This action might not be possible to undo. Are you sure you want to continue?

Diﬀusion equations

2.1 Separation of variables: intervals

Diﬀusion equation is a linear partial diﬀerential equation, since the functions related to u in the

equations (u

t

and ∆u) are both linear. Recall for the linear ordinary diﬀerential equations:

dP

dt

= kP, and D

d

2

P

dx

2

= kP, (2.1)

it is well-known that e

kt

and Acos(

k/Dtx) + Bsin(

**k/Dx) are the solutions of the respective
**

equations. Indeed in the second case, the exponential function e

kitx/D

is a complex solution. Thus

we can make a lucky guess for the solution of the diﬀusion equation:

∂P

∂t

= D

∂

2

P

∂x

2

, (2.2)

that the solution is an exponential function

P(t, x) = e

at+bx

, (2.3)

for some constants a, b, which could be complex numbers. By substitution, we ﬁnd that the function

in (2.3) is a solution of (2.2) if a = Db

2

, and depending on the values of b, we obtain three families

of solutions:

P(t, x) = e

Db

2

t

e

bx

, b > 0, (2.4)

P(t, x) = e

Db

2

t

e

−bx

, b > 0, (2.5)

P(t, x) = e

−Db

2

t

e

−bix

, b > 0. (2.6)

The ﬁrst two families of solutions are not reasonable solutions, since when x → ∞ (or −∞)

P(t, x) →∞, which implies an unlimited growth at x = ∞ (or −∞). The third family of solutions

are complex, but their real and imaginary parts are both solutions of (2.2) too:

P(t, x) = e

−Db

2

t

cos(bx), and P(t, x) = e

−Db

2

t

sin(bx). (2.7)

19

20 CHAPTER 2. DIFFUSION EQUATIONS

These solutions are little more “reasonable” as they are bounded as x → ∞, but still they do not

satisfy natural boundary conditions P, P

x

→0 as x →∞.

Nevertheless solutions with above forms are solutions of diﬀusion equations, and we notice that

they are in form of P(t, x) = U(t)V (x). This motivates us to consider the solutions which is

separable on the two variables t and x:

u(t, x) = U(t)V (x). (2.8)

With the form of u in (2.8), the diﬀusion equation (2.2) becomes

U

′

(t)V (x) = DU(t)V

′′

(x), and

U

′

(t)

U(t)

= D

V

′′

(x)

V (x)

= Dk,

where k is a constant independent of both t and x. Thus U satisﬁes an equation:

U

′

(t) = DkU(t), t > 0, (2.9)

and

U(t) = C

1

e

Dkt

; (2.10)

and V satisﬁes

V

′′

(x) = kV (x). (2.11)

The solutions of (2.11) are C

2

e

√

kx

with any complex number k. Thus we obtain solution of (2.2):

Ce

Dkt

e

√

kx

, (2.12)

just like the three families of solutions we obtain by making lucky guess. To obtain more speciﬁc

solutions, we must take the boundary and initial conditions into considerations.

We start with the one-dimensional diﬀusion equation on an interval with length L:

∂P

∂t

= D

∂

2

P

∂x

2

, x ∈ (0, L). (2.13)

Here the individuals of the the species inhabits a one dimensional patch (0, L), and the diﬀusion

coeﬃcient of the population is D > 0. To get a deﬁnite solution of the problem, we need to have

an initial population distribution:

P(0, x) = P

0

(x), x ∈ (0, L), (2.14)

and a boundary condition. We assume that the exterior is hostile so homogeneous Dirichlet bound-

ary condition is satisﬁed:

P(t, 0) = P(t, L) = 0. (2.15)

We look for solution with form P(t, x) = U(t)V (x). Then we obtain (2.9) and (2.11), and (2.15)

implies that

V (0) = V (L) = 0. (2.16)

(2.11) and (2.16) become a boundary value problem for an ordinary diﬀerential equation, which is

a special case of Sturm-Liouville problems. We shall discuss this kind of problems in a separate

section next.

2.2. EIGENVALUES AND EIGENFUNCTIONS 21

2.2 Eigenvalues and eigenfunctions of boundary value problems

We consider the Dirichlet boundary value problem:

y

′′

= ky, x ∈ (0, L), y(0) = y(L) = 0. (2.17)

We can see that y(x) = 0 is always a solution of the equation. So we shall look for solutions other

than the trivial zero solution. From elementary ODE knowledge, we know that the general solution

of y

′′

= ky is one of the following three forms:

c

1

e

−

√

kt

+c

2

e

√

kt

, if k > 0, (2.18)

c

1

+c

2

t, if k = 0, (2.19)

c

1

cos(

√

−kt) +c

2

sin(

√

−kt), if k < 0. (2.20)

One important characteristic of a boundary value problem is that it may not have a solution, while

the initial value problem always has a solution and the solution is unique if all terms in the equation

are diﬀerentiable. For (2.17), there is no solution when k > 0. Indeed if there is a solution, it must

be y(t) = c

1

e

−

√

kt

+c

2

e

√

kt

for some constants c

1

and c

2

, and we can solve the constants by

0 = y(0) = c

1

+c

2

, 0 = y(L) = c

1

e

−

√

kL

+c

2

e

√

kL

.

But this system of equations is unsolvable: if one gets c

1

= −c

2

from the ﬁrst equation and plugs

it into the second one, one has (e

−

√

kL

−e

√

kL

)c

2

= 0, and one must have c

2

= 0 and c

1

= 0. Thus

(2.17) has only zero solution when k > 0. It is also easy to check that (2.17) has only zero solution

when k = 0. So the only meaningful case is when k < 0.

In this case, (2.17) may still have only the zero solution, but for some number k < 0, it can

have a solution y(t) which is not the zero solution. (Note that from the linear principle, if y(t) is a

solution of (2.17), so is c · y(t) for any constant c, and if y

1

(t) and y

2

(t) are solutions of (2.17), so

is y

1

(t) +y

2

(t).) So for k which (2.17) has non-zero solutions, we call these numbers eigenvalues,

and the solutions of (2.17) eigenfunctions. (In fact, historically the solutions of (2.17) were ﬁrst

called eigenvalues and eigenfunctions.)

Now let’s search for the eigenvalues and eigenfunctions of (2.17): the eigenfunction must be

y(t) = c

1

cos(

√

−kt) + c

2

sin(

√

−kt) for some constants c

1

and c

2

; by the boundary conditions, we

have

0 = y(0) = c

1

, 0 = y(L) = c

1

cos(

√

−kL) +c

2

sin(

√

−kL);

so c

1

= 0 and if c

2

is not zero, sin(

√

−kL) = 0; thus

√

−kL = mπ, the eigenvalue k must be

k = −

m

2

π

2

L

2

, (2.21)

and the eigenfunctions associated with k = −

m

2

π

2

L

2

are c sin

mπt

L

. Therefore (2.17) has a

sequence of eigenvalues: k

m

= −

m

2

π

2

L

2

, and the set of eigenfunctions are the sine functions which

22 CHAPTER 2. DIFFUSION EQUATIONS

can “ﬁt in” the interval (0, L). When the interval is (0, π) (so L = π), the expression of eigen-pairs

is much simpler:

k

m

= −m

2

, f

m

(t) = sin(mt). (2.22)

Back to (2.13) and (2.15), we ﬁnd a sequence of solutions:

u

m

(t, x) = exp

−D

m

2

π

2

L

2

t

sin

mπx

L

, m is a positive integer. (2.23)

Because of the linear principle, the general solution of (2.13) is

u(t, x) =

∞

¸

m=1

c

m

exp

−D

m

2

π

2

L

2

t

sin

mπx

L

, t > 0, x ∈ (0, L). (2.24)

Example 2.1.

y

′′

+ky = 0, y(0) +y

′

(0) = y(π) = 0. (2.25)

This problem is same as previous example except the boundary conditions. So we still consider

the cases of k < 0, k = 0 and k > 0. When k < 0, y(t) = c

1

e

−

√

−kt

+ c

2

e

√

−kt

, from the boundary

conditions, we get y(0)+y

′

(0) = c

1

+c

2

−

√

−kc

1

+

√

−kc

2

= 0 and y(π) = c

1

e

−

√

−kπ

+c

2

e

√

−kπ

= 0,

So (c

1

, c

2

) satisﬁes the equations

1 −

√

−k 1 +

√

−k

e

−

√

−kπ

e

√

−kπ

·

c

1

c

2

=

0

0

(2.26)

The equation has non-zero solution only if det

1 −

√

−k 1 +

√

−k

e

−

√

−kπ

e

√

−kπ

**= 0. The determinant equals
**

to (1 −p)e

pπ

−(1 +p)e

−pπ

if we denote p =

√

−k, thus det = 0 is equivalent to e

2pπ

=

1 +p

1 −p

. The

functions f

1

(p) = e

2pπ

and f

2

(p) =

1 +p

1 −p

have no common points when p > 0 (in fact, they only

intersect at p = 0. thus e

2pπ

=

1 +p

1 −p

is not possible for p > 0, and there is no eigenvalue such that

k < 0. When k = 0, y(t) = c

1

+c

2

t, from the boundary conditions, we get y(0) +y

′

(0) = c

1

+c

2

= 0

and y(π) = c

1

+c

2

π = 0, since the determinant of the matrix

1 1

1 π

**is not zero, then only solution
**

for (c

1

, c

2

) is (0, 0). So 0 is not an eigenvalue either.

For k > 0, y(t) = c

1

cos(

√

kt) + c

2

sin(

√

kt). The boundary conditions: y(0) + y

′

(0) = c

1

+

c

2

√

k = 0, y(π) = c

1

cos(

√

kπ) +c

2

sin(

√

kπ) = 0. The matrix is

1

√

k

cos(

√

kπ) sin(

√

kπ)

, and the

determinant of the matrix is 0 if sin(

√

kπ) −

√

k cos(

√

kπ) = 0, or tan(πp) = p for p =

√

k. The

graphs of f

1

(p) = tan(πp) and f

2

(p) = p intersect at a sequence of points, p

m

, m = 1, 2, 3, · · · , and

0.5 < p

1

< 1.5, 1.5 < p

2

< 2.5, m − 0.5 < p

m

< m + 0.5. k

m

= p

2

m

are the eigenvalues of the

problem (2.25), and the eigenfunctions are y

m

(t) =

√

k

m

cos(

√

k

m

t) −sin(

√

k

m

t).

