8

Downloaded 11/22/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.
php
CHAPTER
VIII
A One Phase Stefan Problem
1. Introduction A one phase Stefan problem is the description, typically, of the melting of a body of ice maintained at 0C in contact with a region of water. In general, the Stefan problem refers to the change of phase in a thermoconductive medium when energy, that is, heat, is contributed to the system. Unknown are (i) the temperature distribution of the water as a function of space and time and (ii) the free boundary consisting of the ice-water interface. The temperature distribution is required to solve the heat equation in the aqueous region and energy is conserved across the interface. By limiting ourselves to the case of one space dimension we shall be able to explore many of the features of this problem while avoiding much of the technical complication. This will be especially true in our investigation of the free boundary by means of the Legendre transform. Let T > 0 and s o > 0 be given. This is the classical Stefan problem.
Problem 1.1. To find a function O(x, t) and a curve F: t = s(x), x> s o , such that
+ O, = 0
in {(x, t) : s(t) <t < T}, for (x, t) e I',
0=0 O x s'(x) = k
I4k]
Downloaded 11/22/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
INTRODUCTION
279 O(x, 0) = h(x) 0(0, t) = g(t) for 0 < x < so , for 0 < t < T,
where h(x) > 0 is the initial temperature, g(t) > 0 is the heat contributed at time t, and k > 0 is the heat of fusion. The function O(x, t) is interpreted as the temperature of the water at the point x at time t and the curve F represents the interface between the ice and the water. At time t the water occupies the subset defined by {x : s(x) < t}, where we understand that s(x) = 0 when x < s o . Note that O(x, t) = min O = 0 for (x, t) e F 1 so O x < 0 on F. Hence s'(x) >_ 0, (1.1)
which means that the curve F is monotone. This is the property of the one phase problem which permits its transformation without difficulty to a variational inequality. At this point we transform Problem 1.1 to a variational inequality. Let R > s o > 0. Later we shall exhibit this R as a function of the initial data k, and so on. Set D = (0, R) x (0, T). Our variational inequality will be for u(x, t) defined by t)
u(x,
_f
1
O(x, T) di,
s(x) <_ t < T, s o < x < R, 0<t<s(x), so <x_<R (1.2)
s(x)
u(x,t)=0, u(x, t)
= J0
O(x, T) dT,
0<t< T,
0 < x < so .
We compute the differential equation satisfied by u assuming that 0 and s are smooth: u(x, t)
=f
O x(x, r) di s'(x)O(x, s(x)) O x(x, i) di, s(x) < t < T, s o < x < R.
s(x)
=J
Also u(x, t) =
s(x)
s(x)
JO
t
xx
(x, i) di s'(x)O x(x, s(x))
= f 0(x,
s(x)
T)
di + k
=0(x,t)+k=u,(x,t)+k,
s(x)<t<T, so<x<R.
280
VIII STEFAN PROBLEM
In the same manner, u(x, t) = u,(x, t) h(x), So let us define

f (x)
and
0 < t < T, 0 <x<s o .
= h(x),
0 <x<s o
s o <x<R
(1.3)
1' (t) = f g(r) dT,

o
0 < t < T,
(1.4)
where we recall that k > 0 and g and h are smooth functions positive in [0, T] and [0, s o ), respectively. With K the set of nonnegative functions in L 2 (D), the conditions on u may be expressed as the alternatives
( u.S+u,)(v u) f(v u)=0

u >0
or
for vcK
( u,, +u,)(v u) f(v u)=kv >_0

u=0
for vel(.
This leads to the variational inequality, for T > 0 given: Problem 1.2. To find u E L 2 (0, T ; H 2 (0, R)) n K such that
U,E K, ( u +
U,)(v U) ^ f (v
U)
Q.C. for v E K, for for for

0 <t<T, 0 <t<T,
x=0,
u =>/'
u=0
u=0
x=R,
t=0, 0<x<R.
Here we have used the notation L 2 (0, T; H 2 (0, R)) to denote the functions
u(x, t) satisfying
Jo 0
T R
(u 2 + ux + us X) dx dt <
oo.
EXISTENCE AND UNIQUENESS OF THE SOLUTION
281
2. Existence and Uniqueness of the Solution
A solution to Problem 1.2 will be found by approximation with a suitable penalty function. The particular choice of penalty is intended to simplify the existence and regularity proofs. For E > 0 we choose #,(t) e C(l) with the properties
13 e (t)=O
if
t>_ e (2.1)
# (0) = 1
1
/3E(t) > 0
and
fl (t) < 0, oo <t < E
Next choose a sequence J(x) of smooth functions in [0, R] which are uniformly bounded and decrease tof (x) defined by (1.3) as r --i 0. Finally choose q(x) e C O (0) to satisfy 1 x = q 0 for 0 < x < As o for as o < x
and 0 < ij(x) < 1, x e R. For T > 0 and E > 0 given, consider the initial boundary value problem Problem 2.1. To find u(x, t), (x, t) e D, such that
u,x +u,+kf e(u) =fE in D, u =erf, t=0, 0<x<R, u=i+ u=0,

E,
(2'2)
(2.3)
0<t<T, x=0, 0<t<T, x=R,
where k> 0 is the constant of Problem 1.1 and i is defined by (1.4).

