Lecture 2 Laplace and heat equations

invariance
mean value equality
maximum principle,
(higher order) derivative estimates and smoothing e¤ect
Harnack inequality
Liouville
strong maximum principle for general elliptic and parabolic equations

Laplace equation 4u = 0
complex analysis in even d: u = Re z k ; z k ; ez ; z13 ez2 ;
algebraic n-d u = k (x1 ; ; x2 )
radial
n 1 1
4u = @r2 u + @ r u + 2 4S n 1 u
r r
n 1
urr + r ur = 0
rn 1 urr + (n 1) rn 2 ur = 0 or (rn 1 ur )r = 0
ur = rnc 1
c
u= ; ln j(x1 ; x2 )j ; or jx1 j
rn 2
Comparison principle
two solutions cannot touch each other

f igure

4u = 0
D2 u1 > D2 u2 ) 0 = 4u1 4u2 > 0 !

two solutions can cross each other

f igure

In contrast to wtt wxx = 0; w1 = x2 + t2 ; w2 = 0:

Invariance for Harmonic functions, solutions to 4u = 0
u (x + x0 )
u (Rx)
u (tx)
RMK. Equations don’t know/care which coordinates they are in.
Ru + v; au; where 4v = 0
u (x y) ' (y) dy
0
November 2, 2014

1
 u(x+"e) u(x)
"
! De u; so is Dk u
u(R"x) u(x)
"
! D u = xi uj xj ui
u((1+")x) u(x)
"
! Du (x) x = rur ; so are r@r (rur ) = rur + r2 urr ; r3 urrr ;
2 n
jxj u jxjx2 Kelvin transformation
jxj2
1 x
RNK. “Kelvin”transformation for the heat equation ut 4u = 0; tn=2 e 4t u t
; t1 :
Mean value equality
Recall the divergence formula (the fundamental theorem of calculus)
Z Z D E
div V~ dx = V~ ; dA:
@
R
V~ = Du; then 0 = @ u dA:
R R
V~ = vDu; then hDv; Dui + v 4 u = @ vu dA:
R R
V~ = uDv; then hDu; Dvi + u 4 v = @ uv dA:
Z Z
v4u u4v = vu uv dA:
@

Mean value case. Now 4u = 0 in B1 ; take v = jxj2 n

; = B1 nB" ;
B1 nB" f igure

R
we then have 0 = @
vu uv dA; or
0
z
Z }| { Z Z Z
(2 n) (2 n)
vu dA = uv dA = u n 1 dA u n 1 dA: (*)
@(B1 nB" ) @(B1 nB" ) @B1 r @B" r
Z Z
1 "!0 R
We get udA = u n 1 dA ! j@B1 j u (0) : So u (0) = j@B1 1 j @B1 udA:
@B1 @B" "
1 1 def
RMK. In hindsight one just takes v = = :
(n 2) j@B1 j jxjn 2
Also Z
1
u (0) = udA:
j@Br j @Br
R1 R1R
Take a weight function j@Br j ; u (0) jB1 j = 0 u (0) j@Br j dr = 0 @Br udAdr =
R R
B1
udx: So u (0) = jB11 j B1 (0) udx:
Also Z
1
u (0) = udx:
jBr j Br (0)

2
 RMK1. Tracing the signRof 4u; one gets mean value inequalities for superharmonic
R
functions 4u 0 : u (0) u and subharmonic functions 4u 0 : u (0) u:
RMK2. “ all the women are strong, all the men are good-looking, and all the
children are above average.”–A Prairie Home Companion with Garrison Keillor.
Application 1. Strong maximum principle (No toughing).

4u1 = 4u2 = 0
u1 u2 ; \ =00 at 0

then Z
1
0 = u1 (0) u2 (0) = (u1 u2 ) dx 0:
jBr j Br

It follows that u1 u2 :
Application 2. Smooth e¤ect and derivative estimates. R R1
Take radial weight ' (y) = ' (jyj) 2 C01 (Rn ) such that 1 = ' (y) dy = 0 ' (r) j@Br j dr:
Then
Z Z 1Z
u (y) ' (x y) dy = u (y) ' (x y) dAdr
Rn 0 @Br (x)
Z 1 Z
= u (x) ' (r) j@Br j dr = u (x) ' (y) dy
0
= u (x) :
R
Consequence u (x) = Rn u (y) ' (x y) dy is smooth for continuous initial u (y) ; and
Z Z
k k k
D u (0) = u (y) Dx ' (x y) dy = ( 1) u (y) Dyk ' (x y) dy:

Thus, (say support of ' is in B1 )

Dk u (0) C (k; n; ') kukL1 (B1 )

and also the point-to-point version for u 0

Z
Dk u (0) C (k; n; ') udx = C (k; n; ') u (0)
B1

Scaled version ( C(k;n;')kukL1 (B )

R
k Rn+k
D u (0) C(k;n;')kukL1 (B )
R
Rk
and also
C (k; n; ')
Dk u (0) u (0) provided u 0:
Rk
That is the larger the domain, the ‡atter the harmonic graph.
Liouville. Every bounded or even one side bounded entire harmonic function
in Rn is constant. Similarly every entire polynomial growth harmonic function is a
polynomial.

3
 Application 3. Harnack inequality–a quantitative version of the strong maxi-
mum principle.
eg. Consider positive harmonic functions r2 n ; x1 r n on fx1 > 0g :

r2 n
; x1 r n
graph f igure

eg. In general for 4 u = 0; u > 0 in B1 (0) ; we have

Z Z
1 1 jB1 j 1
u (x) = udx udx = u (0) = u (0) :
B1 jxj B1 jxj (x) B1 jxj B1 (0) B1 jxj (1 jxj)n

RNK. As those two examples suggest, from estimating the kernel of Poisson repre-
sentation, we have a sharper comparison

(1 jxj) 2
u (0) u (x) 1 u (0) :
2n 1 (1 jxj)n

Harnack. Suppose 4 u = 0; u > 0 in Br (x0 ) : Then we have

sup u 3n inf u:
Br=4 (x0 ) Br=4 (x0 )

Discrete way. In fact

4 circle …gure B1 ; B1=4 ; B1=4 (xmax ) ; B3=4 (xmin )

Z
1
max u = u (xmax ) = udx
B1=4 B1=4 B1=4 (xmax )
Z
1
udx
B1=4 B3=4 (xmin )
= 3n u (xmin ) = 3n min u:
B1=4

Continuous way. Suppose 4 u = 0; u > 0 in B2 (0) : Then we have

max u C (n) min u:

B1 (0) B1 (0)

