O ct.

2 2

Heat Equation M aximum principles In this lecture we will discuss the maximum principles and uniqueness of solution for the heat equations. 1 . Maximum principles. The heat equation also enjoys maximum principles as the Laplace equation, but the details are slightly dierent. Recall that the domain under consideration is T [ 0 , T) ; { 0 } ( [ 0 , T] ) . T (1)

We have the following strong maximum principle. T solves Theorem 1 . ( Maximum principles of the heat equation) Assume u C12 ( T ) C in T. i. ( Weak maximum principle) Then max u = max u.
T T

ut

u= 0

( 2)

( 3)

ii. ( Strong maximum principle) Furthermore, if is connected and there exists a point ( x 0 , t 0 ) T such that u ( x 0 , t 0 ) = max u , t0 . 1 then u is constant in
T

( 4)

Remark 2. Intuitively, the maximum principles can be explained by the following observation. Recall that u( x , t) =
Rn

( x y , t s ) u( x , s ) d y

( 5)

for any s < t , and the nonnegative function satises

Rn

( x y, t s) dy = 1 ,

( 6)

it is clear that u ( x , t ) at time t is obtained from averaging u ( y , s ) at an earlier time s , and therefore the maximum should be decreasing. We present two proofs of this theorem. The rst is through application of the maximum principle of the Laplace equation, the second is through establishing a mean value property. Note that the rst method only yields the weak maximum principle, that is the maximum inside is bounded by that on the boundary, instead of the strong maximum principle, that is the maximum can only be attained at the boundary, unless the function is a constant. Proof 1 Maximum principles of the Laplace equation. Theorem 3. ( Weak maximum principle) Let Rn be open and bounded. Let u C12 ( T ) T , and satisfy C0 in T, we then have ut
T

( 7) ( 8)

sup u = sup u.
T

t ca n no t be re pla ced b y T, a s ca n be see n b y the fo l lo wing co n side ra tio n. Le t u so lve the hea t e q ua tio n 1 . No te tha t 0 with initia l va lue M (co nsta nt) a n d (side ) bo unda ry va lue g ( x ) M. No w if u t 0 a s t 0 , we ca n e xte nd u to t < 0 b y se tting u M.

(If T < , the supreme can be replaced b y maximum). Remark 4. The connection to the theory of Laplace equation can be seen by writing the inequality into u ut. ( 9)

T T for all T < T . Fix one such T . By compactness of Proof. If suces to prove the conclusion for we know that the maximum is attained at some point ( x 0 , t 0 ) . 1 . First consider the case u > ut Now if ( x 0 , t 0 )

in T .

( 1 0)

T , at ( x 0 , t 0 ) , we have 0 because otherwise u ( x 0 , t 0 ) > u ( x 0 , t 0 ) for small enough; 0 because x 0 maximizes u on { t 0 } .

a) u t ( t 0 ) b)

u( x0 , t0 )

But the two conclusions contradict each other. 2 2. For the general case u we consider an auxiliary function We easily check and can apply 1 . Finally letting

ut ,

(11) ( 1 2) ( 1 3)

v ( x , t ) = u ( x , t ) t. v vt

0 we obtain the conclusion.

From weak maximum principle one immediately obtains the uniqueness for nonhomogeneous initial/ boundary-value heat equation when the domain is bounded. Theorem 5. ( Weak maximum principle for = Rn ) Suppose ut u 0 in T ; u( x , t) M e | x | in T for M , > 0; sup u
T
2

u ( x , 0) = g ( x ) ,

( 1 4) ( 1 5)

then

sup g.
Rn

Remark 6. The reason why we need a bound on the growth rate at innity is the following. Recall that the eect of the heat operator is to obtain solutions at later times from averaging solutions at earlier 2 2 times via a kernel of the form e | x | / 4 . Note that the kernel decays at innity as e c | x | . Therefore, if the 2 solution grows faster than its inverse rate e | x | , then it is possible to input extra stu from innity and thus violates uniqueness. Here is an example of how one can store energy at and then input it through averaging by the heat kernel, and thus obtain a nonzero solution with zero initial data. Set g( n ) ( t ) 2 n x ( 1 6) u( x , t) ( 2 n) ! 0 with g( t) e 0

1 tk

t > 0, k > 1 . t = 0.

( 1 7)

2 . T he reason we use T < T here is to conclude u t t = T is not known.

