2

2.1

Heat Equation
Derivation

Ref: Strauss, Section 1.3. Below we provide two derivations of the heat equation, ut kuxx = 0 k > 0. (2.1)

This equation is also known as the diusion equation. 2.1.1 Diusion

Consider a liquid in which a dye is being diused through the liquid. The dye will move from higher concentration to lower concentration. Let u(x, t) be the concentration (mass per unit length) of the dye at position x in the pipe at time t. The total mass of dye in the pipe from x0 to x1 at time t is given by
x1

M (t) =
x0

u(x, t) dx.

Therefore,

dM = dt

x1

ut (x, t) dx.
x0

By Ficks Law,

dM = ow in ow out = kux (x1 , t) kux (x0 , t), dt where k > 0 is a proportionality constant. That is, the ow rate is proportional to the concentration gradient. Therefore,
x1

ut (x, t) dx = kux (x1 , t) kux (x0 , t).

x0

Now dierentiating with respect to x1 , we have ut (x1 , t) = kuxx (x1 , t). Or, ut = kuxx . This is known as the diusion equation. 2.1.2 Heat Flow

We now give an alternate derivation of (2.1) from the study of heat ow. Let D be a region in Rn . Let x = [x1 , . . . , xn ]T be a vector in Rn . Let u(x, t) be the temperature at point x,

time t, and let H(t) be the total amount of heat (in calories) contained in D. Let c be the specic heat of the material and its density (mass per unit volume). Then H(t) =
D

cu(x, t) dx.

Therefore, the change in heat is given by dH = dt cut (x, t) dx.

D

Fouriers Law says that heat ows from hot to cold regions at a rate > 0 proportional to the temperature gradient. The only way heat will leave D is through the boundary. That is, dH = u n dS. dt D where D is the boundary of D, n is the outward unit normal vector to D and dS is the surface measure over D. Therefore, we have cut (x, t) dx =
D D

u n dS.

Recall that for a vector eld F , the Divergence Theorem says F n dS =

D D

F dx.

(Ref: See Strauss, Appendix A.3.) Therefore, we have cut (x, t) dx =

D D

( u) dx.

This leads us to the partial dierential equation cut = ( u).

If c, and are constants, we are led to the heat equation ut = ku, where k = /c > 0 and u =
n i=1

uxi xi .

2.2
2.2.1

Heat Equation on an Interval in R

Separation of Variables

Consider the initial/boundary value problem on an interval I in R, x I, t > 0 ut = kuxx u(x, 0) = (x) xI u satises certain BCs. In practice, the most common boundary conditions are the following: 2

(2.2)

1. Dirichlet (I = (0, l)) : u(0, t) = 0 = u(l, t). 2. Neumann (I = (0, l)) : ux (0, t) = 0 = ux (l, t). 3. Robin (I = (0, l)) : ux (0, t) a0 u(0, t) = 0 and ux (l, t) + al u(l, t) = 0. 4. Periodic (I = (l, l)): u(l, t) = u(l, t) and ux (l, t) = ux (l, t). We will give specic examples below where we consider some of these boundary conditions. First, however, we present the technique of separation of variables. This technique involves looking for a solution of a particular form. In particular, we look for a solution of the form u(x, t) = X(x)T (t) for functions X, T to be determined. Suppose we can nd a solution of (2.2) of this form. Plugging a function u = XT into the heat equation, we arrive at the equation XT kX T = 0. Dividing this equation by kXT , we have T X = = . kT X for some constant . Therefore, if there exists a solution u(x, t) = X(x)T (t) of the heat equation, then T and X must satisfy the equations T = kT X = X for some constant . In addition, in order for u to satisfy our boundary conditions, we need our function X to satisfy our boundary conditions. That is, we need to nd functions X and scalars such that X (x) = X(x) xI (2.3) X satises our BCs. This problem is known as an eigenvalue problem. In particular, a constant which 2 satises (2.3) for some function X, not identically zero, is called an eigenvalue of x for the given boundary conditions. The function X is called an eigenfunction with associated eigenvalue . Therefore, in order to nd a solution of (2.2) of the form u(x, t) = X(x)T (t) our rst goal is to nd all solutions of our eigenvalue problem (2.3). Lets look at some examples below. Example 1. (Dirichlet Boundary Conditions) Find all solutions to the eigenvalue problem X = X X(0) = 0 = X(l). 3 0<x<l (2.4)

Any positive eigenvalues? First, we check if we have any positive eigenvalues. That is, we check if there exists any = 2 > 0. Our eigenvalue problem (2.4) becomes X + 2X = 0 X(0) = 0 = X(l). The solutions of this ODE are given by X(x) = C cos(x) + D sin(x). The boundary condition X(0) = 0 = C = 0. The boundary condition X(l) = 0 = sin(l) = 0 = = Therefore, we have a sequence of positive eigenvalues n = with corresponding eigenfunctions Xn (x) = Dn sin n x . l n l
2

0<x<l

n l

n = 1, 2, . . . .

Is zero an eigenvalue? Next, we look to see if zero is an eigenvalue. If zero is an eigenvalue, our eigenvalue problem (2.4) becomes X =0 X(0) = 0 = X(l). The general solution of the ODE is given by X(x) = C + Dx. The boundary condition X(0) = 0 = C = 0. The boundary condition X(l) = 0 = D = 0. Therefore, the only solution of the eigenvalue problem for = 0 is X(x) = 0. By denition, the zero function is not an eigenfunction. Therefore, = 0 is not an eigenvalue. Any negative eigenvalues? Last, we check for negative eigenvalues. That is, we look for an eigenvalue = 2 . In this case, our eigenvalue problem (2.4) becomes X 2X = 0 X(0) = 0 = X(l). 4 0<x<l 0<x<l

The solutions of this ODE are given by X(x) = C cosh(x) + D sinh(x). The boundary condition X(0) = 0 = C = 0. The boundary condition X(l) = 0 = D = 0. Therefore, there are no negative eigenvalues. Consequently, all the solutions of (2.4) are given by n = n l
2

Xn (x) = Dn sin

n x l

n = 1, 2, . . . .

Example 2. (Periodic Boundary Conditions) Find all solutions to the eigenvalue problem X = X X(l) = X(l), X (l) = X (l). l < x < l (2.5)

Any positive eigenvalues? First, we check if we have any positive eigenvalues. That is, we check if there exists any = 2 > 0. Our eigenvalue problem (2.5) becomes X + 2X = 0 X(l) = X(l), X (l) = X (l). The solutions of this ODE are given by X(x) = C cos(x) + D sin(x). The boundary condition X(l) = X(l) = D sin(l) = 0 = D = 0 or = The boundary condition X (l) = X (l) = C sin(l) = 0 = C = 0 or = Therefore, we have a sequence of positive eigenvalues n = with corresponding eigenfunctions Xn (x) = Cn cos n n x + Dn sin x . l l 5 n l
2

l < x < l

n . l n . l

Is zero an eigenvalue? Next, we look to see if zero is an eigenvalue. If zero is an eigenvalue, our eigenvalue problem (2.5) becomes X =0 X(l) = X(l), X (l) = X (l). The general solution of the ODE is given by X(x) = C + Dx. The boundary condition X(l) = X(l) = D = 0. The boundary condition X (l) = X (l) is automatically satised if D = 0. Therefore, = 0 is an eigenvalue with corresponding eigenfunction X0 (x) = C0 . Any negative eigenvalues? Last, we check for negative eigenvalues. That is, we look for an eigenvalue = 2 . In this case, our eigenvalue problem (2.5) becomes X 2X = 0 X(l) = X(l), X (l) = X (l). The solutions of this ODE are given by X(x) = C cosh(x) + D sinh(x). The boundary condition X(l) = X(l) = D sinh(l) = 0 = D = 0. The boundary condition X (l) = X (l) = C sinh(l) = 0 = C = 0. Therefore, there are no negative eigenvalues. Consequently, all the solutions of (2.5) are given by n = n 2 l 0 = 0 Xn (x) = Cn cos X0 (x) = C0 . n n x + Dn sin x l l n = 1, 2, . . . l < x < l l < x < l

Now that we have done a couple of examples of solving eigenvalue problems, we return to using the method of separation of variables to solve (2.2). Recall that in order for a function of the form u(x, t) = X(x)T (t) to be a solution of the heat equation on an interval I R which satises given boundary conditions, we need X to be a solution of the eigenvalue problem, X = X xI X satises certain BCs for some scalar and T to be a solution of the ODE T = kT. We have given some examples above of how to solve the eigenvalue problem. Once we have solved the eigenvalue problem, we need to solve our equation for T . In particular, for any scalar , the solution of the ODE for T is given by T (t) = Aekt for an arbitrary constant A. Therefore, for each eigenfunction Xn with corresponding eigenvalue n , we have a solution Tn such that the function un (x, t) = Tn (t)Xn (x) is a solution of the heat equation on the interval I which satises our boundary conditions. Note that we have not yet accounted for our initial condition u(x, 0) = (x). We will look at that next. First, we remark that if {un } is a sequence of solutions of the heat equation on I which satisfy our boundary conditions, than any nite linear combination of these solutions will also give us a solution. That is,
N

u(x, t)
n=1

un (x, t)

will be a solution of the heat equation on I which satises our boundary conditions, assuming each un is such a solution. In fact, one can show that an innite series of the form

u(x, t)
n=1

un (x, t)

will also be a solution of the heat equation, under proper convergence assumptions of this series. We will omit discussion of this issue here. 2.2.2 Satisfying our Initial Conditions

We return to trying to satisfy our initial conditions. Assume we have found all solutions of our eigenvalue problem. We let {Xn } denote our sequence of eigenfunctions and {n } denote our sequence of eigenvalues. Then for each n , we have a solution Tn of our equation for T . Let u(x, t) = Xn (x)Tn (t) = An Xn (x)ekn t .
n n

Our goal is to choose An appropriately such that our initial condition is satised. In particular, we need to choose An such that u(x, 0) =
n

An Xn (x) = (x).

In order to nd An satisfying this condition, we use the following orthogonality property of eigenfunctions. First, we make some denitions. For two real-valued functions f and g dened on , f, g =

f (x)g(x) dx

is dened as the L2 inner product of f and g on . The L2 norm of f on is dened as ||f ||2 2 () = f, f = L |f (x)|2 dx.

We say functions f and g are orthogonal on Rn if f, g =

f (x)g(x) dx = 0.

We say boundary conditions are symmetric if [f (x)g(x) f (x)g (x)]|x=a = 0 for all functions f and g satisfying the boundary conditions. Lemma 3. Consider the eigenvalue problem (2.3) with symmetric boundary conditions. If Xn , Xm are two eigenfunctions of (2.3) with distinct eigenvalues, then Xn and Xm are orthogonal. Proof. Let I = [a, b].
b b x=b

n
a

Xn (x)Xm (x) dx =
a b

Xn (x)Xm (x) dx
x=b

=
a

Xn (x)Xm (x) dx Xn (x)Xm (x)|x=a

b

=
a

Xn (x)Xm (x) dx + [Xn (x)Xm (x) Xn (x)Xm (x)]|x=a

b

x=b

= m
a

Xn (x)Xm (x) dx,

using the fact that the boundary conditions are symmetric. Therefore,
b

(n m )
a

Xn (x)Xm (x) dx = 0, 8

but n = m because the eigenvalues are assumed to be distinct. Therefore,

b

Xn (x)Xm (x) dx = 0,
a

as claimed. We can use this lemma to nd coecients An such that An Xn (x) = (x).
n

In particular, multiplying both sides of this equation by Xm for a xed m and integrating over I, we have Am Xm , Xm = Xm , , which implies Am = Xm , . Xm , Xm

Example 4. (Dirichlet Boundary Conditions) In the case of Dirichlet boundary conditions on the interval [0, l], we showed earlier that our eigenvalues and eigenfunctions are given by n = n l
2

Xn (x) = sin

n x l

n = 1, 2, . . . .

Our solutions for Tn are given by Tn (t) = An ekn t = An ek(n/l) t . Now let u(x, t) =
n=1
2

Xn (x)Tn (t) =
n=1

An sin

n 2 x ek(n/l) t . l

Using the fact that Dirichlet boundary conditions are symmetric (check this!), our coefcients Am are given by Am = Xm , = Xm , X m
l 0

sin (mx/l) (x) dx

l 0

sin2 (mx/l) dx

2 l

sin
0

m x (x) dx. l

Therefore, the solution of (2.2) on the interval I = [0, l] with Dirichlet boundary conditions is given by

u(x, t) =
n=1

An sin

n 2 x ek(n/l) t l n x (x) dx. l

where An = 2 l
l

sin
0

Example 5. (Periodic Boundary Conditions) In the case of periodic boundary conditions on the interval [l, l], we showed earlier that our eigenvalues and eigenfunctions are given by cos n x 2 n l n = 1, 2, . . . n = , Xn (x) = sin n x l l 0 = 0, X0 (x) = C0 . Therefore, our solutions for Tn are given by Tn (t) = An e
kn t

An ek(n/l) A0

2t

n = 1, 2, . . . n = 0.

Now using the fact that for any integer n 0, un (x, t) = Xn (x)Tn (t) is a solution of the heat equation which satises our periodic boundary conditions, we dene

u(x, t) =
n

Xn (x)Tn (t) = A0 +
n=1

An cos

n n x + Bn sin x l l

ek(n/l) t .

Now, taking into account our initial condition, we want

u(x, 0) = A0 +
n=1

An cos

n n x + Bn sin x l l

= (x).

It remains only to nd coecients satisfying this equation. From our earlier discussion, using the fact that periodic boundary conditions are symmetric (check this!), we know that eigenfunctions corresponding to distinct eigenvalues will be orthogonal. For periodic boundary conditions, however, we have two eigenfunctions for each positive eigenvalue. Specically, cos n x and sin n x are both eigenfunctions corresponding to the eigenvalue l l n = (n/l)2 . Could we be so lucky that these eigenfunctions would also be orthogonal? By a straightforward calculation, one can show that
l

cos
l

n n x sin x dx = 0. l l

They are orthogonal! This is not merely coincidental. In fact, for any eigenvalue of (2.3) with multiplicty m (meaning it has m linearly independent eigenfunctions), the eigenfunctions may always be chosen to be orthogonal. This process is known as the Gram-Schmidt orthogonalization method. The fact that all our eigenfunctions are mutually orthogonal will allow us to calculate coecients An , Bn so that our initial condition is satised. Using the technique described

10

above, and letting f, g denote the L2 inner product on [l, l], we see that A0 = An = 1 1, = 1, 1 2l
l

(x) dx
l l

cos(nx/l), 1 = cos(nx/l), cos(nx/l) l

cos
l l

n x (x) dx l n x (x) dx. l

sin(nx/l), 1 Bn = = sin(nx/l), sin(nx/l) l

sin
l

Therefore, the solution of (2.2) on the interval I = [l, l] with periodic boundary conditions is given by

u(x, t) = A0 +
n=1

An cos

n n x + Bn sin x l l

ek(n/l)

2t

where A0 = An = 1 2l 1 l
l

(x) dx
l l

cos
l l

n x (x) dx l n x (x) dx. l

1 Bn = l

sin
l

2.2.3

Fourier Series

In the case of Dirichlet boundary conditions above, we looked for coecients so that

(x) =
n=1

An sin

n x . l

We showed that if we could write our function in terms of this innite series, our coecients would be given by the formula An = 2 l
l

sin
0

n x (x) dx. l

For a given function dened on (0, l) the innite series

n=1

An sin

n x l

where An

2 l

sin
0

n x (x) dx l

is called the Fourier sine series of . Note: The notation just means the series associated with . It doesnt imply that the series necessarily converges to . 11

In the case of periodic boundary conditions above, we looked for coecients so that

(x) = A0 +
n=1

An cos

n n x + Bn sin x l l

(2.6)

We showed that in this case our coecients must be given by 1 A0 = 2l An = Bn = 1 l 1 l

l

(x) dx
l l

cos
l l

n x (x) dx l n x (x) dx. l

(2.7)

sin
l

For a given function dened on (l, l) the series

A0 +
n=1

An cos

n n x + Bn sin x l l

where An , Bn are dened in (2.7) is called the full Fourier series of . More generally, for a sequence of eigenfunctions {Xn } of (2.3) which satisfy certain boundary conditions, we dene the general Fourier series of a function as
n

An Xn (x) where An

Xn , . Xn , Xn

Remark. Consider the example above where we looked to solve the heat equation on an interval with Dirichlet boundary conditions. (A similar remark holds for the case of periodic or other boundary conditions.) In order that our initial condition be satised, we needed to nd coecients An such that

(x) =
n=1

An sin

n x . l

We showed that if could be represented in terms of this innite series, then our coecients must be given by 2 l n x (x) dx. An = sin l 0 l While this is the necessary form of our coecients and dening

u(x, t)
n=1

An sin

n 2 x ek(n/l) t l

for An dened above is the appropriate formal denition of our solution, in order to verify that this construction actually satises our initial/boundary value problem (2.2) with Dirichlet boundary conditions, we would need to verify the following. 12

1. The Fourier sine series of converges to (in some sense). 2. The innite series actually satises the heat equation. We will not discuss these issues here, but refer the reader to convergence results in Strauss as well as the notes from 220a. Later, in the course, we will prove an L2 convergence result of eigenfunctions. Complex Form of Full Fourier Series. It is sometimes useful to write the full Fourier series in complex form. We do so as follows. The eigenfunctions associated with the full Fourier series are given by n n cos x , sin x l l for n = 0, 1, 2, . . .. Using deMoivres formula, ei = cos + i sin , we can write n einx/l + einx/l cos x = l 2 n einx/l einx/l sin x = . l 2i Now, of course, any linear combination of eigenfunctions is also an eigenfunction. Therefore, we see that n n x + i sin x = einx/l l l n n cos x i sin x = einx/l l l cos are also eigenfunctions. Therefore, the eigenfunctions associated with the full Fourier series can be written as einx/l n = . . . , 2, 1, 0, 1, 2, . . . . Now, lets suppose we can represent a given function as an innite series expansion in terms of these eigenfunctions. That is, we want to nd coecients Cn such that

(x) =
n=

Cn einx/l .

As described earlier, eigenfunctions corresponding to distinct eigenvalues will be orthogonal, as periodic boundary conditions are symmetric. Therefore, we have
l l

einx/l eimx/l dx = 0

for m = n.

13

For eigenfunctions corresponding to the same eigenvalue, we need to check the L2 inner product. In particular, for the eigenvalue n = (n/l)2 , we have two eigenfunctions: einx/l and einx/l . By a straightforward calculation, we see that
l l l l

einx/l einx/l dx = 0 einx/l einx/l dx = 2l.

for n = 0

Therefore, our coecients Cn would need to be given by 1 Cn = 2l

l l

(x)einx/l dx.

Consequently, the complex form of the full Fourier series for a function dened on (l, l) is given by

n=

Cn e

inx/l

1 where Cn = 2l

l l

(x)einx/l dx.

2.3
2.3.1

Fourier Transforms
Motivation

Ref: Strauss, Section 12.3 We would now like to turn to studying the heat equation on the whole real line. Consider the initial-value problem, ut = kuxx , < x < (2.8) u(x, 0) = (x). In the case of the heat equation on an interval, we found a solution u using Fourier series. For the case of the heat equation on the whole real line, the Fourier series will be replaced by the Fourier transform. Above, we discussed the complex form of the full Fourier series for a given function . In particular, for a function dened on the interval [l, l] we dene its full Fourier series as

n=

Cn e

inx/l

1 where Cn = 2l

l l

(x)einx/l dx.

Plugging the coecients Cn into the innite series, we see that

n=

1 2l

l l

(y)einy/l dy einx/l .

Now, letting k = n/l, we can write this as

n=

1 2

l l

(y)ei(yx)k dy 14

. l

The distance between points k is given by k = /l. As l +, we can think of k dk and the innite sum becoming an integral. Roughly, we have 1 2

(y)ei(yx)k dy dk.

Below we will use this motivation to dene the Fourier transform of a function and then show how the Fourier transform can be used to solve the heat equation (among others) on the whole real line, and more generally in Rn . 2.3.2 Denitions and Properties of the Fourier Transform |f (x)| dx < +.

We say f L1 (Rn ) if

Rn

For f L1 (Rn ), we dene its Fourier transform at a point Rn as f () 1 eix f (x) dx. (2)n/2 Rn

We dene its inverse Fourier transform at the point Rn as f () 1 eix f (x) dx. (2)n/2 Rn

Remark. Sometimes the constants in front are dened dierently. I.e. - in some books, the 1 1 Fourier transform is dened with a constant of (2)n instead of (2)n/2 , and in accordance the 1 inverse Fourier transform is dened with a constant 1 replacing the constant (2)n/2 above. Theorem 6. (Plancherels Theorem) If u L1 (Rn ) L2 (Rn ), then u, u L2 (Rn ) and ||u||L2 (Rn ) = ||||L2 (Rn ) = ||u||L2 (Rn ) . u In order to prove this theorem, we need to prove some preliminary facts. Claim 7. Let Then f () = Proof. f () = 1 2 eix e |x| dx n/2 (2) Rn 1 2 = eix1 1 e x1 dx1 n/2 (2) 15 f (x) = e |x| . 1 2 e|| /4 . n/2 (2 )
2

eixn n e xn dxn .

Therefore, we just need to look at

eix e x dx.

Completing the square, we have x2 ix = = Therefore, we have

x2 + x+

ix i 2

+
2

i 2 +

i 2
2

i 2
2

eix e x dx =

e [x+(i/2 )] e

2 /4

dx.

Now making the change of variables, z = [x + (i/2 )], we have

e [x+(i/2 )] e

2 /4

dx = e

2 /4

e z dz

where is the line in the complex plane given by z C:y =x+ i ,x R . 2 e z dz

R
2

Without loss of generality, we assume > 0. A similar analysis works if < 0. Now e z dz = lim

R+

where R is the line segment in the complex plane given by R z C:z =x+ i , |x| R . 2

Now dene 1 , 2 and 3 as shown in the picture below. That is, R R R 1 {x R : |x| R} R 2 R 3 R z C : z = x + iy, x, y R, x = R, 0 y 2 2 .

z C : z = x + iy; x, y R; x = R, 0 y

Im(z)

/2

2 R R Re(z)

R
16

From complex analysis, we know that e z dz = 0

C
2

where C is the closed curve given by C = R 1 2 3 traversed in the counter-clockwise R R R direction. Therefore, we have e z dz =
R R
2

e z dz

where the integral on the right-hand side is the line integral given by R = 3 1 2 R R R traversed in the direction shown. Therefore, e z dz = lim

R+

e z dz.
R

But, as R +, e z dz 0 e z dz
1 R
2 2

j R

forj = 2, 3

e x dx.