It is well known that Problem 2.1 admits a classical solution u = u e ; cf. Friedman [2]. Lemma 2.2. Let u E , e > 0, denote the solution to Problem 2.1 in D. Then there is an E o > 0 such that
0< Vit`(x,t)<K where K > 0 is independent oft, 0 < r < e.
in D,
282
VIII STEFAN PROBLEM
Proof. We differentiate (2.2) with respect to t and apply the maximum principle. For w = au/at, we obtain
w, + k/3E(u E )w = 0
in D,
w=f, +E ^?xx w =tii',

w=0,
kfl(u,,), t
x=0, x=R.
(2.4)
Since #,(u,) > 0 in D we infer from the maximum principle that
mi (O"
we deduce that
min w, 0) < w(x, t) < max(max w, 0 in D, (2.5) D ??n
where 3D denotes the parabolic boundary of
D, i.e.,
aD {( x, T): 0 < x < R }.

First we evaluate the left-hand member of this expression. Recalling that fE >_fand
0 _< U E(x, 0) = Erl(X) < E
for 0 <_ x <_ 3Sp,
w(x, 0) ? f (x) + 6q..(x), = h(x) + eij (x) > 0,
0 < x < 3s 0 0 < x < as o
fore < E o , s o sufficiently small, because h(x) is positive in [0, s o ). On the other hand w(x,0)= fg(x) k$ E(0)>_h(x) +k >_0 and w(x, 0) = f( x) kf3 E(0) > k k = 0 Hence w(x, 0) >_ 0 On the vertical sides of the boundary,
w(0,t)= 0'(t) =g(t) >0
for
as o <x<s o
for s o < x < R.
for 0 < x < R.
for 0 <t <T
and w(R, t) = 0 Consequently, min(min a' w, 0) = 0. for 0 < t < T.
283
We now inspect the right-hand inequality of (2.5). We need only note that
I J + Erl xx kfE( u E) I < sup I E I + e sup I nxx I + k < K 1 ,

when t = 0, for a K 1 independent of e < E o to show that max(max w, 0 < max(max g, K 1 ) = K. ap n The assertion of the lemma now follows from (2.5). Q.E.D. Lemma 2.3. Let u E , 0 < e < e o , denote the solution to Problem 2.1. Then
II f E(UE) Il c-(n) < 1
for 0
< E < C0 .
Proof. From the previous lemma, 8u,/8t >_ 0 in D so u E(x, t) >_ 0 in D by (2.3). Recalling our choice of /3 E in (2.1), 1 < /3 E(u e(x, t)) < 0 if (x, t) E D. Q.E.D.
At this point we address the existence and uniqueness of a solution to the variational inequality. Theorem 2.4. There exists a unique solution u to Problem 1.2. It enjoys the properties that u, ux, ut, u and u>_0, uE --> u and
uE --p u
L(D) in D. 1 <p < x,
u,>_0
Let u E denote the solution to Problem 2.1. Then as e -+ 0 weakly in H 1 "P(D),
weakly in H 2 "(0, R), 1 <p < oo,
for each t, 0 < t < T o , and hence u E --+ u uniformly in D and u i,, - u x uniformly in (0, R) for each t e (0, T). Proof. In view of the preceding lemmas, the solution u E , 0 < r < E o of Problem 2.1 satisfies
,
IIUEJ1L -cn) = I!f u E, kY( u C)II Lc (D) 0<e<eo. <IIfJII +K +k <Cl,
284
VIII STEFAN PROBLEM
Hence u(x, t) is a family of bounded functions, because the u E are uniformly bounded by Lemma 2.2. Therefore luIll ,n, =
n n independent of E, 0 < E < s o . This implies the existence of a sequence -+ 0 and a function u e H 1 "(D), 1 < p < oo, such that
1)
ff
IuEZ dx dt +
ff
IueXl dx dt +
f f Iu
IP dx dt < C 2
u . -+ u
in H" (D)
(2.6)
and uniformly in D. In particular, u >- 0 and satisfies the boundary conditions of Problem 1.2. In addition, a subsequence of the u E converges weakly to u in H 2 '(0, R), < 1 p < oo, for each fixed t, 0 < t < T. Since, however, u E - -* u uniformly in D, it follows that the entire sequence u E , -+ u weakly in H 2 ' "( 0, R), 0 < t < T. In particular, u XX E L (D). To complete the proof that u is a solution to Problem 1.2, let v E L(D) satisfy v >_ 6 > 0. Multiplying (2.2) by v - u E and noting that /3 E(v) = 0 if E < 6, we see that E < 6. ( uXX + u,,)(v u) k [fE( v) YE(uE)](v uE) = fg( v ue),
Integrating with respect to x e [0, R] and observing that