2.3. FOURIER SERIES 23

2.3 Fourier series

To completely solve (2.13)-(2.15), we need to determine the constants c

m

in (2.24) by setting t = 0

in (2.24):

u

0

(x) =

∞

¸

m=1

c

m

sin

mπx

L

, x ∈ (0, L). (2.27)

The expression in the right-hand side of (2.24) is the Fourier series of u

0

(x) on (0, L). Fourier

series is widely used in mathematical physics and signal process. All sine and cosine functions are

considered to be regular smooth waves, and Fourier series convert any noise (or signal) into a sum

of more regular waves, and that is a big help for the signal processing. Here we collect a few facts

of Fourier series.

Theorem 2.2. Let f(x) and f

′

(x) be piecewise continuous functions on the interval [0, L]. Then

f(x) can be expanded in either a pure sine series

f(x) =

∞

¸

m=1

b

m

sin

mπx

L

, (2.28)

where b

m

=

2

L

L

0

f(x) sin

mπx

L

**dx, m = 1, 2, · · · , or a pure cosine series
**

f(x) =

a

0

2

+

∞

¸

m=1

a

m

cos

mπx

L

, (2.29)

where a

m

=

2

L

L

0

f(x) cos

mπx

L

dx, m = 0, 1, 2, · · · .

The Fourier series is similar to the Taylor series

f(x) =

¸

m=0

f

(n)

(0)

n!

x

n

, (2.30)

but instead of expanding the function in polynomials, Fourier series expands a function in sine or

cosine functions. But Taylor series converges to the original function only in a small neighborhood

of x = 0 (though some time it is large), and the neighborhood depends on the function f. Fourier

series converges to the original functions for all functions and on the whole interval (0, L). (But it

may not converge to f(x) at x = 0 and x = L.)

Theorem 2.2 shows all sound wave can be decomposed into several regular waves plus a small

noise. The approximation of Fourier series is sine or cosine Fourier polynomials:

f(x) ≈

N

¸

m=1

b

m

sin

mπx

a

, f(x) ≈

a

0

2

+

N

¸

m=1

a

m

cos

mπx

a

. (2.31)

24 CHAPTER 2. DIFFUSION EQUATIONS

Now we can complete the solution formula of (2.13)-(2.15):

u(t, x) =

∞

¸

m=1

c

m

exp

−D

m

2

π

2

L

2

t

sin

mπx

L

,

where c

m

=

2

L

L

0

u

0

(x) sin

mπx

L

dx, m ≥ 1, t > 0, x ∈ (0, L).

(2.32)

Example 2.3.

u

t

= u

xx

, t > 0, x ∈ (0, π),

u(t, 0) = u(t, π) = 0,

u(0, x) = sin x.

(2.33)

From the solution formula (2.32), we get

u(t, x) =

∞

¸

m=1

c

m

e

−m

2

t

sin(mx), and c

m

=

2

π

π

0

sin xsin(mx)dx. (2.34)

However

π

0

sin xsin(mx)dx = 0 for any m ≥ 2, and

π

0

sin

2

xdx = π/2. Therefore the solution is

simply

u(t, x) = e

−t

sin x. (2.35)

Example 2.4. For the homogeneous Neumann boundary value problem

u

t

= Du

xx

, t > 0, x ∈ (0, L),

u

x

(t, 0) = u

x

(t, L) = 0,

u(0, x) = u

0

(x),

(2.36)

the solution is

u(t, x) =

c

0

2

+

∞

¸

m=1

c

m

exp

−D

m

2

π

2

L

2

t

cos

mπx

L

,

where c

m

=

2

L

L

0

u

0

(x) cos

mπx

L

dx, m ≥ 0, t > 0, x ∈ (0, L).

(2.37)

The derivation of this formula is left as exercise.

For these examples, the asymptotic behavior of the solution can be read from the formulas. For

homogeneous Dirichlet problem (2.13)-(2.15), the asymptotic limit is

lim

t→∞

u(t, x) = 0, and the asymptotic proﬁle is u(t, x) ∼ c

1

exp

−

π

2

t

L

2

sin

πx

L

. (2.38)

On the other hand, for homogeneous Neumann problem (2.36), the asymptotic limit is

lim

t→∞

u(t, x) =

c

0

2

, and the asymptotic proﬁle is u(t, x) ∼

c

0

2

+c

1

exp

−

π

2

t

L

2

cos

πx

L

. (2.39)

2.4. RELAXATION TO EQUILIBRIUM SOLUTIONS 25

The constant c

0

above is (2/L)

L

0

u

0

(x)dx, thus the limit is

1

L

L

0

u

0

(x)dx, (2.40)

the average value of the initial value u

0

(x). The limits of both cases are also solutions of diﬀusion

equation, and in fact they are solutions which are independent of time t. Such solutions are

equilibrium solutions. For a general reaction-diﬀusion equation

∂u

∂t

= d∆u +f(x, u), t > 0, x ∈ Ω,

u(0, x) = u

0

(x), x ∈ Ω,

B(u) = 0, t > 0, x ∈ ∂Ω,

(2.41)

where B(u) is the boundary condition, the equilibrium solutions satisfy

d∆u +f(x, u) = 0, x ∈ Ω,

B(u) = 0, x ∈ ∂Ω.

(2.42)

A solution u

1

(x) of (2.42) is a solution of (2.41) in the sense that the function U

1

(t, x) = u

1

(x) is a

solution of (2.41) with u(0, x) = u

1

(x). It is not just a coincidence that for two examples we have

considered so far, the limit is always an equilibrium solution. We will show in next section that the

limit of a diﬀusion equation is usually an equilibrium solution.

Equilibrium solutions of diﬀusion equation are easier to ﬁnd, and that will help us understand

the limit of solution when the solution itself is not so easy to ﬁnd. The equilibrium equation for

(2.13)-(2.15) is

u

′′

= 0, u(0) = u(L) = 0. (2.43)

The general solution for u

′′

= 0 is u(x) = ax + b, and from the boundary conditions, we obtain

u(x) = 0 is the only equilibrium solution. The equilibrium equation for (2.36) is

u

′′

= 0, u

′

(0) = u

′

(L) = 0. (2.44)

For this problem u(x) = b for any b ∈ R is an equilibrium solution. However the limit in (2.36)

is the average value of the initial value u

0

(x) only. Why does the solution of diﬀusion equation

selects this particular equilibrium solution, but not the others? We will answer this question in

next section.

2.4 Relaxation to equilibrium solutions

We have obtained analytic solutions of 1-d diﬀusion equation in previous sections. The solutions

are in a form of inﬁnite series. However in general, even such formulas cannot be obtained when

the spatial dimension is higher and the domain is arbitrary. (Later we will formulas of solutions

of diﬀusion equations for some special cases.) Here we use some integral formulas to show some

qualitative properties of solutions of general diﬀusion equation.

26 CHAPTER 2. DIFFUSION EQUATIONS

As a calculus warmup, we consider the diﬀusion equation with no-ﬂux boundary condition

∂u

∂t

= D∆u, t > 0, x ∈ Ω,

u(0, x) = u

0

(x), x ∈ Ω,

∇u(t, x) · n(x) = 0, t > 0, x ∈ ∂Ω.

(2.45)

From the biological modelling point of view, the species only randomly moves in Ω, the growth rate

is zero, and the boundary is “sealed”. Thus the total population should not change since no one

moves in or out, and no one is born or dead. Indeed that is the basic character of the equation,

and it also explains why the limit of the solution of (2.36) is the average value of the initial value

function.

Proposition 2.5. Suppose that u(t, x) is the solution of (2.45), then

Ω

u(t, x)dx =

Ω

u

0

(x)dx. (2.46)

Proof. Integrating the left hand side of the equation in (2.45) over Ω, we obtain

Ω

∂u

∂t

(t, x)dx =

d

dt

Ω

u(t, x)dx

, (2.47)

and integrating the right hand side, we have

D

Ω

∆u(t, x)dx = D

∂Ω

∇u(t, x) · n(x)ds = 0, (2.48)

from the divergent theorem and the no-ﬂux boundary condition. Therefore from (2.47) and (2.48),

we ﬁnd that

d

dt

Ω

u(t, x)dx

= 0, (2.49)

which implies (2.46).

Next we show that the solution u(t, x) always tends to a limit constant c, and because of (2.46),

that constant must be

1

|Ω|

Ω

u

0

(x)dx. (2.50)

A rigorous proof of this fact needs many more advanced mathematical knowledge, so we only sketch

the ideas here. We multiply the equation in (2.45) by the function u(t, x),

u

∂u

∂t

= Du∆u. (2.51)

Similar to the proof of Proposition 2.5, we integrate the left hand side of (2.51):

Ω

u

∂u

∂t

dx =

Ω

∂

∂t

1

2

u

2

dx =

d

dt

Ω

1

2

u

2

dx =

1

2

d

dt

Ω

[u(t, x)]

2

dx, (2.52)

2.4. RELAXATION TO EQUILIBRIUM SOLUTIONS 27

and the right hand side:

D

Ω

u∆udx = D

∂Ω

u(∇u · n)ds −D

Ω

∇u · ∇udx. (2.53)

Here we use Green’s identity (a consequence of divergence theorem):

Ω

u∆vdx =

∂Ω

u(∇v · n)ds −

Ω

∇u · ∇vdx. (2.54)

From the boundary condition,

∂Ω

u(∇u · n)ds = 0, thus from (2.52) and (2.53), we ﬁnd

1

2

d

dt

Ω

[u(t, x)]

2

dx = −D

Ω

|∇u(t, x)|

2

dx. (2.55)

Let E(t) =

Ω

[u(t, x)]

2

dx. Then (2.55) shows that E

′

(t) < 0 for all t > 0 and obviously E(t) > 0

for all t > 0. Thus lim

t→∞

E(t) ≥ 0 exists, and as t →∞, E

′

(t) →0, so

lim

t→∞

Ω

|∇u(t, x)|

2

dx = 0. (2.56)

(2.56) implies lim

t→∞

|∇u(t, x)| = 0 for any x ∈ Ω, thus the limit of u(t, x) when t → ∞ must

be a function with zero gradient, thus the limit function must be a constant u(x) = c. Since

c|Ω| =

Ω

u(x)dx = lim

t→∞

Ω

u(t, x)dx =

Ω

u

0

(x)dx, then c =

Ω

u

0

(x)dx/|Ω|. (Notice that we

implicitly assume u(t, x) has a limit u(x) when t → ∞. While such a limit indeed exists, the

proof of its existence is beyond the scope of this notes, that is why this is only a sketch of proof.)