Therefore, e z dz =

e x dx.

Consequently, we have

eix e x dx = e =e =e =

2 /4

e z dz

2 /4

e x dx
e2 dx ex

2 /4 2 /4

Therefore, we have 1 f () = (2)n/2 = as claimed. 17 e1 /4

2

en /4

1 2 e|| /4 , n/2 (2 )

Claim 8. Let w(x) = u v(x) =

That is, let w be the convolution of u and v. Then w() = u v() = (2)n/2 u()v(). Proof. By denition, w() = 1 (2)n/2 1 = (2)n/2 1 = (2)n/2 1 = (2)n/2

u(x y)v(y) dy L1 (Rn ).

Rn Rn Rn Rn

eix w(x) dx eix

Rn

u(x y)v(y) dy dx

Rn

ei(xy) u(x y) dx eiy v(y) dy

(2)n/2 u()eiy v(y) dy

= u()

= u()(2)n/2 v() = (2)n/2 u()v().

eiy v(y) dy
n

Now we use these two claims to prove Plancherels Theorem. Proof of Theorem 6. By assumption, u L1 (Rn ) L2 (Rn ). Let v(x) u(x) w(x) u v(x). Therefore, v L1 (Rn ) L2 (Rn ) and w L1 (Rn ) L(Rn ). First, we have v() = 1 (2)n/2 1 = (2)n/2 1 = (2)n/2 1 = (2)n/2 = u(). Therefore, using Claim 8, w() = (2)n/2 u()v() = (2)n/2 u()u() = (2)n/2 |u|2 . 18

R R R R

eix v(x) dx
n

eix u(x) dx
n

eiy u(y) dy
n

eiy u(y) dy

Next, we use the fact that if f, g are in L1 (Rn ), then f , g are in L (Rn ), and, moreover,

Rn

f (x)g(x) dx =

Rn

f ()g() dx.

(2.9)

This fact can be seen by direct substitution, as shown below,

Rn

f (x)g(x) dx = = =
2

Rn Rn Rn

f (x)

1 eix f (x) dx g() d (2)n/2 Rn f ()g() d.

1 eix g() d dx n/2 (2) Rn

Therefore, letting f (x) = e |x| and letting g(x) = w(x) as dened above, substituting f and g into (2.9) and using Claim 7 to calculate the Fourier transform of f , we have

Rn

e || w() d =

1 2 e|x| /4 w(x) dx. n/2 Rn (2 ) 0+ . First,

(2.10)

Now we take the limit of both sides above as lim +

0
2

Rn

e || w() d =

Rn

w() d.

(2.11)

Second, we claim

1 2 e|x| /4 w(x) dx = (2)n/2 w(0). n/2 0 (2 ) Rn We prove this claim as follows. In particular, we will prove that lim + 1 2 e|x| /4 w(x) dx w(0) as 0+ . n/2 (4 ) Rn First, we note that 1 2 e|x| /4 dx = 1. n/2 (4 ) Rn This follows directly from the fact that

(2.12)

(2.13)

ez dz =

Therefore, 1 1 2 2 e|x| /4 w(x) dx w(0) = e|x| /4 [w(x) w(0)] dx. n/2 n/2 (4 ) (4 ) Rn Rn Now, we will show that for all > 0 there exists an > 0 such that 1 2 e|x| /4 [w(x) w(0)] dx < (4 )n/2 Rn 19

for 0 < < , thus, proving (2.12). Let B(0, ) be the ball of radius about 0. (We will choose suciently small below.) Now break up the integral above into two pieces, as follows, 1 1 2 2 e|x| /4 [w(x) w(0)] dx e|x| /4 [w(x) w(0)] dx n/2 n/2 (4 ) (4 ) Rn B(0,) 1 2 + e|x| /4 [w(x) w(0)] dx n/2 (4 ) Rn B(0,) I + J. First, for term I, we have 1 (4 )n/2 e|x|
B(0,)
2 /4

[w(x) w(0)] dx |w(x) w(0)|L (B(0,)) < , 2

1 2 e|x| /4 dx n/2 (4 ) Rn

for suciently small, using the fact that w C(Rn ) and (2.13). Now for xed small, we look at term J, 1 2 e|x| /4 [w(x) w(0)] dx n/2 (4 ) Rn B(0,) 1 2 e|x| /4 |w(x)| dx n/2 (4 ) Rn B(0,) 1 2 + e|x| /4 |w(0)| dx n/2 (4 ) Rn B(0,) 1 2 e|x| /4 |w(x)| dx n/2 (4 ) L (Rn B(0,)) Rn 2n/2 2 e|x| /8 dx n/2 (8 ) Rn B(0,) 1 1 2 2 2 C e|| /4 + Ce /8 e|x| /8 dx n/2 n/2 (4 ) (8 ) Rn 1 2 2 e|| /4 + Ce /8 C n/2 (4 ) < 2 + e
2 /8

|w(0)|

for

suciently small, using the fact that for a xed = 0, lim +

0

1 2 2 e|| /4 = 0 = lim e /8 . n/2 0+ (4 ) I +J <

Therefore, we conclude that for chosen suciently small, and, thus, (2.12) is proven. 20

Now combining (2.11) and (2.12) with (2.10), we conclude that

Rn

w() d = (2)n/2 w(0).

Now w() = (2)n/2 |u|2 and w(0) = u v(0) = Rn u(x)v(0 x) dx = Rn u(x)u(x) dx = 2 Rn |u| dx. Therefore, we conclude that (2)n/2 or ||u||L2 = ||u||L2 , as claimed. A similar technique can be used to show that ||||L2 = ||u||L2 . u

Rn

|u|2 d = (2)n/2

Rn

|u|2 dx,

Dening the Fourier Transform on L2 (Rn ). For f L1 (Rn ), that is, f such that

Rn

|f (x)| dx < +

it is clear that the Fourier transform is well-dened, i.e. - the integral converges. If f / 1 n L (R ), the integral may not converge. Here we describe how we dene the Fourier transform of a function f L2 (Rn ) (but which may not be in L1 (Rn )). Let f L2 (Rn ). Approximate f by a sequence of functions {fk } such that fk L1 (Rn ) L2 (Rn ) and ||fk f ||L2 0 as k +. By Plancherels theorem, ||fk fj ||L2 = ||fk fj ||L2 = ||fk fj ||L2 0 as k, j +. Therefore, {fk } is a Cauchy sequence in L2 (Rn ), and, therefore, converges to some g L2 (Rn ). We dene the Fourier transform of f by this function g. That is, f g. Other Properties of the Fourier Transform. Assume u, v L2 (Rn ). Then (a)

Rn

u(x)v(x) dx =

Rn

u()v() d.

(b) xii u() = (ii )i u().

(c) u v() = (2)n/2 u()v(). (d) u = u 21

Proof of (a). Let C. By Plancherels Theorem, we have ||u + v||L2 = ||u + v||L2 = Now using the fact that |a + b|2 = (a + b) (a + b) = |a|2 + ab + ab + |b|2 , we have

Rn

|u + v|2 dx =

Rn

|u + v|2 d

Rn

|u|2 + u(v) + u(v) + |v|2 dx =

Rn

|u|2 + u(v) + u(v) + |v|2 d,

which implies

Rn
Letting = 1, we have

u(v) + u(v) dx =

Rn Rn Rn

u(v) + u(v) d.

Rn
Letting = i, we have

uv + uv dx =

uv + uv d.

(2.14)

Rn

iuv iuv dx =

iuv iuv d.

(2.15)

Multiplying (2.15) by i, we have

Rn

uv + uv dx =

Rn

uv + uv d.

(2.16)

Adding (2.14) and (2.16), we have 2

Rn

uv dx = 2

Rn

uv d,

and, therefore, property (a) is proved. Proof of (b). We assume u is smooth and has compact support. We can use an approximation argument to prove the same equality for any u such that xii u L2 (Rn ). Using the denition of Fourier transform and integrating by parts, we have xi u() = 1 eix xi u(x) dx n/2 (2) Rn 1 = ii eix u(x) dx (2)n/2 Rn ii = eix u(x) dx (2)n/2 Rn = ii u(). 22

Property (b) follows by applying the same argument to higher derivatives. Proof of (c). Property (c) was already proved in Claim 8 in the case when u and v are in L1 (Rn ). One can use an approximation argument to prove the result in the general case. Proof of (d). Fix z Rn . Dene the function v (x) eixz |x| . Therefore, v () = 1 eix v (x) dx (2)n/2 Rn 1 2 = eix eixz e |x| dx n/2 (2) Rn 1 2 = eix(z) e |x| dx n/2 (2) Rn = f ( z) where By Claim 7, we know that f () = Therefore, we have f (x) e |x| . 1 2 e|| /4 . n/2 (2 )
2 2

1 2 e|z| /4 . n/2 (2 ) Now using (2.9), with f = u and g = v , we have v () =

Rn
= Taking the limit as

u()v () d =
2

Rn

u()eiz || d =

1 2 u(x)e|xz| /4 dx n/2 (2 ) Rn
2

Rn

u(x)v (x) dx

0+ , we have lim +
0

u()eiz || d =

u()eiz d
n

and

1 2 u(x)e|xz| /4 dx = (2)n/2 u(z), n/2 0 (2 ) Rn using (2.12). Therefore, we conclude that lim + (2)n/2 u(z) = which implies u(z) = as desired.

eiz u() d,

1 eiz u() d = u(z), (2)n/2 Rn

23

2.4
2.4.1

Solving the Heat Equation in Rn

Using the Fourier Transform to Solve the Initial-Value Problem for the Heat Equation in Rn

Consider the initial-value problem for the heat equation on R, ut = kuxx u(x, 0) = (x). x R, t > 0

Applying the Fourier transform to the heat equation, we have ut (, t) = k uxx (, t) = ut = k(i)2 u = k 2 u Solving this ODE and using the initial condition u(x, 0) = (x), we have u(, t) = ()ek t . Therefore, u(x, t) = = = = =
2 2

1 eix u(, t) d (2)1/2 R 1 2 eix ()ek t d 2 1 1 2 eix eiy (y) dy ek t d 2 2 1 1 2 (y) ei(yx) ek t d dy 2 2 1 (y)f (y x) dy 2
2 2 1 ez /4kt . 2kt

where f () = ek t . By Claim 7, f () = ekt implies f (z) = f (y x) = 1 (xy)2 /4kt e , 2kt

Therefore,

and consequently, the solution of the initial-value problem for the heat equation on R is given by 1 2 u(x, t) = (2.17) (y)e(xy) /4kt dy 4kt for t > 0. We can use a similar analysis to solve the problem in higher dimensions. Consider the initial-value problem for the heat equation in Rn , ut = ku u(x, 0) = (x). 24 x Rn , t > 0 (2.18)

Employing the Fourier transform as in the 1-D case, we arrive at the solution formula u(x, t) = for t > 0. Remark. Above, we have shown that if there exists a solution u of the heat equation (2.18), then u has the form (2.19). It remains to verify that this solution formula actually satises the initial-value problem. In particular, looking at the the formulas (2.17) and (2.19), we see that the functions given are not actually dened at t = 0. Therefore, to say that the solution formulas actually satisfy the initial-value problems, we mean to say that limt0+ u(x, t) = (x). Theorem 9. (Ref: Evans, p. 47) Assume C(Rn ) L (Rn ), and dene u by (2.19). Then 1. u C (Rn (0, )). 2. ut ku = 0 for all x Rn , t > 0. 3.
(x,t)(x0 ,0) x0 ,x n ,t>0

1 2 (y)e|xy| /4kt dy (4kt)n/2 Rn

(2.19)

lim

u(x, t) = (x0 ).
|x|2

1 Proof. 1. Since the function tn/2 e 4kt is innitely dierentiable, with uniformly bounded derivatives of all orders on Rn [, ) for all > 0, we are justied in passing the derivatives inside the integral inside and see that u C (Rn (0, )). 1 2. By a straightforward calculation, we see that the function H(x, t) (4kt)n/2 e|xy| /4kt satises the heat equation for all t > 0. Again using the fact that this function is innitely dierentiable, we can justify passing the derivatives inside the integral and conclude that
2

ut (x, t) ku(x, t) =

Rn

[(Ht kx H)(x y, t)](y) dy = 0.

since H(x, t) solves the heat equation. 3. Fix a point x0 Rn and > 0. We need to show there exists a > 0 such that |u(x, t) (x0 )| < for |(x, t) (x0 , 0)| < . That is, we need to show 1 2 e|xy| /4kt (y) dy (x0 ) < n/2 (4kt) Rn

25

for |(x, t) (x0 , 0)| < where is chosen suciently small. The proof is similar to the proof of (2.12). In particular, using (2.13), we write 1 2 e|xy| /4kt (y) dy (x0 ) n/2 (4kt) Rn 1 2 e|xy| /4kt [(y) (x0 )] dy . = n/2 (4kt) Rn

(2.20)

Let B(x0 , ) be the ball of radius about x0 . We look at the integral in (2.20) over B(x0 , ). Using the fact that is continuous, we see that 1 (4kt)n/2 e|xy|
B(x0 ,)
2 /4kt

[(y) (x0 )] dy |(y) (x0 )|L (B(x0 ,)) <

for chosen suciently small. For this choice of , we look at the integral in (2.20) over the complement of B(x0 , ). For y Rn B(x0 , ), |x0 x| < , we have 2 1 < |y x| + |y x0 |. 2 2 Therefore, on this piece of the integral, we have |y x0 | < 2|y x|. Therefore, this piece of the integral is bounded as follows, |y x0 | |y x| + |x0 x| < |y x| + 1 C 2 2 e|xy| /4kt [(y) (x0 )] dy n/2 e|x0 y| /16kt dy. n/2 (4kt) t Rn B(x0 ,) Rn B(x0 ,) We claim the right-hand side above can be made arbitrarily small by taking t arbitrarily small. We prove this in the case of n = 1. The proof for higher dimensions is similar. By making a change of variables, we have C C 2 e(x0 y) /16kt dy = 1/2 1/2 t t RB(x0 ,) =C

e ey /16kt dy
2 2

/ t

ez dz 0 as t 0+ .

In summary, take suciently small such that the piece of the integral in (2.20) over B(x0 , ) is bounded by /2. Then, choose < /2 suciently small such that

C Part (3) of the theorem follows.

ez dz < . 2

2.4.2

The Fundamental Solution

Consider again the solution formula (2.19) for the initial-value problem for the heat equation in Rn . Dene the function 1 2 e|x| /4kt t > 0 n/2 H(x, t) (4kt) (2.21) 0 t<0 26

As can be veried directly, H is a solution of the heat equation for t > 0. In addition, we can write the solution of (2.18) as u(x, t) = [H(t) ](x) =

Rn

H(x y, t)(y) dy.

As H has these nice properties, we call H(x, t) the fundamental solution of the heat equation. Lets look at the fundamental solution a little more closely. As mentioned above, H itself a solution of the heat equation. That is, Ht kH = 0 x Rn , t > 0.

What kind of initial conditions does H satisfy? We notice that for x = 0, limt0+ H(x, t) = 0. However, for x = 0, limt0+ H(x, t) = . In addition, using (2.13), we see that

Rn

H(x, t) dx = 1

for t > 0. Therefore, limt0+ Rn H(x, t) dx = 1. What kind of function satises these properties? Well, actually, no function satises these properties. Intuitively, the idea is that a function satisfying these properties represents a point mass located at the origin. This object is known as a delta function. We emphasize that the delta function is not a function! Instead, it is part of a group of objects called distributions which act on functions. We make these ideas more precise in the next section, and then will return to discussing the fundamental solution of the heat equation.

2.5

Distributions

Ref: Strauss, Section 12.1. 2.5.1 Denitions and Examples

We begin by making some denitions. We say a function : Rn R has compact support if 0 outside a closed, bounded set in Rn . We say is a test function if is an innitely dierentiable function with compact support. Let D denote the set of all test functions. We say F : D R is a distribution if F is a continuous, linear functional which assigns a real number to every test function D. We let (F, ) denote the real number associated with this distribution. Example 10. Let g : R R be any bounded function. We dene the distribution associated with g as the map Fg : D R which assigns to a test function the real number

(Fg , ) =

g(x)(x) dx.

That is, Fg :

g(x)(x) dx. 27

One particular example is the Heaviside function, dened as H(x) = 1 0 x0 x < 0.

Then, the distribution FH associated with the Heaviside function is the map which assigns to a test function the real number

(FH , ) =
0

(x) dx.

That is, FH :

(x) dx.

Example 11. The Delta function 0 (which is not actually a function!) is the distribution 0 : D R which assigns to a test function the real number (0). That is, (0 , ) = (0). Remarks. (a) We sometimes write

Rn

0 (x)(x) dx = (0),

however, this is rather informal and not accurate because 0 (x) is not a function! It should be thought of as purely notational. (b) We can talk about the delta function centered at a point other than x = 0 as follows. For a xed y Rn , we dene y : D R to be the distribution which assigns to a test function the real number (y). That is, (y , ) = (y).

2.5.2

Derivatives of Distributions

We now dene derivatives of distributions. Let F : D R be a distribution. We dene the derivative of the distribution F as the distribution G : D R such that (G, ) = (F, ). for all D. We denote the derivative of F by F . Then F is the distribution such that (F , ) = (F, ) for all D.

28

Example 12. For g a C 1 function, we dene the distribution associated with g as Fg : D R such that (Fg , ) = g(x)(x) dx.

Therefore, integrating by parts, we have

(Fg , ) =

g(x) (x) dx

g (x)(x) dx.

By denition, the derivative of Fg , denoted Fg is the distribution such that (Fg , ) = (Fg , ) for all D. Therefore, Fg is the distribution associated with the function g . That is,

(Fg , ) = (Fg , ) =

g (x)(x) dx.

Example 13. Let H be the Heaviside function dened above. Let FH : D R be the distribution associated with H, discussed above. Then the derivative of FH , denoted FH must satisfy (FH , ) = (FH , ) for all D. Now

(FH , ) =
0

(x) dx
b

= lim

b b

(x) dx
0

x=b = lim (x)|x=0

= (0). Therefore, (FH , ) = (FH , ) = (0). That is, the derivative of the distribution associated with the Heaviside function is the delta function. We will be using this fact below. 2.5.3 Convergence of Distributions

Let Fn : D R be a sequence of distributions. We say Fn converges weakly to F if (Fn , ) (F, ) for all D.

29

2.5.4

The Fundamental Solution of the Heat Equation Revisited

In this section, we will show that the fundamental solution H of the heat equation (2.21) can be thought of as a solution of the following initial-value problem, Ht kH = 0 H(x, 0) = 0 . x Rn , t > 0 (2.22)

Now, rst of all, we must be careful in what we mean by saying that H(x, 0) = 0 . Clearly, the function H dened in (2.21) is not even dened at t = 0. And, now were asking that it equals a distribution? We need to make this more precise. First of all, when we write H(x, 0) = 0 , we really mean = in the sense of distributions. In addition, we really mean to say that limt0+ H(x, t) = 0 in the sense of distributions. Lets state this more precisely now. Let FH(t) be the distribution associated with H(t) dened by (FH(t) , ) =

Rn

H(x, t)(x) dx.

Now to say that limt0+ H(x, t) = 0 in the sense of distributions, we mean that
t0+

lim (FH(t) , ) = (0 , ) = (0).

(2.23)

Therefore, to summarize, by saying that H is a solution of the initial-value problem (2.22), we really mean that H is a solution of the heat equation for t > 0 and (2.23) holds. We now prove that in fact, our fundamental solution dened in (2.21) is a solution of (2.22) in this sense. Showing that H satises the heat equation for t > 0 is a straightforward calculation. Therefore, we only focus on showing that the initial condition is satised. In particular, we need to prove that (2.23) holds. We proceed as follows.
t0+

lim (FH(t) , ) = lim +

t0

1 2 e|x| /4kt (x) dx. n/2 t0 (4kt) Rn But, now in (2.12), we proved that this limit is exactly (0)! Therefore, we have proven (2.23). Consequently, we can think of our fundamental solution as a solution of (2.22). The beauty of this formulation is the following. If H is a solution of (2.22), then dene = lim + u(x, t) [H(t) ](x) =

Rn

H(x, t)(x) dx

Rn

H(x y, t)(y) dy.

The idea is that u should be a solution of (2.18). We give the formal argument below. u(x, 0) = [H(0) ](x) = In addition, ut kx u = =

Rn

H(x y, 0)(y) dy = (x).

Rn Rn

Ht (x y, t)(y) dy

Rn

kx H(x y, t)(y) dy

[Ht (x y, t) kx H(x y, t)](y) dy = 0. 30

Note: These calculations are formal in the sense that we are ignoring convergence issues, etc. To verify that this formulation actually gives you a solution, you need to deal with these issues. Of course, for nice initial data, we have shown that this formulation does give us a solution to the heat equation.

2.6
2.6.1

Properties of the Heat Equation

Invariance Properties ut = kuxx , < x < .

Consider the equation The equation satises the following invariance properties, (a) The translate u(x y, t) of any solution u(x, t) is another solution for any xed y. (b) Any derivative (ux , ut , uxx , etc.) of a solution is again a solution. (c) A linear combination of solutions is again a solution. (d) An integral of a solution is again a solution (assuming proper convergence.) (e) If u(x, t) is a solution, so is the dilated function u( ax, at) for any a > 0. This can be proved by the chain rule. Let v(x, t) = u( ax, at). Then vt = aut and vx = And, therefore, vxx = 2.6.2 Properties of Solutions a aux .

auxx = auxx .

1. Smoothness of Solutions. As can be seen from the above theorem, solutions of the heat equation are innitely dierentiable. Therefore, even if there are singularities in the initial data, they are instantly smoothed out and the solution u(x, t) C (Rn (0, )). 2. Domain of Dependence. The value of the solution at the point x, time t depends on the value of the initial data on the whole real line. In other words, there is an innite domain of dependence for solutions to the heat equation. This is in contrast to hyperbolic equations where solutions are known to have nite domains of dependence.

31

2.7

Inhomogeneous Heat Equation

Ref: Strauss, Sec. 3.3; Evans, Sec. 2.3.1.c. In this section, we consider the initial-value problem for the inhomogeneous heat equation on some domain (not necessarily bounded) in Rn , ut ku = f (x, t) u(x, 0) = (x). x , t > 0 (2.24)

We claim that we can use solutions of the homogeneous equation to construct solutions of the inhomogeneous equation. 2.7.1 Motivation ut + au = f (t) u(0) = , where a is a constant. By multiplying by the integrating eat , we see the solution is given by u(t) = eat +
0 t

Consider the following ODE:

ea(ts) f (s) ds.

