem 4.1. Fix t (0, T). Subtract from each other the two
equations (4.3) written for the two solutions, and obtain
_

[v
1
(x, t) v
2
(x, t)](x, t) dx
__
Q
t
(v
1
v
2
)[

+a(x, t) ] dxd
=
_

[v
01
(x) v
02
(x)](x, 0) dx +
__
Q
t
[f(v
1
) f(v
2
)]dxd , (4.5)
where we set
a(x, t) =
(v
1
(x, t)) (v
2
(x, t))
v
1
(x, t) v
2
(x, t)
, v
1
(x, t) ,= v
2
(x, t) ,
a(x, t) = 0 , v
1
(x, t) = v
2
(x, t) .
Due to the denition of we have
0 a(x, t) 1 , in Q
T
. (4.6)
28 DANIELE ANDREUCCI
Next choose =

, where for each > 0,

is the solution of

+ (a

(x, t) +)

= 0 , in Q
t
, (4.7)
(x, t) = (x) , x , (4.8)

n
= 0 , on (0, t). (4.9)
Here C

0
(), [(x)[ 1, and a

(Q
T
) satises
0 a

1 , a.e. in Q
T
; |a

a|
2
. (4.10)
Some relevant properties of

are collected in Lemma 4.1 below. Note that (4.8)

is the initial value for the reverse parabolic problem solved by

. By virtue of
(4.15) we have as 0
__
Q
t
[

[ dxd
___
Q
t
dxd
_
1/2
___
Q
t
(

)
2
dxd
_
1/2

2t|D|
2,
[[
1/2
0 ,
as well as (using (4.10))
__
Q
t
[a

a[[

[ dxd
___
Q
t
[a

a[
2

dxd
_
1/2
___
Q
t
(

)
2
dxd
_
1/2

2|D|
2,
0 .
Moreover, using (4.7) in (4.5), and (4.16), we get
_
(t)
[v
1
v
2
]dx =
__
Q
t
[a a

[v
1
v
2
] dxd +
_

[v
01
v
02
]

(x, 0) dx
+
__
Q
t
[f(v
1
) f(v
2
)]

dxd (|v
1
|

+|v
2
|

)
__
Q
t
[[a a

[ +][

[ dxd
+
_

[v
01
v
02
[ dx +
__
Q
t
[v
1
v
2
[ dxd .
As 0 this yields
_
(t)
[v
1
v
2
]dx
_

[v
01
v
02
[ dx +
t
_
0
_
()
[v
1
v
2
[ dxd . (4.11)
Choose now =
n
, where for n

n
(x) sign
_
v
1
(x, t) v
2
(x, t)
_
, a.e. x .
On letting n in (4.11) we obtain
_
(t)
[v
1
v
2
[ dx
_

[v
01
v
02
[ dx +
t
_
0
_
()
[v
1
v
2
[ dxd .
The statement now follows simply invoking Gronwalls lemma.
5. EXISTENCE OF WEAK SOLUTIONS 29
Lemma 4.1. Let C

(Q
T
), 0 <
0
, where and
0
are given
constants. Let C

0
(). Then there exists a unique solution C

(Q
T
) of

t
= 0 , in Q
T
, (4.12)
(x, 0) = (x) , x , (4.13)

n
= 0 , on (0, T), (4.14)
such that for all 0 < t < T
__
Q
t
(
2

+()
2
) dxd +
_
(t)
[D[
2
dx (
0
+ 1)
_

[D[
2
dx. (4.15)
Moreover
||

||

. (4.16)
Proof. The existence of a unique solution C

(Q
T
) to (4.12)(4.14) is a
classical result. Let us multiply (4.12) by , and integrate by parts over Q
t
,
for an arbitrarily xed t (0, T). We nd
__
Q
t
[[
2
dxd =
__
Q
t

dxd =
__
Q
t
D

Ddxd
=
1
2
_

[D(x, 0)[
2
dx
1
2
_

[D(x, t)[
2
dx.
Using again (4.12) we obtain
__
Q
t

dxd =
__
Q
t

2
[[
2
dxd

0
2
_

[D(x, 0)[
2
dx.
Thus, satises the integral estimate (4.15). The bound in (4.16) is an obvious
consequence of the maximum and boundary point principles of Appendix A.
4.4. Exercises.
4.1. Prove that
0 [ 0] =

0 [ 0]
where

is the solution to (4.7)(4.9). Use this fact to prove the comparison

result
v
01
v
02
in = v
1
v
2
in Q
T
,
where v
1
and v
2
are as in Theorem 4.1, provided f

0.
4.2. If C

(), but , C

0
(), as in Lemma 4.1 need not be even C
1
(Q
T
).
Why?
5. Existence of weak solutions
We apply here the ideas of [8], though we approximate the Stefan problem with
smooth parabolic problems, rather than discretizing it in time.
Theorem 5.1. There exists a weak solution v to (3.2)(3.4), in the sense of
Denition 3.1, satisfying
|v|

|v
0
|

. (5.1)
30 DANIELE ANDREUCCI
The proof of this existence result relies on an approximation procedure. Namely,
we approximate (3.2)(3.4) with a sequence of smoothed problems; the solutions
to these problems in turn approach a solution to the original Stefan problem.
We need a sequence of smooth constitutive functions
n
C

(R
N
) approximat-
ing , such that
1
n

n
(s) 1 , s R;
n
, uniformly in R. (5.2)
Clearly we may assume
u
v

n
(v)
1
v
u
E
n
(u)
1
Figure 3. The approximating functions
n
and E
n
.
(s)
n
(s) (s) +
1
n
, s R; (s) =
n
(s) = s , s < 0 . (5.3)
Dene E
n
as the inverse function of
n
. Then
E
n
(s) E(s) , s R; E(s) = E
n
(s) = s , s < 0 . (5.4)
Let us also introduce a sequence v
0n
C

0
() approximating the initial data as
in
v
0n
v
0
, a.e. in ; |v
0n
|

|v
0
|

. (5.5)
For each n there exists a unique solution v
n
C

(Q
T
) to
v
nt

n
(v
n
) = f(v
n
)
n
, in Q
T
, (5.6)

n
(v
n
)
n
= 0 , on (0, T), (5.7)
v
n
(x, 0) = v
0n
(x) , in , (5.8)
where
n
C

0
() is such that

n
(x) = 1 , dist(x, ) >
1
n
; 0
n
(x) 1 , x .
Let us denote u
n
=
n
(v
n
), and rewrite (5.6) as
E
n
(u
n
)
t
u
n
= f(v
n
)
n
, in Q
T
. (5.9)
Owing to the maximum principle (see Theorem 1.2 of Appendix A), we have
|u
n
|

|v
n
|

|v
0
|

+T|f|

=: M . (5.10)
5. EXISTENCE OF WEAK SOLUTIONS 31
5.1. The energy inequality. Multiply (5.9) by u
n
and integrate by parts.
Note that
E
n
(u
n
)
t
u
n
=

t
u
n
_
0
E

n
(s)s ds ,
and that, if k > 0, taking into account (5.2),
k
2
2

k
_
0
E

n
(s)s ds E
n
(k)k (k + 1)k .
If k < 0 we simply have
k
_
0
E

n
(s)s ds =
k
2
2
.
Therefore we obtain after standard calculations, for each t (0, T)
1
2
_

u
n
(x, t)
2
dx +
__
Q
t
[Du
n
[
2
dxd

1
2
_

[
n
(v
0n
)[([
n
(v
0n
)[ + 1) dx +
__
Q
t
f(v
n
)
n
u
n
dxd

[[
2
(|v
0
|

+ 1)
2
+|f|

M[[T .
Note that both terms on the leftmost side of this estimate are positive. Drop-
ping either one, taking the supremum in time, and collecting the two bounds so
obtained, we get
sup
0<t<T
_

u
n
(x, t)
2
dx +
__
Q
T
[Du
n
[
2
dxd C , (5.11)
where C > 0 is a constant depending on the data of the problem, but not on n.
It follows that we may extract a subsequence, still labelled by n, such that
u
n
u, Du
n
Du, weakly in L
2
(Q
T
). (5.12)
Moreover we have
|u|