4
Indeed by the point-to-point version of gradient estimate, jDu (x)j =u (x) C (n) in
B1 (0) : We measure the ratio between umax and umin in B1 (0) by integration
Z xmax Z xmax
u (xmax ) ue
log = d log u (x) = (xmin + t (xmax xmin )) dt
u (xmin ) xmin xmin u
Z xmax
jDuj
(xmin + t (xmax xmin )) dt jxmax xmin j C (n) :
xmin u

Then
max u e2 C(n) min u:
B1 (0) B1 (0)

Consequences ; for example one sided Liouville for entire harmonic functions.
RMK. Harnack inequality is in fact a quantitative version of the strong maximum
principle. It measures how much the minimum leaps when moving inside, or after
‡ipping around, how much the maximum drops when moving inside. For example,
to move inside B1=4 from B1 ;
n
min (u m1 ) 3 max (u m1 )
B1=4 B1=4

or
n
m1=4 m1 + 3 M1=4 m1 :
The ‡ip version is
n
min (M1 u) 3 max (M1 u)
B1=4 B1=4

or
n
M1 M1=4 + 3 M1 m1=4 :
(This should be Moser’s observation: subtracting the leap from the drop, one has
oscillation decay of the “harmonic”function.)

Co-area way for solid mean value equality.

Recall the derivation of the “solid" mean value formula for harmonic functions
Z Z
u4v v4u= uv vu
BR @BR

1 1 1
Set v = jxjn 2 Rn 2
(n 2) j@B1 j
| {z }
cn

Z Z
u (0) = uv dA = u jD j dA
1
@BR = =l
cn Rn 2

1 1
where = cn jxjn 2 :

5
 As R goes from 1 to 0;the level of runs from 1=cn to 1; we seek a weight
R 1=c
satisfying 1 = 1 n w (l) dl; where w (l) = c ( l) to be determined.
Z 1=cn Z
u (0) = w (l) u jD j dAdl
1 =l
Z
= u w (l) jD j2 dx;
B1

dl
where we used change of variable or co-area formula: dx = dvol = dA jD j

…gure gradient length=dl/ jD j

2
Now let w (l) jD j = 1= jB1 j ; then
2
1 cn
w (l) = Rn 1
:
jB1 j (n 2)
At R = 1 and l = 1=cn ; we have
2
1 cn
c (1=cn ) = :
jB1 j (n 2)
The integral for the weight implies
1+
1 1
c = 1:
1+ cn
cn
Thus (1 + ) = jB1 j(n 2)2
= n= (n 2) ; then

n n=(n 2)
2(n 1)
w (l) = (cn ) ( l) n 2 :
n 2
R1 r=( cn l) 1=(n 2)
n
n R 1=cn 2(n 1)
RMK. Old weight way, 1 = 0 nrn 1 dr = n
cn
2 n
2
1
( l) n 2 dl:
And pleasantly w (l) jD j2 = 1= jB1 j!

As the curved Laplace operator is variable coe¢ cients

1 p ij 1 p ij
4g = p @i gg @j = g ij @ij + p @i gg @j ;
g g

we need to deal with general elliptic operator

X X
L= aij (x) @ij + bi (x) @i

where elliptic means

(aij (x)) I > 0:

6
Weak Maximum Principle. Assume u; v 2 C 2 ( ) \ C ;
Max. If Lu 0 in , then max u = max@ u:
Min. If Lu 0 in , then min u = min@ u:
Comparison. If Lu Lv in and u v on @ ; then u v in :
Proof. It su¢ ces to handle the max case, the other cases follow by applying the
…rst case to u and u v respectively.
Strict case. Lu > 0:
Suppose u attains its maximum at an interior point x ; then Du (x ) = 0 and
2
D u (x ) 0: But
X X
Lu (x ) = aij (x ) Dij u (x ) + bi (x ) Di u (x )
= T r aij (x ) D2 u (x8 ) 0:

This contradicts Lu (x ) > 0: Thus max u = max@ u:

General case. Lu 0:
We do analysis by approximating it with strict inequalities. Let

u" = u + "eKx1 :

Then
Lu" = Lu + "eKx1 a11 K 2 + b1 K "eKx1 K2 kbkL1 K > 0
provided constant K is large enough. By the strict case

max u + "eKx1 = max u + "eKx1 :

@

Take " ! 0; we …nd max u = max@ u:

Weak Maximum Principle with c 0: Assume u; v 2 C 2 ( ) \ C and

c (x) 0 in
Max. If Lu + c (x) u 0 in , then max u max@ u+ :
Min. If Lu + c (x) u 0 in , then min u min@ u :
Comparison. If Lu + c (x) u Lv + c (x) v in and u v on @ ; then u v
in :
Proof. Again it su¢ ces to handle the …rst case. We just ignore the negative piece
of u; and reduces it to the weak maximum principle without zeroth order term. Let
+
= fx 2 : u (x) > 0g :

We have
+
Lu c (x) u 0 in :
By the maximum principle without zeroth order term

max u = max u = max u+ :

+ @ + @

It follows that
max u max
+
u = max u+ :
@

7
C-eg with c > 0:

u = cos x
u00 + u = 0 in = ( =2; =2)
max u = 1 0 = max u:
@

C-eg for necessity of the positive part in max@ u+ :

u = chx
00
u u = 0 in = ( 1; 1)
max u = 1 ch1 = max u:
@

C-eg for necessity of the negative part in min@ u :

u = chx
00
u u = 0 in = ( 1; 1)
min u = 1 ch1 = min u:
@

Strong Maximum Principle (E. Hopf). Assume u 2 C 2 ( ) \ C and is

open and connected.
Max. If Lu 0 in and u attains its global maximum at an interior point,
then u is constant.
Min. If Lu 0 in and u attains its global minimum at an interior point, then
u is constant.
It su¢ ces to prove only one case, say minimum one; the maximum case follows by
considering u: The inconsistency: if “strict”minimum occurs, then gradient at the
point is not zero. Hopf achieved it by constructing a barrier.
eg. Consider two positive harmonic functions

u = xy in fx > 0; y > 0g ;
n p o
u = Im z 3 = r3 sin (3 ) = 3yx2 y 3 in y > 0; y 3x :
Observe the inner normal derivatives of the two along the boundary except the corner
points are all strictly positive. This is not an accident, indeed it is the content of
the Hopf boundary lemma: For every positive “harmonic” function vanishing at a
boundary point, on which an interior ball touches, the inner normal derivative is the
strictly positive at this boundary point satisfying the interior ball condition.
Proof. Step1. Set-up.
Suppose u attains its global minimum m = min u at an interior point. Set

m = fx 2 : u (x) = mg :

If u (x) is not constant m; then cm = fx 2 : u (x) > mg is non-empty. Next we

choose a ball touching the boundary of m and inside : There exists an interior

8
point y 2 cm such that = dist (y; @ m ) < dist (y; @ ) : Enlarge the ball centered
at y until it hits the boundary @ m ; say at x0 ; before hitting @ : The largest radius
is :
…gure
For simplicity of notation, we assume y = 0: What we have now is

B 0 (0) c
m and x0 2 @B (0) \ m:

Step2. Hopf boundary lemma.