In other words, the solution u(t) is the propagator eat applied to the initial data, plus the propagator convolved with the nonlinear term. In other words, if we let S(t) be the operator which multiplies functions by eat , we see that the solution of the homogeneous problem, ut + au = 0 u(0) = is given by S(t) = eat . Further, the solution of the inhomogeneous problem is given by
t

u(t) = S(t) +
0

S(t s)f (s) ds.

We claim that this same technique will allow us to nd a solution of the inhomogeneous heat equation. Being able to construct solutions of the inhomogeneous problem from solutions of the homogeneous problem is known as Duhamels principle. We show below how this idea works. Suppose we can solve the homogeneous problem, ut ku = 0 u(x, 0) = (x). x (2.25)

That is, assume the solution of (2.25) is given by uh (x, t) = S(t)(x) for some solution operator S(t). We claim that the solution of the inhomogeneous problem (2.24) is given by
t

u(x, t) = S(t)(x) +
0

S(t s)f (x, s) ds.

32

At least formally, we see that

t

ut ku = [t k]S(t)(x) + [t k]
0 t

S(t s)f (x, s) ds

= 0 + S(t t)f (x, t) +

0

[t k]S(t s)f (x, s) ds

= S(0)f (x, t) = f (x, t). Below, we show how we use this idea to construct solutions of the heat equation on Rn and on bounded domains Rn . 2.7.2 Inhomogeneous Heat Equation in Rn

Consider the inhomogeneous problem, ut ku = f (x, t) u(x, 0) = (x). x Rn , t > 0 (2.26)

From earlier, we know the solution of the corresponding homogeneous initial-value problem ut ku = 0 u(x, 0) = (x) is given by uh (x, t) = x Rn , t > 0 (2.27)

Rn

H(x y, t)(y) dy.

That is, we can think of the solution operator S(t) associated with the heat equation on Rn as dened by S(t)(x) =

Rn

H(x y, t)(y) dy.

Therefore, we expect the solution of the inhomogeneous heat equation to be given by

t

u(x, t) = S(t)(x) +
0

S(t s)f (x, s) ds

t

Rn

H(x y, t)(y) dy +
0

Rn

H(x y, t s)f (y, s) dy ds.

Now, we have already shown that uh (x, t) = Rn H(x y, t)(y) dy satises (2.27), (at least for nice functions ). Therefore, it suces to show that
t

up (x, t)
0

Rn

H(x y, t s)f (y, s) dy ds

(2.28)

satises (2.26) with zero initial data. If we can prove this, then u(x, t) = uh (x, t) + up (x, t) will clearly solve (2.26) with initial data . In the following theorem, we prove that up satises the inhomogeneous heat equation with zero initial data. 33

2 Theorem 14. Assume f C1 (Rn [0, )) (meaning f is twice continuously dierentiable in the spatial variables and once continuously dierentiable in the time variable) and has compact support. Dene u as in (2.28). Then 2 1. u C1 (Rn (0, )).

2. ut (x, t) ku(x, t) = f (x, t) for all x Rn , t > 0. 3.

(x,t)(x0 ,0) x n ,t>0

lim

u(x, t) = 0 for all x0 Rn .

Proof. 1. Since H has a singularity at (0, 0), we cannot justify passing the derivatives inside the integral. Instead, we make a change of variables as follows. In particular, letting y = x y and s = t s, we have
t 0 t

Rn

H(x y, t s)f (y, s) dy ds =

0

Rn

H(y, s)f (x y, t s) dy ds.

2 For ease of notation, we drop the notation. Now by assumption f C1 (Rn [0, )) and H(y, s) is smooth near s = t > 0. Therefore, we have t t

t
0

Rn

H(y, s)f (x y, t s) dy ds =
0

Rn

H(y, s)t f (x y, t s) dy ds H(y, t)f (x y, 0) dy

+ and
t t

Rn

xi xj

Rn

H(y, s)f (x y, t s) dy ds =
0

Rn

H(y, s)xi xj f (x y, t s) dy ds.

2 Therefore, u C1 (Rn (0, )).

2. Now we need to calculate ut ku. Using the same change of variables as above, we have
t

[t kx ]
0 t

Rn

H(y, s)f (x y, t s) dy ds

=
0 t

Rn Rn

H(y, s)[t kx ]f (x y, t s) dy ds +

Rn

H(y, t)f (x y, 0) dy H(y, t)f (x y, 0) dy.

=
0

H(y, s)[s ky ]f (x y, t s) dy ds +

Rn

Now, we would like to integrate by parts to put the derivatives on H as we know H is a solution of the heat equation. However, we know H has a singularity at s = 0.

34

To avoid this, we break up the integral into the intervals [0, ] and [ , t]. Therefore, we write
t

[t kx ]u = +

Rn
0

H(y, s)[s ky ]f (x y, t s) dy ds

Rn

H(y, s)[s ky ]f (x y, t s) dy ds

I + J + K. First, for J , we have

0

Rn

H(y, t)f (x y, 0) dy

Rn

H(y, s)[s ky]f (x y, t s) dy ds (||ft ||L + k||f ||L )

0

Rn

H(y, s) dy ds

C using (2.13).

Rn

H(y, t) dy C ,

For I , using the assumption that f has compact support, we integrate by parts as follows,
t

Rn

H(y, s)[s ky ]f (x y, t s) dy ds
t

= =0+ = Therefore,

Rn

[s ky ]H(y, s)f (x y, t s) dy ds
s=t

Rn Rn

H(y, s)f (x y, t s) dy
s=

H(y, )f (x y, t ) dy

Rn

H(y, t)f (x y, 0) dy

Rn

H(y, )f (x y, t ) dy K.

I +K = Now,
0

Rn

H(y, )f (x y, t ) dy.

ut ku = lim [I + J + K] + = lim +
0

= lim +

1 2 e|y| /4k f (x y, t ) dy n/2 0 (4k ) Rn = f (x, t), 35

Rn

H(y, )f (x y, t ) dy

using the same technique we used to prove part (3) of Theorem 9. 3. To prove that the limit as t 0+ of our solution u(x, t) is 0, we use the fact that
t

|u(x, t)|L (Rn ) =

Rn
t

H(y, s)f (x y, t s) dy ds
t 0

|f |L (Rn [0,t]) C
0

= Ct,

1 2 e|y| /4ks dy ds n/2 (4ks) Rn

Rn

H(y, s) dy ds

using (2.13). Therefore, as t 0+ , u(x, t) 0, as claimed.

2.7.3

Inhomogeneous Heat Equation on Bounded Domains

In this section, we consider the initial/boundary value problem for the inhomogeneous heat equation on an interval I Rn , x I R, t > 0 ut kuxx = f (x, t) u(x, 0) = (x) xI u satises certain BCs t > 0. Using Duhamels principle, we expect the solution to be given by
t

u(x, t) = S(t)(x) +
0

S(t s)f (x, s) ds

where uh (x, t) = S(t)(x) is the solution of the homogeneous equation and S(t) is the solution operator associated with the homogeneous problem. As shown earlier, the solution of the homogeneous problem with symmetric boundary conditions, x I, t > 0 ut kuxx = 0 u(x, 0) = (x) xI u satises symmetric BCs t>0 is given by u(x, t) =
n=1

An Xn (x)ekn t

where Xn are the eigenfunctions and n the corresponding eigenvalues of the eigenvalue problem, X = X xI X satises symmetric BCs, and the coecients An are dened by An = Xn , , Xn , Xn 36

where the inner product is taken over I. Therefore, the solution operator associated with the homogeneous equation is given by

S(t) =
n=1

An Xn (x)ekn t

where An =

Xn , . Xn , X n ,

Therefore, we expect the solution of the inhomogeneous equation to be given by

t

u(x, t) = S(t)(x) +
0

S(t s)f (x, s) ds

n t t

=
n=1

An Xn (x)e

+
0 n=1

Bn (s)Xn (x)en (ts) ds

where Bn (s)

Xn , f (s) . Xn , Xn

In fact, for nice functions and f , this formula gives us a solution of the inhomogeneous initial/boundary-value problem for the heat equation. We omit proof of this fact here. We consider an example. Example 15. Solve the inhomogeneous initial/boundary value problem for the heat equation on [0, l] with Dirichlet boundary conditions, x [0, l], t > 0 ut kuxx = f (x, t) u(x, 0) = (x) x [0, l] (2.29) u(0, t) = 0 = u(l, t) t>0 The solution of the homogeneous problem with initial data , x [0, l], t > 0 ut kuxx = 0 u(x, 0) = (x) x [0, l] u(0, t) = 0 = u(l, t) t>0 is given by uh (x, t) =
n=1

An sin

n 2 x ek(n/l) t , l

where

2 l n An = sin x (x) dx. l 0 l Therefore, the solution of the inhomogeneous problem with zero initial data is given by
t

up (x, t) =
0 n=1

Bn (s) sin 37

n 2 x ek(n/l) (ts) ds l

where

2 Bn (s) = l

sin
0

n x f (x, s) dx. l

Consequently, the solution of (2.29) is given by n 2 u(x, t) = x ek(n/l) t + An sin l n=1 where An and Bn (s) are as dened above.
t

Bn (s) sin
0 n=1

n 2 x ek(n/l) (ts) ds l

2.7.4

Inhomogeneous Boundary Data

In this section we consider the case of the heat equation on an interval with inhomogeneous boundary data. We will use the method of shifting the data to reduce this problem to an inhomogeneous equation with homogeneous boundary data. Consider the following example. Example 16. Consider ut kuxx = 0 u(x, 0) = (x) u(0, t) = g(t) u(l, t) = h(t). 0<x<l 0<x<l

(2.30)

We introduce a new function U(x, t) such that U(x, t) = 1 [(l x)g(t) + xh(t)] . l

Assume u(x, t) is a solution of (2.30). Then let v(x, t) u(x, t) U(x, t). Therefore, vt kvxx = (ut kuxx ) (Ut kUxx ) = Ut 1 = [(l x)g (t) + xh (t)] . l Further, v(x, 0) = u(x, 0) U (x, 0) = (x) v(0, t) = u(0, t) U(0, t) = 0 v(l, t) = u(l, t) U(l, t) = 0. 1 [(l x)g(0) + xh(0)] l

38

Therefore, v is a solution of the inhomogeneous heat equation on the interval [0, l] with homogeneous boundary data, 0<x<l vt kvxx = 1 [(l x)g (t) + xh (t)] l v(x, 0) = (x) 1 [(l x)g(0) + xh(0)] 0<x<l l v(0, t) = 0 = v(l, t). We can solve this problem for v using the technique of the previous section. Then we can solve our original problem (2.30) using the fact that u(x, t) = v(x, t) + U(x, t).

2.8

Maximum Principle and Uniqueness of Solutions

In this section, we prove what is known as the maximum principle for the heat equation. We will then use this principle to prove uniqueness of solutions to the initial-value problem for the heat equation. 2.8.1 Maximum Principle for the Heat Equation

First, we prove the maximum principle for solutions of the heat equation on bounded domains. Let Rn be an open, bounded set. We dene the parabolic cylinder as T (0, T ]. We dene the parabolic boundary of T as T T T .

t T T

We now state the maximum principle for solutions to the heat equation. Theorem 17. (Maximum Principle on Bounded Domains) (Ref: Evans, p. 54.) Let be an open, bounded set in Rn . Let T and T be as dened above. Assume u is suciently 2 smooth, (specically, assume u C1 (T ) C(T )) and u solves the heat equation in T . Then, 1. max u(x, t) = max u(x, t). (Weak maximum principle)
T T

39

2. If is connected and there exists a point (x0 , t0 ) T such that u(x0 , t0 ) = max u(x, t),
T

then u is constant in t0 . (Strong maximum principle) Proof. (Ref: Strauss, p. 42 (weak); Evans, Sec. 2.3.3 (strong)) Here, we will only prove the weak maximum principle. The reader is referred to Evans for a proof of the strong maximum principle. We will prove the weak maximum principle in the case n = 1 and = (0, l). Let M = max u(x, t).
T

We must show that u(x, t) M throughout T . Let

> 0 and dene

v(x, t) = u(x, t) + x2 . We claim that v(x, t) M + l2 throughout T . Assuming this, we can conclude that u(x, t) M + (l2 x2 ) on T and consequently, u(x, t) M on T . Therefore, we only need to show that v(x, t) M + l2 throughout T . By denition, we know v(x, t) M + l2 on T (that is, on the lines t = 0, x = 0 and x = l). We will prove that v cannot attain its maximum in T , and, therefore, v(x, t) M + l2 throughout T . We prove this as follows. First, by denition of v, we see v satises vt kvxx = ut k(u + x2 )xx = ut kuxx 2 k = 2 k < 0. (2.31)

Inequality (2.31) is known as the diusion inequality. We will use (2.31) to prove that v cannot achieve its maximum in T . Suppose v attains its maximum at a point (x0 , t0 ) T such that 0 < x0 < l, 0 < t0 < T . This would imply that vt (x0 , t0 ) = 0 and vxx (x0 , t0 ) 0, and, consequently that vt (x0 , t0 ) kvxx (x0 , t0 ) 0, which contradicts the diusion inequality (2.31). Therefore, v cannot attain its maximum at such a point. Suppose v attains its maximum at a point (x0 , T ) on the top edge of T . Then vx (x0 , T ) = 0 and vxx (x0 , T ) 0, as before. Furthermore, vt (x0 , T ) = lim +
0

v(x0 , T ) v(x0 , T ) 0.

Again, this contradicts the diusion inequality. But v must have a maximum somewhere in the closed rectangle T . Therefore, this maximum must occur somewhere on the parabolic boundary T . Therefore, v(x, t) M + l2 throughout T , and, we can conclude that u(x, t) M throughout T , as desired. 40

We now prove a maximum principle for solutions to the heat equation on all of Rn . Without having a boundary condition, we need to impose some growth assumptions on the behavior of the solutions as |x| +. Theorem 18. (Maximum Principle on Rn ) (Ref: Evans, p. 57) Suppose u (suciently smooth) solves ut ku = 0 x Rn (0, T ) u= x Rn and u satises the growth estimate u(x, t) Aea|x| for constants A, a > 0. Then
2

x Rn , 0 t T

Rn [0,T ]

sup u(x, t) = sup (x, t).

Rn

Proof. (For simplicity, we take k = 1, but the proof works for arbitrary k > 0.) Assume 4aT < 1. Therefore, there exists > 0 such that 4a(T + ) < 1. Fix y Rn and > 0. Let v(x, t) u(x, t)
|xy|2 e 4(T + t) . (T + t)n/2

Using the fact that u is a solution of the heat equation, it is straightforward to show that v is a solution of the heat equation, vt v = 0, x Rn (0, T ].

Fix r > 0 and let U B(y, r), UT = B(y, r) (0, T ]. From the maximum principle, we know max v(x, t) = max v(x, t).
UT T

We will now show that maxT v(x, t) (x). First, look at (x, t) on the base of T . That is, take (x, t) = (x, 0). Then v(x, 0) = u(x, 0)
|xy|2 e 4(T + ) (T + )n/2 u(x, 0) = (x).

41

Now take (x, t) on the sides of T . That is, choose x Rn such that |x y| = r and t such that 0 t T . Then v(x, t) = u(x, t)
r2 e 4(T + t) (T + t)n/2 r2 2 Aea|x| e 4(T + t) (T + t)n/2 r2 2 Aea(|y|+r) e 4(T + ) . (T + )n/2

By assumption, 4a(T + ) < 1. Therefore, 1/4(T + ) = a + for some > 0. Therefore, v(x, t) Aea(|y|+r) (4(a + ))n/2 e(a+)r sup (x)
2 2

Rn

for r suciently large. Therefore, we have shown that v(x, t) supRn (x) for (x, t) T , and, thus, by the maximum principle max v(x, t) = max v(x, t),
UT T

we have v(x, t) supRn (x) for all (x, t) U T . We can use this same argument for all y Rn , to conclude that v(y, t) sup (x)

Rn

for all y Rn , 0 t T as long as 4aT < 1. Then using the denition of v and taking the limit as 0+, we conclude that u(y, t) sup (x)

Rn

for all y Rn , 0 t T as long as 4aT < 1. If 4aT 1, we divide the interval [0, T ] into subintervals [0, T1 ], [T1 , 2T1 ], etc., where 1 T1 = 8a and perform the same calculations on each of these subintervals. 2.8.2 Using the Maximum Principle to Prove Uniqueness

We now use the maximum principles proven above to prove uniqueness of solutions to the heat equation. Theorem 19. (Uniqueness on bounded domains) Let be an open, bounded set in Rn and dene T and T as above. Consider the initial/boundary value problem, ut ku = f u=g x T x T (2.32)

Assume f and g are continuous. Then there exists at most one (smooth) solution of (2.32).

42

Proof. Suppose there exist two smooth solutions u and v. Let w u v. Then w is a solution of wt kw = 0 w=0 x T x T .

(2.33)

By the weak maximum principle, max w(x, t) = max w(x, t) = 0.

T T

Therefore, w(x, t) 0 on T , which implies u(x, t) v(x, t) on T . Next, let w = v u. Therefore, w is also a solution of (2.33), and, by the weak maximum principle, max w(x, t) = max w(x, t) = 0.
T T

Consequently, w(x, t) 0 on T , which implies v(x, t) u(x, t) on T . Therefore, we have u(x, t) v(x, t) (x, t) T v(x, t) u(x, t) (x, t) T . Consequently, we conclude that u = v for in T , as desired. Theorem 20. (Uniqueness on Rn ) (Ref: Evans, p. 58.) Consider the initial-value problem ut ku = f u(x, 0) = g x Rn (0, T ) x Rn . (2.34)

Assume f and g are continuous. Then there exists at most one (smooth) solution of (2.34) satisfying the growth estimate 2 |u(x, t)| Aea|x| . Proof. Assume u, v are two solutions. Let w u v and let w = v u. Now apply the maximum principle on Rn to show that u v. Remark. There are, in fact, innitely many solutions of ut ku = 0 u(x, 0) = 0 x Rn (0, T ) x Rn .
2

(Ref: John, Chapter 7.) All of those solutions other than u(x, t) 0 grow faster than e|x| as |x| +.

43

2.8.3

Energy Methods to prove Uniqueness

We now present an alternate technique to prove uniqueness of solutions to the heat equation on bounded domains. Theorem 21. Let be an open, bounded set in Rn . Let T > 0. Let T , T be the parabolic cylinder and parabolic boundary dened earlier. Consider the initial/boundary value problem, ut ku = f u(x, t) = g (x, t) T (x, t) T . (2.35)

There exists at most one (smooth) solution u of (2.35). Proof. Suppose there exist two solutions u and v. Let w = u v. Then w solves wt kw = 0 w=0 Let E(t)
2

(x, t) T (x, t) T .

(2.36)

w2 (x, t) dx.

Now, E(0) = w (x, 0) dx = 0. We claim that E (t) 0 and, therefore, E(t) = 0 for 0 t T . By using the fact that w is a solution of (2.36) and integrating by parts, we see that E (t) = 2

wwt dx ww dx

= 2k = 2k

| w|2 dx + 2k

w dS(x)

= 2k

| w|2 dx 0,

using the fact that w = 0 on T and k > 0. Therefore, E(t) = 0 for 0 t T . Using the assumption that w is smooth, this implies w 0 in T , and, therefore, u v in T .

44

Nonhomogeneous Heat Equation Dirichlet Boundary Conditions

ut (x, t) = kuxx (x, t) + F (x, t), 0 < x < , t > 0 u(0, t) = 0, u( , t) = 0 u(x, 0) = (x) For the nonhomogeneous problem, instead of looking for a solution in the form

(1)

u(x, t) =
n=1

cn ekn t sin

nx

as we would if F (x, t) 0, we look for

u(x, t) =
n=1

un (t) sin

nx

(2)

To this end we rst nd {Fn (t)} so that

F (x, t) =
n=1

Fn (t) sin

nx

(3)

and {cn } so that (x) =

cn sin
n=1

nx

(4)

by setting Fn (t) = and cn = 2

0

2
0

F (x, t) sin

nx

dx

(5)

(x) sin

nx

dx

(6)

Substituting (2), (3), (5) into (1) and collecting all terms on one side, we obtain (we also rename the index to j)

j=1

d uj (t) kj uj (t) Fj (t) sin dt

jx

=0

(7)

and substituting (4), (6) into (2) with t = 0 we have

[uj (0) cj ] sin

j=1

jx

= 0.

(8)

nx and integrate Now we use orthogonality. Multiply both sides of (7) and (8) by sin from 0 to . Then we obtain a system of initial value problems for ordinary dierential equations d u (t) kn un (t) = Fn (t) dt n , n = 1, 2, . (9) un (0) = cn , The rst equation in (9) is rst order linear and is solved as follows: Multiply both sides by ekn t to obtain d un (t)ekn t = ekn t Fn (t). dt Now integrate both sides from 0 to t to get
t

un (t)ekn t = un (0) +
0

ekn s Fn (s) ds.

Using the second equation in (9) we have

t

un (t)e

kn t

= cn +
0

ekn s Fn (s) ds.

Multiply both sides by ekn t to obtain

t

un (t) = cn ekn t +
0

ekn (ts) Fn (s) ds

(10)

Thus we have obtained the formula

u(x, t) =
n=1

cn e

kn t

sin

nx

+
n=1 0

ekn (ts) Fn (s) ds sin

nx

(11)

Example 1. As an explicit example consider = 1 and for the initial condition and forcing terms take (x) = 0 and F (x, t) = f (x), i.e., suppose that F (x, t) is independent of t. In this case (11) becomes
t

u(x, t) =
n=1 0

ekn (ts) Fn (s) ds sin (nx)

with Fn (s) = 2
0

F (x, s) sin (nx) dx = 2

0

f (x) sin (nx) dx.

and we notice that in this example Fn (s) is independent of s and, indeed, Fn (s) is nothing more than the nth Fourier coecient of f . So we dene fn Fn (s)

Now, again since Fn is independent of s, we have

t t

e
0

kn (ts)

Fn (s) ds =
0

ekn (ts) fn ds ekn t 1 kn (12)

= fn e

kn t

= fn Now recall that n = (n)2 so we can write u(x, t) = 1 k

1 ekn t . kn

n=1

fn 1 ekn t sin (nx) (n)2

(13)

Note that the steady state solution (x) is obtained from the following limit. 1 (x) = lim u(x, t) = t k

n=1

fn sin (nx) . (n)2

Clearly, (0) = (1) = 0 and dierentiating both ends of the above equation twice we get

k (x) =
n=1

fn sin (nx) = (x).