M , (5.13)
and
sup
0<t<T
_

u(x, t)
2
dx +
__
Q
T
[Du[
2
dxd C . (5.14)
We may as well assume
v
n
v , weakly in L
2
(Q
T
), (5.15)
but note that, due to the nonlinear nature of our problem (i.e., the fact that E
and f are not linear functions), weak convergence is not enough to pass to the
limit in the weak formulation of the approximating problem, i.e., in
__
Q
T
v
n

t
+Du
n
D dxdt =
_

v
0n
(x)(x, 0) dx+
__
Q
T
f(v
n
)
n
dxdt , (5.16)
32 DANIELE ANDREUCCI
where is any function out of W
2
1
(Q
T
) with (x, T) = 0. For example we do
not know that v is an admissible enthalpy for u. We must therefore obtain some
stronger kind of convergence for the sequence v
n
, so that, e.g., f(v
n
) converges
to f(v).
However we do have some compactness in suitable integral norms for the sequence
u
n
, due to (5.11). More specically, let h R
N
be any given vector with length
0 < [h[ < , and let k = h/[h[. Setting

= x [ dist(x, ) > ,
we calculate, by a standard argument,
T
_
0
_

[u
n
(x +h, t) u
n
(x, t)[ dxdt =
T
_
0
_

|h|
_
0
Du
n
(x +sk) k ds

dxdt

T
_
0
_

|h|
_
0
[Du
n
(x +sk)[ ds dxdt =
T
_
0
|h|
_
0
_

[Du
n
(x +sk)[ dxds dt

T
_
0
|h|
_
0
_

[Du
n
(x)[ dxds dt = [h[
T
_
0
_

[Du
n
(x)[ dxdt C[h[ . (5.17)
We have used the fact that x + sk for all x

and all 0 < s < [h[, and

(5.11).
5.2. The BV estimate. Introduce a cut o function C

() such that
(x) 1 , dist(x, ) > 4 ; (x) 0 , dist(x, ) < 2 ;
[D[

; [[

2
;
where does not depend on . We may also assume that n > 1/, so that
n

in .
In this Subsection we drop the index n, for ease of notation. Therefore we write
for
n
, v for v
n
, and so on. For a given h R
N
dene the testing function
(x, t) = sign

_
(v(x +h, t)) (v(x, t))
_
(x) ,
where sign

(R) is a smooth approximation of sign, such that

sign

(s) sign(s) , s R, sign

0
(we set sign(0) = sign

(0) = 0). Let us denote

h
(x, t) = (x h, t) .
If [h[ < , which we assume from now on, both and
h
vanish in a neighbour-
hood of (0, T).
Multiply (5.6) by and integrate by parts, obtaining
__
Q
t
v

dxd +
__
Q
t
D(v) Ddxd =
__
Q
t
f(v)dxd . (5.18)
5. EXISTENCE OF WEAK SOLUTIONS 33
On performing the same operation with
h
, we obtain
__
Q
t
v

(x, )(x h, ) dxd +

__
Q
t
D(v(x, )) D(x h, ) dxd
=
__
Q
t
f(v(x, ))(x h, ) dxd . (5.19)
Let us change the integration variable in (5.19),
x h y .
The domain of integration stays the same (i.e., Q
t
) because
supp
h
Q
T
, supp Q
T
.
Still denoting the new variable by x we obtain
__
Q
t
v

(x +h, )(x, ) dxd +

__
Q
t
D(v(x +h, )) D(x, ) dxd
=
__
Q
t
f(v(x +h, ))(x, ) dxd . (5.20)
On subtracting (5.18) from (5.20), and explicitly calculating D, we arrive at
__
Q
t
[v

(x +h, ) v

(x, )]

(x, ) dxd
+
__
Q
t
[D(v(x +h, )) D(v(x, ))[
2
sign

(. . . ) dxd
+
__
Q
t
[D(v(x +h, )) D(v(x, ))] D sign

(. . . ) dxd
=
__
Q
t
[f(v(x +h, )) f(v(x, ))](x, ) dxd . (5.21)
The second integral in (5.21) may be dropped, since it is non negative. Next note
v(x +h, )
>
=
<
v(x, ) (v(x +h, ))
>
=
<
(v(x, )) ,
as =
n
is strictly increasing. Thus, as 0
sign

_
(v(x +h, )) (v(x, ))
_
sign
_
(v(x +h, )) (v(x, ))
_
= sign
_
v(x +h, ) v(x, )
_
=: (x, ) .
34 DANIELE ANDREUCCI
Finally, we may take the limit 0 in (5.21), after dropping the second integral
as we said, to nd
__
Q
t
[v

(x +h, ) v

(x, )](x, )(x) dxd

+
__
Q
t
[D(v(x +h, )) D(v(x, ))] D(x, ) dxd

__
Q
t
[f(v(x +h, )) f(v(x, ))](x, ) dxd . (5.22)
The rst integral in (5.22) equals
__
Q
t

[v

(x +h, ) v

(x, )[(x) dxd

=
_

[v(x +h, t) v(x, t)[(x) dx

_

[v
0
(x +h) v
0
(x)[(x) dx.
The second integral in (5.22) equals
__
Q
t
D[(v(x +h, )) (v(x, ))[ D dxd
=
__
Q
t
[(v(x +h, )) (v(x, ))[ dxd ,
so that, according to (5.17), its absolute value is majorised by

2
__

(0,t)
[(v(x +h, )) (v(x, ))[ dxd
C

2
[h[ .
The third and last integral in (5.22) is bounded simply by taking into account
the Lipschitz continuity of f. Collecting these estimates we nd
_

[v(x +h, t) v(x, t)[(x) dx

_

[v
0
(x +h) v
0
(x)[(x) dx
+
C

2
[h[ +
t
_
0
_

[v(x +h, ) v(x, )[(x) dxd . (5.23)

By Gronwalls lemma we conclude that
_

[v(x +h, t) v(x, t)[(x) dx

e
t
__

[v
0
(x +h) v
0
(x)[(x) dx +
C

2
[h[
_
(5.24)
5. EXISTENCE OF WEAK SOLUTIONS 35
Assume now that the original initial data v
0
satises for all h R
N
, [h[ < ,
> 0,
_

[v
0
(x +h) v
0
(x)[ dx C

[h[ . (5.25)
We may therefore assume the approximating initial data v
0n
satisfy a similar
inequality (see Lemma 1.1 of Appendix B). This and (5.24) allow us to conclude
that for all 0 < t < T, h R
N
, n 1,
_

[v
n
(x +h, t) v
n
(x, t)[ dx C

[h[ . (5.26)
Standard (and trivial) results in functional analysis imply that (5.26) is equiva-
lent, since v
n
C

(Q
T
), to
_

[Dv
n
(x, t)[ dx C

. (5.27)
Here and above, C

denotes a constant depending on , but not on n.