Nevertheless, we put our ﬁndings into another proposition:

Proposition 2.6. Suppose that u(t, x) is the solution of (2.45), then u(t, x) tends to a constant

function u(x) =

1

|Ω|

Ω

u

0

(x)dx.

Similar result can also be shown for Dirichlet problem, in which the limit function is u = 0.

From this property of diﬀusion equation, we can get a conclusion that the diﬀusion equation will

make the initial value ﬂatter and bring it eventually to an equilibrium state without spatial pattern.

This can also be indicated from the diﬀerential equation itself. Let the average of the initial value

be c. For the points where u(t, x) > c, the function is likely to be convex down, with ∆u < 0, thus

the function value will decrease over these points; and for the points where u(t, x) < c, the opposite

occurs. This is called smothering eﬀect of diﬀusion. In the following we use a Maple program to

demonstrate the smothering eﬀect and the asymptotic tendency to the equilibrium. We consider

u

t

= 4u

xx

, t > 0, x ∈ (0, 4),

u(t, 0) = u(t, 4) = 0,

u(0, x) = f(x),

(2.57)

28 CHAPTER 2. DIFFUSION EQUATIONS

where f(x) is given by the formula f(x) =

100, x ∈ [1, 3],

0, x ∈ [0, 1)

¸

(3, 4].

. From (2.32), we get the

formula of the solution of (2.57):

u(t, x) =

∞

¸

m=1

c

m

exp

−

m

2

π

2

4

t

sin

mπx

4

,

where c

m

=

1

2

4

0

f(x) sin

mπx

4

dx, m ≥ 1, t > 0, x ∈ (0, 4).

(2.58)

The coeﬃcients c

m

in the Fourier series can easily be calculated, but we will let Maple to do that.

Here is the Maple program diffusion.mws and some of the results:

> restart;

> with(plots): n:=’n’:

> f:=x->100*Heaviside(x-1)-100*Heaviside(x-3);

> plot(f(x), x=0..4);

> c:=n->0.5*int(f(x)*sin(n*Pi*x/4),x=0..4): c(n);

200.0

cos(1/4 n Pi) - cos(3/4 n Pi)

n Pi

>uapprox1:=(x,t)->sum(c(n)*sin(n*Pi*x/4)*exp(-4*n^2*t), n=1..30);

uapprox1 :=(x,t) - >

30

¸

n=1

c(n) sin(1/4 n Pi x) exp(-4 n

2

t)

> tvals:=seq(0.07/20*i,i=0..20):

> toplot:=[seq(uapprox1(x,t),t=tvals)]:

> plot(toplot,x=0..4);

> densityplot(uapprox1(x,t),x=0..4,t=0..0.2,colorstyle=HUE);

> animate(uapprox1(x,t), x=0..4, t=0..0.1,frames=50);

This is an animation generated by approximating the solution by the partial sum with n = 30 of

the series solution. The approximation is accurate since the remainder is an exponential decaying

function with smaller exponent. Note that the initial condition f(x) can be written as the diﬀerence

of two Heaviside functions, which is deﬁned as

H(x) =

1, x ≥ 0,

0, x < 0.

(2.59)

It is an interesting experiment to observe the changes of the spatial patterns of u(t, x) when t gets

forward. Here are a few snap shots of the animation:

2.4. RELAXATION TO EQUILIBRIUM SOLUTIONS 29

0

20

40

60

80

100

1 2 3 4

x

0

20

40

60

80

100

1 2 3 4

x

Figure 2.1: (a) f(x); (b) partial sum of Fourier series of f(x)

0

20

40

60

80

100

1 2 3 4

x

0

20

40

60

80

100

1 2 3 4

x

Figure 2.2: (a) u(0.002, x); (b) u(0.010, x)

Because of the truncation of the series solution, u(t, x) above indeed is the solution of (2.57)

with initial data

¯

f(x) =

30

¸

m=1

c

m

sin

mπx

4

dx, c

m

=

200

mπ

cos

mπ

4

−cos

mπ

4

. (2.60)

Figure 2.1 is a comparison of f and

¯

f. From the last section, we know that when t → ∞,

u(t, x) →0 exponentially fast. However the road to distinction is worth a more careful looking—it

can be dismantled into the following stages:

1. Forming a plateau From t = 0 (Figure 2.1(b)) to t = 0.002 (Figure 2.2 (a)), we can observe

that the initial data has been smoothed by the diﬀusion equation, and all the small wiggles

in u(0, x) are gone at t = 0.002. On the other hand, although u(0, x) is not a constant for

x ∈ (1, 3), u(0.002, x) is almost a constant on an interval slightly smaller than (1, 3). So in

the initial stage of the evolution of

¯

f under diﬀusion, a plateau (or ﬂat core) of height 100

forms, with a sharp but smooth drop-oﬀ to 0 near x = 1 and x = 3.

30 CHAPTER 2. DIFFUSION EQUATIONS

0

20

40

60

80

100

1 2 3 4

x

0

20

40

60

80

100

1 2 3 4

x

Figure 2.3: (a) u(0.020, x); (b) u(0.040, x)

0

20

40

60

80

100

1 2 3 4

x

0

20

40

60

80

100

1 2 3 4

x

Figure 2.4: (a) u(0.100, x); (b) u(0.600, x)

2. Erosion of the plateau From t = 0.002 (Figure 2.2 (a)) to t = 0.01 (Figure 2.2 (b)), the ﬂat

core keeps shrinking though the top of the core is still at the level near 100. At t = 0.01, the

erosion ﬁnally dissolves the whole ﬂat top, and the maximum point of u(t, ·) at x = 2 looks

more like a regular local maximum point. The graph of u(0.01, x) is close to a bell-shape.

In that time interval, the inﬂection points of u(t, ·) are still near x = 1 and x = 3, but at

t = 0.01, the interface between u = 0 and u = 100 is no longer as sharp as when t = 0.002.

3. Bell getting round In the next stage, smoothing eﬀect brings the concave part of the graph

down, and the convex part of the graph up. And at the same time, the convex part is shrinking

as population loss via boundary gets bigger. At t = 0.04, the graph becomes almost concave

for all x, and the bell-shape becomes an arch. (see Figure 2.3)

4. Collapse of the arch In the ﬁnal stage, the arch is close to the leading term of the series:

c

1

exp(−π

2

t/4) sin(πx/4), and an exponential collapse is obvious. After t = 1, the graph can

hardly be seen. (see Figure 2.4)

2.5. ONE DIMENSIONAL CHEMICAL MIXING PROBLEM 31

2.5 One dimensional chemical mixing problem

Here we consider a simpliﬁed version of the chemical mixing problem which is discussed in Chapter

1: A tube with length L contains salt water. The cross-section (yz-direction) of the tube is so small

so we can assume that the concentration of the salt water is same for any point on a cross-section.

Let the concentration of the salt water in the tube be c(t, x) kg/m

3

, t > 0 and 0 < x < L. We

assume that the diﬀusion of the salt is one-dimensional in the tube, and the velocity of ﬂuid is

ignorable, thus c(t, x) satisﬁes the diﬀusion equation:

∂c

∂t

= D

∂

2

c

∂x

2

. (2.61)

At the left hand side of the tube, a salt water solution with constant concentration C

0

kg/m

3

enters

the tube at a rate of R

0

m

3

/s, and on the right hand side of the tube, the mixed solution is removed

at the rate of R

0

m

3

/s. We also assume that the area of the cross section of the tube is A. Let A

1

and A

2

be the cross sections at x = 0 and x = L correspondingly. From the Fick’s law, the total

in-ﬂux at x = 0 is

A

1

J · (−n)ds =

A

1

[−Dc

x

(t, 0)] · (1)ds = −DAc

x

(t, 0) = C

0

R

0

, (2.62)

and the total out-ﬂux at x = L is

A

2

J · nds =

A

1

[−Dc

x

(t, L)] · (1)ds = −DAc

x

(t, 0) = C(t, L)R

0

, (2.63)

thus we obtain the boundary conditions

∂c(t, 0)

∂x

= −

C

0

R

0

DA

, and

∂c(t, L)

∂x

= −

c(t, L)R

0

DA

. (2.64)

To simplify the notations, we deﬁne

B =

C

0

R

0

DA

, and E =

R

0

DA

. (2.65)

Thus the initial-boundary value problem is

∂c

∂t

= D

∂

2

c

∂x

2

, t > 0, x ∈ (0, L),

∂c(t, 0)

∂x

= −B,

∂c(t, L)

∂x

+Ec(t, L) = 0,

c(t, x) = c

0

(x), t > 0, x ∈ (0, L).

(2.66)

Before attempting to solve the equation, we could ﬁnd the equilibrium solutions, which satisfy

equation:

D

∂

2

c

∂x

2

= 0, x ∈ (0, L),

∂c(0)

∂x

= −B,

∂c(L)

∂x

+Ec(L) = 0.

(2.67)

32 CHAPTER 2. DIFFUSION EQUATIONS

From a simple calculation, we ﬁnd the unique equilibrium solution:

c

∗

(x) =

B

E

+B(L −x) = C

0

+

C

0

R

0

DA

(L −x), (2.68)

which is a linear positive function with negative slope on [0, L], and c

∗

(L) = C

0

. To solve the

equation, we deﬁne b(t, x) = c(t, x) −c

∗

(x). Then b(t, x) satisﬁes

∂b

∂t

= D

∂

2

b

∂x

2

, t > 0, x ∈ (0, L),

∂b(t, 0)

∂x

= 0,

∂b(t, L)

∂x

+Ec(t, L) = 0,

b(t, x) = c

0

(x) −c

∗

(x), t > 0, x ∈ (0, L).

(2.69)

The reason we consider (2.69) instead of (2.66) is that the boundary conditions in (2.69) is homoge-

neous, thus we can use the technique of separation of variables. We assume that b(t, x) = U(t)V (x).