[/E( v )
/' E(uE)](v u E ) k 0,
we obtain
(-u + uE,)(v - u) dx
fo f (v - u ) dx.
E
Taking E -^ 0, owing to the weak convergence we get
R(-
uXX + u,)(v - u) dx >-
R f (v
- u) dx.
(2.7)
By approximation, (2.7) holds for all v >- 0, v E L 2 (0, R). It follows that u is a solution to Problem 1.2. We can easily check that the solution is unique in L 2 (0, T; H 2 (0, R)). This implies, in particular, that the entire sequence u, converges weakly to u in H'' (D) and in H 2 ' ( 0, R) for each t, 0 < t < T. That u, >_ 0 follows from Lemma 2.2. Q.E.D.
Corollary 2.5. Let u denote the solution to Problem (1.2) and define S2(t) = {x E (0, R) : u(x, t) > 0}, Then S2(t) r- 1(t') for t < t', t, t' e [0, T]. 0 < t < T.
2 EXISTENCE AND UNIQUENESS OF THE SOLUTION
285
Proof. This is obvious because u is Lipschitz with u, > 0. Q.E.D. We shall now show that given T > 0, R > 0 may be chosen so that u = 0 in a neighborhood of x = R. This implies the existence of a free boundary. To begin we prove a comparison theorem analogous to our weak maximum principle of Chapter II.
Lemma 2.6. Suppose that f < f and Ii < t and that u, u are solutions of Problem 1.2 for f, b and f, i/i, respectively. Then
u<
in D.
Proof. For v e Il*, the set of nonnegative H'(D) functions obeying the boundary conditions of Problem 1.2, the integrated form of the problem is valid. Namely,
ff
D
[u x (v u)x + (v u)] dx dt >
ff
D
f (v u) dx dt,
v e K.
In this expression we may take v = max(u, u) E III* since i' < >i. Hence
#0ju il) x + (u )] dx dt >_ ff f (u 6) dx dt, A A

A = {(x, t) : u(x, t) > u(x, t)}.
(2.8)
Further observe that u s), + u, = f a.e. in A since u > u > 0 there. Consequently, for = max(u u, 0), which is zero on D,
JJ [u ( + u,(] dx dt =
x
ff
D
ff dx dt.
(2.9)
Writing (2.8) in terms of , we see that
ff [Ox^x + ] dx dt >_
D
ff
D D
fI dx dt.
(2.10)
Adding (2.9) and (2.10) and again using the definition of (, we obtain
ff
D
C2 dx dt
((x,T) 2 dx>
JJ (ff)Cdxdt>0.
Hence C = 0 a.e. in D, i.e., meas A = 0. Q.E.D.
286
VIII STEFAN PROBLEM
We may now construct an explicit solution of the Stefan problem which we employ to dominate our solution u. Denote by Nti, the set
NM = {x : Ix) < M(t + 1) 1 / 2 ),

where M is a positive constant to be determined. Set (x, t) = F
and
112 ( (t + 1)
b(x,t)= x -M(t+ l) 112 ,

where
x >_0, t >_0,
F(z)=C
with C, C' satisfying
2Ce -M2 / 4 =
k,
e2f4 -^d^ -
C'
F(M) = C
(2.11)
e - ; 2/4
dC - C' = 0.
(2.12)
It is easy to check that (x, t) and the curvet defined by cb(x, t) = 0 is a solution to the classical Stefan Problem 1.1 with h(x) = F(x),
and
g(t) = F(0),
0 < x < M,
< t < T.
Hence determines a solution u to Problem 1.2 via (1.2). Note that the heat of fusion corresponding to and r is the same k as that pertaining to u.
Now
= e "' / --,00 as M C2
2 4
C'
k _ _ f e (MZ t î4 < - (n + 2M 2 M

2
1)e (M2- " 2 ^' < o0
is bounded as M -^ oo. Hence we may find M so large,that
h(x) >_ h(x)
and
0(t)
>_ g(t).
287
Now we apply Lemma 2.6 to conclude

in D. >u
Since u(x, t) = 0 for x >_ M(t + 1) 1 / 2 ,

0<t<T.
S2(t)c{x:0<x< M(t +1) 1 / 2 ),
In sum:
Proposition 2.7. Given T > 0 we may choose R > 0 so that the solution u to Problem 1.2 vanishes in a neighborhood of x = R.
Proof. We choose M as above and R > M(T + 1) 1 / 2 . Q.E.D.
It will always be assumed that R has been chosen so large that the conclusion of Proposition 2.7 is satisfied. Set S2 = {(x, t) : u(x, t) > 0),
S2(t) = {x : (x, t) E S2}, 1 = 8S2 D.
, -
0< t < T,
(2.13)
Theorem 2.8. Let u be a solution of Problem 1.2. Then F defined by (2.13) admits the representation
F: x=a(t),
0<t<T,
where a is a continuous increasing function oft with s o = a(0) < a(t) for t > 0. Proof. We first show that S2(t) is connected, which implies that a(t) is well defined. The interval (0, s o ) c SZ(t) for all t because u, _> 0. Suppose that (x i , x 2 is a component of the open set S2(t) which does not contain (0, s o ). Then
)
u(x, t) = k u,(x, t) < 0
in (x 1 , x 2 ),
so, by the maximum principle,

t) <
u(x,
max(u(xl, t), u(x2, t)) = 0,
xl <x <
x2.
This is a contradiction; hence, S2(t) is connected. We define

a(t) = sup{x : x a(0) = so.
E Q(t)},
0< t < T,
288
VIII STEFAN PROBLEM
Since f (t) c f (t') for t < t', it is obvious that 6 is monotone. Suppose now that
x1 =
Q(t) < urn 6(t ) =