Remark 5.1. In general, the L
1
estimate (5.26) does not imply for an integrable
function v the existence of the gradient Dv in the sense of Sobolev. This is a
marked dierence with the case of similar L
p
estimates, with p > 1. Instead,
estimates like (5.26) imply that v is a function of bounded variation, or BV
function, whence the title of this subsection.
5.3. Compactness of v
n
in L
1
. In order to obtain the desired compactness
of the sequence v
n
, in L
1
loc
(Q
T
), we still have to complement (5.26) with a
similar estimate involving translations in time rather than in space. This bound
will be achieved as a consequence of a theorem by Kruzhkov, which we state and
prove in Appendix B. Essentially, the result states that if we already know some
regularity of the solution to a parabolic equation at each time level, we may infer
some (lesser) regularity in the time variable. The regularity in space is in our
case guaranteed by (5.26). The remarkable input of the theorem is that strong
continuity in an integral norm is a consequence of a notion of weak continuity.
Let g = g(x), g C
1
0
(

). On multiplying (5.6) by g and integrating over

between t and t +s, 0 < t < t +s < T, standard calculations give
_

g(x)[v
n
(x, t +s) v
n
(x, t)] dx =
t+s
_
t
_

Dg D
n
(v
n
) dxd
+
t+s
_
t
_

f(v
n
)
n
g dxd . (5.28)
36 DANIELE ANDREUCCI
Note that

t+s
_
t
_

Dg D
n
(v
n
) dxd

t+s
_
t
_

n
(v
n
)[Dg[[Dv
n
[ dxd
|

n
|

|Dg|

s sup
t<<t+s
_

[Dv
n
(x, )[ dx C

|Dg|

s ,
owing to estimate (5.27). Clearly C

does not depend on n. Therefore

g(x)[v
n
(x, t +s) v
n
(x, t)] dx

_
|g|

+|Dg|

_
s , (5.29)
for all g C
1
0
(

). Then, Kruzhkovs Theorem 2.1 of Appendix B (see also

Remark 2.1 there) implies that
_

[v
n
(x, t +s) v
n
(x, t)[ dx C

s , 0 < s <
2
, (5.30)
for all 0 < t < t + s < T, and for all > 0. Combining (5.30) with (5.26) we
obtain that the sequence v
n
is pre-compact in L
1
(

(0, T)). By means of

usual diagonal procedures, we may extract a subsequence (still labelled by n)
such that
v
n
v , a.e. in Q
T
. (5.31)
It follows
u
n
=
n
(v
n
) (v) = u, a.e. in Q
T
, (5.32)
where clearly u must equal the weak limit in (5.12).
For any testing function as in Denition 3.1, we may therefore take the limit
n in (5.16), and obtain (3.7). By construction, v and u satisfy the relevant
regularity requirements. By the same token, (v) = u, i.e., v E(u).
5.4. Removing the extra assumption on v
0
. We have so far proven exis-
tence of solutions under the extra regularity assumption (5.25), which is certainly
satised, for example, if the initial data are in C
1
(). To extend the proof to the
case when v
0
is merely assumed to be a bounded function (so that (5.25) does not
necessarily hold), consider rst a sequence of smoothed initial data v
j
0
C
1
(),
such that
v
j
0
v
0
, in L
1
(), |v
j
0
|

|v
0
|

. (5.33)
Consider the sequence v
j
of weak solutions, corresponding to these initial data,
to the Stefan problem. These are therefore solutions in the sense of Denition 3.1.
Note that, perhaps extracting a subsequence we may still assume that v
j
, u
j
=
(v
j
), Du
j
converge weakly, because estimates (5.10), (5.11) are in force for them,
uniformly on j.
Due to Theorem 4.1, and to (5.33), the sequence v
j
actually is a Cauchy sequence
in L
1
(Q
T
). Again, perhaps extracting a subsequence we may assume that
v
j
v , a.e. in Q
T
. (5.34)
6. A COMPARISON RESULT 37
Hence, we may take the limit in the integral equation satised by v
j
, i.e.,
__
Q
T
v
j

t
+Du
j
D dxdt =
_

v
j
0
(x)(x, 0) dx +
__
Q
T
f(v
j
)dxdt ,
nally proving that v solves the original Stefan problem.
5.5. Exercises.
5.1. Prove in detail that (5.14) follows from (5.11).
5.2. A much simpler proof of existence of weak solutions is available when f does
not depend on v (but, e.g., on u). In this case, indeed it is enough to prove strong
convergence for the sequence u
n
. To obtain the needed compactness estimate,
which actually amounts to a bound of |u
nt
|
2
uniform in n, multiply (5.9) by u
nt
,
and use the properties of E
n
. Go over the details of this approach; e.g., prove
that the weak limit v of v
n
is an admissible enthalpy for the strong limit u of u
n
.
5.3. Assume that (5.25) is replaced by
_

[v
0
(x +h) v
0
(x)[ dx C

([h[) ,
for a non decreasing non negative continuous function : R R, (0) = 0,
such that (s) s, 0 < s < 1. Prove that the solution v satises for [h[ < < 1,
_

[v(x +h, t) v(x, t)[ dx C

([h[) , a.e. t (0, T).

6. A comparison result
Theorem 6.1. Let v
1
0
, v
2
0
L

(), and f
1
, f
2
C

(R) satisfy (3.5). Denote

by v
1
, v
2
the corresponding weak solutions to the Stefan problem. If v
1
0
v
2
0
in
, and f
1
f
2
in R, then v
1
v
2
in Q
T
.
Proof. As we already know a result of uniqueness of the solution, we may prove
the statement by approximation. By the same token, we may assume the initial
data are smooth, e.g., in C
1
(). Let
v
i
n

n=1
, i = 1 , 2 ,
be the two sequences constructed in Section 5 as solutions to the problems (5.6)
(5.8), with v
0n
[f] replaced with v
i
0n
[f
i
] respectively.
Subtract from each other the PDE solved by v
1
n
and v
2
n
, and multiply the resulting
equation by

= sign
+

(
n
(v
2
n
)
n
(v
1
n
)) ,
where sign
+

is a smooth approximation of the function sign

+
(s) =
(0,+)
(s),
with sign
+

0. Integrating by parts over Q

t
we obtain
__
Q
t
_
(v
2
n
v
1
n
)

+ sign
+

(u
2
n
u
1
n
)[D(u
2
n
u
1
n
)[
2

dxd
=
__
Q
t
_
f
2
(v
2
n
) f
1
(v
1
n
)

dxd .
38 DANIELE ANDREUCCI
On dropping the quadratic term above, and letting 0 we get (recall that
sign
+
(u
2
n
u
1
n
) = sign
+
(v
2
n
v
1
n
), see the discussion in Subsection 5.2)
_