Then

U

′

(t)

U(t)

= D

V

′′

(x)

V (x)

= Dk, (2.70)

U(t) = e

Dkt

, and V (x) satisﬁes

V

′′

−kV = 0, x ∈ (0, L), V

′

(0) = 0, V

′

(L) +EV (L) = 0. (2.71)

This is a half-Neumann and half-Robin boundary condition. We use the method in Section 2.2 to

solve (2.71).

If k > 0, then V (x) = c

1

e

−

√

kx

+c

2

e

√

kx

. From the boundary conditions, we get

det

−

√

k

√

k

(E −

√

k)e

−

√

kL

(E +

√

k)e

√

kL

= 0, (2.72)

and k must satisfy

e

2

√

kL

=

√

k −E

√

k +E

. (2.73)

But the function on the left is greater than 1 when k > 0, and the one on the right is always less

than 1 when k > 0. So (2.73) has no solution, thus there are no positive eigenvalues. Similarly

k = 0 is not eigenvalue either.

If k < 0, V (x) = c

1

cos(

√

−kx) + c

2

sin(

√

−kx). From the boundary conditions, we get c

2

= 0

and c

1

[cos(

√

−kL) −

√

−k sin(

√

−kL)] = 0. Thus k satisﬁes

cot(

√

−kL) =

√

−k. (2.74)

From the graph of the functions f

1

(x) = cot(Lx) and f

2

(x) = x, there are exactly one intersection

point x

m

of f

1

and f

2

in ((m−1)π/L, mπ/L) for any positive integer m > 0. Then this k

m

= −x

2

m

is the m-th eigenvalue of (2.71), and the corresponding eigenfunction is cos(

√

−k

m

x). Thus the

solution of (2.69) is

b(t, x) =

∞

¸

m=1

b

m

e

−Dx

2

m

t

cos(x

m

x), where x

m

∈ (

(m−1)π

L

,

mπ

L

) satisﬁes cot(x

m

L) = x

m

. (2.75)

2.6. CRITICAL PATCH SIZE AND BIFURCATIONS 33

Hence the solution is the sum of the equilibrium solution and an exponential decaying Fourier

series, and the solution converges to the equilibrium solution exponentially fast. It shows that in

equilibrium state, the exiting solution also has the concentration C

0

, same as the incoming solution,

thus the amounts of salt entering and exiting the tube are balanced, but the concentration of solution

in the tube is higher than C

0

, and the concentration at x = 0 is the highest.

The example in this section can be thought as a model of a continuously polluted river. The

left end point x = 0 is the source of the pollution where the pollutant enters the river at a constant

rate, and the right end point x = L is where the river merges to a bigger river or the ocean. Our

mathematical result shows that the concentration of pollutant in a continuously polluted river is

even higher than that in the source. As a consequence, we can see that at the conﬂuence of two

polluted rivers, the pollution can be most serious, since the concentration there is much higher than

the ones of either branches. We should be cautious that we ignore the eﬀect of drifting here, which

could make a big diﬀerence for rivers.

2.6 Critical patch size and bifurcations

The simplest population growth model is the Malthus model dP/dt = aP for a positive constant

a > 0. In this section we consider Dirichlet boundary value problem of diﬀusive Malthus equation

∂P

∂t

= D

∂

2

P

∂x

2

+aP, t > 0, x ∈ (0, L),

u(t, 0) = u(t, L) = 0,

u(0, x) = u

0

(x), x ∈ (0, L).

(2.76)

We have seen that population are at a loss from the hostile boundary, but on the other hand, the

population will increase through linear reproduction. So the loss and gain are in a competition.

Let’s see what factor will determine the outcome of this battle.

Equation (2.76) can be solved in a similar way as (2.13)-(2.15). In fact, the form of the solution

only changes slightly: adding a term e

at

to reﬂect the exponential growth. The solution of (2.76)

is

u(t, x) =

∞

¸

m=1

c

m

exp

at −D

m

2

π

2

L

2

t

sin

mπx

L

,

where c

m

=

2

L

L

0

u

0

(x) sin

mπx

L

dx, m ≥ 1, t > 0, x ∈ (0, L).

(2.77)

However the fate of the solution may be diﬀerent: since at is added to the exponent, now the

exponent may be a positive one, then u(t, x) will have a part which is exponentially increasing. In

fact, when the index m in the sum increases, the exponent gets more negative, thus whether u(t, x)

is exponentially increasing is determined by the ﬁrst exponent coeﬃcient a −(Dπ

2

/L

2

). We have

when a >

Dπ

2

L

2

or L >

D

a

π, lim

t→∞

u(t, x) = ∞,

when a <

Dπ

2

L

2

or L <

D

a

π, lim

t→∞

u(t, x) = 0.

(2.78)

34 CHAPTER 2. DIFFUSION EQUATIONS

If we assume that the growth rate a and D are intrinsic parameters determined by the nature of

species and environment, then the size of the habitat L will play the deciding role here. The number

L

0

=

**D/aπ is called the critical patch size, as the population cannot survive if the habitat is too
**

small (the outgoing ﬂux wins over the growth), but the population will not only survive but thrive

to an exponential growth if the habitat is large enough (growth outpaces the emigration).

We say that a bifurcation occurs at L = L

0

as the equilibrium solution u = 0 changes stability

type from stable for L < L

0

to unstable for L > L

0

. We notice that at L = L

0

, the equation has

other equilibrium solutions u(x) = sin(πx/L), the eigenfunction associated with ﬁrst eigenvalue

k

1

= −π

2

/L

2

. We will have more discussion on bifurcation problems when the growth rate is

nonlinear. From (2.78), we can also use a or D as bifurcation parameter instead of L, sometime

that is more convenient since we do not need to change the domain, but we get equivalent results.

The concept of critical patch size is related to habitat fragmentation in ecological studies.

Habitat fragmentation

1

is the breaking up of a continuous habitat, ecosystem, or land-use type

into smaller fragments, which is considered to be one of several spatial processes in land transfor-

mation. It is commonly used in relation to the fragmentation of forests. Habitat fragmentation is

mainly caused by human activities such as logging, conversion of forests into agricultural areas and

suburbanization, but can also be caused by natural processes such as ﬁre. From our mathematical

result in this section, if the the fragmentation of habitat limits the spatial movement of the plants

and animals, then the species can become extinct.

2.7 Separation of variables: rectangles

Most of this chapter, we have exclusively considered one dimensional spatial domain, for the sim-

plicity of the mathematical analysis. However any interesting application happens in 2-D or 3-D

spaces. In this section, we discuss the separation of variables in a two-dimensional rectangle.

We ﬁrst consider a linear diﬀusion reaction equation:

∂u

∂t

= D

∂

2

u

∂x

2

+

∂

2

u

∂y

2

+λu, t > 0, (x, y) ∈ R = (0, a) ×(0, b),

∇u · n = 0, (x, y) ∈ ∂R,

u(0, x, y) = u

0

(x, y), (x, y) ∈ R,

(2.79)

where D, λ, a, b > 0, and ∂R is the boundary of rectangle R. From the method of separation of

variables, we assume that

u(t, x, y) = U(t)V (x, y). (2.80)

Then similar to one-dimensional case, we obtain

U

′

(t) = (Dk +λ)U(t), (2.81)

1

From TEMS, Terrestrial Ecosystem Monitoring Sites, http://www.fao.org/gtos/tems/

2.7. SEPARATION OF VARIABLES: RECTANGLES 35

and

∂

2

V

∂x

2

+

∂

2

V

∂y

2

= kV, t > 0, (x, y) ∈ R = (0, a) ×(0, b),

∇V · n = 0, (x, y) ∈ ∂R.

(2.82)

For (2.82), we need a further separation of variables:

V (x, y) = W(x)Z(y), (2.83)

Then we obtain that

W

′′

(x)

W(x)

+

Z

′′

(y)

Z(y)

= k, (2.84)

so both W

′′

(x)/W(x) and Z

′′

(y)/Z(y) must be constants. On the other hand, the Neumann

boundary condition implies that W

′

(0) = W

′

(a) = 0 and Z

′

(0) = Z

′

(b) = 0. Therefore W and Z

satisfy

W

′′

(x) = k

1

W(x), x ∈ (0, a), W

′

(0) = W

′

(a) = 0, (2.85)

Z

′′

(y) = k

2

Z(y), y ∈ (0, b), Z

′

(0) = Z

′

(b) = 0, (2.86)

and

k = k

1

+k

2

. (2.87)

Both (2.85) and (2.86) are familiar eigenvalue problems in one dimensional space which have been

studied in Chapter 2. The eigenvalues and eigenfunctions of (2.85) are

k

1n

= −

n

2

π

2

a

2

, W

n

(x) = cos

nπx

a

, n = 0, 1, 2, · · · , (2.88)

and the eigenvalues and eigenfunctions of (2.86) are

k

2m

= −

m

2

π

2

b

2

, Z

m

(y) = cos

mπy

b

, m = 0, 1, 2, · · · . (2.89)

Therefore the eigenvalues and eigenfunctions of (2.82) are

k

n,m

= −

n

2

π

2

a

2

−

m

2

π

2

b

2

, V

n,m

(x, y) = cos

nπx

a

· cos

mπy

b

, (2.90)

where n, m = 0, 1, 2, · · · . Thus we ﬁnd that the solution of (2.79) is

u(t, x, y) =

∞

¸

n=0,m=0

c

n,m

e

(Dk

n,m

+λ)t

cos

nπx

a

· cos

mπy

b

, (2.91)

where c

n,m

can be determined by the initial conditions.

The spatial patterns of the eigenfunctions for rectangle are much more complex than those

of one-dimensional. Here we consider a square R = (0, π) × (0, π), then the ﬁrst ﬁve distinctive

eigenvalues are (see the contour graph of eigenfunctions below)

k

0,0

= 0, k

0,1

= k

1,0

= −1, k

1,1

= −2, k

2,0

= k

0,2

= −4, k

2,1

= k

1,2

= −5, (2.92)

36 CHAPTER 2. DIFFUSION EQUATIONS

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

Figure 2.5: (From left to right) (a) V

0,0

= 1; (b) V

0,1

= cos y; (c) V

1,0

= cos x; (d) V

1,1

= cos x· cos y.