'
x2
t'-.t >c
for some t, 0 < t < T. In this event, the rectangle

Q= {(x,i):x 1 +c<x<x 2 e,t
T
<T}cQ,
for E > 0 small, and u xx +u,= k

u(x,t) =0
in Q, for xl +fi <x <x2 +. for x l + E < x < X 2 E.
By the regularity theory for parabolic equations, u is smooth in Q so

t) = u x(x, t) = u X .,(x, t) = 0
u(x,
From the equation, u,(x, t) = k for x l + E < x < x 2 E, whence u,(x, i) 15(x 1 + x 2 ), t). This contradicts the terms of <0 in a neighborhood in Q of (our existence theorem, Theorem 2.4. In the same way we may show that u is continuous from below. Finally we point out that Q(t) > s o for t > 0. Suppose not and that Q(t o ) = s o for some t o > 0. By monotonicity of Q, u(t) = so for 0 < t < t o . Consider the rectangle
Q = {(x, 0: so E < x < s, 0< t < t0}
for c> 0, small, and observe that

-0 u,.,+u,=f=h >
u = 0 u >0
in Q, for x =s o , 0 <t<t o , for x =s o

E,
0 <t<t o .
Since u X is continuous in D and u attains its minimum at x = s o , u X (s o , t) = 0 for 0 < t < t o . in Q,

(2.14)
Introduce the solution v(x, t) to the Dirichlet problem

v=0 on
v,.,+v,=h
a p Q,
where 8,Q denotes the parabolic boundary of Q. Since h > 0,

v >minv =0 apQ
in Q.
SMOOTHNESS PROPERTIES OF THE SOLUTION
289
Hence by the Hopf-Friedman maximum principle,

vx (s o , t) < 0,
0 < t < t o . (2.15)
On the other hand, u v is a solution to the heat equation in Q with u v > 0 on B p Q. Hence, since u and v are not identical, (u v)x(s0, t) < 0 because u v attains its minimum there. Combining this with (2.14) and
(2.15) we obtain
0 = u x (s o , t) < v x (s o , t) < 0,
the desired contradiction. Q.E.D.
0< t < t o ,
It is useful to point out that F is a Lipschitz curve in the coordinates obtained from (x, t) by rotation through ir/4. Hence a function C E H I (S2) admits a trace in L 2 (F).
3. Smoothness Properties of the Solution
To discuss the free boundary F of our Stefan problem it is useful to know that u xx (x, t) and u x,(x, t) are continuous in S2 near F. One feature of the development is the proof that u, E H'(D n {t >_ t o > 0}) which proceeds by use of the penalizations. Another is the application of a weak maximum principle and Bernstein's method to show that u, is Lipschitz in 1 near F.
Lemma 3.1. Let u E , 0 < e < E o denote the solution to Problem 2.1. Then
,
az uE
x at and
z
(x, t) < C
for
0 < x < s 0 /6, 0 < t < T
xz (x, t) < C
for
0 < x < s 0 /6, 0 < t < T,
where C> 0 is independent
of e, 0 < a < a o .
Proof. From the initial conditions (2.4),
u E (x, 0) = e
for 0 < x <
s0 /3,
`3Z;
VIII STEFAN PROBLEM
so, since u E,(x, t) > 0, in D, U,(x, t) > E in 0 < x < s 0 /3,

T, 0
<t<
T.
Hence /3 E(u E(x, t)) = 0 in 0 < x < s 0/3, 0 < t < equation
and u, is a solution to the
UExx
+U =f,
0<x<s 0 /3, 0<t<T.
The conclusion now follows from the standard Schauder theory of the heat equation (Friedman [2]). Q.E.D. Lemma 3.2. Let u E , 0 < E < s o , denote the solution to Problem 2.1. Then
t) 2 dx dt < C
where C> 0 is independent of s, 0 < e < E o . Proof. For the duration of the proof, we set u = u , 3 = /J E , and w = u t . Differentiate Eq. (2.3) with respect to t and multiply by w to obtain
E
ww xx + ww, + k$'(u)w 2 = 0
in D,
which, integrated over (0, R) for fixed t, 0 < t < T, yields rR Id j.R r R owwxxdx+ 2 dx +k J o /3'(u)w 2 dx =0. J w 2dt
(3.1)
The first term may be integrated by parts. We observe that
fo
,wwxx dx =
w 2 dx + w(0, t)w x(0, t)
w(R, t)w x(R, t)
= JR wX dx + w(0, t)w (0, t)

x 0
because w(R, t) = u,(R, t) = 0. By the previous lemma
I w(0, t)wx(0, t) I <_ I ^Gt(t)uext(0, t) I <_ const = C 1 .