Thus we see that satises k (x) + (x) = 0 which is the steady state solution. Example 2. As another example consider = 1 and for the initial condition and forcing terms take (x) = 0 and F (x, t) = f (x) sin(t). In this case (11) becomes
t

u(x, t) =
n=1 0

ekn (ts) Fn (s) ds sin (nx)

with Fn (s) = 2(sin(s))

0

F (x, s) sin (nx) dx = 2

0 1

f (x) sin(s) sin (nx) dx

and we dene fn by fn = 2
0

f (x) sin (nx) dx

so that Fn (s) = fn sin(s).

For this example we have

t t

ekn (ts) Fn (s) ds =

0 0

ekn (ts) sin(s)fn ds

t

= fn ekn t
0

ekn s) sin(s) ds

(14) (15) (16)

(see calculation below) = fn ekn t cos(t) kn sin(t) . (1 + (kn )2 )

In the above we need to use a special form of integration by parts which we carry out here
t t

ekn s) sin(s) ds =
0 0

ekn s) ( cos(s)) ds
t t

= ekn s) ( cos(s))
0 0

(kn )ekn s) ( cos(s)) ds

t

= ekn t) cos(t) + 1 kn
0

ekn s) (sin(s)) ds
t t

= e

kn t)

cos(t) + 1 kn e

kn s)

sin(s) + kn
0 0 t

ekn s) sin(s) ds

= e Thus we have
t

kn t)

cos(t) + 1 kn e

kn t)

sin(t) (kn )

2 0

ekn s) sin(s) ds.

(1 + (kn )2 )
0

ekn s) sin(s) ds = ekn t) cos(t) + 1 kn ekn t) sin(t) ekn t) cos(t) + 1 kn ekn t) sin(t) . (1 + (kn )2 )

so that
0

ekn s) sin(s) ds = Finally, then

t

ekn t)
0

ekn s) sin(s) ds =

cos(t) kn sin(t) + ekn t) . (1 + (kn )2 )

With this we can write the solution as

u(x, t) =
n=1

cos(t) kn sin(t) + ekn t) (1 + (kn )2 ) 4

fn sin (nx)

(17)

Notice that in this case the steady state is not independent of time. Namely we have

v(x, t) =
n=1

cos(t) kn sin(t) (1 + (kn )2 )

fn sin (nx)

which is a 2 periodic funtion of t, Example 3. Let us consider an example with non-homogeneous boundary conditions. We are interested in the system ut (x, t) = kuxx (x, t), 0 < x < , t > 0, u(0, t) = 0 (t), u( , t) = 1 (t) u(x, 0) = (x) (18) (19) (20)

Our method will involve transforming the problem (18)-(20) into a problem in the form (1). To this end we look for functions v0 (x) and v1 (x) satisfying v0 (x) = 0, v0 (0) = 1, v0 ( ) = 0, We have v1 (x) = 0, v1 (0) = 0, v1 ( ) = 1.

v0 (x) = 0, v0 (0) = 1, v0 ( ) = 0, implies v0 (x) = c1 x + c2 1 = v0 (0) = c2 c2 = 1 and 0 = v0 ( ) = c1 + 1 c1 = 1/ . Thus v0 (x) = 1 x/ . Now we seek u(x, t) in the form Thus and

v1 (x) = 0, v1 (0) = 0, v1 ( ) = 1, implies v1 (x) = c1 x + c2 0 = v1 (0) = c2 c2 = 0

1 = v1 ( ) = c1

c1 = 1/ .

v1 (x) = x/ .

u(x, t) = v(x, t) + v0 (x)0 (t) + v1 (x)1 (t) Then we have 0 = ut (x, t) uxx (x, t) = (vt (x, t) vxx (x, t)) + (v0 (x)0 (t) + v1 (x)1 (t))

Further 0 = u(x, 0) = v(x, 0) + v0 (x)0 (0) + v1 (x)1 (0) So we have v(x, 0) = (v0 (x)0 (0) + v1 (x)1 (0)) (x). 5

For the boundary conditions we have 0 (t) = u(0, t) = v(0, t) + v0 (0)0 (t) + v1 (0)1 (t) = v(0, t) + 0 (t), v(0, t) = 0, and 1 (t) = u( , t) = v( , t) + v0 ( )0 (t) + v1 ( )1 (t) = v( , t) + 1 (t), v( , t) = 0. Thus we arrive at the following Initial Boundary Value Problem for v(x, t) vt (x, t) = vxx (x, t) + F (x, t), v(0, t) = 0, v( , t) = 0, v(x, 0) = (x), F (x, t) = ((1 x/ )0 (t) + x/ (x)1 (t)) . We can apply the formula (11), which we list again here for v(x, t) solving (21) ,

(21)

v(x, t) =
n=1

cn e

kn t

sin

nx

+
n=1 0

ekn (ts) Fn (s) ds sin

nx

(22)

Recall that from (5) and (6) Fn (s) = and cn = Thus we need Fn (s) = 2
0

2
0

F (x, s) sin

nx

dx

2
0

(x) sin

nx

dx.

((1 x/ )0 (s) + x/ (x)1 (s)) sin 2

0

nx

dx

= 0 (s) 1 (s) and cn = 2

0

(1 x/ ) sin x/ sin
0

nx

dx

nx

dx

(x) sin 2
0

nx

dx nx dx

((1 x/ )0 (0) + (x/ )1 (0)) sin 2

0

= 0 (0) 1 (0) 2

(1 x/ ) sin (x/ ) sin nx

nx

dx

dx

Therefore, in order to obtain a nal result we need formulas for 2

0

(1 x/ ) sin

nx

dx, and

2
0

(x/ ) sin

nx

dx.

We introduce the change of variables = x/ to obtain 2

0

(1 x/ ) sin

nx

dx = 2
0

(1 ) sin(n) d = 2
0

(1 )
1

cos(n) n

= 2 (1 ) 1 1 n n

cos(n) n
1

0 0

cos(n) d n

=2

cos(n) d
0 1 0

1 1 sin(n) =2 n (n)2 = and 2

0

2 . n
1 1

(x/ ) sin

nx

dx = 2
0

sin(n) d = 2
0

cos(n) n cos(n) d n

=2

cos(n) n

+
0 1 0

=2

cos(n) 1 + n n

cos(n) d
0 1 0

=2

(1)(n+1) 1 + sin(n) n (n)2

2(1)(n+1) . n

From this we obtain the Fourier Sine series for (1 x/ ) and x/ on (0, ): 2 (1 x/ ) = and 2 x/ =

sin

nx

n=1

n
nx

n=1

(1)n sin n 7

In addition, we have Fn (s) = 0 (s) and cn = 0 (0) 2(1)n 2 + 1 (s) n n

2 2(1)n + 1 (0) . n n

Let us consider a very special case. We consider the case where 0 (t) = T0 and 1 (t) = T1 are constants. This is the problem we rst considered in class when we started this chapter. In this case 0 = 0, 1 = 0 and 0 (0) = T0 , 1 (0) = T1 , so we have Fn (s) = 0, and n = 1, 2, ,

2 ((1)n T1 T0 ) . n Thus our solution v(x, t) of (21) is given by cn =

v(x, t) =
n=1

2 ((1)n T1 T0 ) kn t nx e sin n

so we can write our desired solution u(x, t) of (18)-(20) as

u(x, t) =
n=1

2 ((1)n T1 T0 ) kn t nx e sin + (1 x/ )T0 + (x/ )T1 n

VARIATIONS OF THE SOLUTION TO A STOCHASTIC HEAT EQUATION By Jason Swanson University of Wisconsin-Madison December 30, 2005
We consider the solution to a stochastic heat equation. This solution is a random function of time and space. For a xed point in space, the resulting random function of time, F (t), has a nontrivial quartic variation. This process, therefore, has innite quadratic variation and is not a semimartingale. It follows that the classical It o calculus does not apply. Motivated by heuristic ideas about a possible new calculus for this process, we are led to study modications of the quadratic variation. Namely, we modify each term in the sum of the squares of the increments so that it has mean zero. We then show that these sums, as functions of t, converge weakly to Brownian motion.

1. Introduction. Consider the solution to the stochastic heat equation ut = 1 uxx + W (t, x) 2

with boundary conditions u(0, x) = 0, where W (t, x) is a two-dimensional white noise. We can write the solution as
t

u(t, x) =
R 0
2

p(t r, x y) dW (r, y),

where p(t, x) = (2t)1/2 ex /2t is the heat kernel. For xed x R, it will be shown that F (t) = u(t, x) has a nontrivial quartic variation. In particular, F is not a semimartingale. We therefore cannot dene a classical stochastic integral using F as our integrator.
This work was supported in part by the VIGRE grants of both University of Washington and University of Wisconsin-Madison. AMS 2000 subject classications: Primary 60F17; secondary 60G15, 60G18, 60H05, 60H15 Keywords and phrases: quartic variation, quadratic variation, stochastic partial dierential equations, stochastic integration, long-range dependence, iterated Brownian motion, fractional Brownian motion, self-similar processes

JASON SWANSON

The results in this article are motivated by the following heuristic construction of a stochastic integral with respect to a quartic variation process. This construction is the idea of Chris Burdzy and was communicated to me during my time as a graduate student. Let F be a continuous, centered, nite quartic variation process, and let g be a smooth function. Write g(x + h) = where R(x, h) = 1 4!
0

1 (j) g (x)hj + R(x, h), j! j=0

h

(h t)4 g (5) (x + t) dt.

In particular, if M = sup |g (5) (t)|, where the supremum is taken over all t between x and x + h, then (1.1) |R(x, h)| M |h|5 /5!.

Taking two dierent values of h yields (1.2) g(x + h1 ) g(x + h2 ) = 1 (j) g (x)(hj hj ) + R(x, h1 ) R(x, h2 ). 1 2 j! j=0
4

Now x n, let t = n1 and tj = jt. Fix t 0 and let N = nt/2 . Then

N

(1.3)

g(F (t)) = g(F (0)) +

j=1

{g(F (t2j )) g(F (t2j2 ))}

+ g(F (t)) g(F (t2N )). Dene F j = F (tj ) F (tj1 ) and substitute x = F (t2j1 ), h1 = F 2j , and h2 = F 2j1 into (1.2), so that (1.3) becomes
N

g(F (t)) = g(F (0)) +

j=1

g (F (t2j1 ))(F (t2j ) F (t2j2 )) g (F (t2j1 ))(F 2 F 2 ) 2j 2j1

(1.4)

1 2

N j=1

+ 1 + 2 + 3 + 4 ,

VARIATIONS OF STOCHASTIC HEAT EQ

where 1 = 2 = 3 =
j=1

1 N 1 N g (F (t2j1 ))F 3 g (F (t2j1 ))F 3 2j 2j1 3! j=1 3! j=1 1 N (4) 1 N (4) g (F (t2j1 ))F 4 g (F (t2j1 ))F 4 2j 2j1 4! j=1 4! j=1
N N

R(F (t2j1 ), F 2j )
j=1

R(F (t2j1 ), F 2j1 )

4 = g(F (t)) g(F (t2N )). By continuity, 4 0. If F has nite quartic variation V (4) (t), then under suitable assumptions on g, 3 0 by (1.1). Moreover, under the right conditions on F , we might expect both terms in 2 to converge to 1 4!
t 0

1 g (4) (F (s)) d[ V (4) (s)], 2

nt

so that 2 0. Finally, if j=1 F 3 converges uniformly to zero, then j under the right conditions we would hope to have 1 0. Hence, if we wish to dene a Stratonovich-type integral with respect to F by
t N

(1.5)
0

g (F (s)) dF (s) = lim

g (F (t2j1 ))(F (t2j ) F (t2j2 )),

j=1

then we must investigate the convergence of the second summation on the right-hand side of (1.4). Let us simplify things for the moment and assume that g = 1. In this case, we must consider the sum
N

(1.6)
j=1

(F 2 F 2 ) 2j 2j1

If EF 2 t1/2 , then each of the random variables, F 2 F 2 , has j 2j 2j1 an approximate mean of zero and an approximate variance of t. If these terms were also independent, then we would expect (1.6), as a function of t, to converge weakly as t 0 to a Brownian motion, suggesting that the integral dened by (1.5) obeys an It-type rule of the form o
t

g(F (t)) = g(F (0)) +

0

g (F (s)) dF (s) +

1 2

g (F (s))dB(s),
0

JASON SWANSON

where B is the limit of (1.6). For the process F (t) = u(t, x), the terms in (1.6) are certainly not independent. In fact, F exhibits a form of long-range dependence, in the sense that the correlation between increments decays only polynomially. To prove that (1.6) converges to Brownian motion, we rst attempted to simplify the problem by instead considering the sum
nt

(1.7)
j=1

F 2 sgn(F j ), j

whose limit we called the signed quadratic variation of F . In [19], it was shown that (1.7) converges to Brownian motion. Using the same techniques, we show here a general result that includes the convergence of both (1.6) and (1.7) as corollaries. It should be remarked that the original quartic variation process that motivated the above heuristic integral construction was iterated Brownian motion. In general, iterated diusions are related to a certain class of fourthorder parabolic dierential equations. (See [1], [3], [6], and [12].) Unlike iterated Brownian motion, however, the process we are considering is a Gaussian process with a stochastic integral representation. We make heavy use of both of these properties in the proofs. Another quartic variation process with these two properties is fractional Brownian motion with Hurst parameter H = 1/4. It therefore seems reasonable that the methods and results in this paper would extend to fractional Brownian motion as well. For a survey of dierent approaches to stochastic integration against fractional Brownian motion, see [8]. 2. The Quartic Variation of F . Dene the Hilbert space H = L2 (R2 ) and construct a centered Gaussian process, I(h), indexed by h H, such 2 that E[I(g)I(h)] = gh. Recall that p(t, x) = (2t)1/2 ex /2t . and for a xed pair (t, x), let htx (r, y) = 1[0,t] (r)p(t r, x y) H. Then
t

(2.1)

F (t) = u(t, x) =
R 0

p(t r, x y) dW (r, y) = I(htx ).

Since F is a centered Gaussian process, its law is determined by its covariance function, which is given in the following lemma. We also derive some needed estimates on the increments of F . Lemma 2.1. (2.2) For all s, t [0, ), 1 EF (s)F (t) = (|t + s|1/2 |t s|1/2 ). 2

VARIATIONS OF STOCHASTIC HEAT EQ

If 0 s < t, then (2.3) E|F (t) F (s)|2 2(t s) 1 3/2 |t s|2 . t

For xed t > 0, dene tj = jt, and let F j = F (tj )F (tj1 ). If i, j N with i < j, then (2.4) E[F i F j ] + t 1 t2 , ji 2 (ti + tj )3/2

where j = 2 j j 1 j + 1. Proof. For (2.2), we may assume s t. By (2.1),

s

E[F (s)F (t)] =

R 0 s

p(t r, x y)p(s r, x y) dr dy (2)1 (t r)(s r) (2)1 (t r)(s r) exp

R

=
0 s

(x y)2 (x y)2 2(t r) 2(s r) (x y)2 (t + s 2r) 2(t r)(s r)

dy dr dy dr.

=
0

exp
R

Since (2)1/2 exp{(x y)2 /2c} dy = 1 E[F (s)F (t)] = 2

s

c, we have

1 1 dr = (|t + s|1/2 |t s|1/2 ), t + s 2r 2

which veries the formula. For (2.3), let 0 s < t. Then (2.2) implies (2.5) Thus, E|F (t) F (s)|2 2(t s) 1 = | t + s 2t + 2s| 1 ( t s)2 = , t + s + 2t + 2s 1 E|F (t) F (s)|2 = ( t + s 2t + 2s + 2t 2s).

JASON SWANSON

which gives E|F (t) F (s)|

2

2(t s) ( t s)2 (1 + 2) t = |t s|2 . (1 + 2) t ( t + s)2

Hence, (2.6) E|F (t) F (s)|2 1 2(t s) |t s|2 , (1 + 2)t3/2

which proves (2.3). Finally, for (2.4), x i < j. Observe that for any k i, E[F (tk )F i ] = E[F (tk )F (ti ) F (tk )F (ti1 )] 1 = ( tk + ti tk ti tk + ti1 + tk ti1 ) 2 t = ( k + i k i k + i 1 + k i + 1). 2 Thus, E[F i F j ] = E[F (tj )F i ] E[F (tj1 )F i ] = which simplies to (2.7) E[F i F j ] = t (j+i1 + ji ). 2 t ( j+i 2 j+i1+ ji j+i1+ ji+1 j i),

ji1+

j+i2

The strict concavity of x x implies that > 0 for all k N. Also, if k we write k = f (k 1) f (k), where f (x) = x + 1 x, then for each k 2, the Mean Value Theorem gives k = |f (k )| for some [0, 1]. Since |f (x)| x3/2 /4, we can easily verify that for all k N, (2.8) 0 < k 1 . 2 k 3/2

VARIATIONS OF STOCHASTIC HEAT EQ

Since j + i 1 (j + i)/2, we have E[F i F j ] + t ji 2 1 t = 2 2 ((j + i)/2)3/2 1 2t , (i + j)3/2

and this proves (2.4). 2 By (2.2), the law of F (t) = u(t, x) does not depend on x. We will therefore assume that x = 0. Note that (2.6) implies (2.9) 1/2 t EF 2 2 t j for all j 1. In particular, since F is Gaussian, we have E|F (t) F (s)|4n Cn |t s|n for all n. By the Kolmogorov-Centsov Theorem (see, for example, Theorem 2.2.8 in [14]), F has a modication which is locally Hlder contino uous with exponent for all (0, 1/4). We will henceforth assume that we are working with such a modication. Also note that (2.7) and (2.8) together imply 2t2 2 t = E[F i F j ] < 0 (tj ti )3/2 (j i)3/2 for all 1 i < j. In other words, the increments of F are negatively correlated and we have a polynomial bound on the rate of decay of this correlation. For future reference, let us combine these results into the following single inequality: for all i, j N, 2 t (2.10) |E[F i F j ]| , |i j|3/2 where we have adopted the notation xr = (x 1)r . In fact, with a little more work, we have the following general result. Lemma 2.2. (2.11) For all 0 s < t u < v, 2 |t s||v u| . |u s| v t

|E[(F (v) F (u))(F (t) F (s))]|

Proof. Fix 0 s < t. For any r > t, dene 1 f (r) = E[F (r)(F (t) F (s))] = ( r + t r t r + s + r s). 2 Then 1 f (r) = 2 2 r+s r+t rs rt . (r + t)(r + s) (r t)(r s)

JASON SWANSON

Since

|t s| rs rt , r t |r s| (r t)(r s)

we have
v 1 |t s| |E[(F (v) F (u))(F (t) F (s))]| dr r t |r s| 2 u 1 1 |t s| v dr rt 2 |u s| u 2 |t s| = ( v t u t) |u s| 2 |t s||v u| |u s| v t

whenever 0 s < t u < v. Theorem 2.3. Let

V (t) =
j=1

|F (tj t) F (tj1 t)|4 ,

where = {0 = t0 < t1 < t2 < } is a partition of [0, ), that is, tj and || = sup(tj tj1 ) < . Then lim E sup V (t)
0tT

||0

6 t

=0

for all T > 0. Proof. Since V is monotone, it will suce to show that V (t) 6t/ in L2 for each xed t. In what follows, C is a nite, positive constant that may change value from line to line. Let N = max{j : tj < t}. For each j, let 2 F j = F (tj ) F (tj1 ), j = EF 2 , and tj = tj tj1 . Note that j
N N 4 (F 4 3j ) + j j=1 j=1 4 3j

V (t) = |F (t) F (tN )|4 +

6 6 tj + (tN t).

By (2.3), E|F (t) F (tN )|8 C|t tN |2 0 and

4 3j

6 2 tj = 3 j +

2tj 2 j

2tj C C 5/2 7/4 3/2 tj 3/4 tj . tj tj

VARIATIONS OF STOCHASTIC HEAT EQ

Thus,
N

(2.12)
j=1

4 3j

6 tj

C||3/4

tj

3/4 j=1 tj

t which tends to zero as || 0 since 0 x3/4 dx < . To complete the proof, we will need the following fact about Gaussian random variables. Let X1 , X2 be mean zero, jointly normal random variables 2 with variances j . If = (1 2 )1 E[X1 X2 ], then

(2.13)

4 4 4 4 E[X1 X2 ] = 1 2 (244 + 722 + 9).

Applying this in our context, let ij = (i j )1 E[F i F j ] and write

N 2 N N 4 4 |E[(F 4 3i )(F 4 3j )]| i j i=1 j=1 N N 4 4 |E[F 4 F 4 ] 9i j |. i j i=1 j=1

E
j=1

(F 4 j

4 3j )

= Then by (2.13), we have

N 2 4 (F 4 3j ) j j=1 N

N 4 4 i j 2 ij

C
i=1 j=1 N N

=C
i=1 j=1 N N

2 2 i j |E[F i F j ]|2

C
i=1 j=1

ti

1/2

tj

1/2

|E[F i F j ]|2 .

By Hlders inequality, |E[F i F j ]|2 ti o show that

N 2 N

1/2

tj , so it will suce to

1/2

ti
i=1 j=i+2

1/2

tj

1/2

|E[F i F j ]|2 0

10

JASON SWANSON

as || 0. For this, suppose j > i + 1. By (2.11), |E[F i F j ]|2 Ct2 t2 i j |tj1 ti1 |2 |tj ti | Cti tj ti
1/2 1/2 5/4

|tj1 ti1 |1/2 |tj ti |1/4

3/4

C|| Hence,
N 2 N

tj

1/2

|tj1 ti |3/4

ti
i=1 j=i+2

1/2

tj

1/2

|E[F i F j ]|2
N 2 N

C||3/4
i=1 j=i+2

|tj1 ti |3/4 ti tj , y|3/4 dx dy < . 2

which tends to zero as || 0 since

t t 0 0 |x

3. Main Result. Let us now specialize to the uniform partition. That is, for xed n N, let t = n1 , tj = jt, and F j = F (tj ) F (tj1 ). We nt wish to consider sums of the form j=1 gj (F j ), where {gj } is a sequence of random functions. We will write these functions in the form gj (x) = 1 2 2 j hj (j x), where j = EF 2 . j Assumption 3.1. Let {hj (x) : x R} be a sequence of independent stochastic processes which are almost surely continuously dierentiable. Assume there exists a constant L such that Ehj (0)2 L and Ehj (0)2 L for all j. Also assume that for each j, (3.1) |hj (x) hj (y)| Lj |x y|

for all x, y R, where EL2 L. Finally, assume that j (3.2) and (3.3) |Ehi (X)hj (Y )| L|| Ehj (X) = 0

whenever X and Y which are independent of {hj } and are jointly normal with mean zero, variance one, and covariance = EXY .