We construct a subsolution v such that u v near x0 and vr (x0 ) < 0: This
=at x0
forces ur (x0 ) < 0 or Du (x0 ) 6= 0; a contradiction to interior minimum u (x0 ) at x0 :
First take
2 2
w (x) = e Kjxj e K
with large constant K to be determined later. We compute
h
2 X X i
Lw = e Kjxj aij 4K 2 xi xj 2K ij + bi ( 2Kxi )
> 0 in B (0) nB =2 (0)

provided K is taken large enough.

RMK. Another barrier is jxj K :
Notice u > m in B 0 (0) : Then we can take small enough " so that v (x) = "w (x)+
u (x0 ) stays below u (x) on the boundary of the annulus

u (x) v (x) on @B =2 (0) [ @B (0) :

By the construction
Lu 0 Lv in B (0) nB =2 (0) :
From the comparison principle (weak maximum principle)

u (x) v (x) in B (0) nB =2 (0) :

Remember u (x0 ) = v (x0 ) ; taking radial derivatives, we get

K 2
ur (x0 ) vr (x0 ) = "K e > 0:

This contradicts the fact that Du (x0 ) = 0 at minimum interior point x0 :

Therefore, u (x) m:
Strong maximum principle with c 0: Assume u 2 C 2 ( ) \ C :
Max. If Lu + cu 0 in and u attains its nonnegative global maximum at an
interior point, then u is constant.
Min. If Lu + cu 0 in and u attains its nonpositive global minimum at an
interior point, then u is constant.
Proof. Again it su¢ ces to justify only one case, say the minimum one.
Now
m = min u 0:

9
We go over the same two steps as in proof of the strong maximum principle without
2 2
c term. Then for v = " e Kjxj e K +m

Lv + cv
hX X i
Kjxj2
=e aij 4K 2 xi xj 2K ij + bi ( 2Kxi ) + c
+ cm
hX X i
Kjxj2
e aij 4K 2 xi xj 2K ij + bi ( 2Kxi ) + c
> 0 in B (0) nB =2 (0)

provided K is taken large enough. Then

Lu + cu 0 Lv + cv in B (0) nB =2 (0) ;
u (x) v (x) on @B =2 (0) [ @B (0) :

The remaining argument goes through.

Strong maximum principle with no sign restriction on c but with one sign
restriction on solution. Assume u 2 C 2 ( ) \ C and the arbitrary sign of c (x) in
:
Max. If Lu + cu 0 in and max u = 0 attains at an interior point, then
u 0:
Min. If Lu + cu 0 in and min u = 0 attains at an interior point, then u 0:
Proof. Again, we only need to handle one case, say the minimum case. Note that
u 0; then splitting c = c+ c ; we have

Lu c u c+ u 0:

We are back to the above strong maximum principle with c 0:

C-eg to c 0:

u = cos x
00
u + u = 0 in ( =2; =2)
u (0) is an interior global maximum.

C-eg to nonpositive minimum.

u = chx
00
u u=0
positive u (0) = 1 is an interior global minimum.

Heat equation ut 4u = 0:
Invariance for caloric functions, solutions to ut 4u = 0

10
 u (x + x0 ; t + t0 )
u (Rx; t)
u (sx; s2 t)
RMK. Equations don’t know/care which coordinates they are in.
Ru + v; au; where vt 4v = 0
u (x y; t s) ' (y; s) dyds
u(x+"e;t) u(x;t)
"
! De u; u(x;t+")" u(x;t) ! Dt u; so is Dxk Dtm u
u(R"x;t) u(x;t)
"
! D u = xi uj xj ui
u((1+")x;(1+")2 t) u(x;t)
"
! Du (x) x+2ut = rur +2ut ; so are (r@r + 2@t ) (rur + 2ut ) ;
jxj2
1
tn=2
e 4t u xt ; t1 Kelvin transformation
Examples.
all harmonic functions are caloric, x1 x2 x3 ; 1= jxjn 2 ; Re = Im e3x1 +4x2 +i5x3 :
t + jxj2 =2n; t2 + t jxj2 =2n + (x41 + + x2n ) =24n:
2 2
Re = Im ei x j j t ; e x+j j t
RMK. e x ; Re = Im ei( x+j jt) ; u ( x + j j t) are wave functions. But ex1 +t are both
caloric and wave functions.
caloric polynomials
jxj2
1
(x; t) = tn=2
e 4t ; then

x 1
(x; t) Dx Dtl ; is a caloric polynomial of degree j j + l:
t t
eg.
1 x1 x2 x3 jxj2
n=2 jxj x1 x2 x3
2
D123 (x; t) = e 4t ! ( t) e 4t
( 2t)3 ( x ; 1 )
tn=2 8
t t

(x;t) x1 x2 x3
! ( 1)n=2
8

1 x31 + 6tx1
jxj2
3
n=2 jxj x1 + 6tx1
2
D111 (x; t) = e 4t ! ( t) e 4t
tn=2 8t3 ( xt ; t1 ) 8
(x;t) x3 + 6tx1
! ( 1)n=2 1 :
8
!
1 jxj2 n=2 jxj2 jxj2 n 1
Dt (x; t) = n=2
e 4t + 2 ! ( t)n=2 e 4t t + jxj2
t t 4t 2 4
( xt ; t1 )
(x;t) n 1
! ( 1)n=2 t + jxj2
2 4
Self-similar way
self similar solution u (x; t) = t
1
v r
p
t
to ut 4u = \ (x; t)00 or 0

11
 1 1 r 0 1 r
ut = t +1 v+ t 2t3=2
v = t +1 v p
2 t
v0
1 n 1
p
4u = t
@rr + r
@r v r= t
1 1 00 n 1p 1 0 1 n p1 0
= t t
v + r t
v = t +1 v 00 + r= t
v
r n 1 0 r 1pr 0 r r
v 00 p
t
+ r
p
v p
t
+ 2 t
v p
t
+ v p
t
=0
t
Set = prt ; then
v 00 ( ) + n 1 v 0 ( ) + 12 v 0 ( ) + v ( ) = 0
0 1 n 0
( n 1v0) + v + n 1v = 0
2
| {z }
0
=( 12 n v ) when =n=2
* Let = n=2
2 0 2
n 1 0 1 n
v + = c or v 0 + 2 v = c= n 1 or ve 4
2
v = ce 4 = n 1

2 2 R 2
Then v ( ) = c20 e 4 + c e 4 e 4 = n 1 d and 3
Z 2
1 6 0 7
jxj2 jxj2
u (x; t) = tn=2 4c e 4t + ce 4t e 4 = n 1d p
5
=r= t
| {z }
After some testing for the fundamental solution, we …nd c = 0:
Mean value equality
Recall the derivation of the “solid" mean value formula for harmonic functions
Z Z
u4v v4u= uv vu
BR @BR