(v
2
n
v
1
n
)
+
(x, t) dx
__
Q
t
_
f
2
(v
2
n
) f
1
(v
1
n
)

n
sign
+
(v
2
n
v
1
n
) dxd .
We have performed an integration in time, and used v
1
0n
v
2
0n
. But, since
f
1
f
2
,
[f
2
(v
2
n
) f
1
(v
1
n
)] sign
+
(v
2
n
v
1
n
) = [f
2
(v
2
n
) f
2
(v
1
n
)] sign
+
(v
2
n
v
1
n
)
+ [f
2
(v
1
n
) f
1
(v
1
n
)] sign
+
(v
2
n
v
1
n
) (v
2
n
v
1
n
)
+
.
Therefore _

(v
2
n
v
1
n
)
+
(x, t) dx
__
Q
t
(v
2
n
v
1
n
)
+
dxd .
Finally, an application of Gronwalls lemma concludes the proof.
APPENDIX A
Maximum principles for parabolic equations
1. The weak maximum principle
Let Q
T
be a bounded open set of R
N+1
, contained in R
N
(0, T), where T > 0.
Definition 1.1. We denote by Q

T
the parabolic interior of Q
T
, that is the set
of all points ( x,

t) with the property

> 0 : B

( x,

t) t <

t Q
T
.
Here B

( x,

t) denotes the (N+1)-dimensional ball with radius and center ( x,

t).
Dene also the parabolic boundary
p
Q
T
of Q
T
, as

p
Q
T
= Q
T
Q

T
.
The set
p
Q
T
is the parabolic analogue of the boundary of Q
T
, i.e., roughly
speaking, the region where initial and boundary data should be prescribed for
parabolic problems set in Q
T
(see Figure 1).
For example, if Q
T
= (0, L) (0, T) then Q

T
= (0, L) (0, T].
Obviously we have Q
T
Q

T
Q
T
. In general
p
Q
T
is not a closed set.
t
x
A B
C D
E
Q
T
Figure 1. The dashed lines and the point E belong to the para-
bolic interior, but the points A, B, C, D, as well as the solid lines,
belong to the parabolic boundary.
In the following we denote
Lu = u
t
a
ij
(x, t)u
x
i
x
j
+b
i
(x, t)u
x
i
+c(x, t)u,
L
0
u = u
t
a
ij
(x, t)u
x
i
x
j
+b
i
(x, t)u
x
i
.
39
40 DANIELE ANDREUCCI
Throughout this Appendix we employ the summation convention, and assume
that
u C
2,1
(Q

T
) C(Q
T
) , a
ij
, b
i
, c C(Q

T
) ;
a
ij
(x, t)
i

j
[[
2
, for all R
N
, (x, t) Q

T
.
Here > 0 is a given constant. We also assume
N

i , j=1
|a
ij
|

=: A < ,
N

i=1
|b
i
|

=: B < , |c|

=: C < .
Lemma 1.1. Assume that L
0
u( x,

t) < 0, where ( x,

t) Q

T
. Then u can not
attain a local maximum at ( x,

t).
Proof. Recalling the denition of Q

T
, we have, reasoning by contradiction,
L
0
u( x,

t) = u
t
( x,

t) a
ij
( x,

t)u
x
i
x
j
( x,

t) +b
i
( x,

t)u
x
i
( x,

t)
= u
t
( x,

t) a
ij
( x,

t)u
x
i
x
j
( x,

t) .
Note that, if ( x,

t) is a point of local maximum, then u

t
( x,

t) 0. By the same
token, a
ij
u
x
i
x
j
0 at ( x,

t), as we show below. This leads of course to an

inconsistency, as L
0
u( x,

t) < 0 by assumption.
To prove the assertion a
ij
u
x
i
x
j
0 at ( x,

t), we change spatial coordinates den-

ing y = x+(x x), and v(y, t) = u(x(y), t), where = (
ij
) is an NN matrix
such that (a
ij
( x,

t))
t
coincides with the diagonal matrix diag (
1
, . . . ,
N
) (we
may assume without loss of generality that (a
ij
) is symmetric). Then
a
ij
u
x
i
x
j
= a
ij
v
y
h
y
k

hi

kj
=
h
v
y
h
y
h
, at ( x,

t).
Note that v(,

t) attains a local maximum at y = x, so that v

y
h
y
h
0 for all h.
We also take into account that
h
> 0 for all h, since (a
ij
) is positive denite.
The result immediately follows.
Theorem 1.1. (Weak Maximum Principle) Let L
0
u 0 in Q

T
. Then
max
Q
T
u = sup

p
Q
T
u. (1.1)
Proof. Let us dene
v = (u M)e
t
, M = sup

p
Q
T
u, > 0 .
If u M is positive at some (x, t) Q

T
, then v attains a positive maximum
somewhere in Q

T
, say at ( x,

t). Indeed v 0 on
p
Q
T
. We calculate
v
t
= u
t
e
t
v , v
x
i
= u
x
i
e
t
, v
x
i
x
j
= u
x
i
x
j
e
t
.
Therefore
L
0
v( x,

t) = e

t
L
0
u( x,

t) v( x,

t) v( x,

t) < 0 .
Upon recalling Lemma 1.1, this inconsistency concludes the proof.
2. THE STRONG MAXIMUM PRINCIPLE 41
1.1. More general operators. Let us consider here more general operators of
the form L.
Theorem 1.2. Let Lu f in Q

T
, f C(Q
T
). Dene
t
_
0
|f
+
(, )|

d =: H(t) , 0 < t < T . (1.2)

Then
max
Q
T
u
+
e
C

T
_
sup

p
Q
T
u
+
+H(T)
_
, (1.3)
where C

= |c

|
,Q
T
.
Proof. Dene for a constant > 0 to be chosen
w(x, t) = e
t
u(x, t) mH(t) ,
where
m := sup

p
Q
T
u
+
.
By denition of m, and since > 0, we have w 0 on
p
Q
T
. Moreover
L
0
w = e
t
L
0
u e
t
u |f
+
(, t)|

ce
t
u +e
t
f
+
(x, t) e
t
u |f
+
(, t)|

[c +]e
t
u.
Therefore we have L
0
w < 0 where w > 0 (and hence u > 0 too), provided we
select = C

+ for any arbitrarily xed > 0. Thus, if w attains a positive

maximum in Q

T
, we arrive at an inconsistency with Lemma 1.1. We conclude
that w 0 in Q
T
, and we recover (1.3) on letting 0.
Note that according to our theorem above, if u 0 on
p
Q
T
, and f 0 in Q
T
,
then u 0 in Q
T
, regardless of the sign of c.
Remark 1.1. All the results of this section still hold if the parabolicity constant
is equal to 0. This is not the case for the results in next two sections. See also
Section 4.
2. The strong maximum principle
Let ( x,

t) Q

T
. Dene the set o( x,

t) as the set of all (x, t) Q

T
with the
property
(x, t) can be connected to ( x,

t) by a polygonal contained in Q
T
,
along which t is increasing, when going from (x, t) to ( x,

t).
A polygonal is a connected curve made of a nite number of straight line seg-
ments. Essentially, the strong maximum principle asserts that if L
0
u 0 in Q
T
,
and u attains its maximum at a point ( x,

t) of Q

T
, then u is constant in o( x,

t).
Our rst result is a weaker version of this principle.
The proof we present here was taken from [2]
1
. The strong maximum principle
for parabolic equations was rst proven in [13].
1
The author of [2] quotes a course of D. Aronson (Minneapolis) as source of the proof.
42 DANIELE ANDREUCCI
t
x
( x,

t)
S( x,

t)
Q
T
Figure 2. The set o( x,

t) as dened in the text.