0 at the maximizer. S ince the equation is solved on T only, u t at

One can verify that u satises the equation by formally dierentiating inside the summation. With some calculation 3 , one can further show that the series on the RHS is majorized by the Taylor expansion of 1 x2 1 1 k exp t t> 0 . ( 1 8) U ( x , t) = t 2 0 t 0 for some > 0 . Note that U ( x , t ) cannot be bounded by M e x for any constants M , .
1 . 4
2

Proof. First we divide ( 0 , T ) into subintervals with size r < 1 interval. From now on we assume T < 4 . Consider the auxiliary function

It suces to prove the claim on a sub-

where is chosen such that T + < It is easy to verify that The strategy is the show rst

1 v ( x , t) u( x , t) e ( 4 ( T + t) ) n/ 2
1 . 4

| x| 2 4(T+ t)

( 1 9)

vt v ( x , t)

0. sup g
Rn

( 20) ( 21 )

for any ( x , t ) and then letting 0 . We rst notice that, on the sphere | x y | = R , we have v ( x , t) Me
( | x | + R) 2

1 e ( 4 ( T + t) ) n/ 2

R2 4( T + t)

sup g
Rn

( 22)

once R is big enough. Now apply the weak maximum principle on the domain BR ( 0 , T ) , we obtain v ( x , t ) sup g . Proof 2 Mean value property. The mean value property for the heat equation turns out to be much more complicated than the one for the Laplace equation. We rst dene the heat ball: Denition 7. ( Heat ball) For xed x Rn , t R , r > 0 , we dene E ( x , t; r) ( y , s ) Rn + 1 s

t, ( x y , t s )

1 . rn

( 23)

Remark 8. We try to gain some idea of what E ( x , t ; r ) looks like. It is dened by

| x y| 4(t s ) 1 e ( 4 ( t s ) ) n/ 2 2

1 . rn

( 24)

Thus in the time direction, we have t where the lower bound is from the fact that e s

t
| x y| 2 4(t s )

r2 4 1 . Furthermore

( 25)

E( x , t; r) { s = 0 } = E( x , t; r) { s = t } = { x } .

( 26)

3 . S ee pp. 2 1 1 2 1 3 in F . J ohn Part ial D ierent ial E quat ions , 4ed. T he constant is determined in P roblem 3 on p. 73 of the same book.

Next, to nd out the correct formula, we need to nd a kernel K ( x y , t s ) such that

E ( x , t; r )

is independent of r . Notice that

K( x y , t s ) d y ds
x y t s 2

( 27)

where Thus we have This implies when

| x y| 1 e 4(t s ) n/ 2 ( 4 ( t s) )

1 rn

1 e 4 n/ 2 (4 (t s ))

( 28) ( 29) ( 30) ( 31 ) ( 32)

t = t / r 2 , s = s / r 2 , x = x / r , y = y / r. K( x y , t s ) rn + 2 d y ds .

E ( x , t; r)

K( x y , t s ) d y ds =

E( x , t ; 1 )

K( x , t ) = K( x , t) rn + 2 x = x/ r , One can somehow 4 verify that

5

t = t/ r2 . | x| dx dt = 1 . t2
2

1 4

( 36)

E( 0, 0; 1 )

Thus nally we have the correct formulation. Theorem 9. ( Mean value property for the heat equation) Let u C12 ( T ) solve the heat equation, then 2 | x y| 1 u y , s u( x , t) = d y d s. ( 37) ( ) 4 rn ( t s) 2 E ( x , t; r ) for each E ( x , t ; r ) T. Proof. Without loss of generality we can set ( x , t ) = ( 0 , 0) and denote E ( 0 , 0; r ) by Er . Dene ( r) 1 rn u ( y, s )
Er

| y| d y ds = s2

u r y , r2 s
E1

|y| d y ds . |s |2

( 38)

In the following we will omit the prime and simply use y and y .
4. I have no idea how to come up with this particular function. 5 . A s the computation is tedious and long, we put it in as a footnote and furthermore only present the main steps. F irst notice E ( 0 , 0; 1 ) T hus the integral becomes
1 /4 0
| x| 1 e 4t n/ 2 ( 4 t) 2

1 ; | x| 2 4

( 2 n t ) log

1 . 4t

( 3 3)

t 2

| x|

2 n t log 4t

| x | 2 d x dt =

n n / 2 2 ( n + 2 ) / 2 n ( n + 2 ) / 2 n n+2 +1
2 n/ 2 n +1 2

1 4

n 2 n

log

1 4t

n 2

+1

d t.