1 1 1
Set v = jxjn 2 Rn 2
(n 2) j@B1 j
| {z }
cn
Z Z
u (0) = uv dA = u jD j dA
1
@BR = =l
cn Rn 2

where = cn1 jxjn1 2 :

As R goes from 0 to 1;the level of runs from 1 to 1=cn ; we seek a weight
R 1=c
satisfying 1 = 1 n w (l) dl; where w (l) = c ( l) to be determined.
Z 1=cn Z
u (0) = w (l) u jD j dAdl
1 =l
Z
= u w (l) jD j2 dx;
B1
dl
where we used change of variable or co-area formula: dx = dvol = dA jD j

…gure gradient “height”=dl/ jD j

12
 Now let w (l) jD j2 = 1= jB1 j ; then
2 1 2
1 cn cn r n 2
=l 1 cn n 1
w (l) = rn 1
= ( cn l) n 2
jB1 j (n 2) jB1 j (n 2)
2 2 2(n 1)
= n (n 2) n 2 j@B1 j n 2 ( l) n 2 :

R1 r=( cn l) 1=(n 2) n n R
1=cn 2(n 1)
RMK. Old weight way, 1 = 0 nrn 1 dr = n 2 n
c n 2
1
( l) n 2 dl:
2
And pleasantly w (l) jD j = 1= jB1 j!
Mean value equality for caloric functions
Z
1 jxj2
u (0; 0) = uq dA:
(4 R2 )n=2 =(4 R2 ) n=2 4t2 jxj2 + 2nt + jxj2
2

…gure: heat sphere

One “solid”version
Z
1 jxj2
u (0; 0) = u dA;
(4 R2 )n=2 (4 R2 ) n=2 4t2

where
1 jx0 xj2
(x0 ; t0 ; x; t) = e 4(t0 t)

[4 (t0 t)]n=2
…gure: heat kernel graphs
Derivation of the hollow version.

Z Z Z
div (uDv) = (div; Dt ) (uDv; 0) = (uDv; 0) ( x ; t ) dA
UT UT @UT
Z Z Z
) div (vDu) = (div; Dt ) (uvDu; 0) = (vDu; 0) ( x ; t ) dA
UT UT @UT
Z Z Z
+) Dt (uv) = (div; Dt ) (0; uv) = (0; uv) ( x ; t ) dA
UT UT @UT

) Z Z
u Dt v + 4v + v Dt u 4u = uv vu + uv t dA:
UT
| {z } | {z } @UT
x x

\ (0;0)00 0

jxj 2
1 1
Take v = [4 ( t)]n=2
e 4t [4 ]n=2
; then
( )
1 jxj2 1 t<0
E= n=2
e 4t
n=2
, jxj2 2nt ln ( t)
[4 ( t)] [4 ]
UT = E \ ft sg :

13
 …gure UT

We have from the Green’s identity

Z Z
0= uv x + uvdA
@E\ft sg Is
| {z }
! u (0; 0)
as s ! 0 ; say for C 0 u

D jD j2
v x =D x =D =
j(D ; Dt )j j(D ; Dt )j
x 2
2
2t jxj2
=r h i = q :
2 2 2 2 2
2 x 2 + 2 n 1 jxj 4t2 jxj + 2nt + jxj
2t 2 t 4t2

Therefore
Z
1 jxj2
u (0; 0) = uq dA
(4 )n=2 =(4 ) n=2
2 2
4t jxj + 2nt + jxj2 2

Z
1 jxj2
u (0; 0) = uq dA
(4 R2 )n=2 =(4 R2 ) n=2
4t2 2
jxj + 2nt + jxj 2 2

Z
jD j2
= u dA
=(4 R2 ) n=2
=l j(D ; Dt )j
Having this sphereR version, let’s have a “solid”mean value formula.
1
Set w (l) s.t. 1 = 1 n=2 w (l) dl;
(4 )
Z 1 Z
jD j2
u (0; 0) = w (l) u dA dl
=l j(D ; Dt )j
n=2
( 41 )
Z
= n=2
uw (l) jD j2 dxdt
( 41 )
Z
2 x 2
= u w (l) dxdt
( 41 )
n=2 | {z } 2t
constant

1 n=2 1 R1 1 n=2 1 1
Let w (l) = l2
; then 1 n=2 w (l) dl = l ( 41 )
n=2 = 1: Thus
4 (4 ) 4

Z
1 jxj2
u (0; 0) = u dxdt
(4 )n=2 ( 41 )
n=2 4t2

14
or Z
1 jx0 xj2
u (x0 ; t0 ) = u dxdt
(4 R2 )n=2 (x0 ;t0 ;x;t) ( 4 1R2 )
n=2
4 (t0 t)2
…gure heat ball

In particular, for u 1
Z Z
1 jxj2 jxj2
1= jxj2 dxdt = dxdt:
(4 )n=2 e 4t 1 4t2 1 4t
2
[4 ( t)]n=2 (4 )n=2 | {z }
say 4 R2 =1

RMK. Other choices of weight

eg. w (l) = c=l2 ; < 1
Z
jxj2
u (0; 0) = c u dxdt;
1 4t2

the kernel is still singular.

Now that we have discovered mean value formulas for harmonic and caloric func-
tions, we could provide a derivative way to verify those formulas (to impress others).
Veri…cation of solutions to 4u = 0 satisfying
Z
1
udx = constant = u (0) jB1 j :
sn Bs (0)

Indeed, we just prove

Z Z Z
d 1 d 1 s
udx = udAds
ds sn Bs (0) ds sn 0 @Bs (0)
Z Z
n 1 n
= ns udx + s udA
Bs (0) @Bs (0)

= 0:

Note
Z Z Z Z
2 2
2n udx = u 4 jxj dx = jxj 4 udx + u@ jxj2 jxj2 @ u dA
Bs Bs Bs @Bs
Z Z Z Z
2 2
= 2s udA + jxj s 4 u dx ( @ u= 4u)
@Bs Bs @Bs Bs
Z
= 2s udA:
@Bs

15
Dividing both sides by sn+1 ; we see changing rate of the average of harmonic function
is zero.
This mean value veri…cation for harmonic functions leads to veri…cation proofs of
mean value formula for caloric functions, minimal surfaces (monotonicity formula),
should also for mean curvature ‡ow.
Veri…cation of solutions to 4u ut = 0 satisfying
Z
1 jxj2
jxj2
u dxdt = constant = (4 )n=2 u (0; 0) :
Rn 1
e 4t R n 4t2
( t)n=2