Therefore, from (3.1) and recalling that /3'(u) > 0,
f w,dx + -d fR w 2 dx<C .
l
This, integrated over (0, T), gives
o J
r R
1 R
, 0
wX dx dt + I [w(x, R) z w(x, 0) 2 ] dx < C 1 T.

2 o
3 SMOOTHNESS PROPERTIES OF THE SOLUTION
291
According to Lemma 2.2, I w(x, t) I < K independent of E, whence

T
fJ
r wz dx dt < C T + RK . Q.E.D.
R
I
0 0
Lemma 3.3. Let u E , 0 < e < E o , denote the solution to Problem 2.1. Then for 0<c <t<T,
Ruext(x, Jo
t) 2 dx +
JrJ
RuEtt(x,
T) 2 dx dT < C,
6
where C> 0 is a constant independent of E, 0 < E < e o .

This lemma is somewhat more difficult. Clearly it leads to the integrability criterion for the solution u we mentioned at the beginning of the paragraph. Incidentally, we use here the assumption that f (t) < 0.
Proof. As before, we set u = u E , fi = R E , and w = u 1 . Differentiating Eq. (2.2) with respect to t and multiplying by w t we obtain
w w, + w, + k/3'(u)w, w = 0 which, integrated over (0, R), yields in D,
Jww,dx+
0
rR
f
x x
w, dx+k$'(u)ww,dx=0.
(3.2)
We begin as usual with the first term, so, rR R

w xx w, dx =
J w w , dx + w (0, t)w,(0, t),

x
0 < t < T,
1d
2 dt
^R wx dx + w x (O, t)w 1 (O, t), o
0 < t < T.
The boundary term at x = R does not appear because w = w, = 0 there. When x = 0, we may use Lemma 3.1 and our assumptions about i to conclude
(
I WX(0, t)w1(0, t) I <_ I wx(0, t) I '11(t) I <_ C2
independent of r, 0 < r < e o . Consequently, by (3.2), 2/ dx + dx + k
w f:
rR
0
J J^'uw
dx < C 2 .
(3.3)
292
VIII STEFAN PROBLEM
To estimate the last term, we write
fo , fo At this point we employ Lemma 2.2 in an essential way to show that w > 0 and /3"(u) < 0 imply
,
^ (u) at w Z dx = fo at [fl (u)w 2 ] dx
s (u)w 3 dx.
fo,fl"(u)w3 dx >_ 0.
Therefore
f
ld
R f (u) at w z dx >_ dx. t 10 /3'(u)w2

-
Using this estimate in (3.3) we obtain

2 dt
R[wx + k(3'(u)w 2 ] dx +
R w^
dx _ < CZ .
o
R
For i < t we integrate over (r, t), which provides us with the estimate
rt rR [w x (x, t) 2 + kf3'(u)w(x, t) 2 ] dx + 2 J^ J
rR
w, dx di'
< C 2 (t
T)
+ I [w x (x,
0
i) 2
+ /3'(u)w(x, t ) 2 ] dx
Now this estimate holds for each i < t, so we may integrate it with respect to T over the interval (0, a), a < t, say, to obtain
r R
J 0
wx(x, t) 2 dx + 2Q
J fo, wi dx di
a
1.
< [w (x, o
R
aJ
C3
t
dx di
t) 2 + kf'(u)w(x, t) 2 ] dx
+ 2Q
J JoR w
a
<_ C 3 +
< C3 <
J Lw (x, t) + kw(x, i) ^^ /i(u(x, i))] dx dT r r r f3(u(x, i)) dx dT w (x, i) dx di + Kk + J f aT JJ