VARIATIONS OF STOCHASTIC HEAT EQ

11

Remark 3.2. We may assume that each Lj is (hj )-measurable. In particular, {Lj } is a sequence of independent random variables. Also, since Ehj (0)2 L, we may assume that (3.4) |hj (x)| Lj (1 + |x|)

for all j. Similarly, since Ehj (0)2 L, we may assume that (3.5) for all j. Lemma 3.3. Let {hj } satisfy Assumption 3.1. Let X1 , . . . , X4 be mean zero, jointly normal random variables, independent of the sequence {hj }, 2 such that EXj = 1 and ij = EXi Xj . Then there exists a nite constant C, that depends only on L, such that
4

|hj (x)| Lj (1 + |x|2 )

(3.6)

E
j=1

hj (Xj ) C |12 34 | +

1 1

2 i2<j 12

max |ij |

whenever |12 | < 1. Moreover,

4

(3.7)

E
j=1

hj (Xj ) C max |1j |.

2j4

Furthermore, there exists > 0 such that

4

(3.8)

E
j=1

hj (Xj ) CM 2

whenever M = max{|ij | : i = j} < . Proof. In the proofs in this section, C will denote a nite, positive constant that depends only on L, which may change value from line to line. Let us rst record some observations. By (3.4), with probability one,
1

(3.9) Also,

|hj (y) hj (x)| =

0

(y x)hj (x + t(y x)) dt

Lj |y x|(1 + |x| + |y x|).

hj (y) hj (x) (y x)hj (x) =

(y x)(hj (x + t(y x)) hj (x)) dt,

12

JASON SWANSON

so by (3.1),
1

(3.10)

|hj (y) hj (x) (y x)hj (x)| |y x|

Lj t|y x| dt
0

Lj |y x|2 for all x, y R. Also, by (3.4) and (3.5), if we dene a stochastic process on Rn by G(x) = n hj (xj ), then G is almost surely continuously dierenj=1 tiable with |j G(x)| C( n Lj )(1 + |x|2n1 ). Hence, j=1
t

|G(y) G(x)| =
0 t

d G(x + t(y x)) dt dt (y x) G(x + t(y x)) dt

(3.11)

=
0

C
j=1

Lj n|y x|(1 + |x|2n1 + |y x|2n1 )

for all x, y Rn . Now let Y1 = (X1 , X2 )T and Y2 = (X3 , X4 )T . If |12 | < 1, then we may dene the matrix A = (EY2 Y1T )(EY1 Y1T )1 . Note that (3.12) |A| C 1 2 12
i2<j

max |ij |.

Let Y2 = Y2 AY1 , so that E Y2 Y1T = 0, which implies Y2 and Y1 are 2 by independent, and dene stochastic processes on R Fij (x) = hi (x1 )hj (x2 ), so that 4 hj (Xj ) = F12 (Y1 )F34 (Y2 ). j=1 Also dene X = (X2 , X3 , X4 )T and c = (12 , 13 , 14 )T . Note that X1 and cX1 are independent. Dene a process on R3 by X F (x) = h2 (x1 )h3 (x2 )h4 (x3 ), so that 4 hj (Xj ) = h1 (X1 )F (X). j=1 Let = E X X T . If M is suciently small, then is invertible, we may 1 c, and we have |a| CM and |aT c| CM 2 < 3/4. Note that dene a = aT c = E|aT X|2 0. Let = (1 aT c)1/2 so that 1 < 2 and (3.13) 1= 1 1 2aT c < < 2aT c. 1 1 + 1

VARIATIONS OF STOCHASTIC HEAT EQ

13

Dene U = (X1 aT X). Note that E[U X T ] = (cT aT ) = 0, so that U and X are independent. Hence, 2 = E(X1 )2 = EU 2 + 2 E|aT X|2 = EU 2 + 2 aT c, so that U is normal with mean zero and variance one. For the proof of (3.6), we have
4

(3.14)

E
j=1

hj (Xj ) = E[F12 (Y1 )F34 (Y2 )] + E[F12 (Y1 )(F34 (Y2 ) F34 (Y2 ))] = EF12 (Y1 )EF34 (Y2 ) + E[F12 (Y1 )(F34 (Y2 ) F34 (Y2 ))]

and (3.15) By (3.11), |F34 (Y2 ) F34 (Y2 )| CL3 L4 |AY1 |(1 + |Y2 |3 + |AY1 |3 ). Note that EL2 L2 = EL2 EL2 L2 . Also, since E|Y2 |2 = E|Y2 |2 + E|AY1 |2 , 3 4 3 4 we see that the components of AY1 are jointly normal with mean zero and a variance which is bounded by a constant independent of {ij }. Hence, Hlders inequality gives o E|F34 (Y2 ) F34 (Y2 )|2 C|A|2 (E|Y1 |4 )1/2 (1 + E|Y2 |12 + E|AY1 |12 )1/2 C|A|2 . By (3.12), (3.16) (E|F34 (Y2 ) F34 (Y2 )|2 )1/2 C 1 2 12
i2<j

EF34 (Y2 ) = EF34 (Y2 ) E[F34 (Y2 ) F34 (Y2 )].

max |ij |.

Hence, by (3.15), (3.17) |EF34 (Y2 )| |EF34 (Y2 )| + C 1 2 12

i2<j

max |ij |.

Note that (3.5) implies E|F12 (Y1 )|2 C. Therefore, using (3.14), (3.15), (3.16), and Hlders inequality, we have o
4

E
j=1

hj (Xj ) |EF12 (Y1 )EF34 (Y2 )| +

C 1 2 12

i2<j

max |ij |.

14

JASON SWANSON

By (3.3), this completes the proof of (3.6). For (3.7), we have

4

E
j=1

hj (Xj ) = Eh1 (X1 )EF (X cX1 ) + E[h1 (X1 )(F (X) F (X cX1 ))].

Since X1 and X cX1 are independent, (3.2) gives

4

(3.18)

E
j=1

hj (Xj ) = E[h1 (X1 )(F (X) F (X cX1 ))].

By Hlders inequality and (3.5), o

4

E
j=1

hj (Xj ) C(E|F (X) F (X cX1 )|2 )1/2 .

By (3.11),
4

|F (X) F (X cX1 )| C
j=2

Lj |cX1 |(1 + |X|5 + |cX1 |5 ).

Hence, E|F (X) F (X cX1 )|2 C|c|2 , which gives

4

E
j=1

hj (Xj ) C|c|,

and proves (3.7). Finally, for (3.8), we begin with an auxiliary result. Note that
4

E X2
j=2

hj (Xj ) = E[X2 h2 (X2 )]EF34 (Y2 ) + E[X2 h2 (X2 )(F34 (Y2 ) F34 (Y2 ))].

By Hlders inequality and (3.5), o

4

E X2
j=2

hj (Xj )

C|EF34 (Y2 )| + C(E|F34 (Y2 ) F34 (Y2 )|2 )1/2 .

VARIATIONS OF STOCHASTIC HEAT EQ

15

If M is suciently small, then |12 | C < 1. Hence, by (3.16), (3.17), and (3.3),
4

E X2
j=2

hj (Xj )

CM.

It now follows by symmetry that

4

(3.19)

E vT X
j=2

hj (Xj )

C|v|M

for any v R3 . Returning to the proof of (3.8), since (3.2) implies Eh1 (U ) = 0 and U and X are independent, we have
4

E
j=1

hj (Xj ) = E[(h1 (X1 ) h1 (U ) (X1 U )h1 (U ))F (X)] + E[(X1 U )h1 (U )F (X)].

By (3.10), |h1 (X1 ) h1 (U ) (X1 U )h1 (U )| L1 |X1 U |2 . By (3.13), |1 | CaT c C|a|M , so that |X1 U | = |(1 )X1 + aT X| C|a|(M |X1 | + |X|). Hence, using Hlders inequality and (3.5), we have o
4

(3.20)

E
j=1

hj (Xj ) C(E|X1 U |4 )1/2 + |E[(X1 U )h1 (U )F (X)]| C|a|2 + |E[(X1 U )h1 (U )F (X)]|.

To estimate the second term, note that E[(X1 U )h1 (U )F (X)] = (1)E[X1 h1 (U )F (X)]+Eh1 (U )E[aT XF (X)]. Therefore, by (3.4), (3.5), (3.13), and (3.19), |E[(X1 U )h1 (U )F (X)]| C|1 | + C|a|M C|a|M. Combining this with (3.20) and recalling that |a| CM completes the proof of (3.8). 2

16

JASON SWANSON

Corollary 3.4. Let {hj } be independent of F and satisfy Assumption 3.1. For k N4 with k1 k4 , dene
4

(3.21)

k =
j=1

1 2 kj hkj (kj F kj ),

2 where j = EF 2 . Let xr = (x 1)r . Then there exists a nite constant j C such that

(3.22) Moreover, (3.23) and (3.24)

|Ek |

Ct2 (k4 k3 )3/2

and

|Ek |

Ct2 . (k2 k1 )3/2

|Ek | C

1 1 + t2 (k4 k3 )3/2 (k2 k1 )3/2 (k3 k2 )3/2

|Ek |

Ct2 , m3

where m = min{ki+1 ki : 1 i < 4}.

1 Proof. Let Xj = kj F kj . By (2.9) and (2.10), we have

|ij | = |E[Xi Xj ]| = Also,

1 1 ki kj |E[F ki F kj ]|

2 . |ki kj |3/2

4 2 kj

|Ek | =
j=1

E
j=1

hkj (Xj ) .

This, together with (3.7) and symmetry, yields (3.22). For (3.23), rst note that Hlders inequality and (3.5) give the trivial o bound |Ek | Ct2 . Hence, we may assume that at least one of k4 k3 and k2 k1 is large. Specically, by symmetry, we may assume that k2 k1 4. In this case, |12 | /4 < 1. Hence, (3.6) gives
4 2 kj j=1 4 4

2 kj |12 34 | +

E
j=1

hkj (Xj )

C
j=1

1 1

2 i2<j 12

max |ij |

and (3.23) is immediate.

VARIATIONS OF STOCHASTIC HEAT EQ

17

As above, we may assume in proving (3.24) that m is large. Therefore, we can assume that M = max{|ij | : i = j} < . Hence, (3.8) implies
4 2 kj j=1 4 4

E
j=1

hkj (Xj )

C
j=1

2 kj M 2 ,

which proves (3.24). Proposition 3.5. (3.25) With notation as in Corollary 3.4, let
nt

Bn (t) =
j=1

1 2 j hj (j F j ).

If {hj } is independent of F and satises Assumption 3.1, then there exists a constant C such that (3.26) E|Bn (t) Bn (s)|4 C nt ns n
2

for all 0 s < t and all n N. The sequence {Bn } is therefore relatively compact in the Skorohod space DR [0, ). Proof. To prove (3.26), observe that
nt 4 1 2 j hj (j F j ) . j= ns +1

E|Bn (t) Bn (s)| = E Let

S = {k N4 : ns + 1 k1 k4 nt }. For k S, dene hi = ki+1 ki and let M = M (k) = max(h1 , h2 , h3 ) m = m(k) = min(h1 , h2 , h3 ) c = c(k) = med(h1 , h2 , h3 ) where med denotes the median function. For i {1, 2, 3}, let Si = {k S : hi = M }. Dene N = nt ( ns + 1) and for j {0, 1, . . . , N }, let j j Si = {k Si : M = j}. Further dene Ti = Tij, = {k Si : m = } and Vi = Vij, , = {k Ti : c = }.

18

JASON SWANSON

Recalling (3.21), we now have

nt 4 1 2 j hj (j F j ) j= ns +1 3

(3.27)

4!
kS

|Ek | 4!
i=1 kSi

|Ek |.

Observe that
N

(3.28)
kSi

|Ek | =
j=0 kS j
i

|Ek |

and
j j

(3.29)
j kSi

|Ek | =
=0 kT i

|Ek | +
=

|Ek |.
j +1 kTi

Begin by considering the rst summation. Suppose 0

j

j and write

|Ek | =
kTi = kVi

|Ek |.

Fix and let k Vi be arbitrary. If i = 1, then j = M = h1 = k2 k1 . If i = 3, then j = M = h3 = k4 k3 . In either case, (3.22) gives |Ek | C 1 j 3/2 t2 C 1 1 + 3/2 t2 . 3/2 ( ) j

If i = 2, then j = M = h2 = k3 k2 and = h3 h1 = (k4 k3 )(k2 k1 ). Hence, by (3.23), |Ek | C 1 1 + 3/2 t2 . 3/2 ( ) j

Now choose i = i such that hi = . With i given, k is determined by ki . Since there are two possibilities for i and N + 1 possibilities for ki , |Vi | 2(N + 1). Therefore,
j j j

|Ek | C(N + 1)
=0 kT i =0 = j

1 1 + t2 ( )3/2 j 3/2 1
3/2

C(N + 1) C(N + 1)t .

=0 2

1 j 1/2

t2

VARIATIONS OF STOCHASTIC HEAT EQ

19

For the second summation, suppose j + 1 j. (In particular, j 1.) In this case, if k Ti , then = m = min{ki+1 ki : 1 i < 4}, so that by (3.24), 1 |Ek | C 3 t2 . Since |Ti | =
j = j =

|Vi | 2(N + 1)j, we have

j

|Ek | C(N + 1)j

j +1 kTi =

1
3

t2

j +1

C(N + 1)j C(N + 1)t .

2

1 dx t2 x3

We have thus shown that kS j |Ek | C(N + 1)t2 . i Using (3.27)-(3.29), we have
nt 4 1 2 j hj (j F j ) j= ns +1 N

C
j=0

(N + 1)t2 = C

nt ns n

which is (3.26). To show that a sequence of cadlag processes {Xn } is relatively compact, it suces to show that for each T > 1, there exist constants > 0, C > 0, and > 1 such that (3.30) MX (n, t, h) = E[|Xn (t + h) Xn (t)| |Xn (t) Xn (t h)| ] Ch for all n N, all t [0, T ] and all h [0, t]. (See, for example, Theorem 3.8.8 in [11].) Taking = 2 and using (3.26) together with Hlders inequality o gives MB (n, t, h) C nt + nh nt n nt nt nh n .

If nh < 1/2, then the right hand side of this inequality is zero. Assume nh 1/2. Then nt + nh nt nh + 1 3h. n n The other factor is similarly bounded, so that MB (n, t, h) Ch2 . 2 Let us now introduce the ltration Ft =
A

dW (r, y) : A R [0, t], m(A) < ,

20

JASON SWANSON

where m denotes Lebesgue measure on R2 . Recall that

t

(3.31)

F (t) =
0 R

p(t r, y) dW (r, y)

so that F is adapted to {Ft }. Also, given constants 0 s t, we have

E[F (t)|F ] =
0 R

p(t r, y) dW (r, y)

and E|E[F (t) F (s)|F ]|2 =

0 R

|p(t r, y) p(s r, y)|2 dy dr.

As in the proof of Lemma 2.1,

0

1 p(t r, y)p(s r, y) dy dr = (|t + s|1/2 |(t u) + (s u)|1/2 ). 2 R

Therefore, using (2.5), we can verify that (3.32) E|E[F (t) F (s)|F ]|2 = E|F (t) F (s)|2 E|F (t ) F (s )|2 . Combined with (2.3), this gives E|E[F (t) F (s)|F ]|2 In particular, (3.33) whenever tj1 . Lemma 3.6. Let Bn be given by (3.25) and assume {hj } is independent of F and satises Assumption 3.1. Fix 0 s < t and a constant . If
n

2|t s|2 . |t |3/2

E|E[F j |F ]|2

2t2 (tj )3/2

lim E|Bn (t) Bn (s)|2 = 2 (t s), Bn (t) Bn (s) |t s|1/2

then as n , where is a standard normal random variable.

VARIATIONS OF STOCHASTIC HEAT EQ

21

Proof. We will prove the lemma by showing that every subsequence has a subsequence converging in law to the given random variable. Let {nj } be any sequence. For each n N, choose m = mn {nj } such that mn > mn1 and mn n4 (t s)1 . Now x n N and let = m(t s)/n. For 0 k < n, dene uk = ms + k , and let un = mt , so that
mt n 1 2 j hj (j F j ) j= ms +1 uk 1 2 j hj (j F j ). k=1 j=uk1 +1

Bm (t) Bm (s) =

For each pair (j, k) such that uk1 < j uk , let F j,k = F j E[F j |Fuk1 t ]. Note that F j,k is Fuk t -measurable and independent of Fuk1 t . We also make the following observation about F j,k . If we dene Gk (t) = F (t + k ) E[F (t + k )|Fk ], where k = uk1 t, then by (3.31),
t+k

Gk (t) =

p(t + k r, y) dW (r, y).

R

Hence, Gk and Fk are independent, and Gk and F have the same law. Since F j,k = F j E[F j |Fk ] = Gk (tj k ) Gk (tj1 k ), it follows that {F j,k } has the same law as {F juk1 }. 2 2 Now dene j,k = EF j,k = juk1 and 2
uk

Zn,k =
j=uk1 +1

j,k hj (j,k F j,k ) 2 1

so that Zn,k , 1 k n, are independent and

n

(3.34) where
n

Bm (t) Bm (s) =
k=1

Zn,k + m ,

uk 1 2 {j hj (j F j ) j,k hj (j,k F j,k )}. 2 1

m =
k=1 j=uk1 +1

22

JASON SWANSON

Since F j,k and F j F j,k = E[F j |Fuk1 t ] are independent, we have

2 (3.35) j = EF 2 = EF j,k +E|F j F j,k |2 = j,k +E|F j F j,k |2 , 2 j 2 which implies that j,k j Ct1/2 . In general, if 0 < a b and x, y R, 2 then by (3.5) and (3.9), 2

|b2 hj (b1 y) a2 hj (b1 y)| (b2 a2 )Lj (1 + |b1 y|2 ) |a2 hj (b1 y) a2 hj (a1 x)| |a|2 CLj |b1 y a1 x|(1 + |b1 y| + |a1 x|). Note that |b1 y a1 x| |b1 a1 ||y| + |a1 ||y x|, and |b1 a1 | = Hence, if = b2 a2 , then |b2 hj (b1 y)a2 hj (a1 x)| CLj (1+|b1 y|2 +|a1 x|)(+|a1 y|+|a||yx|). o Using (2.9), Hlders inequality, and (3.35), this gives
1 2 2 1 E|j hj (j F j ) j,k hj (j,k F j,k )|

b2 a2 b2 a2 . ab(b + a) a3

CE|F j F j,k |2 + Ct1/4 (E|F j F j,k |2 )1/2 . By (3.33), E|F j F j,k |2 Therefore,
n uk

2t2 2t1/2 = . (tj uk1 t)3/2 (j uk1 )3/2

E|m |
k=1 j=uk1

n Ct1/2 = Cm1/2 (j uk1 )3/4 +1 k=1

uk uk1

j 3/4 .
j=1

Since uk uk1 C, this gives E|m | Cm1/2 n1/4 = Cn3/4 m1/4 (t s)1/4 . But since m = mn was chosen so that m n4 (t s)1 , we have E|m | Cn1/4 |t s|1/2 and m 0 in L1 and in probability. Therefore, by (3.34), we need only show that
n

Zn,k |t s|1/2
k=1

VARIATIONS OF STOCHASTIC HEAT EQ

23

in order to complete the proof. For this, we will use the Lindeberg-Feller Theorem (see, for example Theorem 2.4.5 in [10]), which states the following: for each n, let Zn,k , 1 k n, be independent random variables with EZn,k = 0. Suppose
2 (a) n EZn,k 2 , and k=1 (b) for all > 0, limn n 2 k=1 E[|Zn,k | 1{|Zn,k |>} ]

= 0.

Then n Zn,k as n . k=1 To verify these conditions, recall that {F j,k } and {F juk1 } have the same law, so that
uk 4

E|Zn,k |4 = E
j=uk1 +1 uk uk1

j,k hj (j,k F j,k ) 2 1

4 1 2 j hj+uk1 (j F j ) j=1

=E

= E|Bm,k ((uk uk1 )t)|4 , where

mt

Bm,k (t) =
j=1

1 2 j hj+uk1 (j F j ).

Hence, by Proposition 3.5, E|Zn,k |4 C(uk uk1 )2 t2 . Jensens inequality now gives n E|Zn,k |2 Cnt = C(t s), so that k=1 by passing to a subsequence, we may assume that (a) holds for some 0. For (b), let > 0 be arbitrary. Then
n n

E[|Zn,k |2 1{|Zn,k |>} ] 2

k=1

E|Zn,k |4 n t2

k=1 2 2

= C2 n1 (t s)2 , which tends to zero as n . It therefore follows that n Zn,k as n and it remains k=1 only to show that = |t s|1/2 . For this, observe that the Continuous Mapping Theorem implies that |Bm (t) Bm (s)|2 2 2 . By the Skorohod Representation Theorem, we may assume that the convergence is a.s. By Proposition 3.5, the family |Bm (t) Bm (s)|2 is uniformly integrable. Hence,

24

JASON SWANSON

|Bm (t) Bm (s)|2 2 2 in L1 , which implies E|Bm (t) Bm (s)|2 2 . But by assumption, E|Bm (t) Bm (s)|2 2 (t s), so = |t s|1/2 and the proof is complete. 2 Lemma 3.7. Let Bn be given by (3.25) and assume {hj } is independent of F and satises Assumption 3.1, so that by Proposition 3.5, the sequence {Bn } is relatively compact. If X is any weak limit point of this sequence, then X has independent increments. Proof. Suppose that Bn(j) X. Fix 0 < t1 < t2 < < td < s < t. It will be shown that X(t) X(s) and (X(t1 ), . . . , X(td )) are independent. With notation as in Lemma 3.6, let
nt

Zn =
j= ns +2

j,k hj (j,k F j,k ), 2 1

and dene Yn = Bn (t) Bn (s) Zn . As in the proof of Lemma 3.6, Yn 0 in probability. It therefore follows that (Bn(j) (t1 ), . . . , Bn(j) (td ), Zn(j) ) (X(t1 ), . . . , X(td ), X(t) X(s)). Note that F( ns +1)t and Zn are independent. Hence, (Bn (t1 ), . . . , Bn (td )) and Zn are independent, which implies X(t) X(s) and (X(t1 ), . . . , X(td )) are independent. 2 Theorem 3.8. (3.36) Let
nt

Bn (t) =
j=1

1 2 j hj (j F j )

and assume {hj } is independent of F and satises Assumption 3.1. If there exists a constant such that
n

lim E|Bn (t) Bn (s)|2 = 2 (t s)

for all 0 s < t, then Bn B, where B is a standard Brownian motion. Proof. Let {n(j)} be any sequence of natural numbers. By Proposition j=1 3.5, the sequence {Bn(j) } is relatively compact. Therefore, there exists a

VARIATIONS OF STOCHASTIC HEAT EQ

25

subsequence m(k) = n(jk ) and a cadlag process X such that Bm(k) X. By Lemma 3.7, the process X has independent increments. By Lemma 3.6, the increment X(t) X(s) is normally distributed with mean zero and variance 2 |t s|. Also, X(0) = 0 since Bn (0) = 0 for all n. Hence, X is equal in law to B, where B is a standard Brownian motion. Since every subsequence of {Bn } has a further subsequence which converges weakly to B, it follows that Bn B. 2 4. Examples. 4.1. Independent mean zero sign changes. Proposition 4.1. Let {j } be a sequence of independent mean zero ran2 dom variables with Ej = 1. Suppose that the sequence {j } is independent of F . Let
nt

Bn (t) =
j=1

F 2 j . j

Then Bn 6 1 B, where B is a standard Brownian motion. Proof. Let hj (x) = j x2 . Then {hj } satises Assumption 3.1 with Lj = 2|j | and L = 4, and Bn has the form (3.36). Moreover,
nt 2 nt

E|Bn (t) Bn (s)|2 = E

j= ns +1

F 2 j j

=
j= ns +1

EF 4 . j

By (2.12),
n

lim E|Bn (t) Bn (s)|2 = 6 1 (t s). 2

The result now follows from Theorem 3.8.