Indeed we just prove the derivative of the left hand side with respect to R is
zero. In order to take the derivative of the integral, we re-formulate it by the co-area
formula Z Z RZ
jxj2 jxj2 dl
u 2 dxdt = u 2 dA ;
R 4t 0 =l 4t jr j
where
jxj2
1=2
= ( t) e 4nt :
Then we need to prove
" Z #
d 1 jxj2
u 2 dxdt
dR Rn R 4t
Z Z
n 1 jxj2 n jxj2 1
= nR u 2 dxdt + R u 2 dA
R 4t =R 4t jr j
= 0:

We look for a function ' such that 4' + 't = jxj2 =4t2 (in order to apply Green’s
identity). Surprisingly, the log function of the backward heat kernel does it

n jxj2
'= ln ( t) + + n ln R:
2 4t
We proceed with Green’s identity
Z Z
jxj2
u dxdt = u (4' + 't ) dxdt
<R 4t2 <R
Z Z
= ' (4u ut ) dxdt + u' x 'u x + u' t dA
<R =R
Z
= u' x dA (Recall ' = 0 on = R)
=R
Z
R jxj2 1
= u 2 dA;
n =R 4t jr j

16
where we used
D x 2ntx
x ' = D' =
jr j 2t jr j
2
R jxj 1
= on the boundary = R:
n 4t2 jr j
" Z #
d 1 jxj2
This is exactly what we need in showing u 2 dxdt = 0:
dR Rn R 4t
Applications of mean value formulas
App1. Strong max principle:
Let u 2 C12 solution to ut 4u = 0 in UT :
max u only attains at the parabolic boundary of UT ;
otherwise, if max u = u (x0 ; t0 ) ; where (x0 ; t0 ) is an interior or non-parabolic
boundary of UT ; then we have u (x; t) u (x0 ; t0 ) for all (x; t) in the closure of the
connected set of UT \ ft t0 g by chain of downward heat balls.
Def: Parabolic boundary: the closure of those points that cannot center any
backward heat ball inside the domain UT :
Examples of UT
…gure parabolic bdry

Proof of the strong max principle.

Suppose u (x0 ; t0 ) = maxUT u and (x0 ; t0 ) is not a parabolic boundary point, that
is, (x0 ; t0 ) centers a heat ball in UT : By the mean value formula in this ball
Z
1 jx0 xj2
u (x0 ; t0 ) = u (x; t) dxdt
(4 R2 )n=2 (x0 ;t0 ;x;t) (4 R2 ) n=2 4 (t0 t)2
Z
kernel 0 1 jx0 xj2
u (x ;
0 0t ) dxdt
(4 R2 )n=2 (x0 ;t0 ;x;t) (4 R2 ) n=2 4 (t0 t)2
R
kernel=1
= u (x0 ; t0 ) :

Thus u (x; t) u (x0 ; t0 ) in

8 jx0 xj2
9
< e 4(t0 t)
1 =
(x; t)j :
: (t0 t)n=2 Rn ;

RMK. The closure includes the points at horizontal level ft = t0 g ; this is because
such a point (y; t0 ) is the limit of (y; s) as s " t0 ; and the segment (x0 ; t0 ) -(y; s) can
be covered a chain of heat balls.

…gure downward segment

17
Then u (y; t) = lim u (y; s) = lim u (x0 ; t0 ) :
Uniqueness of caloric function on bounded domains
Let u; v be two C12 (UT ) \ C00 UT solutions to wt 4w = 0 in UT ; and u = v on
the parabolic boundary of UT : THEN u v.
RMK. UT including U1 domains like ft > convex (x)g ; say

…gure t jxj4 :

Question. What happens to Rn [0; T ] or R+

1
[0; T ]?
App2. Regularity
Faking space dimension Rn ! Rn+m will lead us to a C12 (even better ones for
larger m) kernel in the mean value formula:

ut (x; t) (4x + 4y ) u (x; t) = 0

Z
jx0 xj2 + jy0 yj2
u (x0 ; t0 ) = u (x; t) dydxdt
(x0 ;y0 ;t0 ;x;y;t) (4 R2 ) n=2 4 (t0 t)2
Z
= u (x; t) K (x0 x; t0 t) dxdt
n=2
(x0 ;t0 ;x;t) (4 R2 )

where Kuptsov kernel

Z
jx0 xj2 + jy0 yj2
K (x0 x; t0 t) = dy
jx0 xj2 +jy0 yj2 4(t t0 )[
(n+m)
2
ln(t0 t) (n+m) ln R] 4 (t0 t)2

is Cx2 and Ct1 for m 5: Then start from L1 function, satisfying the parabolic mean
value formula, we immediately have C12 solution to the heat equation (no need exis-
tence).
We can also get interior estimates via multiple-integrals.
First, by using a di¤erent argument via Green’s identity, we show the solutions
are C 1 in x; t and C ! in x: Recall the fundamental solution is not C ! in t:
Green’s identity
Z Z
u Dt v + 4v + v Dt u 4u = uv x vu x + uv t dA:
UT
| {z } | {z } @UT
\ (0;0)00 0

jx yj2
v= (x; t; y; s) = e 4(t s) = [4 (t s)]n=2

18
 Z
u (x; t) = u (y; s) y (x; t; y; s) dA
@U [0;t] | {z }
C ! in x not in t; C 1 in t
Z
+ u y (y; s) (x; t; y; s) dA
@U [0;t] | {z }
C ! in x not in t; C 1 in t
Z
+ u (y; 0) (x; t; y; 0) dy:
U | {z }
C ! in x;t for t 0

So we conclude u is C ! (x) and C 1 (t) in UT0 UT :

Interior estimates

max Dxk Dtl u C (k; l; K) max Dx2 u + jDt uj C (k; l; K) max juj
C1 C2 C3

or
C (k; l; K)
max Dxk Dtl u max juj
CR Rk+2l C3R

via scaling v (x; t) = u (Rx; R2 t) ; Dxk Dtl v (x; t) = Rk+2l Dxk Dtl uj(Rx;R2 t) :
Liouville Th’m
Global (eternal) solution, say C12 to ut 4u = 0 in Rn ( 1; +1) satisfying

ju (x; t)j A jxjk + jtjl for large jxj + jtj

must be a caloric polynomial of degree less than k + 2l:

Proof.
C (k 0 ; l0 ; K) R!1
Dxk Dtl u (0; 0) k 0 +2l0 A Rk + R2l ! 0
R
0
for k + 2l0 > k + 2l: Note (0; 0) could be anywhere, so u (x; t) is a caloric polynomial
of degree k + 2l:

App3. Harnack inequality

u 0 solution to ut 4u = 0; then

max u C (n) min

+
u:
Cr Cr

Here Cr = Br ( 3r2 ; 2r2 ) and Cr = Br ( r2 ; 0) :

…gure Harnack:

19
One proof is via “fake” dimension mean value formula, it is little “involved” in cal-
culating the positive weight. However, everything is at calculus level.
eg. The following example shows the delay of time is necessary in the Harnack
inequality. Take the heat kernel in Q = B1 (2) ( 1; 1) ; we have

0 < max 0 = min :

Q Q

Uniqueness for Cauchy problem with constraints in Rn R+ :

Max Principle: Let u be C12 (Rn (0; T )) \ C (Rn [0; T ]) solution to

ut 4u = 0
:
u (x; 0) = g (x)
2
Suppose ju (x; t)j Aeajxj in Rn [0; T ] : Then

ju (x; t)j sup g (x) in Rn [0; T ] :

Rn

Proof. We only need to prove u (x; t) supRn g (x) , M for subcaloric solution
2
ut 4u 0 with sub quadratic-exponential growth u (x; t) Aeajxj :
Step1.
…gure t-direction thin domain:

For any > 0; set

jxj2
1
e ( )
1
4 8a t
v=M+ n=2
1
8a
t
vt 4v = 0 ut 4u
1
v u on @BR 0; for R large
16a
v u on Rn f0g :

RMK. Invariance
p
way to construct this barrier:
jxj2 1 jxj2
* p1t e 4t 99K p1 t e 4t still sol.
* the time shift sol v is quadratic-exponential growth at t = 0; it goes to 1 as
1
t % 8a for every x:
It follows from the maximum principle on the bounded domain,
jxj2
1 1
e ( ) in B
1
4 8a t
u (x; t) M+ n=2 R 0; :
1
t 16a
8a

20
 1
So for any …xed (x0 ; t0 ) with t0 16a
; we have
jx0 j2
1 !0
e ( ) !
1
4 8a t0
u (x0 ; t0 ) M+ n=2
M:
1
8a
t0
1 2 2
Step2. The above argument works equally well on 16a ; 16a ; ; 3
16a 16a
; ; still
[0; T ] :
Corollary. The Cauchy problem with growth constraint

ut 4u = f
in Rn (0; T )
u (x; 0) = g (x)
2
with ju (x; t)j Aeajxj in Rn [0; T ] ; has at most one solution.
Proof. The di¤erence of any two solutions satis…es the condition in the max
2
principle with g = 0 and di¤erence is less 2Aeajxj ; so the di¤erence is 0:

ut 4u = 0
eg. The caloric function, or a solution to 2
u (x; 0) = eajxj
Integral way to construct the barrier:
Z
1 jx yj2 2
u (x; t) = n=2
e 4t eajyj dy
(4 t) Rn
Z p
1 jyj2 +aj2 ty xj
2
= n=2
e dy
Rn
1 jxj2
= n=2
e1 4at

(1 4at)
is
1 1 jxj2
* a > 0 unique in Rn [0; 4a ) with constraint ju (x; t)j (1 4at)n=2
e1 4at and
2
grows faster than eajxj for t > 0;
2
* a < 0 uniqueness in Rn [0; 1) with constraint ju (x; t)j e100jxj and grows
2
faster than eajxj for t > 0:
The message: the growth/decay rate is not preserved precisely.

ut 4u = 0 in Rn [0; 1)
Nonuniqueness of Cauchy problem
u (x; 0) = 0
Tikhonov’s counterexample.
Idea of construction:
* along t = 0; position u (x; 0) alone determines all the derivatives (if analytic).
* along x = 0; position u (0; t) and velocity ux (0; t) determines all the derivatives
in x; (if analytic in x):
Now we solve a “real”Cauchy problem along the t-axis

u (0; t) = g (t)
ux (0; t) = 0

21
 ux (0; t) = 0; uxxx (0; t) = uxt (0; t) = 0; ; Dx2k+1 u (0; t) = Dtk ux (0; t) = 0
uxx (0; t) = ut (0; t) = g (t) ; uxxxx (0; t) = utt (0; t) = g 00 (t) ;
0
; Dx2k u (0; t) =
Dtk u (0; t) = Dtk g (t) :
Assuming u is C ! in terms of x; then
X
1
g (k) (t)
u (x; t) = g (t) + x2k :
k=1
(2k)!

Technical realization:
1
e t t>0
graph for g (t) = need > 1:
0 t 0

How to control the derivatives

1st Direct try.
g (t) = e t
g 0 = e t ht 1
i
1 2
g 00 = e t ( t ) ( + 1) t 2
h i
1 3
g 000 = e t ( t ) + + ( + 1) ( + 2) t 3

h i
1 k
g (k) = e t ( t ) + ( + 1) ( + 2) ( +k 1) t k

t 1 k
e k! ( t )
g (k) k! 1 k k!
(2k)!
e t
(2k)!
( t ) ; and (2k)! = 2k 1 3 5 1 (2k 1) 1
22k k!
P e t 1 k t 1
x2
1 x2 1 t!0+
ju (x; t)j g (t) + 1 k=1 k! 22k ( t ) x2k = e t e 4 =e t
+ 4 t +1 !
1:

2nd complex try

Observe e t is analytic when t > 0

…gure complex plan z = t + is

1
R
g(z)
g (t) = 2 i
dz
k!
t
Rzg(z)
g (k) (t) = 2 i (z t)k+1
dz
1
Now by continuity of z at 1; Re z 2
for jz 1j ; where = (1=2) < 1;
1
then Re (tz) 2
t for jtz tj t
1
or Re z 2
t for jz tj t

22
 and
1
z
e = eRe z
e 2
t
for jz tj t:
So Z 1 1
(k) k! g (z) k! e 2 t k! e 2
t
g (t) dz 2 t=
2 jz tj= t (z t)k+1 2 ( t)k+1 ( t)k
and
X
1
1 k! e 2 t
1

ju (x; t)j g (t) + x2k

k=1
(2k)! ( t)k
1
X1
1 x2
k
t t
e +e 2

k=1
k! t
1
t + xt
2
< 1 for all (x; t) with t > 0;
e 2 provided > 1:
! 0 as t ! 0 + for each …xed x;

Then the Tikhonov’s series converges, we’ve constructed a “super”quadratic-exponential

caloric function such that ut uxx = 0 and u (x; 0) = 0; u (x; t) is not identically 0
for t > 0:
RMK. Choosing
1 1

g (t) = e t (1 t) t>0 ;
0 t 0 or t 1
…gure complex plan z = t + is for this g

we have another Tikhonov solution/caloric function vanishes before 0 or after 1:

…gure Tikhonov double sided vanishing

RMK. Let v = u2 ; where u is the above Tikhonov caloric function, then vt 4v =

2uut 2u 4 u 2 jDuj2 = 2 jDuj2 0: This v is non-negative sub-caloric function
vanishing at t = 0 and t = 1; yet doesn’t vanish identically between (0; 1) :

Nonanalytic, yet smooth solution in Rn [0; 1)

eg.
ut uxx = 0
4
u (x; 0) = e x

23
has the bounded solution/quadratic-exponential growth solution
Z
1 jx yj2 4
u (x; t) = 1=2
e 4t e jyj dy
(4 t) R1
Z p 4
1
e y e jx 2 tyj dy
2
= 1=2
R1

which is C 1 (R1 [0; 1)) ; but NOT C ! at t = 0:

In fact, if u (0; t) is analytic in terms of t near t = 0; then

X
1
u (0; t) = ak tk ;
k=0

where

1 k 1 1 X1
( x4 )
m
ak = Dt u (0; 0) = Dx2k u (0; 0) = Dx2k
k! k! k! m=0
m!
x=0
2k=4m 1 (4m)!
= > m!:
(2m)! m!