( 3 4)

where we have used p olar coordinates and the formula ( n ) = s = 4 t and using the formulas z ( z ) =
1 0

for the volume of n -dimension balls. N ow setting

t 1

log

1 t

z 1

dt ;

( z + 1 ) = z ( z ) ,

( 35)

we will see after careful calculations that most terms cancel out and what remains is 4 .

thus we compute ( r) = =
i= 1

d dr
n

u r y , r2 s
E1 2

( yi u ) yi 1
n Er i = 1

| y| | y| 2 + 2 r ( s u ) s d y d s | s| 1 | y| 2 + rn + 1 | s|
2

| y| d y ds | s| 2

rn + 1 A + B. Now we introduce and note that = 0 on Er . Now we have B =

( y i u ) yi

2 ( s u )
Er

| y| s

( 39)

2 n | y| log ( 4 s ) + + n log r , s 2

( 40)

1 rn + 1 1 rn + 1
Er

4 ( s u )
i= 1

yi ( y i ) d y d s
n

4 n us + 4
Er i= 1

u s yi yi d y d s.

( 41 )

Now integrate by parts w. r. t. s , we have B = = = 1 rn + 1 1 rn + 1 1 rn + 1

Er Er n

4 n us + 4 4 n us + 4 4 n us

u yi yi s d y d s
i= 1

u y i yi

n | y| 2 s 4 s2

d y ds ( 42)

Er

2n s

u yi yi d y d s A.

Consequently, using the equation, we have ( r) = A + B 1 = n+ 1 r = = 0. Thus we have ( r ) = lim ( t ) = u ( 0 , 0)

t 0 t 0

Er

4n
Er

2n s

u y i yi d y d s n u y yi d y d s 2s i ( 43)

1 rn + 1

4 n u yi yi

lim

1 tn

Et

| y| d y ds s2

= 4 u ( 0 , 0) .

( 44)

From the mean value property we immediately have the strong maximum principle: Proof. ( of the strong maximum principle) Suppose there is ( x 0 , t 0 ) T such that u ( x 0 , t 0 ) = max T u , then by picking r small enough so that E ( x 0 , t 0 ; r ) T , and using the mean value property, we conclude that u is constant inside E ( x 0 , t 0 ; r ) . Next for any ( y0 , s 0 ) T such that the line segment connecting x 0 , y0 is in T , we can show that u ( y0 , s 0 ) = u ( x 0 , t 0 ) whenever s 0 < t 0 by covering the line segment connecting ( y0 , s 0 ) and ( x 0 , t 0 ) with the heat balls. Finally, since is connected, any y0 can be connected from x 0 via nitely many line segments. And therefore u ( y, s ) = u ( x 0 , t 0 ) for all y , s < t 0 .

Remark 1 0. It turns out there is also a version of mean value property involving only the surface integral over E ( x , t ; r ) . The idea is to compute the integral of ( x y , T t ) r n ( u t u ) over E ( x , t ; r ) via integration by parts. This will give a representation formula for u ( x , t ) , which turns out to be u( x , t) = 1 rn u ( y, s )
E ( x , t ; r )

| x y| d S. 2 ( t s)

( 45)

See J. Jost Partial Dierential Equations , pp. 81 83 for details. The strong maximum principle can also be proved using this formula. See Theorem 4. 1 . 3 on p. 86 of J. Jost Partial Dierential Equations . 2. Uniqueness. As in the elliptic case, uniqueness is established through maximum principles. Theorem 1 1 . ( Uniqueness on bounded domains) Let g C ( T ) , f C ( T ) . Then there exists at T of the initial/boundary value prob lem most one solution u C12 ( T ) C ut u= f in T ; u= g on T . ( 46)

In the case = Rn , we have to add extra condition. Theorem 1 2. ( Uniqueness when = Rn ) Suppose g C ( Rn ) , f C ( Rn [ 0 , T ] ) . Then there exists at most one solution u C12 ( Rn ( 0 , T ] ) C ( Rn [ 0 , T ] ) of the initial value problem satisfying the growth estimate for constants M , > 0 . ut u = f in Rn ( 0 , T ) ; | u( x , t) | u = g on Rn { t = 0 } M e | x |
2

( 47) ( 48)

Remark 1 3. There are in fact innitely many physically incorrect solutions, which do not satisfy the growth bound, to the zero initial value problem. See Chapter 7 of F. John Partial Dierential Equations . Exercises.
2 T be a solution to E xercise 1 . L et u C1 ( T) C

ut where g

u= 0

in T ;

u= 0

on [ 0 , T ] ;

u= g

on { t = 0 } ,

( 49)

0 . T hen u is p ositive everywhere within T if g is p ositive somewhere on .