Lemma 2.1. Let L
0
v 0 in P
T
= B

(0) (0, T), v C

2,1
(P

T
) C(P
T
), where
> 0, T > 0 and we assume
v(x, t) M , [x[ = , 0 < t < T , (2.1)
v(x, 0) < M , [x[ , (2.2)
for a given constant M. Then
v(x, T) < M , [x[ < . (2.3)
Proof. We have by continuity
v(x, 0) < M
4
, [x[ , (2.4)
for a suitable > 0, which we x from now on subject to this constraint.
Let us dene, for a > 0 to be chosen,
w(x, t) = M (
2
[x[
2
)
2
e
t
v(x, t) .
Then
L
0
w(x, t) = L
0
[(
2
[x[
2
)
2
e
t
] L
0
v(x, t)
L
0
[(
2
[x[
2
)
2
e
t
] = e
t
_
(
2
[x[
2
)
2
4(
2
[x[
2
)a
jj
+ 8a
ij
x
i
x
j
4b
i
x
i
(
2
[x[
2
)
_
. (2.5)
Here we employ the summation convention even for the term a
jj
and we under-
stand the coecient a
ij
, b
i
to be calculated at (x, t). We aim at proving that
the quantity . . . in last formula above is non negative, for a suitable choice of
> 0. Introduce a parameter (0, 1), and distinguish the cases:
(i) If [x[ , then
. . . (
2
[x[
2
)
_

2
(1
2
) 4A4B

0 ,
provided

2
(1
2
) 4A4B 0 . (2.6)
2. THE STRONG MAXIMUM PRINCIPLE 43
(ii) If > [x[ > , then
. . .
_
4A
2
(1
2
) + 8[x[
2
4B
3
(1
2
)

4
_
A
2
(1
2
) + 2
2

2
B
3
(1
2
)

= 4
2
_
2
2
A(1
2
) B(1
2
)

0 ,
provided
2
2
A(1
2
) B(1
2
) 0 . (2.7)
We may rst select so as (2.7) is satised, and then choose so that (2.6) is
satised too. Having xed in this fashion the values of and , we proceed to
observe that
L
0
w 0 , in P
T
.
Moreover on the parabolic boundary of P
T
we have
w(x, t) = M v(x, t) 0 , on [x[ = ;
w(x, 0) = M (
2
[x[
2
)
2
v(x, 0)
M
4
v(x, 0) 0 , in [x[ ;
we have made use of (2.4). Therefore w 0 in P
T
owing to the weak maximum
principle.
Especially
w(x, T) = M (
2
[x[
2
)
2
e
T
v(x, T) 0 ,
and we nally prove our claim, i.e., for [x[ < ,
v(x, T) M (
2
[x[
2
)
2
e
T
< M .

Theorem 2.1. (Strong Maximum Principle) Let L

0
u 0 in Q

T
. If ( x,

t)
Q

T
, and
max
Q
T
u = u( x,

t) ,
then
u(x, t) = u( x,

t) , for all (x, t) o( x,

t). (2.8)
Proof. Let us proceed by contradiction. Assume a point (x
1
, t
1
) o( x,

t) exists
such that u(x
1
, t
1
) < u( x,

t) =: M, and consider a polygonal (which must exist

by denition of o( x,

t))

n
i=1
(1 )(x
i
, t
i
) +(x
i+1
, t
i+1
) [ [0, 1] ,
where (x
n+1
, t
n+1
) = ( x,

t), and t
i
< t
i+1
, for i = 1, . . . n. We are going to prove
that
u(x
i
, t
i
) < M = u(x
i+1
, t
i+1
) < M ,
which obviously leads us to the contradiction
u( x,

t) = u(x
n+1
, t
n+1
) < M = u( x,

t) .
We may assume without loss of generality i = 1, i + 1 = 2. Let us switch to
dierent space coordinates:

j
= x
j
x
j
1
(x
j
2
x
j
1
)
t t
1
t
2
t
1
, j = 1 , . . . , N .
44 DANIELE ANDREUCCI
Thus
(x
1
, t
1
) (0, t
1
) , (x
2
, t
2
) (0, t
2
) .
Also dene the function
v(, t) = u(x(, t), t) .
In the change to the new variables, the set Q
T
is transformed to an open set
which certainly contains the closure of the cylinder
E

= (, t) [ [[ < , t
1
< t < t
2
,
provided > 0 is suitably chosen. Possibly redening we may assume (by
continuity)
v(, t
1
) < M , [[ ,
From now on, let be xed in this way.
We have in E

L
0
v(, t) := v
t
(, t) a
ij
(x(, t), t)v

j
(, t) +

b
i
(, t)v

i
(, t) 0 . (2.9)
where

b
i
(, t) = b
i
(x(, t), t)
x
i
2
x
i
1
t
2
t
1
.
Note that

L
0
is an operator satisfying the same assumptions as L
0
. More specif-
ically
n

i=1
|

b
i
|

B := B +N
[x
2
x
1
[
(t
2
t
1
)
.
Finally,
v(, t) M , [[ = , t
1
< t < t
2
,
follows from the denition of M. Therefore we may apply Lemma 2.1 to conclude
that
u(x
2
, t
2
) = v(0, t
2
) < M ,
as claimed.
2.1. More general operators. Lemma 2.1 and Theorem 2.1 still hold, if L
0
in
their statements is replaced with the more general operator L, provided c(x, t)M
0 in Q
T
(here M = u( x,

t) in Theorem 2.1).
We sketch here the changes needed in the proof of the Lemma, the proof of the
Theorem being essentially the same:
The calculation in (2.5) should be replaced with

Lw(, t) =

L[M (
2
[[
2
)
2
e
(tt
1
)
]

Lv(, t)
cM

L[(
2
[[
2
)
2
e
(tt
1
)
] C(
2
[[
2
)
2
e
(tt
1
)
+T
= e
(tt
1
)
_
( C)(
2
[[
2
)
2
4(
2
[[
2
)a
jj
+ 8a
ij

j

_
b
i

x
i
2
x
i
1
t
2
t
1

4
i
(
2
[[
2
)
_
,
where T denotes the the last term in the chain of inequalities in (2.5). Here we
used the fact that cM 0. It is clear that, additionally assuming e.g., > 2C,
the proof can be continued as above, taking into account Theorem 1.2.
3. HOPFS LEMMA (THE BOUNDARY POINT PRINCIPLE) 45
3. Hopf s lemma (the boundary point principle)
Definition 3.1. We say that a point ( x,

t)
p
Q
T
has the property of the
spherical cap if there exists an open ball B
r
(x
0
, t
0
) such that
( x,

t) B
r
(x
0
, t
0
) , B
r
(x
0
, t
0
) t <

t Q
T
,
with x
0
,= x.
In the following we denote by C
r
( x,

t) a cap B
r
(x
0
, t
0
) t <

t as the one
appearing in Denition 3.1. Note that if ( x,

t) has the property of the spherical

cap, then there exist innitely many such caps.
Remark 3.1. If ( x,

t) has the property of the spherical cap, the N-dimensional

open set G := Q

T
t =

t has the usual property of the sphere at x. Indeed,
it contains the N-dimensional sphere B
r
(x
0
, t
0
) t =

t, which however touches
the boundary of G at x.
On the other hand, examples can be easily given where ( x,

t)
p
Q
T
has the
property of the spherical cap, but fails to have the property of the sphere; see
Figures 3 and 4.
t
y
x
Q
1
T
Q
2