So
m!1
a2m t2m > m!t2m ! 1 for any …xed t > 0:
Then the series diverges, u (x; t) cannot be analytic in t at (0; 0) :

Nonexistence of nonnegative solution to Cauchy problem

ut uxx = 0
4
u (x; 0) = ex
First note the representation
Z
1 jx yj2 y 4
1=2
e e dy = 1:
(4 t) R1

Overheated, the nonnegative solution blows up once time starts.

Now the proof.
2
…gure for ex and gk

4
gk (x) 2 C01 (Bk+1 ) and gk (x) = ex on Bk : The bounded C12 solution to

ut uxx = 0
u (x; 0) = gk

24
is Z
1 y2
p
uk (x; t) = 1=2
e gk x 2 ty dy:
R1
y2
p
…gure for e and gk R 2 ty

For each …xed k and say, 0.9, there exists Rk = R (k; 0:9) large so that

0 uk ( R; t) 0:9 for 0 t 1:

This is because 2
1 k+1 R p
k 1 4
uk ( Rk ; t) 1=2
ep : 2 t ek
t
Then as u is nonnegative, uk 0:9 + u on the parabolic boundary of the cylinder
BRk [0; 1] : The maximum principle implies uk 0:9+u in BRk [0; 1] : In particular

uk (0; 1) 0:9 + u (0; 1) for all k:

But uk (0; 1) goes to +1; as k goes to +1: A contradiction!

In fact, u (0; l) is forced to be 1 for all small l > 0:
Question: Existence of sign-changing solutions? Answer: YES.

As the mean curvature ‡ow operator is variable coe¢ cients

1 p ij 1 p ij
4g @t = p @i gg @j = g ij @ij + p @i gg @j ;
g g

we need to deal with general parabolic operator

X X
L= aij (x; t) @ij + bi (x; t) @i @t

where -parabolic means

(aij (x; t)) I > 0:
Notation T = (0; T ] Rn R1 ; the parabolic boundary of T consists of
two parts, the bottom and lateral side

@p T = f0g [ @ [0; T ] :

Note that the top side fT g is considered as interior of the “parabolic” domain
T :
Weak Maximum Principle. Assume u; v 2 C 2;1 ( T ) \ C T ;
Max. If Lu 0 in , then max T u = max@p u:
Min. If Lu 0 in , then min T u = min@p u:
Comparison. If Lu Lv in T and u v on @p ; then u v in T :

25
 Proof. It su¢ ces to handle the max case, the other cases follow by applying the
…rst case to u and u v respectively.
Strict case. Lu > 0:
Suppose u attains its maximum at an interior point (x ; t ) ; then Du (x ; t ) = 0;
D2 u (x ) 0; and @t u (x ; t ) 0 (not necessarily @t u (x ; t ) = 0). But
X X
Lu (x ; t ) = aij (x ; t ) Dij u (x ; t ) + bi (x ; t ) Di u (x ; t ) @t u (x ; t )
= T r aij (x ; t ) D2 u (x ; t ) 0:

This contradicts Lu (x ; t ) > 0: Thus max T u = max@p T u:

General case. Lu 0:
We do analysis by approximating it with strict inequalities. Let

u" = u "t

(or as in the elliptic case u" = u + "eKx1 ). Then

Lu" = Lu + " > 0:

By the strict case

max (u "t) = max (u "t) :
T @p T

Take " ! 0; we …nd max u = max@ u:

Weak Maximum Principle with c 0: Assume u; v 2 C 2;1 ( T ) \ C T and

c (x; t) 0 in T
Max. If Lu + c (x; t) u 0 in T , then max T u max@p T u+ :
Min. If Lu + c (x; t) u 0 in T , then min T u min@p T u :
Comparison. If Lu + c (x; t) u Lv + c (x; t) v in T and u v on @p T ; then
u v in T :
Proof. Again it su¢ ces to handle the …rst case. And we can just assume max T u >
0:We repeat the above argument for no c term. Note max T (u "t) > 0 for small
enough ": This maximum cannot attain inside T : Otherwise we have two inconsistent
inequalities
L (u "t) + c (u "t) " c"t > 0
and also at the interior maximum point

L (u "t) + c (u "t) c (u "t) 0:

Thus the positive maximum can only happen at the parabolic boundary

max (u "t) = max (u "t) :

T @p T

Let " go to 0; we have

max u = max u = max u+ :
T @p T @p T

26
 C-eg with c > 0:

u = cos xet
u00 + 2u @t u = 0 in T = ( =2; =2) (0; 1]
max u = e 1 = max u:
1 @p 1

C-eg for necessity of the negative part in min@ u :

u = chx e t
u00 2u @t u = 0 in = ( 9; 9) (0; 1]
min u = e 1 1 = min u:
1 @p 1

And chxe t gives us a counterexample for necessity of the positive part in max@p T u+ :
Strong Maximum Principle (Nirenberg). Assume u 2 C 2;1 ( T ) \ C T
and T is connected.
Max. If Lu 0 in T and u attains its global maximum at an interior point
(xM ; tM ), then u is constant in tM = [0; tM ] :
Min. If Lu 0 in and u attains its global minimum at an interior point
(xm ; tm ), then u is constant tm = [0; tm ] :
It su¢ ces to prove only one case, say minimum one; the maximum case follows
by considering u: The inconsistency: if “strict”minimum occurs, then either space
gradient (Step2.1) or time derivative (Step2.2) at the point is not zero. Nirenberg
constructed a Hopf like barrier to achieved.
Proof. Set-up.
Suppose u attains its global minimum m = min T u at an interior point (xm ; tm ).
Set
m
tm = f(x; t) 2 tm : u (x; t) = mg :