4.2. The signed variations of F . In this subsection, we adopt the notation xr = |x|r sgn(r). We begin by showing that the signed cubic variation of F is zero. Proposition 4.2. If Zn (t) = compacts in probability.
nt j=1

F 3 , then Zn (t) 0 uniformly on j

Proof. Note that xn 0 in DR [0, ) if and only if xn 0 uniformly on compacts. Hence, we must show that Zn 0 in probability in DR [0, ), for which it will suce to show that Zn 0.

26

JASON SWANSON

Note that
nt 2 nt nt

E|Zn (t) Zn (s)| = E

j= ns +1

F 3 j

i= ns +1 j= ns +1

|E[F 3 F 3 ]|. i j

To estimate this sum, we use the following fact about Gaussian random variables. Let X1 , X2 be mean zero, jointly normal random variables with 2 variances j . If = (1 2 )1 E[X1 X2 ], then
3 3 3 3 E[X1 X2 ] = 1 2 (62 + 9).

Applying this in our context, let ij = (i j )1 E[F i F j ], so that

nt nt 3 3 i j |ij | i= ns +1 j= ns +1 nt nt 2 2 i j |E[F i F j ]|. i= ns +1 j= ns +1

E|Zn (t) Zn (s)|2 C

=C Using (2.10), this gives

nt nt

E|Zn (t) Zn (s)| C

i= ns +1 j= ns +1

t.

t t

t |i j|3/2

nt ns n

Hence, Zn (t) 0 in probability for each xed t. Moreover, taking = 1 in (3.30), this shows that MZ (n, t, h) = 0 when nh < 1/2, and MZ (n, t, h) Ch t Ch3/2 when nh 1/2. Therefore, {Zn } is relatively compact and Zn 0. 2 Lemma 4.3. Let X1 , X2 be mean zero, jointly normal random variables 2 with EXj = 1 and = E[X1 X2 ]. Let K(x) = 6 2 x 1 x2 + (1 + 2x2 ) sin1 (x),

2 2 where sin1 (x) [/2, /2]. Then E[X1 X2 ] = K(). Moreover, for all 2 2 x [1, 1], we have |K(x) 8x/| 2|x|3 , so that |E[X1 X2 ]| 5||.

VARIATIONS OF STOCHASTIC HEAT EQ

27

Proof. Dene U = X1 and V = (1 2 )1/2 (X2 X1 ), so that U and V are independent standard normals. Then X1 = U and X2 = V + U , where = 1 2 , and
2 2 E[X1 X2 ] =

1 2 1 = 2 4 =

[u(v + u)]2 e(u

2 0 2 0 0

2 +v 2 )/2

du dv
2 /2

[cos ( sin + cos )]2 r5 er

dr d

[cos2 ( tan + )]2 d.

If a = tan1 (/), then we can write

2 2 E[X1 X2 ] =

/2 a

[cos2 ( tan + )]2 d

a /2

8 =

8
a

[cos2 ( tan + )]2 d

[cos2 ( tan )]2 d

a /2

/2

[cos2 ( tan + )]2 d.

By symmetry, we can assume that 0, so that a 0. Then

2 2 E[X1 X2 ] = a 8 a 32 cos4 tan d [cos2 ( tan + )]2 d a /2 32 1 16 a = cos4 a cos4 d 4 0 a 16 + 2 cos4 (1 tan2 ) d. 0

Using a = sin1 () and the formulas cos4 (1 tan2 ) d = ( + sin cos + 2 sin cos3 )/4 cos4 d = (3 + 3 sin cos + 2 sin cos3 )/8,
2 2 we can directly verify that E[X1 X2 ] = K(). To estimate K, note that K C (1, 1) with

8 ( 1 x2 + x sin1 (x)) 8 K (x) = sin1 (x). K (x) =

28

JASON SWANSON

Since K is increasing, K(x) 8 1 x x2 K (|x|). 2

But for y [0, /2], we have sin y 2y/. Letting y = x/2 gives sin1 (x) x/2 for x [0, 1]. We therefore have K (|x|) 4|x|, so that |K(x)8x/| 2|x|3 . 2 Proposition 4.4. Lemma 2.1. If Let K be dened as in Lemma 4.3, and i as in
nt

Bn (t) =
j=1

F 2 sgn(F j ), j
i=0 K(i /2)

then Bn B, where 2 = 6 1 4 1 standard Brownian motion. Proof. Let hj (x) = h(x) = x2 , so that
nt

> 0 and B is a

Bn (t) =
j=1

1 2 j hj (j F j ).

Since h is continuously dierentiable and h (x) = 2|x| is Lipschitz, {hj } satises (3.1). Moreover, if X and Y are jointly normal with mean zero, variance one, and covariance = EXY , then Eh(X) = 0 and |Eh(X)h(Y )| 5|| by Lemma 4.3. Hence, {hj } satises Assumption 3.1. By Proposition 3.5, (4.1) E|Bn (t) Bn (s)|4 C nt ns n
2

for all 0 s < t and all n N. By Theorem 3.8, the proof will be complete once we establish that is well-dened, strictly positive, and (4.2)
n

lim E|Bn (t) Bn (s)|2 = 2 (t s)

for all 0 s < t. By (2.8), i /2 (0, 1] for all i. Thus, by Lemma 4.3,

0<
i=1

K(i /2)

i=1

1 1 i + 2 i 2 2 i=1

i +
i=1

1 4

3 i . i=1

VARIATIONS OF STOCHASTIC HEAT EQ

29
i=1 i

Since i = f (i1)f (i), where f (x) = f (0) = 1. Moreover, by (2.8),

3 i i=1

x + 1 x, we have that
1

i=1

1 1 = ( 2 i3/2 )3 2 2

i=1

1 i9/2

1 1+ 2 2

1 x9/2

9 2 dx = . 28

Thus,

4 9 2 3 K(i /2) + < , 112 2 i=1

which gives 6 1 4 1 K(i /2) > 0, so that is well-dened and i=0 strictly positive. Now x 0 s < t. First assume that s > 0. Then
nt 2

E|Bn (t) Bn (s)| = E

j= ns +1 nt

F 2 j
nt j1

=
j= ns +1 nt

EF 4 j

+2
j= ns +2 i= ns +1 nt j1

E[F 2 F 2 ] i j

=
j= ns

4 6 t +1

K(ji /2)t + Rn
j= ns +2 i= ns +1

where
nt

Rn =
j= ns +1

EF 4 j
nt

6 t E[F 2 F 2 ] + i j 2 K(ji /2)t .

j1

+2
j= ns +2 i= ns +1

Observe that
nt j= ns

6 6 t = +1

nt ns n

6 (t s)

and
nt j1 nt j ns 1

K(ji /2)t = (4.3)

j= ns +2 i= ns +1 j= ns +2 N j i=1

K(i /2)t 1 K(i /2), n

=
j=1 i=1

30

JASON SWANSON

where N = nt ns 1. Thus,
nt j1 N N

K(ji /2)t = (4.4)

j= ns +2 i= ns +1 i=1 j=i N

1 K(i /2) n

=
i=1

i N K(i /2). n n
N i=1 iK(i /2)

Since < , it follows that Since N/n (t s), this gives

nt j= ns

i=1 K(i /2)

n1

0 as n .

4 6 t +1

nt

j1

K(ji /2)t 2 (t s)
j= ns +2 i= ns +1

and it suces to show that Rn 0 as n . Now, by Lemma 4.3, F i 2 2 E[F 2 F 2 ] = i j E i j i

2 where i = EF 2 and i

F j j

= 2 2 K(ij ) i j

ij = E

F i i

F j j

= (i j )1 E[F i F j ].

2 2 Dene a = i j , b = K(ij ), c = 2t/, and d = K(ji /2). By (2.10), we have |a| Ct. By (2.10) and (2.3),

2 2 |a c| = i j

2t +

2t 2 i

1 2t C 3/2 t5/2 . t
i

By Lemma 4.3, |K(x) K(y)| C|x y|, so that |d| C and |b d| C|ij +ji /2|. Rewriting this latter inequality, we have the |bd| is bounded above by 1 C E[F i F j ] + i j Observe that i j 1 i j = 2t j 2t i + j

t ji 2

t ji 2

1 i j

2t

2t i 2 i + i + j

2t 2 j

VARIATIONS OF STOCHASTIC HEAT EQ

31

so that by (2.10) and (2.3)

2 2 i j

1 i j
3/2

2t

1
3/2 ti

t2 .

Hence, by (2.4), |a||b d| Cti E[F 2 F 2 ] + i j

t5/2 . We therefore have

2 K(ji /2)t = |ab cd| |a||b d| + |d||a c| Cti

3/2

t5/2 .

Since ti > s > 0, this shows that

nt j1

E[F 2 F 2 ] + j i
j= ns +2 i= ns +1

2 K(ji /2)t 0.

Combined with (2.12), this shows that Rn 0. We have now proved (4.2) under the assumption that s > 0. Now assume s = 0. Let (0, t) be arbitrary. Then by Hlders inequality and (4.1), o |E|Bn (t)|2 2 t| = |E|Bn (t) Bn ()|2 2 (t ) + 2E[Bn (t)Bn ()] E|Bn ()|2 2 | |E|Bn (t) Bn ()|2 2 (t )| + C( t + ). First let n , then let 0, and the proof is complete. 4.3. Centering the squared increments. Proposition 4.5. Let i be dened as in Lemma 2.1. If
nt

Bn (t) =
j=1

2 F 2 j , j 2 i=0 i

then Bn B, where 2 = 4 1 + 2 1 Brownian motion.

and B is a standard

Proof. Let hj (x) = x2 1. Then {hj } clearly satises (3.1) and (3.2). For jointly normal X and Y with mean zero and variance one, E(X 2 1)(Y 2 1) = 22 , so {hj } also satises (3.3). Since
nt

Bn (t) =
j=1

1 2 j hj (j F j ),

32

JASON SWANSON

it will suce, by Theorem 3.8, to show that (4.5) By Proposition 3.5, E|Bn (t) Bn (s)|4 C nt ns n
2 n

lim E|Bn (t) Bn (s)|2 = 2 (t s).

for all 0 s < t and all n N. Hence, as in the proof of Proposition 4.4, it will suce to prove (4.5) for s > 0. Assume s > 0. Then
nt 2 2 F 2 j j j= ns +1 nt nt 4 2j + 2 j= ns +1 j= ns +2 i= ns +1 nt j1

E|Bn (t) Bn (s)|2 = E

= 4 = where
nt

2|EF i F j |2 2 +
j1 2 ji t + Rn , j= ns +2 i= ns +1

nt ns n

Rn =
j= ns +1

4 2j nt

4 t
j1

+4
j= ns +2 i= ns +1

|EF i F j |2

t 2 . 2 ji

By (2.4), (2.8), and (2.10), |EF i F j |2 As in (4.3) and (4.4),

nt j1 t 1 (t s) . 3/2 3/2 (j i) i i=1 +1

t 2 Ct5/2 ji 3/2 . 2 s (j i)3/2

j= ns +2 i= ns

Together with (2.12), this shows that Rn 0. Hence, E|Bn (t) Bn (s)|2 and the proof is complete.
4 2 2 (t s) + (t s) i , i=1

VARIATIONS OF STOCHASTIC HEAT EQ

33

Corollary 4.6.

If
nt

Bn (t) =
j=1

F 2 j

2n t
2 i=0 i

then Bn B, where 2 = 4 1 + 2 1 Brownian motion. Proof. Note that Bn (t) = and by (2.3),
ns

and B is a standard

2 ( nt nt) + F 2 j n j=1

nt

2t ,

sup
0st j=1

F 2 j

2t

ns

nt

j=1

F 2 j

2 j

j=1

2 j

2t 1 j 3/2 . 2

The result now follows from Proposition 4.5. 4.4. Alternating sign changes. Proposition 4.7.

nt j=1

Let i be dened as in Lemma 2.1. If

nt

Bn (t) =
j=1

F 2 (1)j , j
i 2 i=0 (1) i

then Bn B, where 2 = 4 1 + 2 1 standard Brownian motion. Proof. Let

nt

> 0 and B is a

Yn (t) =
j=1

2 (1)j (F 2 j ) j

and
nt

An (t) =
j=1

2 (1)j j ,

34

JASON SWANSON

so that Bn = Yn + An . Note that Yn is of the form (3.36) with hj (x) = (1)j (x2 1). As in the proof of Proposition 4.4, is well-dened and strictly positive. Using the methods in the proof of Proposition 4.5, we have that Yn B. To complete the proof, observe that
nt/2 0st

sup |An (s)| 2nt +

j=1

2 2 |2j 2j1 |,

and by (2.3),
2 2 |2j 2j1 |

t . (2j 1)3/2 2

Hence, An 0 uniformly on compacts. Corollary 4.8. Let be as in Proposition 4.7. Dene

nt/2

Bn (t) = and
nt/2

1 j=1

(F 2 F 2 ) 2j 2j1

In (t) =
j=1

F (t2j1 )(F (t2j ) F (t2j2 )).

Then the sequence {(F, Bn , In )} is relatively compact and any weak limit point, (F , B, I), satises 1 I = F 2 B, 2 2 d where F = F and B is a standard Brownian motion. Proof. Note that
nt

Bn (t)
j=1

F 2 (1)j F 2nt , j

so that by Proposition 4.7, Bn B. Also note that by (1.4), F (t)2 = 2In (t) + Bn (t) + n (t), where n (t) = F (t)2 F (t2N )2 and N = nt/2 . Since n (t) 0 uniformly on compacts, it follows that {(F, Bn , In )} is relatively compact and any d weak limit point, (F , B, I), satises F 2 = 2I + B, where F = F and B is a standard Brownian motion. 2

VARIATIONS OF STOCHASTIC HEAT EQ

35

It is conjectured that in Corollary 4.8, if (F , B, I) is a limit point, then F are independent. In particular, the limit point is unique. Exploring and B this conjecture, as well as generalizing Corollary 4.8 to functions of F other than F 2 , will be the subject of future work. Acknowledgements. Part of this material appeared in my doctoral dissertation and I would like to thank my advisors, Chris Burdzy and ZhenQing Chen, for their helpful advice. Regarding the generalization of my dissertation work, I gratefully acknowledge the helpful ideas of Sona Zaveri Swanson. I would also like to thank Bruce Erickson and Yaozhong Hu for helpful discussions and feedback, as well as Tom Kurtz for all of his support and guidance. This work was done while supported by VIGRE, for which I thank the NSF, University of Washington, and University of WisconsinMadison. REFERENCES
[1] Hassan Allouba and Weian Zheng, Brownian-time processes: the PDE connection and the half-derivative generator. The Annals of Probability, 29(4) (2001), 17801795. [2] Elisa Al`s, Olivier Mazet, and David Nualart, Stochastic Calculus with Respect o to Fractional Brownian Motion with Hurst Parameter Lesser Than 1/2. Stochastic Processes and their Applications, 86 (2000), 121139. [3] Chris Burdzy and Andrzej Mdrecki, It Formula for an Asymptotically 4-Stable a o Process. The Annals of Applied Probability, 6(1) (1996), 200217. [4] Philippe Carmona, Laure Coutin, and Grard Montseny, Stochastic Integration with e Respect to Fractional Brownian Motion. Annales de lInstitut Henri Poincar. Probe abilits et Statistique, 39(1) (2003), 2768. e [5] Laure Coutin and Zhongmin Qian, Stochastic analysis, rough path analysis and fractional Brownian motions. Probability Theory and Related Fields, 122 (2002), 108140. [6] R. Dante DeBlassie, Iterated Brownian Motion in an Open Set. The Annals of Applied Probability, 14(3) (2004), 15291558. u [7] Laurent Decreusefond and A. S. Ustnel, Stochastic Analysis of the Fractional Brownian Motion. Potential Analysis, 10(2) (1999), 177214. [8] Laurent Decreusefond, Stochastic Integration with Respect to Fractional Brownian Motion. In Theory and Applications of Long-Range Dependence, Birkhuser Boston a (2003), 203226. [9] Tyrone E. Duncan, Yaozhong Hu, and Bozenna Pasik-Duncan, Stochastic Calculus for Fractional Brownian Motion I. Theory. SIAM Journal of Control and Optimization, 38(2) (2000), 582612. [10] Richard Durrett, Probability: Theory and Examples, Second Edition. Duxbury Press, 1996. [11] Stewart N. Ethier and Thomas G. Kurtz, Markov Processes: Characterization and Convergence. Wiley-Interscience, 1986. [12] Tadahisa Funaki, Probabilistic Construction of the Solution of Some Higher Order Parabolic Dierential Equations. Proceedings of the Japan Society, Series A, 55 (1979), 176179. [13] Helge Holden, Bernt Oksendal, Jan Ube, and Tusheng Zhang, Stochastic Partial o

36

JASON SWANSON

Dierential Equations: A Modeling, White Noise Functional Approach. Birkhuser, a 1996. [14] Ioannis Karatzas and Steven E. Shreve, Brownian Motion and Stochastic Calculus. Springer, 1991. [15] Terry J. Lyons, Dierential equations driven by rough signals. Revista Matemtica a Iberoamericana, 14(2) (1998), 215310. [16] Benoit B. Mandelbrot and John W. Van Ness, Fractional Brownian Motions, Fractional Noises and Applications. SIAM Review, 10(4) (1968), 442437. [17] David Nualart, The Malliavin Calculus and Related Topics. Springer-Verlag, 1995. [18] Nicolas Privault, Skorohod stochastic integration with respect to non-adapted processes on Wiener space. Stochastics and Stochastics Reports, 65(1-2) (1998), 1339. [19] Jason Swanson, Variations of Stochastic Processes: Alternative Approaches. Doctoral Dissertation, University of Washington, 2004.
Mathematics Department University of Wisconsin-Madison 480 Lincoln Dr. Madison, WI 53706-1388 url: www.math.wisc.edu/swanson

CHAPTER 2

The Diusion Equation

In this chapter we study the one-dimensional diusion equation u 2u = 2 + p(x, t), t x which describes such physical situations as the heat conduction in a one-dimensional solid body, spread of a die in a stationary uid, population dispersion, and other similar processes. In the last section we will also discuss the quasilinear version of the diusion equation, known as, the Burgers equation u u 2u +u 2 = p(x, t) t x x which arises in the context of modelling the motion of a viscous uid as well as trac ow. We begin with a derivation of the heat equation from the principle of the energy conservation. 2.1. Heat Conduction Consider a thin, rigid, heat-conducting body (we shall call it a bar) of length l. Let (x, t) indicate the temperature of this bar at position x and time t, where 0 x l and t 0. In other words, we postulate that the temperature of the bar does not vary with the thickness. We assume that at each point of the bar the energy density per unit volume is proportional to the temperature, that is (x, t) = c(x)(x, t), (2.1.1)

where c(x) is called heat capacity and where we also assumed that the mass density is constant throughout the body and normalized to equal one. Although the body has been assumed rigid, and with constant mass density, its material properties, including the heat capacity, may vary from one point to another.
23

24

2. THE DIFFUSION EQUATION

To derive the homogeneous heat-conduction equation we assume that there are no internal sources of heat along the bar, and that the heat can only enter the bar through its ends. In other words, we assume that the lateral surface of the bar is perfectly insulated so no heat can be gained or lost through it. The fundamental physical law which we employ here is the law of conservation of energy . It says that the rate of change of energy in any nite part of the bar is equal to the total amount of heat owing into this part of the bar. Let q(x, t) denote the heat ux that is the rate at which heat ows through the body at position x and time t, and let us consider the portion of the bar from x to x+ x. The rate of change of the total energy of this part of the bar equals the t ow is to the right. In order to obtain the equation describing the heat conduction at an arbitrary point x we shall consider the limit of (2.1.2) as x 0. First, assuming that x, invoking the the integrand c(z)(z, t) is suciently regular, we are able to dierentiate inside the integral. Second, dividing both sides of the equation by Mean-Value Theorem for Integrals, and taking c(x) q = t x x 0 we obtain the equation (2.1.3)
x+ x

total amount of heat that ows into this part through its ends, namely c(z)(z, t)dz = q(x +
x

x, t) + q(x, t).