Figure 3. Every point of

p
Q
1
T
t > 0 has the spherical cap
property. This fails for the points on the vertical edges of Q
2

.
A version of Hopfs lemma for parabolic equations was rst proven in [19], [7].
Theorem 3.1. (Hopfs lemma) Let L
0
u 0 in Q

T
. Let ( x,

t)
p
Q
T
have the
property of the spherical cap. If
u(x, t) < u( x,

t) , for all (x, t) C

r
( x,

t), (3.1)
then
u
e
( x,

t) < 0 , (3.2)
46 DANIELE ANDREUCCI
where e R
N+1
is any direction such that
( x,

t) +se C
r
( x,

t) , for 0 < s < (e), (3.3)

and we also assume that the derivative in (3.2) exists.
Proof. First, let us invoke the strong maximum principle to prove that the
maximum value u( x,

t) may not be attained in the parabolic interior of C

r
( x,

t).
Namely, we obtain in this fashion the additional piece of information that
u(x, t) < u( x,

t) , for all (x,

t) B
r
(x
0
, t
0
).
Then, for each xed e as in (3.3), we may nd a spherical cap C

such that its

closure C

is contained in C
r
( x,

t)

( x,

t) and
u(x,

t) < u( x,

t) , for all (x, t) C

, (x, t) ,= ( x,

t), (3.4)
( x,

t) +se C

, for 0 < s <

(e). (3.5)
Well keep the notation C
r
( x,

t) in the following for a cap satisfying (3.4), (3.5).

Let us consider the barrier function
w(x, t) = exp
_
([x x
0
[
2
+[t t
0
[
2
)
_
expr
2
,
where > 0 is to be chosen, and (x
0
, t
0
), r are as in Denition 3.1. Thus
1 > w > 0 in B
r
(x
0
, t
0
), w = 0 on B
r
(x
0
, t
0
), and
w
t
(x, t) = 2(t t
0
)exp
_
([x x
0
[
2
+[t t
0
[
2
)
_
,
w
x
i
(x, t) = 2(x
i
x
i
0
)exp
_
([x x
0
[
2
+[t t
0
[
2
)
_
,
w
x
i
x
j
(x, t) =
_
2
ij
+ 4(x
i
x
i
0
)(x
j
x
j
0
)
2
_
exp . . . .
Therefore we have
L
0
w(x, t) = 2exp . . .
_
(t t
0
) +a
ii
2a
ij
(x
i
x
i
0
)(x
j
x
j
0
) b
i
(x
i
x
i
0
)

2exp . . .
_
(t
0
t) +A2[x x
0
[
2
+B[x x
0
[

. (3.6)
Dene
= C
r
( x,

t) [x x[ < ,
where the positive number is selected so as, for (x, t) ,
[x x
0
[ [ x x
0
[ [x x[ r sin
1
2
r sin .
Here is the angle between the t axis and the straight line joining (x
0
, t
0
), ( x,

t).
Note that (0, /2] as a consequence of x
0
,= x, according to Denition 3.1;
see also Figure 4. Hence, in we have
L
0
w(x, t) 2exp . . .
_
r +A
1
2
r
2
sin
2
+Br

0 ,
provided we nally choose so that
2(A+r +Br)
r
2
sin
2

.
Dene for a positive number to be chosen presently,
v(x, t) = u(x, t) +w(x, t) , (x, t) .
Note that

p
= S
1
S
2
, S
1
B
r
(x
0
, t
0
) , S
2
[x x[ = C
r
( x,

t) .
3. HOPFS LEMMA (THE BOUNDARY POINT PRINCIPLE) 47
t
( x,

t)

p
Q
T
C
r
( x,

t)
[x x[ =

(x
0
, t
0
)
Figure 4. and C
r
( x,

t), in the case t

0
>

t.
Then, on S
1
w = 0 and thus
v(x, t) = u(x, t) u( x,

t) .
On S
2
, taking into account (3.4),
u(x, t) u( x,

t) ,
for a suitable > 0. It follows that on S
2
too
v(x, t) = u(x, t) +w(x, t) u( x,

t) + u( x,

t) ,
if we choose . Moreover
L
0
v = L
0
u +L
0
w L
0
u 0 , in .
The weak maximum principle yields
v(x, t) u( x,

t) , in .
On the other hand, v( x,

t) = u( x,

t), so that
v
e
( x,

t) 0 .
Therefore
u
e
( x,

t) =
v
e
( x,

t)
w
e
( x,

t) 2( x x
0
,

t t
0
) e exp . . . < 0 .

A typical application of Theorem 3.1 is the following: assume u satises L

0
u = 0
in Q
T
and attains its maximum at a point ( x,

t)
p
Q
T
, having the spherical
cap property. Unless u is identically constant in a portion of Q
T
, by the strong
maximum principle, u is strictly less than its maximum value in Q
T
. Therefore
we are in a position to apply Hopfs lemma, and prove
u
n
> 0, if n is the spatial
outer normal to Q
T
at ( x,

t) (as dened in Subsection 2.1 of Chapter 2).

48 DANIELE ANDREUCCI
3.1. More general operators. Theorem 3.1 still holds, if L
0
in the statement
is replaced with the more general operator L, provided c(x, t) 0 in Q
T
, and
u( x,

t) 0. We sketch here the changes needed in the proof:

The operator L
0
is to be replaced everywhere with L.
Estimate (3.6) is substituted with the relation
Lw(x, t) 2exp . . .
_
c
2
(tt
0
)+a
ii
2a
ij
(x
i
x
i
0
)(x
j
x
j
0
)b
i
(x
i
x
i
0
)

2exp . . .
_
A+
C
2

1
2
r
2
sin
2
+r +Br

0 ,
which is valid in as above, for a suitable selection of > 1. The proof is
concluded as above.
3.2. Maximum estimates in problems with boundary conditions involv-
ing the spatial gradient. Assume u solves the problem, to be complemented
with initial and additional boundary data,
L
0
u = 0 , in G(0, T), (3.7)
u
n
= 0 , on G(0, T), (3.8)
where G R
N
, and n denotes the outer spatial normal to G(0, T). Then u
can not attain either its maximum or its minimum on , owing to Hopfs lemma
(unless it is identically constant in some open set).
Assume now (3.8) is replaced with
u
n
= h(x, t)u +k(x, t) , on G(0, T).
Here h > 0 and k are continuous functions. Assume ( x,

t) is a point of
maximum for u. Then
0 <
u
n
( x,

t) = h( x,

t)u( x,

t) +k( x,

t) ,
implying
u( x,

t) <
k( x,

t)
h( x,

t)
.
Note however that a similar, but not necessarily strict, inequality can be proven
trivially without invoking Hopfs lemma. Analogous estimates hold at points of
minimum.
4. Maximum principle for weak solutions
In this section we proceed formally, with the purpose of exhibiting the ideas
behind a possible extension of the maximum principle to weak solutions of
u
t
div a(x, t, u, Du) 0 , in Q
T
= G(0, T), (4.1)
where
a(x, t, u, Du) Du 0 .
Assume that G(0, T) = S
1
S
2
, with
u(x, t) M , on S
1
,
a(x, t, u, Du) n 0 , on S
2
,
4. MAXIMUM PRINCIPLE FOR WEAK SOLUTIONS 49
in a sense suitable for the calculations showed below, and also that
u(x, 0) M , x G.
Here n is, as above, the outer spatial normal.
Multiply (formally) (4.1) by (u M)
+
, and integrate by parts, obtaining, on
dropping the non negative term involving a Du,
_
G
(u(x, t) M)
2
+
dx 2
t
_
0
_
G
(u M)
+
a(x, t, u, Du) nd d
+
_
G
(u(x, 0) M)
2
+
dx,
for all t (0, T). The last integral equals 0, because of the assumed bound on
the initial data. The surface integral is non positive: indeed, on S
1
we have
(u M)
+
= 0, while on S
2
it holds
(u M)
+
a(x, t, u, Du) n 0 .
Therefore we get
_
G
(u(x, t) M)
2
+
dx 0 , for all 0 < t < T,
i.e., u M in G(0, T).
APPENDIX B
A theorem by Kruzhkov
We present here a result of [10], which is instrumental in our proof of existence
of weak solutions to the Stefan problem, in the modied version quoted in [8].
1. Mollifying kernels
Let be a mollifying kernel, i.e.,
C