If u (x; t) is not constant m; then tm n m tm = f(x; t) 2 tm : u (x; t) > mg is non-

empty. Next we choose an (n+1)-dim ball touching the boundary of m tm and its
lower part t tm inside tm : There exists an interior point Y = (y; s) 2 tm n m tm
such that = dist Y; @ m tm < dist (Y; @p tm ) : Enlarge the (n+1)-dim ball centered
at Y until it hits the boundary @ m tm ; say at (x0 ; t0 ) ; before hitting @p tm : We also
assume the …nal ball hits the boundary only at (x0 ; t0 ) ; say, by moving the center
and shrinking the radius a bit. The …nal radius is :

…gure

For simplicity of notation, we assume (y; s) = (0; 0) : What we have now is

m m
B (0; 0) \ ft tm g tm n tm and (x0 ; t0 ) 2 @B (0; 0) \ tm :

We need to handle two cases separately: Case x0 6= 0 and Case x0 = 0:

Step2.1 Case x0 6= 0 (Hopf boundary lemma).
We take another ball B 0 (x0 ; t0 ) centered at (x0 ; t0 ) with radius 0 = jx0 0j =2:

27
 We construct a subsolution v such that u v near (x0 ; t0 ) and vr (x0 ; t0 ) <
=at (x0 ;t0 )
0: This forces ur (x0 ; t0 ) < 0 or Dx u (x0 ; t0 ) 6= 0; a contradiction to interior minimum
u (x0 ; t0 ) at (x0 ; t0 ) :
First take 2
w (x) = e K (jxj +t ) e K
2 2

with large constant K to be determined later. We compute

hX X i
K (jxj2 +t2 ) 2
Lw = e aij 4K xi xj 2K ij + bi ( 2Kxi ) + 2Kt
> 0 in B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g

provided K is taken large enough.

Notice u > m inside B (0; 0) : Then we can take small enough " so that v (x) =
"w (x) + u (x0 ) stays below u (x) on the parabolic boundary of the intersection

B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g ;

that is
@B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g
and
B (0; 0) \ @B 0 (x0 ; t0 ) \ ft t0 g
By the construction

Lu 0 Lv in B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g :

From the comparison principle (weak maximum principle valid even in this non
straight lateral boundary case)

u (x; t) v (x; t) in B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g :

Remember u (x0 ; t) = v (x0 ; t0 ) ; taking radial derivatives, we get

K 2
ur (x0 ; t0 ) vr (x0 ; t0 ) = 2"K jx0 0j e > 0:

This contradicts the fact that Dx u (x0 ; t0 ) = 0 at minimum interior point (x0 ; t0 ) :
Therefore, u (x) m in [0; tm ] :
Step2.2 Case x0 = 0 (Time-like Hopf boundary lemma).
We construct a di¤erent subsolution v such that u v near (x0 ; t0 ) and
=at (x0 ;t0 )
vt (x0 ; t0 ) < 0: This forces ut (x0 ; t0 ) < 0; in turn, Lu > 0 at the minimum point
(x0 ; t0 ) ; a contradiction to the equation Lu 0:
First take
w (x; t) = jxj2 K (t t0 )

28
for a large constant K to be determined later. This function w vanishes on the
paraboloid PK = (x; t) : jxj2 + K (t t0 ) = 0 which is inside B (0; 0) \ ft tm g ;
now just B (0; 0) when t is close to t0 :We compute
X X
Lw = aij 4 + bi ( 2xi ) + K
> 0 in B (0; 0)
provided K is taken large enough.
Notice u > m inside B (0; 0) : Then we can take small enough " so that v (x; t) =
"w (x; t)+m stays below u (x; t) on the boundary @B (0; 0) portion under the paraboloid
Pk : Certainly v (x; t) is below u (x; t) on the paraboloid PK intersecting B (0; 0) ; as
v vanishes and u m there. By the construction
Lu 0 < Lv in P B = B (0; 0) \ jxj2 + K (t t0 ) < 0 :
It follows that max (v u) cannot happen at interior points of P B: Otherwise, L (v u)
0 or Lv Lu there. As v u 0 on the boundary of P B; we can only have

u (x; t) v (x; t) in B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g :

Remember u (x0 ; t) = v (x0 ; t0 ) ; taking time derivatives, we get
ut (x0 ; t0 ) vt (x0 ; t0 ) = K < 0:
It follows that Lu K > 0 at (x0 ; t0 ) : This contradicts the equation Lu 0:
Therefore, u (x) m in [0; tm ] :
Remark. Note that we can only draw the constant conclusion below tm : Other-
wise, Step2.2 fails for ball B (0; 0) touching (x0 ; t0 ) above (for example, no compar-
2
ison principle). More e¤ectively, recall the heat kernel = t 1=2 e x =4t vanishes in
B1 (2) ( 1; 0]; but become positive for positive time.
Strong maximum principle with c 0: Assume u 2 C 2 ( T ) \ C T :
Max. If Lu + cu 0 in T and u attains its nonnegative global maximum at
an interior point (xM ; tM ), then u is constant in tM = [0; tM ] :
Min. If Lu + cu 0 in and u attains its nonpositive global minimum at an
interior point (xm ; tm ), then u is constant in tm = [0; tM m ] :
Proof. Again it su¢ ces to justify only one case, say the minimum one.
Now
m = min u 0:
We go over the same two steps as in proof of the strong maximum principle without
2
c term. The di¤erence is in Step2.1, for v = " e K (jxj +t ) e K
2 2
+m

Lv + cv
2
hX X i
= e K (jxj +t )
2
a ij 4K 2 xi xj 2K ij + bi ( 2Kxi ) + c
+ cm
hX X i
K (jxj2 +t2 )
e aij 4K 2 xi xj 2K ij + bi ( 2Kxi ) + c
> 0 in B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g

29
provided K is taken large enough. Then

Lu + cu 0 Lv + cv in B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g ;
u (x; t) v (x; t) on @p [B (0; 0) \ B 0 (x0 ; t0 ) \ ft t0 g] :

The remaining argument goes through.

Step2.2 can be adjusted similarly.
Strong maximum principle with no sign restriction on c but with one sign
restriction on solution. Assume u 2 C 2;1 ( T ) \ C T and the arbitrary sign of
c (x; t) in T :
Max. If Lu + cu 0 in and max u = 0 attains at an interior point (x0 ; t0 ),
then u 0 in [0; t0 ] :
Min. If Lu + cu 0 in and min u = 0 attains at an interior point (x0 ; t0 ),
then u 0 in [0; t0 ] :
Proof. Again, we only need to handle one case, say the minimum case. Note that
u 0; then splitting c = c+ c ; we have

Lu c u c+ u 0:

We are back to the above strong maximum principle with c 0:

Remember C-eg to c 0:

u = cos x
u00 + u @t u = 0 in ( =2; =2) (0; 10]
u (0; 1) is an interior global maximum.

C-eg to nonpositive minimum.

u = chx
00
u u @t u = 0 in R1 R1
positive u (0; 0) = 1 is an interior global minimum.

30