a
f 0
T
J
J
rT
o0
'
rT
o
T) 2 dx dt' + Kk Rwx(, f o J
[fJ(u(x, T)) fJ(u(x, 0))] dx
C 3 + wX(x, t) 2 dx dT+ 2RKk.
SMOOTHNESS PROPERTIES OF THE SOLUTION
293
Here we have used the fact that /3' and w are nonnegative and Lemma 2.3. In view of Lemma 3.2, our lemma is proved. Q.E.D.
Theorem 3.4. Let u be the solution to Problem 1.2. Then there is a constant C > 0 such that for each a> 0
u X ,(x, t) 2 dx + J
rt
R
tt
afo u
(x, t) 2 dx di _<
for a < t < T.
Proof. This follows immediately from Lemma 3.3 by the weak convergence of u E, to u, in L". In particular, note that u, e H' (D n {(x, t) : t >_ a}) for each a > 0. Q.E.D.
We now embark on the next phase of our program, the Lipschitz continuity of u, near F.
Lemma 3.5. Let u denote the solution of Problem 1.2 and let F denote the free boundary associated to it. Then for (x o , t o ) e F, there is a neighborhood U of (x o , t o ) such that
ux(x, t) < 0 in U n S2.
Proof. This is elementary. In view of Theorem 2.8, there is a neighborhood U of (x o , t o ) E F such that U c {(x, t) : x > S, t > 0}. Hence f (x) _ k
in U and
u.,z(x,t)=u,(x,t)+k>k>0
in UnQ.
Therefore
u.,(x, t) = u xx (y, t) dy <0
x
in U n S2.
Lemma 3.6. Let F denote the free boundary associated to the solution u of Problem 1.2 and let Q = {(x, t): I x x o I < e, 0 < t o t < S} for a given (x o , t o ) e F and e> 0, S > 0. Suppose that w, O e H' (Q n S2) satisfy
w xx + w t >_ 0 O xx +0, =0
in Q n Q, in QnQ,
w> on
a p ( Q n S2),
where (Q n S2) = 8(Q n K2) {(x, t 0 ): Ix x o I < a} denotes the parabolic boundary of Q n S. Then
w>O
in Qn .
294
VIII STEFAN PROBLEM
The proof of this maximum principle is analogous to our weak maximum principle of Chapter II. We relegate it to an exercise. Lemma 3.7. Let u denote the solution to Problem 1.2 and F its free boundary. Then for each (x o , t o ) E F with t o >_ b > 0, < u j (x, t0) ^ cl(x
x0) 2
c2xu x (x, t)
for (x, t) e Q n Q,
,
where c 1 ,c 2 , and e are positive constants independent of (x o , t o )e I , t o > 6 > 0, and

Q = {(x, t) : Ix x o I < e and It t o l < Q.
Proof. Let us first summarize our knowledge about u t . We know there is a neighborhood Q = {(x, t) : I x x o I < e, I t t o I < e} such that u, e H'(Q) by Theorem 3.4. Since F is a Lipschitz curve, u, admits a trace u,(a(t), t) on F n Q. To see that trace is zero simply note that

u,(x, t) = lim u,(x + , t)

-0
in H'(Q)
and that u, = 0 in Q\S2. Since the trace is a continuous mapping from H I (Q) to L 2 (F n Q), trace r u, = 0. Also note that u, xx + u tt = 0 Let us set
t) =
in Q rn Q.
(3.4)
w(x,
c l (x x 0 ) 2 c 2 xu X(x, t),
(x, t) e Q
n S2,
and choose c l > 0 so large that c l (x
x 0 ) 2 > u t(x, t)
for (x, t) E e(Q n S2).
This is possible, clearly, because u, = 0 on F. Now choose c 2 > 0 so large that

w xx + w, = 2c 1 + c 2 u, in Q n S2 2c 1 + 2kc 2 > 0 in Q n 0. (3.5)
>
According to Lemma 3.5, w(x, t) > 0 in Q n S2, for suitable Q, so

w > u,
on 8 p(Q n 0)
By (3.4), (3.5) we may apply our maximum principle Lemma 3.6, whence
0 < u,(x, t) < w(x, t),
(x, t) e Q n 0.
The conclusion follows. Q.E.D.
3 SMOOTHNESS PROPERTIES OF THE SOLUTION
295
A particular consequence of the lemma is that u, is continuous.

Theorem 3.8. Let u denote the solution to Problem 1.2 and F its free
boundary. Then for each (x o , t o ) e F, there is a neighborhood U of (x 0 , t o ) such
that
u,eH1(U). Proof. It suffices to prove that u, e H' (U n S2). Choose Q as in the previous lemmas so the conclusion of Lemma 3.7 holds, and set 0 = u t . We shall apply the method of S. Bernstein, but to an approximation of O in place of 0 because its smoothness at F is in question (Bernstein [1]). Let a(T) e C O (0) be a mollifier and set
a h() _ (1/h)a(^/h)
Oh(x, t) =
for 0 < h < 1,

t) d^, (x, t) E Q,
ah(x
and
Oh = {(x,t):0<x<u(t)h,0<t<T}.
Observe that
Ohxx +
Ohl
= 0
in '2 h n Q
and
< O h (x, t) < sup O in o h n Q. (3.6)
We now transform the estimate of Lemma 3.7 to an estimate for O hx on F n Q. First note that when (x, t) u Oil, n Q, there is a y with I x y I = h and (y, t) E F. Now by Lemma 3.7 0 < sup O(^, t) < sup O(^, t)
!x - Î<h IY - Î<2h
< sup {c1(Y ly - 9I <2h
^) 2 c2û^(, t)} < c1h,
recalling here that ux (x, t) is a Lipschitz function of x. So we have established that 0 < sup O(^, t) < c l h
Ix-ff <h
for (x, t) e 2S2 h n Q.
(3.7)
296
VIII STEFAN PROBLEM
We compute that
I @hx(x, t) I <
fix-^1 <h
I hx(x - ^) I e(^, t) d
(x, t) E Onh n Q,
< sup t)
Ix-fl<hfix-^J<h
I ahx(x ) I d^,
<_ 1 sup O(^, t)

h Ix
-
lqj!^ CI<hl
I a,,(rl)I drl < c 2 (3.8)
in view of (3.7). Now choose a function 'e C (Q), 0 < C < 1, with = 1 near (x o , t o ) and set
W = Wh = b 2 Ohx + POh, (x, t) E Q n oh.
Applying the heat operator to w we compute