0
(R) , supp [1, 1] , 0 , > 0 on [1/2, 1/2].
Dene

(x) =
1

N

_
[x[

_
, x R
N
.
On multiplying, if required, by a positive constant, we may assume
_
R
N

(x) dx =
_
R
N
([x[) dx = 1 , for all > 0.
Let v L
1
loc
(R
N
). Dene for all x R
N
,
v

(x) =
_
R
N
v(y)

(x y) dy =
_
R
N
v(x z)

(z) dz . (1.1)
Let be a bounded open set of R
N
with a Lipschitz continuous boundary. Dene
for 1 > > 0

= x [ dist(x, ) > .
In the following

will denote a modulus of continuity, that is a continuous, non

decreasing, non negative function

: R R such that

(0) = 0. The notation

emphasizes the possible dependence of this function on .
Our rst result is not actually needed in the proof of the main estimate Theo-
rem 2.1, but it was quoted in the proof of Theorem 5.1 of Chapter 2.
Lemma 1.1. Let v L
1
(), and assume that for each 0 < < 1, h R
N
, [h[ <
_

[v(x +h) v(x)[ dx

([h[) . (1.2)
Then, for all 0 < < , [h[ < ,
_

2
[v

(x +h) v

(x)[ dx

([h[) . (1.3)
Note that, to keep our notation formally coherent with (1.1), we extend in (1.3)
v to v = 0 outside of . However, the integrals dening v

(x+h) and v

(x) there
are actually calculated over (check this).
51
52 DANIELE ANDREUCCI
Proof. By denition of v

,
_

2
[v

(x +h) v

(x)[ dx =
_

_
R
N
v(y)[

(x +h y)

(x y)] dy

dx
=
_

_
R
N

(y)[v(x +h y) v(x y)] dy

dx

_
R
N

(y)
_ _

2
[v(x +h y) v(x y)[ dx
_
dy
=
_
R
N

(y)
_ _

2
y
[v(z +h) v(z)[ dz
_
dy .
Here we use the standard notation
Gy = z [ z +y G .
Recall that

(y) = 0 if [y[ . Therefore in last integral we may assume, as

< by assumption,

2
y
2

.
Hence the last integral above is bounded by
_
R
N

(y)

([h[) dy

([h[) .

Let us dene the sign function

sign(s) =
_

_
1 , s > 0 ,
0 , s = 0 ,
1 , s < 0 .
Clearly, as 0,
v(x)[sign(v(x))]

v(x) sign(v(x)) = [v(x)[ , a.e. x R

N
.
Next lemma gives a measure of the speed of convergence in this limiting relation.
Lemma 1.2. In the same assumptions of Lemma 1.1, for all 0 < < ,
_

[[v(x)[ v(x)[sign(v(x))]

[ dx 2

() . (1.4)
Proof. For all x, y R
N
[[v(x)[ v(x) sign(v(y))[ = [[[v(x)[ [v(y)[] [v(x) v(y)] sign(v(y))[
2[v(x) v(y)[ .
2. THE MAIN ESTIMATE 53
Thus
_

[[v(x)[ v(x)[sign(v(x))]

[ dx
=
_

_
R
N
_
[v(x)[ v(x) sign(v(y))

(x y) dy

dx

_
R
N
2[v(x) v(y)[

(x y) dy dx = 2
_
R
N
_

[v(x) v(x y)[ dx

(y) dy .
Since

(y) = 0 for [y[ , we may bound above last integral by

2
_
R
N
_

[v(x) v(x y)[ dx

(y) dy 2
_
R
N

([y[)

(y) dy 2

() .

2. The main estimate

Theorem 2.1. Let v L

(Q
T
), and assume that for all 1 > > 0, [h[ we
have
_

[v(x +h, t) v(x, t)[ dx

([h[) , a.e. t (0, T). (2.1)

Assume moreover that for all g C
1
0
(

), and a given C

> 0,

g(x)[v(x, t +s) v(x, t)] dx

s
_
|g|

+|Dg|

_
, (2.2)
a.e. 0 < t < t + s < T. Then, we have a.e. 0 < t < t + s < T, and for all
0 < < ,
_

[v(x, t +s) v(x, t)[ dx

_
s

+ +

()
_
. (2.3)
Here

depends on , C

, |v|

and N.
Proof. Choose in (2.2)
g(x) =

(x) =
_
R
N
(y)

(x y) dy ,
where 0 < < and
(x) =

2+
(x) sign(v(x, t +s) v(x, t)) .
Note that
[Dg[
(N)

;
moreover g(x) = 0 if x

+
. Finally,

(x) = [sign(v(x, t +s) v(x, t))]

, x
3+
.
54 DANIELE ANDREUCCI
Thus, exploiting assumption (2.2),

3+
[v(x, t +s) v(x, t)]

(x) dx

[v(x, t +s) v(x, t)]

(x) dx

3+
[v(x, t +s) v(x, t)]

(x) dx

(N)
s

+()|v|

.
Therefore
_

[v(x, t +s) v(x, t)[ dx

_

3+
[v(x, t +s) v(x, t)[ dx +()|v|

3+
[[v(x, t +s) v(x, t)[ [v(x, t +s) v(x, t)]

(x)[ dx+
+

3+
[v(x, t +s) v(x, t)]

(x) dx

+()|v|

() +
s

+ ,
where we used Lemma 1.2. Indeed, it is easily checked, with the help of (2.1)
that
_

[[v(x +h, t +s) v(x +h, t)[ [v(x, t +s) v(x, t)[[ dx 2

([h[) ,
a.e. 0 < t < t +s < T, and for and h as above.
The proof is concluded.
Remark 2.1. It is clear that estimate (2.3) implies continuity in t of v in the
L
1
(

) norm. In the optimal case when

([h[) = C

[h[, on choosing =

s (for
s <
2
) we obtain
_

[v(x, t +s) v(x, t)[ dx

s . (2.4)
APPENDIX C
The spaces H
m+,
m+
2
(Q
T
)
We dene here a class of Banach spaces of standard use in the regularity theory
of parabolic equations (see [11]). Let Q
T
= (0, T), where R
N
is a
bounded open set.
Fix the integer m 0, and (0, 1). In the following we denote with D
r
t
D
s
x
u,
for r, s N, any derivative of u, taken r times with respect to the time variable,
and s times with respect to space variables. For a given function u : Q
T
R we
introduce the quantities
u)
()
x,Q
T
= sup
(x,t) ,(x