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

Figure 2.6: (From left to right) (a) V

0,2

= cos(2y); (b) V

2,0

= cos(2x); (c) V

1,2

= cos x · cos(2y); (d)

V

2,1

= cos(2x) · cos y.

We notice that many eigenvalues k

m,n

have multiplicity more than 1, for example k

m,n

=

k

n,m

= −(m

2

+n

2

). In such case, the eigenspace is of higher dimensional, so more possible spatial

patterns are generated. For example, k

1,0

= k

0,1

= −1, and the eigenspace is span{cos x, cos y},

thus k

1

cos x +k

2

cos y is a spatial pattern for any k

1

and k

2

. In particular, the patterns generated

by cos x ±cos y are symmetric with respect to one of the diagonal lines.

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

0

0.5 1

1.5 2

2.5 3

y

0

0.5

1

1.5

2

2.5

3

x

Figure 2.7: (From left to right) (a) cos x + cos y; (b) cos y −cos x.

The solution (2.91) is always an exponential growth if λ > 0. Indeed, from the equation, the

no-ﬂux boundary condition prevents the emigration through the boundary, and λ > 0 implies a

positive growth rate, thus the total population in R has an exponential growth. We can consider

corresponding Dirichlet boundary problem:

∂u

∂t

= D

∂

2

u

∂x

2

+

∂

2

u

∂y

2

+λu, t > 0, (x, y) ∈ R = (0, a) ×(0, b),

u(t, x, y) = 0, (x, y) ∈ ∂R,

u(0, x, y) = u

0

(x, y), (x, y) ∈ R.

(2.93)

2.8. STEADY STATES IN RADIALLY SYMMETRIC DOMAINS AND TRANSIT TIMES 37

Similar to the method above, the solution is

u(t, x, y) =

∞

¸

n=1,m=1

b

n,m

e

(Dk

n,m

+λ)t

sin

nπx

a

· sin

mπy

b

, (2.94)

where b

n,m

can be determined by the initial conditions, and k

n,m

is deﬁned in (2.90). Similar to

last section, the persistence/extinction of the population can be determined:

when

λ

Dπ

2

>

1

a

2

+

1

b

2

, lim

t→∞

u(t, x) = ∞,

when

λ

Dπ

2

<

1

a

2

+

1

b

2

, lim

t→∞

u(t, x) = 0.

(2.95)

2.8 Steady states in radially symmetric domains and transit times

Chapter 2 Exercises

1. Find the eigenvalues and eigenfunctions of homogeneous Neumann boundary problem

y

′′

= ky, x ∈ (0, L), y

′

(0) = y

′

(L) = 0. (2.96)

2. Find the eigenvalues and eigenfunctions of periodic boundary problem

y

′′

= ky, x ∈ (0, π), y(0) = y(π), y

′

(0) = y

′

(π). (2.97)

3. Find the eigenvalues and eigenfunctions of Robin boundary problem (b > 0)

y

′′

= ky, x ∈ (0, L), y

′

(0) = by(0), y

′

(L) = −by(L). (2.98)

4. Find the eigenvalues and eigenfunctions of the problem

y

′′

+ 2y

′

+ky = 0, x ∈ (0, π), y(0) = y(π) = 0. (2.99)

5. In Example 2.1, use Maple to ﬁnd the numerical values of the ﬁrst three eigenvalues, and plot

the graphs of the corresponding eigenfunctions.

6. Show that the solution of

u

t

= Du

xx

, t > 0, x ∈ (0, L),

u

x

(t, 0) = u

x

(t, L) = 0,

u(0, x) = u

0

(x), x ∈ (0, L),

(2.100)

is

u(t, x) =

c

0

2

+

∞

¸

m=1

c

m

exp

−D

m

2

π

2

L

2

t

cos

mπx

L

,

where c

m

=

2

L

L

0

u

0

(x) cos

mπx

L

dx, m ≥ 0, t > 0, x ∈ (0, L).

(2.101)

38 CHAPTER 2. DIFFUSION EQUATIONS

7. (a) Find the solution of

u

t

= 4u

xx

, t > 0, x ∈ (0, 1),

u

x

(t, 0) = u

x

(t, 1) = 0,

u(0, x) = f(x),

(2.102)

where f(x) is given by the formula f(x) =

1, x ∈ [0, 0.5],

−1, x ∈ (0.5, 1].

(b) Modify the Maple program in Section 2.4 to simulate the solution of (2.102).

8. Find the solutions of equilibrium equation

u

′′

= 0, x ∈ (0, 1), u(0) + 3u

′

(0) = 5, u(1) −5u

′

(1) = 7. (2.103)

9. Find the solutions of equilibrium equation

u

′′

+ 3u

′

= 0, x ∈ (0, 1), u(0) = 5, u

′

(1) = 4. (2.104)

10. Suppose that u(t, x) is the solution of

∂u

∂t

= D∆u, t > 0, x ∈ Ω,

u(0, x) = u

0

(x), x ∈ Ω,

∇u(t, x) · n(x) +a · u(t, x) = 0, t > 0, x ∈ ∂Ω.

(2.105)

where a > 0. Show that the mean-squared value I(t) =

Ω

[u(t, x)]

2

dx is strictly decreasing

unless u(t, x) ≡ 0.

11. Suppose that u(t, x) is the solution of the convective-diﬀusion equation: (D > 0 and V > 0)

∂u

∂t

= D

∂

2

u

∂x

2

−V

∂u

∂x

, x ∈ (0, π),

u(0, x) = u

0

(x), x ∈ (0, π),

∂u

∂x

(t, 0) =

∂u

∂x

(t, π) = 0.

(2.106)

Show that the total population

1

0

u(t, x)dx is a constant if u

0

(x) = sin x but not a constant

if u

0

(x) = cos(x).

12. Suppose that in the chemical mixing problem in Section 2.5, the salt solution drifts to the

right with velocity V . Then c(t, x) satisﬁes the convective-diﬀusion equation:

∂c

∂t

= D

∂

2

c

∂x

2

−V

∂c

∂x

, (2.107)

with D, V > 0. Suppose that c(t, x) satisﬁes the same initial and boundary conditions in

(2.66).

2.8. STEADY STATES IN RADIALLY SYMMETRIC DOMAINS AND TRANSIT TIMES 39

(a) Find the equilibrium solution of (2.107).

(b) Show that the equilibrium solution is positive, decreasing and concave down, and explain

the eﬀect of the drifting to the geometric properties of the equilibrium solution.

(c) Compare the exiting ﬂux −Dc

x

(t, L) of the equilibrium solution with that of (2.67).

13. Suppose that in the chemical mixing problem in Section 2.5, the salt is consumed in a chemical

reaction at a rate k. Then c(t, x) satisﬁes a linear diﬀusion equation:

∂c

∂t

= D

∂

2

c

∂x

2

−kc, (2.108)

with D, k > 0. Suppose that c(t, x) satisﬁes the same initial and boundary conditions in

(2.66).

(a) Find the equilibrium solution of (2.108). (You can use Maple to ﬁnd the general formula.)

(b) Show that the equilibrium solution is positive, decreasing and concave up, and explain

the eﬀect of the reaction to the geometric properties of the equilibrium solution.

(c) Compare the exiting ﬂux −Dc

x

(t, L) of the equilibrium solution with that of (2.67).

(d) Plot the graphs of the equilibrium solutions of (2.67), (2.107) and (2.108) with same

parameters in Maple.

14. Determine the critical patch for a diﬀusive Malthus equation with Robin boundary conditions:

∂P

∂t

= D

∂

2

P

∂x

2

+aP, t > 0, x ∈ (0, L),

u

x

(t, 0) = bu(t, 0), u

x

(t, L) = −bu(t, L),

u(0, x) = u

0

(x), x ∈ (0, L).

(2.109)

15. Consider the doubly periodic boundary value problem:

∂u

∂t

= D

∂

2

u

∂x

2

+

∂

2

u

∂y

2

+λu, t > 0, (x, y) ∈ R = (0, a) ×(0, b),

u(0, y) = u(a, y), u

x

(0, y) = u

x

(a, y),

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

y

(x, 0) = u

y

(x, b),

u(0, x, y) = u

0

(x, y), (x, y) ∈ R.

(2.110)

Find the eigenvalues and eigenfunctions of the problem, and the series representation of the

solution.

16. Consider a convective-diﬀusion equation

∂u

∂t

=

∂

2

u

∂x

2

+ 4

∂u

∂x

+ 8u, t > 0, x ∈ (0, L),

u(t, 0) = 0, u(t, L) = 0,

u(0, x) = u

0

(x), x ∈ (0, L).

(2.111)

Find the solution of the equation with the following steps:

40 CHAPTER 2. DIFFUSION EQUATIONS

(a) Use separation of variables method to show that if u(t, x) = U(t)V (x) is a solution, then

for some constant k, U and V satisfy

U

′

(t) = kU(t), V

′′

(x) + 4V

′

(x) + 8V (x) = kV (x), V (0) = V (L) = 0.

(b) Find the eigenvalues and eigenfunctions of

V

′′

(x) + 4V

′

(x) + 8V (x) = kV (x), V (0) = V (L) = 0.

(Hint: treat the cases of k > 4, k = 4 and k < 4 separately.)

(c) Find the solution of the equation in a series form.

(Hint: c

n

=

2

L

L

0

e

2x

u

0

(x) sin

nπx

L

dx.)

(d) Determine the critical patch L

0

of the problem.

(e) Describe the population distribution qualitatively when L > L

0

.

(f) Suppose that the species lives in a river with length L, and the convection is due to the

drifting of the river. Explain your mathematical results in this context.