(2.1.2)

We use here commonly acceptable convention that the heat ux q(x, t) > 0 if the

relating the rate of change of temperature with the gradient of the heat ux. We are ready now to make yet another assumption; a constitutive assumption which relates the heat ux to the temperature. Namely, we postulate what is known as Fouriers Law of Cooling, that the heat ows at the rate directly proportional to the (spatial) rate of change of the temperature. If in addition we accept that the heat ows, as commonly observed, from hot to cold we get that q(x, t) = (x) . x (2.1.4)

where the proportionality factor (x) > 0 is called the thermal conductivity. Notice the choice of the sign in the denition of the heat ux guarantees that if

2.1. HEAT CONDUCTION

25

the temperature is increasing with x the heat ux is negative and the heat ows from right to left, i.e., from hot to cold. Combining (2.1.3) and (2.1.4) produces the partial dierential equation c(x) = ((x) ), t x x 0 < x < l, (2.1.5)

governing the heat ow in a inhomogeneous ( is in general point dependent) onedimensional body. However, if the bar is made of the same material throughout, whereby the heat capacity c(x) and the thermal conductivity (x) are point independent, (2.1.5) reduces to 2 = 2, t x where = . c (2.1.7) 0 < x < l, (2.1.6)

This equation is known as the heat equation, and it describes the evolution of temperature within a nite, one-dimensional, homogeneous continuum, with no internal sources of heat, subject to some initial and boundary conditions. Indeed, in order to determine uniquely the temperature (x, t), we must specify the temperature distribution along the bar at the initial moment, say (x, 0) = g(x) for 0 x l. In addition, we must tell how the heat is to be transmitted through the boundaries. We already know that no heat may be transmitted through the lateral surface but we need to impose boundary conditions at the ends of the bar. There are two particularly relevant physical types of such conditions. We may for example assume that (l, t) = (t) (2.1.8)

which means that the right hand end of the bar is kept at a prescribed temperature (t). Such a condition is called the Dirichlet boundary condition. On the other hand, the Neumann boundary condition requires specifying how the heat ows out of the bar. This means prescribing the ux

26

2. THE DIFFUSION EQUATION

q(l, t) = (l)

(l, t) = (t). x

(2.1.9)

at the right hand end. In particular, (t) 0 corresponds to insulating the right hand end of the bar. If both ends are insulated we deal with the homogeneous Neumann boundary conditions. Remark 2.1. Other boundary conditions like the periodic one are also possible. 2.2. Separation of Variables The most basic solutions to the heat equation (2.1.6) are obtained by using the separation of variables technique, that is, by seeking a solution in which the time variable t is separated from the space variable x. In other words, assume that (x, t) = T (t)u(x), (2.2.1)

where T (t) is a x-independent function while u(x) is a time-independent function. Substituting the separable solution into (2.1.6) and gathering the time-dependent terms on one side and the x-dependent terms on the other side we nd that the functions T (t) and u(x) must solve an equation T u = . T u (2.2.2)

The left hand side of equation (2.2.2) is a function of time t only. The right hand side, on the other hand, is time independent while it depends on x only. Thus, both sides of equation (2.2.2) must be equal to the same constant. If we denote the constant as and specify the initial condition (x, 0) = u(x), we obtain that (x, t) = et u(x) (2.2.4) 0 x l, (2.2.3)

solves the heat equation (2.1.6) provided we are able to nd u(x) and such that u = u (2.2.5)

2.2. SEPARATION OF VARIABLES

27

along the bar. This is an eigenvalue problem for the second order dierential
d operator K dt2 with the eigenvalue and the eigenfunction u(x). The
2

particular eigenvalues and the corresponding eigenfunctions will be determined by the boundary conditions that u inherits from . Once we nd all eigenvalues and eigenfunctions we will be able to write the general solution as a linear combinations of basic solutions (2.2.4). Homogeneous Boundary Conditions. Let us consider a simple Dirichlet boundary value problem for the heat conduction in a (uniform) bar held at zero temperature at both ends, i.e., (0, t) = (l, t) = 0, where initially (x, 0) = g(x), 0 < x < l. (2.2.7) This amounts, as we have explained earlier, to nding the eigenvalues and the eigenfunctions of (2.2.5) subject to the boundary conditions u(0) = u(l) = 0. (2.2.8) t 0, (2.2.6)

Notice rst that as evident from the form of the equation (2.2.5) the eigenvalues must be real. Also, it can be easily checked using the theory of second order ordinary linear dierential equations with constant coecients that if 0, then the boundary conditions (2.2.8) yield only the trivial solution u(x) 0. Hence, the general solution of the dierential equation (2.2.5) is a combination of trigonometric functions u(x) = a cos x + b sin x (2.2.9)

where we let = 2 with > 0. The boundary condition u(0) = 0 implies that a = 0. Because of the second boundary condition u(l) = b sin l = 0 (2.2.10)

l must be an integer multiple of . Thus, the eigenvalues and the eigenfunctions of the eigenvalue problem (2.2.5) with boundary conditions (2.2.8) are i = i l
2

ui (x) = sin

i x, l

i = 1, 2, 3, . . . .

(2.2.11)

28

2. THE DIFFUSION EQUATION

The corresponding basic solutions (2.2.4) to the heat equation are i (x, t) = exp i2 2 i t sin x, 2 l l

i = 1, 2, 3, . . . .

(2.2.12)

By linear superposition of these basic solutions we get a formal series

(x, t) =
i=1

ai ui (x, t) =
i=1

exp

i i2 2 t sin x. 2 l l

(2.2.13)

Assuming that the series converges we have a general series solution of the heat equation with the initial temperature distribution

(x, 0) = g(x) =
i=1

ai sin

i x. l

(2.2.14)

This is a Fourier sine series on the interval [0, l] of the initial condition g(x)1. Its coecients ai can be evaluated explicitly thanks to the remarkable orthogonality property of the eigenfunctions. Indeed, it is a matter of a simple exercise on integration by parts to show that
l

sin
0

k n x sin xdx = 0 l l
l

(2.2.15)

only if n = k, and that

k l x= . (2.2.16) l 2 0 Multiplying the Fourier series of g(x) by the k-th eigenfunction and integrating sin2 over the interval [0, l] one gets that ak = 2 l
l

g(x) sin
0

k xdx, l

k = 1, 2, 3, . . . .

(2.2.17)

Example 2.2. Consider the initial-boundary value problem x, 0 x 1, (0, t) = (2, t) = 0, (x, 0) = g(x) = x + 2, 1 x 2, of g(x) are a2k+2 0,
1Fourier

(2.2.18)

for the heat equation for a homogeneous bar of length 2. The Fourier coecients 8 , (2k + 1)2 2

a2k+1 = (1)k

k = 0, 1, 2, . . . .

(2.2.19)

series are introduced and treated extensively in Appendix B

2.2. SEPARATION OF VARIABLES

29

The resulting series solution is

(x, t) = 8
i=0

(1)i (2i + 1)2 2 t exp (2i + 1)2 2 4

sin(i +

)x. 2

(2.2.20)

Notice rst that although the initial data is piecewise dierentiable the solution is smooth for any t > 0. Also, as long as the initial prole is integrable (e.g., piecewise continuous) on [0, 2] its Fourier coecients are uniformly bounded, namely:
2 2

|ak |
0

|g(x) sin kx| dx

0

|g(x)| dx M.

(2.2.21)

Consequently, the series solution (2.2.20) is bounded by an exponentially decaying time series |(x, t)| M
i=0

exp

(2i + 1)2 2 t 4

(2.2.22)

This means that solution decays to the zero temperature prole, a direct consequence of the fact that both ends are hold at zero temperature. This simple example shows that in the case of homogeneous boundary conditions any initial heat distributed throughout the bar will eventually dissipate away. Moreover, as the Fourier coecients in (2.2.20) decay exponentially as t , the solution gets very smooth despite the fact that the initial data was not. In fact, this is an illustration of the general smoothing property of the heat equation. Theorem 2.3. If u(t, x) is a solution to the heat equation with the initial condition such that its Fourier coecients are uniformly bounded, then for all t > 0 the solution is an innitely dierentiable function of x. Also, u(t, x) 0 as t , in such a way that there exists K > 0 such that |u(t, x)| < Ke for all t t0 > 0. The smoothing eect of the heat equation means that it can be eectively used to de-noise signals by damping the high frequency modes. This, however, means also that it is impossible to reconstruct the initial temperature by measuring the temperature distribution at some later time. The heat equation cannot be run backwards in time. There is no temperature distribution at t < 0 which would
2 t/l2

30

2. THE DIFFUSION EQUATION

produce a non-smooth temperature distribution at t = 0. Had we tried to run it backwards, we would only get noise due to the fact that the Fourier coecients grow exponentially as t < 0. The backwards heat equation is ill possed. Inhomogeneous Boundary Conditions. There is a simple homogenization transformations that converts a homogeneous heat equation with inhomogeneous Dirichlet boundary conditions (0, t) = (t), (l, t) = (t), t 0, (2.2.23)

into an inhomogeneous heat equation with homogeneous Dirichlet boundary conditions. Suppose (t) (t) x (2.2.24) l where (x, 0) = g(x). (x, t) is a solution of a homogeneous heat equation if and (x, t) = (x, t) (t) + only if (x, t) satises the inhomogeneous equation 2 2 = x t t l subject to the initial condition (x, 0) = g(x) (0) + where (0, t) = (0, l) = 0. (2.2.27) Note that (x, t) is a solution to the homogeneous heat equation if and only if the Dirichlet boundary conditions are constant. As the homogeneous boundary conditions are essential in being able to superpose basic solutions (eigensolutions) the Fourier series method can be used now in conjunction with the separation of variables to obtain solutions of (2.2.25). Example 2.4. Consider 2 2 = x cos t, t t subject to the initial condition 0 < x < 1, t > 0, (2.2.28) (0) (0) x, l (2.2.25)

(2.2.26)

(x, 0) = x,

(2.2.29)

2.2. SEPARATION OF VARIABLES

31

and the following homogeneous boundary conditions: (0, t) = (1, t) = 0, x t > 0. (2.2.30)

First, let us look for a solution of the homogeneous version of (2.2.28) with the given boundary conditions (2.2.30) using the separation of variables method. To this end the reader can easily show that the eigenfunctions are: ui (x) = sin i x, i = (2i + 1) , 2 i = 0, 1, 2 . . . . (2.2.31)

By the analogy with the form of the solution to the homogeneous heat equation let us suppose a solution of (2.2.28) as a series of eigenfunctions

(x, t) =
i=0

i (t) sin i x.

(2.2.32)

Also, represent the right-hand side of (2.2.28) as a series of eigenfunctions. Namely, write

x cos t =
i=0

bi sin i x cos t,

(2.2.33)

where
1

bi =
0

x sin i xdx =

(1)i . 2 i

(2.2.34)

Substituting the solution (2.2.32) with (2.2.33) for its right hand side we are able to show that the unknown functions i (t) satisfy an inhomogeneous ordinary dierential equation di (1)i + 2 i = cos t. i dt 2 i sin t cos t + 2 ]. 4 1 + i i (1 + 4 ) i (2.2.35)

Using the method of undetermined coecients it is easy to obtain its solution i (t) = Aei t + (1)i [
2

(2.2.36)

From the initial condition (2.2.29) and using (2.2.32) one can calculate that (1)i+1 (4 2 + 1) i i A= . 2 4 i (i + 1) This enables us to construct the solution (2.2.32). (2.2.37)

32

2. THE DIFFUSION EQUATION

Periodic Boundary Conditions. Heat ow in a circular ring is governed by the same homogeneous heat equation as is heat conduction in a rod (2.1.6), however, this time subject to periodic boundary conditions (, t) = (, t), t 0, (2.2.38) x x where < x < is the angular variable, and where we assume that the heat (, t) = (, t), can only ow along the ring as no radiation of heat from one side of the ring to another is permitted2. Beneting from the separation of variables technique we are seeking a solution in the form (x, t) = et u(x). Assuming for simplicity that = 1, we arrive, as before, at the associated eigenvalue problem d2 u + u = 0, u() = u(), u () = u (). (2.2.41) dx2 Its solutions are combinations of trigonometric sine and cosine functions ui (x) = ai cos ix + bi sin ix, with the eigenvalues i = i2 , The resulting innite series solution is 1 (x, t) = a0 + 2 be such that

i = 0, 1, 2, . . . ,

(2.2.42)

i = 0, 1, 2, . . . .

(2.2.43)

ei t [ai cos ix + bi sin ix] .

i=1

(2.2.44)

If we postulate the initial condition (x, 0) = g(x) the coecients ai and bi must 1 g(x) = a0 + 2

[ai cos ix + bi sin ix] ,

i=1

(2.2.45)

2The

heat conduction equation for a heated ring can easily be derived from the two(2.2.39)

dimensional heat equation

2 2 = + 2 t x2 y by rewriting its right hand side in polar coordinates (r, ) 1 = t r r r r + 1 2 , r2 2

(2.2.40)

and assuming that the solution is r independent.

2.3. UNIQUENESS OF SOLUTIONS

33

which is precisely the Fourier series of the initial condition g(x) provided ai = 1

g(x) cos ixdx,

bi =

g(x) sin ixdx,

i = 0, 1, 2, . . . . (2.2.46)

2.3. Uniqueness of Solutions In this section we investigate the uniqueness of solutions to the initial-boundary value problem for the heat equation. To this end let us consider solutions of the homogeneous heat equation 2 = , t x2 with the initial condition and the boundary conditions (0, t) = (t), (l, t) = (t). (2.3.3) 0 < x < l, 0 < t < , (2.3.1)

(x, 0) = g(x),

(2.3.2)

Suppose that 1 and 2 are two solutions of (2.3.1) both satisfying the initial condition (2.3.2) and boundary conditions (2.3.3). As the equation (2.3.1) is linear the function (x, t) 1 2 is also a solution but with the zero initial prole and the homogeneous boundary conditions. Let us multiply (2.3.1) by (x, t) and integrate the resulting equation with respect x on the interval [0, l] to obtain
l 0

dx = t

l 0

2 dx. x2

(2.3.4)

Assuming that (x, t) is regular enough, and integrating the right-hand side by parts we reduce the relation (2.3.4) to 1d 2 dt Let I(t) Then,
t l 0 l 0

dx = x
2

l 0

x 1 2

dx =
0 l 0

dx 0.

(2.3.5)

2 dx 0. x
2

(2.3.6)

I(t) I(0) =
0

dxdt 0.

(2.3.7)

34

2. THE DIFFUSION EQUATION

However, I(0) = 0 implying that I(t) 0. On the other hand according to its denition I(t) 0. Hence, I(t) 0. This is possibly only if (x, t) 0 proving that 1 (x, t) = 2 (x, t) everywhere. Note that the same technique can be used to prove uniqueness of solutions to other boundary value problems as long as = 0 at x = 0 and x = l. x 2.4. Fundamental Solutions The idea of the fundamental solution of a partial dierential equation is an extension of the Greens function method for solving boundary value problems of ordinary dierential equations. To set the stage for further considerations let us briey review the main points of the that method3. Consider a homogeneous boundary value problem for the linear ordinary differential equation L(u) = (x ), u(0) = u(l) = 0, 0 < x < l, (2.4.1)

where L(u) denotes a linear second-order dierential operator actig on the function u(x) dened on [0, l] interval, while (x ) (x) is the (Dirac) delta function at . Note that if the boundary conditions are inhomogeneous we can use the homogenization transformation (2.2.24) to transform the problem into one with the homogeneous boundary conditions and a dierent inhomogeneous right hand side. Let u(x, ) = G(x, ) denote the solution to (2.4.1). This is the Greens function of this particular boundary value problem. Once we found this solution we can use linearity to obtain the general solution of L(u) = f (x), u(0) = u(l) = 0, 0 < x < l, (2.4.2)

in the form of the superposition integral . Indeed, let

l

u(x)
0 3Details

G(x, )f ()d.

(2.4.3)

can be found in Section 1.2

2.4. FUNDAMENTAL SOLUTIONS

35

It is easy to see that u(x) solves the boundary value problem (2.4.2) as
l l

L(u) =
0

Lx (G)(x, )f ()d =
0

(x )f ()d = f (x)

(2.4.4)

and the boundary conditions are satised. Lx (u) denotes here the partial dierential operator induced by L. We will try to use the same idea in the context of the heat equation. Consider rst the initial value problem for the heat conduction in an innite homogeneous bar subjected initially to a concentrated unit heat source applied at a point y. We assume for simplicity that the thermal diusivity = 1. This requires solving the heat equation with the initial condition u(x, 0) = (x y), < x < . (2.4.5)

To avoid any specic boundary conditions but to guarantee the uniqueness of solutions (see Section 2.3) we require the solution to be square integrable at all times, that is

|u(x, t)|2 dx <

for all

t 0.

(2.4.6)

This, in fact, implies that the solution vanishes at innity. Let us now take the complex separable solution to the heat equation u(x, t) = ek t eikx ,
2

(2.4.7)

where, as there are no boundary conditions, there are no restrictions on the choice of frequencies k. Mimicking the Fourier series superposition solution when there are innitely many frequencies allowed we may combine these solutions into a Fourier integral (see Appendix B.5) 1 u(x, t) = 2

ek t eikx y (k)dk

(2.4.8)

to realize, provided we can dierentiate under the integral, that it solves the heat equation. Moreover, the initial condition is also satised as 1 u(x, 0) = 2

eikx y (k)dk = (x y),

(2.4.9)

36

2. THE DIFFUSION EQUATION

where y (k) denotes the Fourier transform of the delta function (x y), that is 1 y (k) = eiky . 2 (2.4.10)

Combining (2.4.9) with (2.4.10) we nd that the fundamental solution of the heat equation is 1 F (x y, t) = 2
(xy)2 1 2 ek t eik(xy) dk = e 4t . 2 t

(2.4.11)

It is worth pointing out here that although the individual component of the Fourier series (2.4.8) are not square integrable the resulting fundamental solution (2.4.11) is. Another interesting derivation of the fundamental solution based on the concept of the similarity transformation can be found in [Kevorkian]. Remark 2.5. It is important to point out here that one of the drawbacks of the heat equation model is - as evident from the form of the fundamental solution - that the heat propagates at innite speed. Indeed, a very localized heat source at y is felt immediately at the entire innite bar because the fundamental solution is at all times nonzero everywhere. With the fundamental solution F (x y, t) at hand we can now adopt the superposition integral formula (2.4.3) to construct the solution to the heat conduction problem of an innite homogeneous bar with the an arbitrary initial temperature distribution u(x, 0) = g(x) as 1 u(x, t) = 2 t

(xy)2 4t

g(y)dy.

(2.4.12)

That is, the general solution is obtained by a convolution of the initial data with the fundamental solution. In other words, the solution with the initial temperature prole g(x) is an innite superposition over the entire bar of the point source solutions of the initial strength

g(y) =

(x y)g(x)dx.

(2.4.13)

2.4. FUNDAMENTAL SOLUTIONS

37

Inhomogeneous Heat Equation for the Innite Bar . The Greens function method can also be used to solve the inhomogeneous heat conduction problem u 2 u 2 = p(x, t), t x < x < , t > 0, (2.4.14a)

where the bar is subjected to a heat source p(x, t) which may vary in time and along its length. We impose the zero initial condition u(x, 0) = 0, (2.4.14b)

and some homogeneous boundary conditions. The main idea behind this method is to solve rst the heat equation with the concentrated source applied instantaneously at a single moment, and to use the method of superposition to obtain the general solution with an arbitrary source term. We therefore begin by solving the heat equation (2.4.14a) with the source term p(x, t) = (x y)(t s). (2.4.15)

It represents a unit heat input applied instantaneously at time s and position y. We postulate the same homogeneous initial and boundary conditions as in the general case. Let u(x, t) = G(x y, t s) (2.4.16)

denote the solution to this problem. We will refer to it as the general fundamental solution or a Greens function. Thanks to the linearity of the heat equation the solution of the general problem is given by the superposition integral
t

u(x, t) =
0

G(x y, t s)p(y, s)dyds,

(2.4.17)

where the forcing term may be also rewritten by the superposition formula as

p(x, t) =
0

p(y, s)(t s)(x y)dyds.

(2.4.18)

If we replace the zero initial condition by u(x, 0) = f (x), then once again due to the linearity of the dierential equation we may write the solution as a combination of a solution to the homogeneous equation with inhomogeneous initial data

38

2. THE DIFFUSION EQUATION

and the solution with the homogeneous initial condition but a nonzero forcing term
t

u(x, t) =

F (x y, t)f (y)dy +
0

G(x y, t s)p(y, s)dyds. (2.4.19)

To nd the general fundamental solution in an explicit form let us take the Fourier transform with respect to variable x of both sides of the dierential equation (2.4.14a) with the forcing term (2.4.15). Using (2.4.10) we nd that du 1 + k 2 u = eiky (t s), dt 2 (2.4.20)

where u(k, t) denotes the Fourier transform of u(x, t), and where k is viewed as a parameter. This is an inhomogeneous rst order linear ordinary dierential equation for the Fourier transform of u(x, t) with the initial condition u(k, 0) = 0 for s > 0.
2t

(2.4.21) we obtain

Using the integrating factor method with the integrating factor ek that

1 2 u(k, t) = ek (ts)iky (t s), (2.4.22) 2 where (t s) is the usual step function. The Greens function is than obtained by the inverse Fourier transform 1 G(x y, t s) = 2 (t s) 2 (t s) 2 (t s)

eikx u(k, t)dk.

(2.4.23)

Using the formula (2.4.11) of the fundamental solution we deduce that

G(x y, t s) = =

eik(xy)+k

2 (ts)

dk .

(2.4.24)

exp

(x y)2 4(t s)

The general fundamental solution (Greens function) is just a shift of the fundamental solution for the initial value problem at t = 0 to the starting time t = s. More importantly, its form shows that the eect of a concentrated heat source applied at the initial moment is the same as that of a concentrated initial temperature.

2.4. FUNDAMENTAL SOLUTIONS

39

Finally, the superposition integral (2.4.17) gives us the solution

t

u(x, t) =
0

p(y, s) 2 (t s)

exp

(x y)2 4(t s)

dsdy

(2.4.25)

of the heat conduction problem for the innite homogeneous bar with a heat source. Heat Equation for the Semi-innite Bar . To illustrate how the Greens function method can be applied in the case of the semi-innite domain we consider the heat equation with the concentrated forcing term u 2 u 2 = (x y)(t), t x boundary conditions u(0, t) = 0,
x

0 x < ,

t>0

(2.4.26a)

and impose the zero initial condition, i.e., u(x, 0) = 0, and the homogeneous lim u(x, t) = 0. (2.4.26b)

As we have remarked earlier the eect of such a concentrated instantaneous heat source is the same as that of the concentrated initial distribution. Thus, the only dierence between this case and the case of the fundamental solution (2.4.5) is the imposition of the boundary condition at x = 0. One possible way to tackle this diculty is to consider in place of this semi-innite problem such an innite domain problem in which the homogeneous boundary condition at x = 0 is permanently satised. Hence, consider the heat conduction problem for an innite homogeneous bar with (t)[(x y) (x + y)] as the forcing term, homogeneous initial condition, and the homogeneous boundary conditions at innities. In other words, in the innite domain we apply a unit strength source at x = y, and simultaneously a negative source of unit strength at x = y. This approach is known as the method of images. Once again, due to the linearity of the heat equation and that of the forcing term, the temperature prole at t > 0 will be the sum of two fundamental solutions F (x y, t) and F (x + y, t) each corresponding to one of the source terms. In particular, due to the skew-symmetry of these solutions the combined solution will always be vanishing at x = 0. Moreover since all the

40

2. THE DIFFUSION EQUATION

boundary conditions of the original problem (2.4.26a) are satised, and since the second source term (t)(x + y) is outside of the original semi-innite domain, the Greens function G(x y, t) F (x y, t) F (x + y, t) (2.4.27)

is the solution of (2.4.26a), where the fundamental solution F is dened by (2.4.11). In conclusion, the solution of the inhomogeneous heat conduction problem for a semi-innite bar u 2 u 2 = p(x, t), 0 x < , t > 0, (2.4.28) t x with the initial condition u(x, 0) = 0 and the homogeneous boundary conditions (2.4.26b) has the form
t 0

u(x, t) =
0

p(y, s) 2 (t s)

exp

(x + y)2 (x y)2 exp 4(t s) 4(t s)

dyds. (2.4.29)

Example 2.6. Suppose that a semi-innite homogeneous bar is initially heated to a unit temperature along a nite interval [a, b], where a > 0. Assume also that at x = 0 the temperature is held at zero (by attaching an innite rod of this temperature) and vanishes at innity. This corresponds to the following initial value problem for the heat equation: u 2u = , t x2 0, if 0 < x < a, 1, if a < x < b, 0, if x > b, (2.4.30a)

u(x, 0) = (x a) (x b) =

with the homogeneous boundary conditions u(0, t) = 0,

x

lim u(x, t) = 0,

t > 0.