Aw + w, = _2{
2
+ 4 ^^x Ohx hxx + [(4) 4, + i] }
< 2{(1 )^ 2 2 + Onx[,u + (x)x ^t (4 /t)^x]} (3.9) <0
for a p > 0 which depends only on C. Hence by the maximum principle, w(x, t) < max w < c2 + (sup ) 2
av(nhnQ)
keeping in mind (3.8) and (3.6). Restricting ourselves to a neighborhood U of (x o , t o ) where C = 1 and passing to the limit as h ^ 0 we conclude that in S2 (- U. I u, x(x, t) I < const As a particular consequence of this fact, u x is a Lipschitz function of t, so, employing the estimate of Lemma 3.7, we deduce that
0 < u,(x, t) < const Jx x o ) 2 + (t t o ) 2
when (x o , t o ) e F and (x, t) E U n S2. To complete the proof of the theorem, introduce the mollification `P h of u, in the time coordinate only and derive a differential inequality analogous to (3.9) for the function
Z = Zh =
' 2qi 2
+ i^Thx
The details are left to the reader. Q.E.D. We conclude this section with Theorem 3.9. Let u denote the solution to Problem
1.2 and r its free
4 THE LEGENDRE TRANSFORM
297
boundary. Then for each (x o , t o ) e F, there is a neighborhood U of (x o , t o ) such that uxx, u xt E C(S2 n U). Proof. Choose U as in the previous theorem. Since u=u,+k
to t,
for (x, t)EU nQ,
u, x is not only continuous but Lipschitz in U n S2. Differentiating with respect u, u^, E L(U n S2),
l u x ^(x, t) u x ^(x', t)1 < C l x x' I , (x, t), (x', t) e Q n SI, with C independent of t. Moreover, u, , a solution of the heat equation in
U rn S2, is continuous there. Finally, let (x, t) E F n U and (x, t') e U n Q. Given e > 0 we estimate
uxt(x, t) ux,(x, t') <_ I ux,(x, t) ux(x e, t)1
+ 1 uxr(x e, t) u[(x E, t )
'
+ I u, (x e, t') ux,(x, t ) < 2Ce + 1 uxr(x e, t) + u x ,(x e, t') J .

'
Hence
t'-.t
lim I ux1(x, t) uX1(x, t')! < 2Ce,
which proves the theorem. Q.E.D.
4. The Legendre Transform
Applying the Legendre transform to the free boundary value problem for u gives rise to a new problem with a smooth "free" boundary but with a highly nonlinear equation, even in the simplest case. Throughout this section, let u denote the solution to Problem 1.2 and F its free boundary. Given (x o , t o ) e F, let U be a neighborhood of (x o , t o ) such that u,x e C'(S2 n U). Since r is a Lipschitz curve it is easy to check that u, u,,, e C(i n U) imply u,, e C'(11 n U) and indeed that u x admits a C' extension into the neighborhood U.
298
VIII STEFAN PROBLEM
Now we introduce the transformation u x(x, t),

i=t, (x, t)E U. (4.1)
This mapping is C' with

i)
a(^,
_ _u xx(x t) Uxt(x, t)
(
8(x, t)
(4.2)
Since u xx(x o , t o ) = k > 0, i)/8(x, t) is nonsingular at (x o , t o ) and maps a neighborhood, say U n S2, 1: 1 onto a region
Gc{(^,t): >
0}.
Recall here that u x <0 in 1 near F. Also, U n F is mapped onto a subset