20

CHAPTER 2. DIFFUSION EQUATIONS

These solutions are little more “reasonable” as they are bounded as x → ∞, but still they do not satisfy natural boundary conditions P, Px → 0 as x → ∞. Nevertheless solutions with above forms are solutions of diﬀusion equations, and we notice that they are in form of P (t, x) = U (t)V (x). This motivates us to consider the solutions which is separable on the two variables t and x: u(t, x) = U (t)V (x). With the form of u in (2.8), the diﬀusion equation (2.2) becomes U ′ (t)V (x) = DU (t)V ′′ (x), and V ′′ (x) U ′ (t) =D = Dk, U (t) V (x) (2.8)

where k is a constant independent of both t and x. Thus U satisﬁes an equation: U ′ (t) = DkU (t), t > 0, and U (t) = C1 eDkt ; and V satisﬁes V ′′ (x) = kV (x). The solutions of (2.11) are C2 e

√ kx √

(2.9) (2.10) (2.11)

**with any complex number k. Thus we obtain solution of (2.2): CeDkt e
**

kx

,

(2.12)

just like the three families of solutions we obtain by making lucky guess. To obtain more speciﬁc solutions, we must take the boundary and initial conditions into considerations. We start with the one-dimensional diﬀusion equation on an interval with length L: ∂P ∂2P =D 2, ∂t ∂x x ∈ (0, L). (2.13)

Here the individuals of the the species inhabits a one dimensional patch (0, L), and the diﬀusion coeﬃcient of the population is D > 0. To get a deﬁnite solution of the problem, we need to have an initial population distribution: P (0, x) = P0 (x), x ∈ (0, L), (2.14)

and a boundary condition. We assume that the exterior is hostile so homogeneous Dirichlet boundary condition is satisﬁed: (2.15) P (t, 0) = P (t, L) = 0. We look for solution with form P (t, x) = U (t)V (x). Then we obtain (2.9) and (2.11), and (2.15) implies that V (0) = V (L) = 0. (2.16) (2.11) and (2.16) become a boundary value problem for an ordinary diﬀerential equation, which is a special case of Sturm-Liouville problems. We shall discuss this kind of problems in a separate section next.

√ √ so c1 = 0 and if c2 is not zero. EIGENVALUES AND EIGENFUNCTIONS 21 2. For (2. we call these numbers eigenvalues.17). if k = 0. L).17) We can see that y(x) = 0 is always a solution of the equation.17). Indeed if there is a solution. (In fact. if y(t) is a solution of (2.) So for k which (2.17) has only zero solution when k = 0.2.17) has only zero solution when k > 0. 0 = y(L) = c1 e− √ kL + c2 e √ kL . y(0) = y(L) = 0.21) mπt . But this system of equations is unsolvable: if one gets c1 = −c2 from the ﬁrst equation and plugs √ √ it into the second one. and one must have c2 = 0 and c1 = 0. One important characteristic of a boundary value problem is that it may not have a solution. so is c · y(t) for any constant c. It is also easy to check that (2. sin( −kL) = 0. one has (e− kL − e kL )c2 = 0. Thus (2. we have √ √ 0 = y(0) = c1 .17) may still have only the zero solution.2 Eigenvalues and eigenfunctions of boundary value problems We consider the Dirichlet boundary value problem: y ′′ = ky.17) has a L m2 π 2 are c sin L2 m2 π 2 . if k < 0.17) eigenfunctions. and the solutions of (2. So we shall look for solutions other than the trivial zero solution. x ∈ (0. by the boundary conditions. and the set of eigenfunctions are the sine functions which L2 . So the only meaningful case is when k < 0. historically the solutions of (2. In this case. and we can solve the constants by 0 = y(0) = c1 + c2 . so is y1 (t) + y2 (t).17). From elementary ODE knowledge.20) c1 + c2 t.2. but for some number k < 0.19) (2.17): the eigenfunction must be y(t) = c1 cos( −kt) + c2 sin( −kt) for some constants c1 and c2 . it can have a solution y(t) which is not the zero solution.17) were ﬁrst called eigenvalues and eigenfunctions. we know that the general solution of y ′′ = ky is one of the following three forms: c1 e− √ kt + c2 e √ kt .18) (2. (Note that from the linear principle. and if y1 (t) and y2 (t) are solutions of (2. Therefore (2. √ √ c1 cos( −kt) + c2 sin( −kt). (2. if k > 0. (2. it must √ √ be y(t) = c1 e− kt + c2 e kt for some constants c1 and c2 . while the initial value problem always has a solution and the solution is unique if all terms in the equation are diﬀerentiable. 0 = y(L) = c1 cos( −kL) + c2 sin( −kL). the eigenvalue k must be k=− and the eigenfunctions associated with k = − sequence of eigenvalues: km = − m2 π 2 . L2 (2. there is no solution when k > 0. (2.) Now let’s √ search for the √ eigenvalues and eigenfunctions of (2.17) has non-zero solutions. thus −kL = mπ.

x) = exp −D m2 π 2 mπx t sin . x) = Example 2. 0). (2.25).13) is u(t.22) Back to (2. conditions. thus e2pπ = 1−p k < 0. 2. 2 L L (2. 1.1. we ﬁnd a sequence of solutions: um (t. L).5 < p1 < 1. The matrix is cos( kπ) sin( kπ) √ √ √ √ determinant of the matrix is 0 if sin( kπ) − k cos( kπ) = 0.26) √ √ 1 −√ −k 1 + −k √ The equation has non-zero solution only if det = 0. c2 ) is (0. the general solution of (2. So we still consider √ √ boundary the cases of k < 0. The to (1 − p)epπ − (1 + p)e−pπ if we denote p = −k.25) ∞ m=1 cm exp −D m2 π 2 mπx t sin .23) Because of the linear principle. The boundary conditions: y(0) + y ′ (0) = c1 + √ √ √ √ k 1 √ √ . So 0 is not an eigenvalue either. 2 L L (2.5. and the c2 k = 0.5 < pm < m + 0. from the boundary conditions. t > 0. √ m = p2 are the eigenvalues of the k m √ √ problem (2. m = 1. The graphs of f1 (p) = tan(πp) and f2 (p) = p intersect at a sequence of points. √ √ For k > 0. since the determinant of the matrix is not zero. and there is no eigenvalue such that intersect at p = 0. thus det = 0 is equivalent to e2pπ = 1−p 1+p have no common points when p > 0 (in fact. . the expression of eigen-pairs is much simpler: km = −m2 . k = 0 and k > 0. y(t) = c1 cos( kt) + c2 sin( kt). y ′′ + ky = 0.5 < p2 < 2. c2 ) satisﬁes the equations √ √ 1 −√ −k 1 + −k c √ · 1 −kπ − −kπ c2 e e = 0 0 (2. When k = 0. or tan(πp) = p for p = k. m is a positive integer. The determinant equals e− −kπ e −kπ √ 1+p .√from the √ √ √ ′ (0) = c +c − −kc + −kc = 0 and y(π) = c e− −kπ +c e −kπ = 0. 3.15). · · · . (2. we get y(0) + y ′ (0) = c1 + c2 = 0 1 1 and y(π) = c1 +c2 π = 0. pm . When the interval is (0.22 CHAPTER 2. and the eigenfunctions are ym (t) = km cos( km t) − sin( km t). π) (so L = π). y(t) = c1 e− −kt + c2 e −kt .5. fm (t) = sin(mt). x ∈ (0. then only solution 1 π for (c1 . y(t) = c1 + c2 t. DIFFUSION EQUATIONS can “ﬁt in” the interval (0. and 0. m − 0. y(0) + y ′ (0) = y(π) = 0. L).13) and (2. When k < 0. they only functions f1 (p) = e2pπ and f2 (p) = 1−p 1+p is not possible for p > 0.5. we get y(0)+y 1 2 1 2 1 2 So (c1 . y(π) = c1 cos( kπ) + c2 sin( kπ) = 0.24) This problem is same as previous example except the boundary conditions.

m = 0. L (2. All sine and cosine functions are considered to be regular smooth waves. (But it may not converge to f (x) at x = 0 and x = L. L (2.13)-(2. 2.15). or a pure cosine series L f (x) = a0 + 2 ∞ m=1 am cos mπx . we need to determine the constants cm in (2.3 Fourier series To completely solve (2. The approximation of Fourier series is sine or cosine Fourier polynomials: N f (x) ≈ bm sin m=1 a0 mπx + . m = 1.30) but instead of expanding the function in polynomials. L). · · · . · · · .2 shows all sound wave can be decomposed into several regular waves plus a small noise. 1.24) is the Fourier series of u0 (x) on (0. x ∈ (0. and that is a big help for the signal processing.28) where bm = f (x) sin 0 mπx dx. and Fourier series convert any noise (or signal) into a sum of more regular waves. L]. L). a (2. Then f (x) can be expanded in either a pure sine series f (x) = 2 L L ∞ m=1 bm sin mπx .29) where am = 2 L L f (x) cos 0 mπx dx. Here we collect a few facts of Fourier series. Theorem 2. Fourier series is widely used in mathematical physics and signal process. L).24) by setting t = 0 in (2. Let f (x) and f ′ (x) be piecewise continuous functions on the interval [0. But Taylor series converges to the original function only in a small neighborhood of x = 0 (though some time it is large). n! (2.24): u0 (x) = ∞ m=1 cm sin mπx . 2. and the neighborhood depends on the function f .27) The expression in the right-hand side of (2.) Theorem 2.2. f (x) ≈ a 2 N am cos m=1 mπx . FOURIER SERIES 23 2. L The Fourier series is similar to the Taylor series f (x) = m=0 f (n) (0) n x .3. Fourier series expands a function in sine or cosine functions. L (2.2. Fourier series converges to the original functions for all functions and on the whole interval (0.31) .

24 CHAPTER 2. DIFFUSION EQUATIONS Now we can complete the solution formula of (2. x) = sin x. m ≥ 1. L).13)-(2. L (2. L). For the homogeneous Neumann boundary value problem ut = Duxx . x ∈ (0. For homogeneous Dirichlet problem (2. For these examples. 2 L L (2. we get u(t.39) t→∞ . and the asymptotic proﬁle is u(t.36). m ≥ 0.37) mπx dx. x ∈ (0.15). (2. u(t. x) ∼ + c1 exp − 2 2 2 L cos πx . π). 2 L L (2. From the solution formula (2. L) = 0. x) ∼ c1 exp − π2t L2 sin πx .15): u(t. 0 (2. x) = c0 π2t c0 . x) = u0 (x). and cm = π 2 π π sin x sin(mx)dx. t > 0.32). x) = e−t sin x. the asymptotic limit is lim u(t. t > 0. Example 2. and sin2 xdx = π/2. x) = π ∞ m=1 2 ut = uxx . π) = 0. u0 (x) sin L t > 0. t > 0.33) cm e−m t sin(mx). 0) = u(t. u0 (x) cos L The derivation of this formula is left as exercise.32) 2 = L mπx dx. x) = where cm c0 + 2 2 = L ∞ m=1 L 0 (2. 0) = ux (t. L). the asymptotic behavior of the solution can be read from the formulas.13)-(2. x ∈ (0.34) However simply 0 sin x sin(mx)dx = 0 for any m ≥ 2. x ∈ (0. x) = where cm ∞ m=1 cm exp −D L 0 mπx m2 π 2 t sin .3.35) 0 u(t. x u(0.4. the solution is u(t. u(0. x) = 0. the asymptotic limit is lim u(t. L (2.36) cm exp −D mπx m2 π 2 t cos . Example 2. Therefore the solution is (2. u (t.38) t→∞ On the other hand. for homogeneous Neumann problem (2. and the asymptotic proﬁle is u(t.