The method of images and the superposition formula (2.4.12) yield the solution 1 u(x, t) = 2 t 1 = erf 2
b

e
a

(xy)2 4t

dy +
a

(xy)2 4t

dy erf xb 2 t + erf

(2.4.31) x+b 2 t ,

xa 2 t

+ erf

x+a 2 t

1 2

2.5. BURGERS EQUATION

41

where the error function

z 2 2 erfz e d. (2.4.32) 0 Note that the error function is odd and that its asymptotic value at innity is 1.

2.5. Burgers Equation In this last section we will study the quasilinear version of the diusion equation u u 2u +u = 0, > 0, (2.5.1) t x x2 to show how the solution methods developed in previous sections for the heat equation may be used to obtain solutions for other equations. Also, Burgers equation is a fundamental example of an evolution equation modelling situations in which viscous and nonlinear eects are equally important. Moreover, it plays somewhat important role in discussing discontinuous solutions (shocks) of the one-dimensional conservation law u u +u = 0, (2.5.2) t x a topic which will not be discussed here (see for example [Knobel], [Smoller]). We start by looking at ways at which the methods for solving the initialboundary value problems of heat equations can be used to solve (2.5.1). The Cole-Hopf Transformation. This is a change of dependent variable w = W (u) which enables us to transform Burgers equation into the linear diusion equation studied already in this wx , (2.5.3) w where wx denotes partial dierentiation. Calculating all derivatives and substiu 2 tuting them into (2.5.1) yields wx ( wxx wt ) w( wxx wt )x = 0. In particular, if w(x, t) solves the diusion equation wxx wt = 0, the function u(x, t) given by (2.5.3) satises Burgers equation (2.5.1). (2.5.5) (2.5.4) chapter. Let

42

2. THE DIFFUSION EQUATION

Initial Value Problem on the Innite Domain. Let us consider the following initial value problem: u u 2u +u = 0, t x x2 u(x, 0) = g(x), < x < , (2.5.6)

and suppose that we are looking for the solutions which satisfy the corresponding diusion equation (2.5.5). According to (2.5.3), the initial condition for the new variable w(x, 0) must be such that g(x)w(x, 0) = 2 wx (x, 0). (2.5.7)

The general solution of this linear ordinary dierential equation for w(x, 0) is w(x, 0) = A exp 1 2
x

g(s)ds ,
0

(2.5.8)

where A is a constant, and where we assume that the integral exists. Hence, we essentially need to solve the following initial value problem for the homogeneous diusion equation with the inhomogeneous initial condition: wxx wt = 0, w(x, 0) = h(x), < x < . (2.5.9)

Its solution has the form of (2.4.12) : 1 w(x, t) = 2 t

(xy)2 4 t

h(y)dy

(2.5.10)

where we replaced t by t Note that the parameter

may be eliminated from

the equation, and so from the solution, by an appropriate scaling of variables. We retain it, however, so we one can later study the asymptotic behavior of solutions when 0. Dierentiating with respect to x and using the Cole-Hopf formula (2.5.3) we compute that u(x, t) = where h(y) exp as the constant A cancels out. 1 2
0 (xy) t (xy)2 4 t

exp

h(y)dy

exp

(xy)2 4 t

h(y)dy

(2.5.11)

g(s)ds ,

(2.5.12)

2.5. BURGERS EQUATION

43

Boundary Value Problem on a Finite Interval . Using separation of variables method, we solve here the following initialboundary value problem: u u 2u +u = 0, t x x2 u(x, 0) = g(x), 0 < x < a, t > 0. (2.5.13a) (2.5.13b)

u(0, t) = u(a, t) = 0, value problem for the diusion equation: wt wxx = 0, w(x, 0) = Ah(x),

After Cole-Hope transformation we obtain the corresponding initial-boundary

0 < x < a, t > 0.

(2.5.14a) (2.5.14b)

wx (0, t) = wx (a, t) = 0, derived (see page 27) as

As the boundary condition are homogeneous the solution w(x, t) can easily be
k 2 a0 k w(x, t) = ak e( a ) t cos + x, 2 a k=1

(2.5.15)

where

k 2A a h(x) cos ak = xdx. (2.5.16) a 0 a From the Cole-Hope transformation formula (2.5.3) one now gets the solution to the initial vale problem (2.5.13) u(x, t) = 2
a a0 2 k=1 kak exp k=1 ak exp

( k )2 t sin k x a a ( k )2 t cos k x a a

(2.5.17)

Example 2.7. Consider Burgers equation ut + uux uxx = 0, with the the piecewise initial condition < x, , (2.5.18)

1, if x < 0 u(x, 0) = 2(x) 1 = 1, if x > 0.

(2.5.19)

The initial condition of the associated diusion equation (2.5.5) may now be obtained from (2.5.7): w(x, 0) = Ae 2
1

|x|

(2.5.20)

44

2. THE DIFFUSION EQUATION

The solution w(x, t) takes the form of (2.4.12) with t replaced by t. Namely, 1 w(x, t) = 2 t
(xy) t

(xy)2 4 t

e 2 dy.

|y|

(2.5.21)

Therefore, the solution of the original initial value problem is given by (2.5.11): u(x, t) = exp
(xy)2 4 t

)e 2 dy
|y|

|y|

exp

(xy)2 4 t

e 2 dy

(2.5.22)

Integrating independently from to 0 and from 0 to , and using the substitution = respectively, we nally obtain u(x, t) = e erfc e erfc
x x

(x y t) , 2 t
xt 2 t xt 2 t x+t erfc 2 t x+t + erfc 2 t

(2.5.23)

where the complimentary error function 2 erfc(z) 1 erf(z) =

z

e d.

(2.5.24)

Solutions for Problems for The 1-D Heat Equation

Problem 4
@2u @u = ; @t @x2

Solve the inhomogeneous heat problem with type I boundary conditions: u (0; t) = 0 = u (1; t) ; u (x; 0) = P" (x)

where t > 0, 0

1, and P" (x) = ( 0

u0 "

if x if x

1 2 1 2

>

" 2 " 2

(5)

Note: you already know the solution (just replace P" (x) with f (x) and write down the solution from class). Using symmetry of P" (x) about 1/2 can be used to simplify the calculation of the Fourier coe cients.

Solution: This is the Heat Problem with Type I homogeneous BCs. The solution we derived in class is, with f (x) replaced by P" (x), u (x; t) =
1 X n=1

un (x; t) =

1 X n=1

Bn sin (n x) e

n2

2t

(6)

where the Bn are the Fourier coe cients of f (x) = P" (x), given by s Z 1 Bn = 2 P" (x) sin (n x) dx
0

Breaking the integral into three pieces and substituting for P" (x) from (5) gives Z 1=2 "=2 Z 1=2+"=2 Z 1 Bn = 2 P" (x) sin (n x) dx + 2 P" (x) sin (n x) dx + 2 P" (x) sin (n x) dx
0

= 0+2 2u0 = " = u0

1=2 "=2

1=2+"=2

1=2+"=2

1=2 "=2

u0 sin (n x) dx + 0 "
1=2+"=2 1=2 "=2 cos n2

cos (n x) n
n 2

cos

(1

") "n =2

(1 + ")

(7)

We apply the cosine rule cos (r s) cos (r + s) = 2 sin r sin s 6

with r = n =2, s = n "=2 to Eq. (7), Bn = 4u0 n n " sin sin "n 2 2

When n is even (and nonzero), i.e. n = 2m for some integer m, B2m = When n is odd, i.e. n = 2m B2m 2u0 sin m sin m " = 0 "m

1 for some integer m,

1

= 2u0 ( 1)m+1

sin ((2m 1) "=2) : (2m 1) "=2

2

(8) is,

(a) The temperature at the midpoint of the rod, x = 1=2, at scaled time t = 1= from (6) and (8), u (x; t) = = For t
1 X

2u0 ( 1)m+1 2u0

2 e(2m 1)

m=1 1 X

sin ((2m 1) "=2) sin (2m (2m 1) "=2 :

1)

(2m 1)2

m=1

sin ((2m 1) "=2) (2m 1) "=2

1= 2 , the rst term gives a good approximation to u (x; t), u 1 1 ; 2 2 u1 1 1 ; 2 2 = 2u0 e sin ( "=2) "=2 :

"=2 To distinguish between pulses with " = 1=1000 and " = 1=2000, note that lim"!0 sin "=2 = 1, 1 and so for smaller and smaller ", the corresponding temperature u 2 ; 12 gets closer and closer to 2u0 =e,

u In particular, u1

1 1 ; 2 2

u1

1 1 ; 2 2

2u0 e

" + 2 3!

2 2

"

1:

1 1 1 ; 2;" = 2 1000

u1

1 1 1 ; 2;" = 2 2000

2u0 sin ( =2000) e =2000 2u0 3:1 10 7 e

sin ( =4000) =4000

Thus it is hard to distinguish these two temperature distributions, at least by measuring the temperature at the center of the rod at time t = 1= 2 . By this time, diusion has smoothed out some of the details of the initial condition. 7

Figure 1: Time temperature proles u (x0 ; t) at x0 = 0:5, 0:4 and 0:1 (from top to bottom). The t-axis is the time prole corresponding to x0 = 0, 1.

(b) Illustrate the solution qualitatively by sketching (i) some typical temperature proles in the u t plane (i.e. x = constant) and in the u x plane (i.e. t = constant), and (ii) some typical level curves u (x; t) = constant in the x t plane. At what points of the set D = f(x; t) : 0 x 1; t 0g is u (x; t) discontinuous? The solution u (x; t) is discontinuous at t = 0 at the points x = (1 ") =2. That said, u (x; t) is piecewise continuous on the entire interval [0; 1]. Thus, the Fourier series for u (x; 0) converges everywhere on the interval and equals u (x; 0) at all points except x = (1 ") =2. The temperature proles (u t plane, u x plane), 3D solution and level curves are shown.

Problem 5

Consider two iron rods (thermal diusivity = 0:15 cm2 sec 1 ) each 20 cm long and with insulated sides, one at a temperature of 100o C and the other at 0o C throughout. The rods 8

Figure 2: Spatial temperature proles u (x; t0 ) at t0 = 0 (dash), 0:001, 0:01, 0:1. The x-axis from 0 to 1 is the limiting temperature prole u (x; t0 ) as t0 ! 1.

are joined end to end in perfect thermal contact, and their free ends are kept at 0o C. Show that the temperature at the interface 10 minutes after contact has been made approximately 36.5o C. Find an upper bound for the error in your answer. Can this method be applied if the rods are made of glass (thermal diusivity = 0:006 cm2 sec 1 )? Solution: The rods are placed end-to-end and treated as one rod with length l = 40 cm. We dene the dimensionless spatial coordinate x = x0 =l. Let u (x; t) be the temperature in the joined rods, for x 2 [0; 1] and t 0. The join is at x = 1=2. The initial temperature distribution in the joined rods is ( 100; 0 x 1=2; u (x; 0) = f (x) = (9) 0; 1=2 x 1: Since the ends of the rod are held at 0o C, the boundary conditions are u (0; t) = 0 = u (1; t). Since there are no sources in the rods, the homogeneous Heat Equation ut = uxx governs the variation in temperature. The problem for u (x; t) is thus the basic Heat Problem with Type I homogeneous BCs and IC f (x). From the derivation in class, we found the solution to be u (x; t) =
1 X n=1

Bn sin (n x) exp 10

n2

Figure 3: Level curves u (x; t) =u0 = C for various values of the constant C. Numbers adjacent to curves indicate the value of C. The line segment (1 ") =2 x (1 + ") =2 at t = 0 is the level curve with C = 1=" = 10. The lines x = 0 and x = 1 are also level curves with C = 0.

11

where Bn = 2

To save time, we note that we only desire the solution at x = 1=2, u (x; t) =
1 X n=1

f (x) sin (n x) dx

(10)

Bn sin

n 2

exp

n2

Since sin (n =2) is zero for even n, the sum is over the odd terms, u (x; t) = =
1 X k=1

B2k B2k

sin

(2k 2
1

1)

exp (2k 1)2

(2k
2

1)2

t (11)

1 X k=1

( 1)k

exp

t :

Substituting the IC (9) into (10) and setting n = 2k 1 gives Z 1=2 (2k 1) 200 1 cos B2k 1 = 200 sin ((2k 1) x) dx = (2k 1) 2 0 Substituting the B2k u
1

200 : (2k 1) (12)

in (12) into the expression (11) for u (1=2; t) gives 1 ;t 2 200 X ( 1)k 1 = exp (2k 1) k=1
1

(2k

1)2

t :

(13)

We are asked to ne the temperature at x = 1=2 after t0 = 10 minutes. This corresponds to a scaled time of t10 = 10 mins l2 = 0:15=402 10 = 0:006=402 10

60 ' 0:056 60 ' 0:002

for iron ( = 0:15 cm2 /s) for glass ( = 0:006 cm2 /s)

Recall in the notes we made the rst term approximation for t 1= 2 ' 0:1, and hence both these values fall under that. To see how the number of terms retained aects the sum, we compute u (1=2; t) from (13) for various numbers of terms. For iron (t = t10 = 0:056), we obtain u 1 ; t10 2 ' 36: 631 (1 term) ' 36: 484 (2 terms) ' 36: 484 (3 or more terms) 12 (14)

In this case, the rst term u1 (1=2; t10 ) does a good job of approximating the series for u (1=2; t10 ). For glass (t = t10 = 0:002), u 1 ; t10 2 ' 62: 4 (1 term) ' 44:7 (2 terms) ' 52:4 (3 terms) ' 49:0 (4 terms) ' 50:4 (5 terms) ' 49:9 (6 terms) ' 50:04 (7 terms) In this case, the convergence is much slower and the rst term u1 (1=2; t10 ) is a poor estimate of u (1=2; t10 ). The upper bound on the error was discussed in 6.2. The approximate error we derived in class is, since the series for u (1=2; t) only has odd terms, u 1 ;t 2 u1 1 ;t 2 Be 3 1 e 2
2t 2t

(15)

where B is the upper bound for B2k 1 for all k = 2; 3; :::. In the notes, we wrote Z 1 1 jBn j 2 jf (x)j dx = 2 100 = 100 2 0 However, we can obtain a better approximation since we have the formula for Bn jB2k 1 j = 200 (2k 1) 200 3 k = 2; 3; :::

Therefore, B = 200= (3 ) and from (15), u 1 ;t 2 u1 1 ;t 2 200 e 3 3 1 e

2t

2t

< 6:1 for t

t10 = 0:056:

This error bound is still not very good - in (14) the error between u (1=2; t10 ) (for iron, t10 = 0:056) and the rst term u1 (1=2; t10 ) is roughly 0:15, much less than 6:1. I have now added a much better estimate to 6.2. It turns out that, since the series u (1=2; t) only has odd terms, 2 2 200 e 9 t 1 Be 9 t 1 ;t u1 ;t = u 2 2 3 1 e 6 2t 1 e 6 2t Now the error for t = t10 = 0:056 is 0:152, which is more in line with (14). 13

Problem 7
@u @2u = ; 0 < x < 1; t>0 @t @x2 @u (1; t) ; t>0 u (0; t) = 0 = @x u (x; 0) = f (x) 0 < x < 1:

Consider the heat problem with dimensionless position and time, ow (16)

Solution: (a) The physical signicance of the condition ux (1; t) = 0 is that the end of the rod at x = 1 is insulated, i.e. the heat (proportional to ux by Fourier law) is zero at x = 1. ux s R1 2 (b) Showing that u (t) = 0 u (x; t) dx is non-increasing in time follows from the derivation in 8.1 of the lecture notes and noting that uux = 0 at x = 0; 1 since u = 0 at x = 0 and ux = 0 at x = 1. (c) Proving that (16) has at most one solution follows the derivation in class. Take two solutions u1 , u2 of (16) and dene v (x; t) = u1 u2 . Then show that the function R1 v (t) = 0 v 2 (x; t) dx is non-increasing as in part (b). (d) To nd a series solution for f (x) = u0 , u0 a constant, we use separation of variables, u (x; t) = X (x) T (t) The PDE in (16) gives the usual T0 X 00 = = X T where is constant since the left hand side is a function of x only and the middle is a function of t only. Substituting (17) into the BCs in (16) gives X (0) = dX (1) = 0 dx (17)

The Sturm-Liouville boundary value problem for X (x) is thus X 00 + X = 0; Let us try < 0. Then the solutions are X (x) = c1 e p
j jx

X (0) =

dX (1) = 0 dx

(18)

p + c2 e j

jx

14

and imposing the BCs gives c1 = c2 = 0, i.e. X (x) must be the trivial solution. For = 0, X (x) = c1 x + c2 and, again, imposing the BCs gives c1 = c2 = 0 and X (x) is the trivial solution. Thus, in order to have a nontrivial solution, must be taken positive. In this case, p p X = c1 sin x + c2 cos x The BC X (0) = 0 implies c2 = 0. The other BC implies 0= p p dX (1) = c1 cos dx > 0 then we must have cos

For a non-trivial solution, c1 must be nonzero. Since which implies the eigenvalues are
n

= 0,

(2n 4

1)2

; (2n

n = 1; 2; 3; ::: 1) 2
nt

and the eigenfunctions are Xn (x) = sin x

For each n, the solution for T (t) is Tn (t) = e u (x; t) = At t = 0,

1 X n=1

Bn sin

. Hence the series solution for u (x; t) is ! (2n 1)2 2 (2n 1) x exp t (19) 2 4
1 X n=1

f (x) = u (x; 0) =

Bn sin

(2n 2

1)

(20)

The orthogonality conditions are found using the identity 2 sin (2n 2 1) x sin (2m 1) x 2 = cos ((m n) x) cos ((1 m n) x)

The last integral follows since 1 m n cannot be zero for any positive integers m, n. Thus, the orthogonality conditions are ( Z 1 1=2 m = n (2n 1) (2m 1) x sin x dx = (21) sin 2 2 0 m 6= n 0 15

Note also that for m; n = 1; 2; 3:::, we have ( Z 1 1 m=n cos ((m n) x) dx = 0 m 6= n 0 Z 1 cos ((1 m n) x) dx = 0
0

Multiplying each side of (20) by sin ((2m 1) x=2), integrating from x = 0 to 1, and applying the orthogonality condition (21) gives Z 1 (2n 1) x f (x) dx (22) Bn = 2 sin 2 0 Substituting f (x) = u0 into (22) gives Z 1 (2n 1) 4u0 Bn = 2u0 sin x dx = 2 (2n 1) 0 Thus, the series solution is u (x; t) =
1 4u0 X n=1

cos

(2n 2

1)

4u0 (2n 1) (23)

1 (2n 1)

sin

(2n 2

1)

x exp

(2n 4

1)2

t :

An approximate solution valid for large times is the rst term, u (x; t) u1 (x; t) = 4u0 sin x exp 2
2

Similar upper bounds on error can be derived as in the notes. Temperature proles (u vs. x) are plotted below for dierent times.

Problem 8

Suppose a chemical is dissolved in water, in some long thin reaction container and let (moles/cm3 ) indicate its concentration. Fick Law in chemistry states that the rate of s diusion of a solute is proportional to the negative gradient of the solute concentration. Assume that the chemical is created, due to a chemical reaction, at a rate g (x; t) (moles/cm3 sec). (a) Derive a PDE describing the distribution of . Formulate appropriate BCs and IC and state all assumptions. (b) Show that the solution to the initial boundary value problem derived in (a) is unique. Solution: The derivation is analogous to that of the Heat Equation with a source. Mass conservation of the reactant is used in place of energy conservation, and Fick Law is used s in place of Fourier Law. s Consider a thin segment from x to x + x of the reaction container, of cross-sectional area A. Let (x; t) be the concentration of the reactant at position x along the container 16

Figure 4: Temperature proles u (x; t0 ) at various times t0 = 0:001, 0:01, 0:1 and 0:7 (from left to right). Dashed line indicates the initial condition. The x-axis is the limit of the solution as t ! 1.

17

and at time t. Analogous to the derivation of the heat equation, conservation of mass gives change of concentration in segment in time reactant in from = left boundary reactant reactant out from + generated : (24) right boundary in segment

The last term in the mass balance equation is just gA x t. Fick Law states that the s reactant in and out from the left and right boundaries is, respectively, tA F0 @ @x ;
x

tA

F0

@ @x

x+ x

where F0 is the chemical diusivity. Therefore, (24) becomes A x (x; t + t) A x (x; t) = tA F0 @ @x tA

x

F0

@ @x

+ gA x t
x+ x

Dividing by A x t and rearranging yields (x; t + Taking the limit t) t (x; t) = F0

@ @x x+ x @ @x x

+ g:

@2 @ = F0 2 + g (25) @t @x We assume the concentration is smooth. For BCs, the ends of the reaction container are closed, so that x = 0 at x = 0; l (Type II homogeneous BCs). Alternatively, we could be supplying or removing reactant at the ends, keeping the concentration xed: = 0 at x = 0; l (Type I inhomogeneous BCs). The IC is (x; 0) = f (x) where f (x) is the initial distribution of reactant. If the container is well mixed, then f (x) = u0 . If there is no reactant initially in the container, then (x; 0) = 0. Whatever the IC, we assume it is smooth. To show uniqueness, we note that given two solutions u1 , u2 , we dene the dierence v (x; t) = u1 u2 , which satises the homogeneous diusion equation
t

t; x ! 0 gives the chemical diusion equation with a source,

= F0

xx

Similarly, for either Type II homogeneous or Type I inhomogeneous BCs on u1 and u2 , the BCs on v (x; t) are homogeneous Type I or II. In either case, we dene the mean concentration as Z
1

v (t) =

v 2 (x; t) dx

and follow the derivation in the lecture notes.

18