,t)Q
T
[u(x, t) u(x

, t)[
[x x

, u)
(m+)
x,Q
T
=

2r+s=m
D
r
t
D
s
x
u)
()
x,Q
T
,
u)
(/2)
t,Q
T
= sup
(x,t) ,(x,t

)Q
T
[u(x, t) u(x, t

)[
[t t

[
/2
,
u)
(
m+
2
)
t,Q
T
=

2r+s=m1,m
D
r
t
D
s
x
u)
(
m+2rs
2
)
t,Q
T
.
The sums above (and below) are extended to all the derivatives D
r
t
D
s
x
u with r,
s as indicated. If m = 0, the only such function is u itself.
Then we dene the norm
[u[
(m+)
Q
T
=

2r+sm
|D
r
t
D
s
x
u|
,Q
T
+u)
(m+)
x,Q
T
+u)
(
m+
2
)
t,Q
T
.
The Banach space of the functions u whose norm [u[
(m+)
Q
T
is nite is denoted by
H
m+,
m+
2
(Q
T
) .
1. Comments
The regularity in time of u H
m+,
m+
2
(Q
T
) is a half of the regularity in
space, in a sense made precise by the denition itself. To illustrate this point,
assume s C([0, T]), and regard s as a function dened over Q
T
. Let us make
explicit the meaning of the statement
s H
m+,
m+
2
(Q
T
) .
If m is even, m = 2k, this is equivalent to: s C
k
([0, T]), and
sup
0<t,t

<T
[s
(k)
(t) s
(k)
(t

)[
[t t

2
< .
If instead m is odd, m = 2k 1, we have: s C
k1
([0, T]), and
sup
0<t,t

<T
[s
(k1)
(t) s
(k1)
(t

)[
[t t

[
1+
2
< .
55
56 DANIELE ANDREUCCI
Clearly, for all m 0,
s H
m+,
m+
2
(Q
T
) = s H
2+m+,
2+m+
2
(Q
T
) . (1.1)
APPENDIX D
Symbols used in text
|f|
p,D
norm of f in L
p
(D), 1 p ; often the indication of the domain
is omitted, when superuous.
(t) the set (x, t) [ x .
[[ the Lebesgue measure of the set .
dist(A, B) standard Euclidean distance between A and B.
B
r
(x) ball with center x and radius r.
x s
0
+ x goes to s
0
from the right.
x s
0
x goes to s
0
from the left.
f(s
0
+) denotes the limit of f(x) as x s
0
+.
f(s
0
) denotes the limit of f(x) as x s
0
.
s
+
positive part of s R, s
+
= max(s, 0).
s

negative part of s R, s

= max(s, 0).

ij
Kroneckers symbol:
ij
= 1 if i = j,
ij
= 0 if i ,= j.
Df spatial gradient of the function f(x, t): Df = (
f
x
1
, . . . ,
f
x
N
).
e
k
k-th unit vector of the standard basis in R
N
.
C(A) class of continuous functions in A. The same as C
0
(A).
C
n
(A) class of functions, continuous in A together with
their derivatives up to order n.
C
n
0
(A) class of the functions in C
n
(A), whose support is
compact and contained in A.
C
2,1
(A) class of functions f, such that f,
f
t
,
f
x
i
, and

2
f
x
i
x
j
are continuous in A for all i, j = 1, . . . , N.
W
2
1
(G) Sobolev space of functions in L
2
(G) whose rst derivatives are in L
2
(G).
W
2
2,1
(G) Sobolev space of functions f W
2
1
(G) whose spatial derivatives

2
f
x
i
x
j
are in L
2
(G) for all i, j = 1, . . . , N.
PDE Partial Dierential Equation/Equations.
57
Bibliography
[1] D. Andreucci. Behaviour of mushy regions under the action of a volumetric heat source.
Math. Methods in the Applied Sciences, 16:3547, 1993.
[2] O. Arena. Partial Dierential Equations of Parabolic Type. Notes of a course held at the
University of Florence (Italy). Unpublished, 1983.
[3] J. R. Cannon. The One-Dimensional Heat Equation, volume 23 of Encyclopedia of Mathe-
matics and Its Applications. Addison-Wesley, Menlo Park, CA U.S.A., 1984.
[4] E. DiBenedetto. Continuity of weak solutions to certain singular parabolic equations. Ann.
Mat. Pura Appl., 130:131176, 1982.
[5] A. Fasano. Free Boundary problems and Their Applications. Notes of a course held at
SASIAM, Bari (Italy). 1989.
[6] A. Fasano and M. Primicerio. New results on some classical parabolic free-boundary prob-
lems. Quart. Appl. Math., 38:439460, 1981.
[7] A. Friedman. Remarks on the maximum principlefor parabolic equations and its applica-
tions. Pacic J. Math., 8:201211, 1958.
[8] I. G. Gotz and B. B. Zaltzman. Nonincrease of mushy regions in a nonhomogeneous Stefan
problem. Quart. Appl. Math., 49:741746, 1991.
[9] S. L. Kamenomostskaya. On Stefan problem. Mat. Sbornik, 53:489514, 1961. (In Russian).
[10] S. N. Kruzhkov. First-order quasilinear equations in several independent variables. Mat.
Sb. (n.s.), 81:228255, 1970. Engl. Transl.: Math. USSR - Sb 10 (1970), 217-234.
[11] O. A. Ladyzhenskaja, V. A. Solonnikov, and N. N. Uralceva. Linear and Quasilinear Equa-
tions of Parabolic Type, volume 23 of Translations of Mathematical Monographs. American
Mathematical Society, Providence, RI, 1968.
[12] A. M. Meirmanov. The Stefan Problem, volume 3 of de Gruyter Expositions in Mathematics.
Walter de Gruyter, Berlin, Germany, 1992.
[13] L. Nirenberg. A strong maximum principle for parabolic equations. Comm. Pure Appl.
Math., 6:167177, 1953.
[14] O. A. Oleinik. A method of solution of the general Stefan problem. Soviet Math. Dokl.,
1:13501354, 1960.
[15] M. Primicerio. Mushy regions in phase-change problems. In K. H. Homann and R. Goren-
o, editors, Applied Nonlinear Functional Analysis: Variational Methods and Ill-Posed
Problems, pages 251269, Frankfurt, Germany, 1983. Verlag Peter Lang.
[16] D. G. Schaeer. A new proof of the innite dierentiability of the free boundary in the
Stefan problem. J. Di. Equations, 20:266269, 1976.
[17] J. Stefan.

Uber die Theorie der Eisbildung. Monatshefte Mat. Phys., 1:16, 1890.
[18] D. Tarzia. A bibliography on moving-free boundary problems for the heat-diusion equa-
tion. The Stefan and related problems. MAT-Serie A, 2, 2000.
http://www.austral.edu.ar/web/investigacion/2(2000).htm.
[19] R. Viborni. On properties of solutions of some boundary value problems for equations of
parabolic type. Doklady Akad. Nauk SSSR (n.s.), 117:563565, 1957. (In Russian).
[20] A. Visintin. Models of Phase Transition, volume 28 of Progress in Nonlinear Dierential
Equations and Their Applications. Birkha user, Boston, MA U.S.A., 1996.
59