40) where B(u) is the boundary condition. even such formulas cannot be obtained when the spatial dimension is higher and the domain is arbitrary.41) u(0. 2. u(0) = u(L) = 0.42) is a solution of (2.) Here we use some integral formulas to show some qualitative properties of solutions of general diﬀusion equation. However in general. the limit is always an equilibrium solution. . The solutions are in a form of inﬁnite series. Equilibrium solutions of diﬀusion equation are easier to ﬁnd. RELAXATION TO EQUILIBRIUM SOLUTIONS L 25 The constant c0 above is (2/L) 0 u0 (x)dx. x ∈ Ω. but not the others? We will answer this question in next section. the average value of the initial value u0 (x). t > 0. and that will help us understand the limit of solution when the solution itself is not so easy to ﬁnd. Such solutions are equilibrium solutions. x ∈ Ω.4 Relaxation to equilibrium solutions We have obtained analytic solutions of 1-d diﬀusion equation in previous sections. x ∈ ∂Ω. ∂t (2. and from the boundary conditions.2. The equilibrium equation for (2. u′ (0) = u′ (L) = 0. (2. B(u) = 0. x) = u0 (x). x ∈ ∂Ω.4. (2.44) For this problem u(x) = b for any b ∈ R is an equilibrium solution.41) in the sense that the function U1 (t. (Later we will formulas of solutions of diﬀusion equations for some special cases.15) is u′′ = 0.41) with u(0.36) is the average value of the initial value u0 (x) only. thus the limit is 1 L L u0 (x)dx.36) is u′′ = 0.43) The general solution for u′′ = 0 is u(x) = ax + b. and in fact they are solutions which are independent of time t. However the limit in (2. x) = u1 (x) is a solution of (2.13)-(2. we obtain u(x) = 0 is the only equilibrium solution. x ∈ Ω. We will show in next section that the limit of a diﬀusion equation is usually an equilibrium solution. It is not just a coincidence that for two examples we have considered so far. Why does the solution of diﬀusion equation selects this particular equilibrium solution. x) = u1 (x).42) A solution u1 (x) of (2. (2. B(u) = 0. 0 (2. u). For a general reaction-diﬀusion equation ∂u = d∆u + f (x. the equilibrium solutions satisfy d∆u + f (x. The equilibrium equation for (2. u) = 0. t > 0. The limits of both cases are also solutions of diﬀusion equation.

u ∂u = Du∆u. we have D Ω ∆u(t. (2.46).46).36) is the average value of the initial value function. DIFFUSION EQUATIONS As a calculus warmup. Therefore from (2. and the boundary is “sealed”. x ∈ Ω. We multiply the equation in (2. we integrate the left hand side of (2. ∂t (2. x)dx = D ∂Ω ∇u(t. Ω (2. x)dx = Ω Ω u0 (x)dx.51): u Ω ∂u dx = ∂t Ω ∂ ∂t 1 2 d u dx = 2 dt Ω 1 2 1d u dx = 2 2 dt [u(t. Ω (2.5.45). so we only sketch the ideas here. the growth rate is zero.47) Ω and integrating the right hand side.45) by the function u(t.26 CHAPTER 2. that constant must be 1 u0 (x)dx.5. Thus the total population should not change since no one moves in or out. and no one is born or dead. x ∈ ∂Ω.45) u(0. x)dx .47) and (2. (2. x) = u0 (x). we obtain ∂u d (t. x) is the solution of (2. x)dx = 0.50) |Ω| Ω A rigorous proof of this fact needs many more advanced mathematical knowledge. x)]2 dx. (2. Suppose that u(t. x) always tends to a limit constant c. Integrating the left hand side of the equation in (2.48) from the divergent theorem and the no-ﬂux boundary condition. (2. ∇u(t.49) dt Ω which implies (2. then u(t. t > 0. we ﬁnd that d u(t. x)dx = ∂t dt u(t. x ∈ Ω. and because of (2. x) · n(x) = 0.52) .45) over Ω.48). the species only randomly moves in Ω. and it also explains why the limit of the solution of (2. x) · n(x)ds = 0. t > 0. Next we show that the solution u(t. we consider the diﬀusion equation with no-ﬂux boundary condition ∂u = D∆u. ∂t (2. From the biological modelling point of view. Indeed that is the basic character of the equation.51) Similar to the proof of Proposition 2.46) Proof. Proposition 2. x).

x) > c. E ′ (t) → 0. and as t → ∞. This can also be indicated from the diﬀerential equation itself. x) is the solution of (2. so t→∞ Ω lim |∇u(t. 4). 1 d 2 dt Let E(t) = Ω · n)ds = 0. we put our ﬁndings into another proposition: Proposition 2. x) < c.56) implies limt→∞ |∇u(t. x)|2 dx. u(0.6. t→∞ Ω u(t.2. u(t.54) From the boundary condition. x) tends to a constant 1 function u(x) = u0 (x)dx. x) when t → ∞ must be a function with zero gradient. Since c|Ω| = Ω u(x)dx = lim implicitly assume u(t. |Ω| Ω Similar result can also be shown for Dirichlet problem. x)]2 dx. thus the function value will decrease over these points. Suppose that u(t. Thus limt→∞ E(t) ≥ 0 exists. We consider ut = 4uxx . we can get a conclusion that the diﬀusion equation will make the initial value ﬂatter and bring it eventually to an equilibrium state without spatial pattern. (2.53). In the following we use a Maple program to demonstrate the smothering eﬀect and the asymptotic tendency to the equilibrium. and for the points where u(t. (2.56) (2.) Nevertheless. Then (2. 0) = u(t. then c = Ω u0 (x)dx/|Ω|. x)|2 dx = 0.57) . then u(t. the proof of its existence is beyond the scope of this notes. x)| = 0 for any x ∈ Ω. in which the limit function is u = 0. For the points where u(t. x)]2 dx = −D |∇u(t. This is called smothering eﬀect of diﬀusion. x)dx = Ω u0 (x)dx. (2. x) = f (x). From this property of diﬀusion equation. with ∆u < 0.53) Here we use Green’s identity (a consequence of divergence theorem): u∆vdx = Ω ∂Ω ∂Ω u(∇u u(∇v · n)ds − Ω ∇u · ∇vdx. 4) = 0. RELAXATION TO EQUILIBRIUM SOLUTIONS and the right hand side: D Ω 27 u∆udx = D ∂Ω u(∇u · n)ds − D Ω ∇u · ∇udx. Let the average of the initial value be c. (Notice that we (2. (2.4.45). thus from (2.52) and (2. that is why this is only a sketch of proof. While such a limit indeed exists. t > 0. x) has a limit u(x) when t → ∞.55) shows that E ′ (t) < 0 for all t > 0 and obviously E(t) > 0 for all t > 0. thus the limit function must be a constant u(x) = c.55) [u(t. we ﬁnd Ω Ω [u(t. the function is likely to be convex down. thus the limit of u(t. x ∈ (0. the opposite occurs.

28 100. animate(uapprox1(x..0.> n=1 c(n) sin(1/4 n Pi x) exp(-4 n2 t) > > > > > tvals:=seq(0. Here are a few snap shots of the animation: ...x=0.t).t=tvals)]: plot(toplot.20): toplot:=[seq(uapprox1(x. x=0.4). 1) (3.. plot(f(x). x ∈ (0.4). 0.1. DIFFUSION EQUATIONS x ∈ [1.07/20*i.i=0. From (2.30). x=0. m ≥ 1.59) It is an interesting experiment to observe the changes of the spatial patterns of u(t. 4).t).58) 1 = 2 mπx dx. x) when t gets forward.4): c(n). .frames=50). 30 uapprox1 :=(x. t=0.. (2. t sin 4 4 (2. c:=n->0. we get the x ∈ [0. CHAPTER 2. 4]. This is an animation generated by approximating the solution by the partial sum with n = 30 of the series solution. Here is the Maple program diffusion. but we will let Maple to do that.4.32).0.t=0.4.57): u(t. where f (x) is given by the formula f (x) = formula of the solution of (2. with(plots): n:=’n’: f:=x->100*Heaviside(x-1)-100*Heaviside(x-3)..2.5*int(f(x)*sin(n*Pi*x/4). 0. densityplot(uapprox1(x. x < 0.cos(3/4 n Pi) n Pi >uapprox1:=(x.x=0.t) .mws and some of the results: > > > > > restart. which is deﬁned as H(x) = 1.. The approximation is accurate since the remainder is an exponential decaying function with smaller exponent.t).t)->sum(c(n)*sin(n*Pi*x/4)*exp(-4*n^2*t). Note that the initial condition f (x) can be written as the diﬀerence of two Heaviside functions.. f (x) sin 4 The coeﬃcients cm in the Fourier series can easily be calculated.x=0. t > 0.. x ≥ 0. x) = where cm ∞ m=1 cm exp − 4 0 mπx m2 π 2 . n=1. 200. 3].colorstyle=HUE).0 cos(1/4 n Pi) .