E c {(, T) : c = 0},
inasmuch as ux = 0 on F. The Legendre transform of u is

v(, t) = Xt +
U(X,
t),
(^, T) E G u 1, (x, t) E n n U.
It enjoys the property that dv=xd + dx+u x dx+u,dt = x d +u,di, or v 4 = x and vz = u^. In particular,
V
ax 1 1
= a^ _ (m /ax) = uxx
on the basis of (4.2). Our free boundary problem is
in f n U, u, = k u =u x =0 on I nU.
,
(4.3)
In terms of v, this boundary value problem becomes
(1/v) + v = k
in G,
on E
(4.4)
v=0
COMMENTS AND BIBLIOGRAPHICAL NOTES
299
The new problem (4.4) is parabolic because its linearized operator L evaluated at (^, i) is
L^= d
th: L(v + t; ) 4 ^
(/v44 2 )
+(v+rt),
+ Ct,
a parabolic operator with continuous coefficients. It follows from the theory of nonlinear parabolic equations (Ladyzhenskaya et al. [1]) that veC(GuE) so F n U: x = v 4(0, 2), I i t o I small, is a C parametrization of a portion of F. We have shown Theorem 4.1. Let u be the solution of Problem 1.2 and F its free boundary. Then F is a C curve. Note that in general IF cannot be analytic.
COMMENTS AND BIBLIOGRAPHICAL NOTES
The classical Stefan problem, for arbitrary dimension, is discussed in Friedman [3]. The treatment of the one phase problem as a variational inequality was proposed in Duvaut [1]. It was further developed and the properties of its solution explored in Friedman and Kinderlehrer [1]. For a general discussion of the free boundary we refer to Kinderlehrer and Nirenberg [2, 3]. An interesting proof of the infinite differentiability of F (N = 1) is due to Schaeffer [1]. Also in this case, it was shown that IF is analytic provided the supplied heat is analytic (Friedman [4]). Recently, Caffarelli and Friedman [2] have shown that the temperature in the N dimensional one phase problem is continuous. The Stefan problem is a paradigm for many variational and quasivariational inequalities of parabolic type. These concern, for example, stopping times, optimal control, and impulse control. We refer to Bensoussan and Lions [1] and Friedman [5] as examples.

8

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

8

Uploaded by

Copyright:

Available Formats

Downloaded 11/22/13 to 190.144.171.70. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.

A One Phase Stefan Problem

in {(x, t) : s(t) <t < T}, for (x, t) e I',

s(x) <_ t < T, s o < x < R, 0<t<s(x), so <x_<R (1.2)

(x, i) di s'(x)O x(x, s(x))

VIII STEFAN PROBLEM

In the same manner, u(x, t) = u,(x, t) h(x), So let us define

0 < t < T, 0 <x<s o .

1' (t) = f g(r) dT,

( u.S+u,)(v u) f(v u)=0

( u,, +u,)(v u) f(v u)=kv >_0

Q.C. for v E K, for for for

EXISTENCE AND UNIQUENESS OF THE SOLUTION

2. Existence and Uniqueness of the Solution

fl (t) < 0, oo <t < E

u,x +u,+kf e(u) =fE in D, u =erf, t=0, 0<x<R, u=i+ u=0,

0<t<T, x=0, 0<t<T, x=R,

where k> 0 is the constant of Problem 1.1 and i is defined by (1.4).

0< Vit`(x,t)<K where K > 0 is independent oft, 0 < r < e.

VIII STEFAN PROBLEM

w=f, +E ^?xx w =tii',

Since #,(u,) > 0 in D we infer from the maximum principle that

min w, 0) < w(x, t) < max(max w, 0 in D, (2.5) D ??n

where 3D denotes the parabolic boundary of

aD {( x, T): 0 < x < R }.

for 0 <_ x <_ 3Sp,

w(x, 0) ? f (x) + 6q..(x), = h(x) + eij (x) > 0,

0 < x < 3s 0 0 < x < as o

for s o < x < R.

for 0 < x < R.

for 0 <t <T

and w(R, t) = 0 Consequently, min(min a' w, 0) = 0. for 0 < t < T.

EXISTENCE AND UNIQUENESS OF THE SOLUTION

I J + Erl xx kfE( u E) I < sup I E I + e sup I nxx I + k < K 1 ,

II f E(UE) Il c-(n) < 1

L(D) in D. 1 <p < x,

Let u E denote the solution to Problem 2.1. Then as e -+ 0 weakly in H 1 "P(D),

weakly in H 2 "(0, R), 1 <p < oo,

IIUEJ1L -cn) = I!f u E, kY( u C)II Lc (D) 0<e<eo. <IIfJII +K +k <Cl,

VIII STEFAN PROBLEM

Integrating with respect to x e [0, R] and observing that

Taking E -^ 0, owing to the weak convergence we get

uXX + u,)(v - u) dx >-

2 EXISTENCE AND UNIQUENESS OF THE SOLUTION

[u x (v u)x + (v u)] dx dt >

#0ju il) x + (u )] dx dt >_ ff f (u 6) dx dt, A A

Writing (2.8) in terms of , we see that

Hence C = 0 a.e. in D, i.e., meas A = 0. Q.E.D.

VIII STEFAN PROBLEM

NM = {x : Ix) < M(t + 1) 1 / 2 ),

b(x,t)= x -M(t+ l) 112 ,

k _ _ f e (MZ t ^i4 < - (n + 2M 2 M

1)e (M2- " 2 ^' < o0

is bounded as M -^ oo. Hence we may find M so large,that

h(x) >_ h(x)

EXISTENCE AND UNIQUENESS OF THE SOLUTION

Now we apply Lemma 2.6 to conclude

Since u(x, t) = 0 for x >_ M(t + 1) 1 / 2 ,

S2(t)c{x:0<x< M(t +1) 1 / 2 ),

u(x, t) = k u,(x, t) < 0

so, by the maximum principle,

max(u(xl, t), u(x2, t)) = 0,

This is a contradiction; hence, S2(t) is connected. We define