x) is not a constant for x ∈ (1. 3). u(t.1: (a) f (x). x) above indeed is the solution of (2. and all the small wiggles in u(0. Forming a plateau From t = 0 (Figure 2. 3). u(0.002. (2.002. However the road to distinction is worth a more careful looking—it can be dismantled into the following stages: 1.4. x) → 0 exponentially fast.1 is a comparison of f and f .57) with initial data 30 f (x) = m=1 cm sin 200 mπ mπ mπx dx. (b) u(0. x). x) is almost a constant on an interval slightly smaller than (1. RELAXATION TO EQUILIBRIUM SOLUTIONS 29 100 100 80 60 40 20 80 60 40 20 0 0 1 2 x 3 4 1 2 x 3 4 Figure 2.002.2 (a)).2: (a) u(0. So in the initial stage of the evolution of f under diﬀusion.60) Figure 2.1(b)) to t = 0. cm = cos − cos 4 mπ 4 4 . we know that when t → ∞. we can observe that the initial data has been smoothed by the diﬀusion equation. a plateau (or ﬂat core) of height 100 forms. u(t. although u(0. with a sharp but smooth drop-oﬀ to 0 near x = 1 and x = 3. From the last section.2. (b) partial sum of Fourier series of f (x) 100 80 60 40 20 100 80 60 40 20 0 1 2 x 3 4 0 1 2 x 3 4 Figure 2. . x) Because of the truncation of the series solution.002 (Figure 2. x) are gone at t = 0. On the other hand.010.

the graph can hardly be seen. x) is close to a bell-shape.4) . (b) u(0.040.002. 3. ·) at x = 2 looks more like a regular local maximum point. And at the same time. (b) u(0. the interface between u = 0 and u = 100 is no longer as sharp as when t = 0. the convex part is shrinking as population loss via boundary gets bigger. the ﬂat core keeps shrinking though the top of the core is still at the level near 100. Erosion of the plateau From t = 0. x). (see Figure 2.600. and the maximum point of u(t. (see Figure 2.002 (Figure 2. and the convex part of the graph up.2 (b)).100. x) 2. ·) are still near x = 1 and x = 3. and an exponential collapse is obvious.04. At t = 0. Collapse of the arch In the ﬁnal stage. x) 100 80 60 40 20 0 100 80 60 40 20 0 1 2 x 3 4 1 2 x 3 4 Figure 2. In that time interval. After t = 1. the graph becomes almost concave for all x. DIFFUSION EQUATIONS 100 80 60 40 20 0 100 80 60 40 20 0 1 2 x 3 4 1 2 x 3 4 Figure 2. smoothing eﬀect brings the concave part of the graph down.01.01 (Figure 2.01. The graph of u(0.3) 4. the erosion ﬁnally dissolves the whole ﬂat top. the inﬂection points of u(t.020. At t = 0. x).3: (a) u(0. and the bell-shape becomes an arch. but at t = 0. the arch is close to the leading term of the series: c1 exp(−π 2 t/4) sin(πx/4). Bell getting round In the next stage.4: (a) u(0.30 CHAPTER 2.01.2 (a)) to t = 0.

We assume that the diﬀusion of the salt is one-dimensional in the tube. From the Fick’s law.63) A2 A1 thus we obtain the boundary conditions ∂c(t. t > 0. the total in-ﬂux at x = 0 is J · (−n)ds = [−Dcx (t. a salt water solution with constant concentration C0 kg/m3 enters the tube at a rate of R0 m3 /s. which satisfy equation: 2 D ∂ c = 0. (2.5 One dimensional chemical mixing problem Here we consider a simpliﬁed version of the chemical mixing problem which is discussed in Chapter 1: A tube with length L contains salt water.62) A1 A1 and the total out-ﬂux at x = L is J · nds = [−Dcx (t.61) At the left hand side of the tube. we could ﬁnd the equilibrium solutions. thus c(t. 0) = C0 R0 . x ∈ (0.5. and =− . (2. We also assume that the area of the cross section of the tube is A. t > 0 and 0 < x < L. and the velocity of ﬂuid is ignorable. t > 0. L). Let the concentration of the salt water in the tube be c(t. ∂x ∂x Thus the initial-boundary value problem is 2 ∂c = D ∂ c . and E = . L)] · (1)ds = −DAcx (t.64) Before attempting to solve the equation. we deﬁne B= C0 R0 R0 . 0) = C(t. L) c(t. x) satisﬁes the diﬀusion equation: ∂2c ∂c = D 2. the mixed solution is removed at the rate of R0 m3 /s.2. ∂x DA ∂x DA To simplify the notations. ONE DIMENSIONAL CHEMICAL MIXING PROBLEM 31 2. ∂t ∂x (2. x) = c0 (x). ∂x ∂x c(t. L)R0 =− . 0) C0 R0 ∂c(t. ∂t ∂x2 ∂c(t. 0) ∂c(t. The cross-section (yz-direction) of the tube is so small so we can assume that the concentration of the salt water is same for any point on a cross-section. x ∈ (0.67) ∂c(L) ∂c(0) = −B. L). Let A1 and A2 be the cross sections at x = 0 and x = L correspondingly.66) . 0)] · (1)ds = −DAcx (t. + Ec(L) = 0. ∂x2 (2. and on the right hand side of the tube. L)R0 . L) = 0. + Ec(t. x ∈ (0. x) kg/m3 . L). DA DA (2. L) = −B. (2.65) (2.

mπ/L) for any positive integer m > 0. Thus the solution of (2. then V (x) = c1 e− det and k must satisfy √ kx + c2 e kx . x) − c∗ (x). L) (2. V (x) = c1 cos( −kx) + c2 sin( −kx).71) This is a half-Neumann and half-Robin boundary condition. If k > 0. E DA (2.73) has no solution. ) satisﬁes cot(xm L) = xm . 0) ∂b(t. We use the method in Section 2. L) = 0. and the corresponding eigenfunction is cos( −km x).69) is homogeneous. We assume that b(t. (E − k)e (E + k)e √ 2 kL √ (2. From the boundary conditions. x) = ∞ m=1 bm e−Dxm t cos(xm x). ∂t ∂x2 ∂b(t. x ∈ (0. we get c2 = 0 √ √ √ and c1 [cos( −kL) − −k sin( −kL)] = 0. Then this km = −x2 m √ is the m-th eigenvalue of (2. (2. (2.74) From the graph of the functions f1 (x) = cot(Lx) and f2 (x) = x. there are exactly one intersection point xm of f1 and f2 in ((m − 1)π/L. ∂x = 0. (2. x ∈ (0. L]. V ′ (L) + EV (L) = 0. L).72) √ k−E =√ . x) satisﬁes 2 ∂b = D ∂ b . and V (x) satisﬁes V ′′ − kV = 0. thus there are no positive eigenvalues. L).69) + Ec(t. t > 0. we get √ √ k − k √ √ − kL √ √kL = 0. x ∈ (0.69) is b(t. Then b(t. (2. we ﬁnd the unique equilibrium solution: c∗ (x) = C0 R0 B + B(L − x) = C0 + (L − x). we deﬁne b(t. L).71). Thus k satisﬁes √ √ cot( −kL) = −k. x) = c(t. So (2.75) L L . where xm ∈ ( 2 (m − 1)π mπ .70) U (t) V (x) U (t) = eDkt . Similarly k = 0 is not eigenvalue either.66) is that the boundary conditions in (2. and c∗ (L) = C0 . ∂x b(t.32 CHAPTER 2.2 to solve (2.69) instead of (2. DIFFUSION EQUATIONS From a simple calculation. √ √ If k < 0. To solve the equation. t > 0.71). The reason we consider (2. thus we can use the technique of separation of variables. x) = U (t)V (x).73) e k+E But the function on the left is greater than 1 when k > 0. V ′ (0) = 0. Then U ′ (t) V ′′ (x) =D = Dk. (2.68) which is a linear positive function with negative slope on [0. From the boundary conditions. and the one on the right is always less than 1 when k > 0. x) = c0 (x) − c∗ (x).

cm exp at − D L 0 m2 π 2 mπx t sin . 2. and the solution converges to the equilibrium solution exponentially fast. t > 0. The solution of (2.15). x) = u0 (x). CRITICAL PATCH SIZE AND BIFURCATIONS 33 Hence the solution is the sum of the equilibrium solution and an exponential decaying Fourier series. t > 0.76) can be solved in a similar way as (2.13)-(2. Let’s see what factor will determine the outcome of this battle. We have Dπ 2 or L > L2 Dπ 2 when a < 2 or L < L when a > D π. Equation (2. then u(t. but on the other hand. x) will have a part which is exponentially increasing. Our mathematical result shows that the concentration of pollutant in a continuously polluted river is even higher than that in the source. x) = ∞. 2 L L (2. thus whether u(t. since the concentration there is much higher than the ones of either branches.78) . same as the incoming solution. x ∈ (0. We should be cautious that we ignore the eﬀect of drifting here. lim u(t. As a consequence. u(0. the pollution can be most serious. x ∈ (0. lim u(t. x ∈ (0. x) = 0. x) is exponentially increasing is determined by the ﬁrst exponent coeﬃcient a − (Dπ 2 /L2 ).76) u(t. and the concentration at x = 0 is the highest. L). L).76) is u(t. the exiting solution also has the concentration C0 .6.2. when the index m in the sum increases. which could make a big diﬀerence for rivers. ∂t ∂x2 (2. In fact.6 Critical patch size and bifurcations We have seen that population are at a loss from the hostile boundary. 0) = u(t. In this section we consider Dirichlet boundary value problem of diﬀusive Malthus equation 2 ∂P = D ∂ P + aP. t→∞ a D π. L) = 0. It shows that in equilibrium state. So the loss and gain are in a competition. The left end point x = 0 is the source of the pollution where the pollutant enters the river at a constant rate. L). The example in this section can be thought as a model of a continuously polluted river. the population will increase through linear reproduction. In fact. thus the amounts of salt entering and exiting the tube are balanced. L However the fate of the solution may be diﬀerent: since at is added to the exponent. the form of the solution only changes slightly: adding a term eat to reﬂect the exponential growth. and the right end point x = L is where the river merges to a bigger river or the ocean. t→∞ a (2. m ≥ 1. we can see that at the conﬂuence of two polluted rivers.77) 2 = L mπx u0 (x) sin dx. now the exponent may be a positive one. the exponent gets more negative. x) = where cm ∞ m=1 The simplest population growth model is the Malthus model dP/dt = aP for a positive constant a > 0. but the concentration of solution in the tube is higher than C0 .

x. or land-use type into smaller fragments. We ﬁrst consider a linear diﬀusion reaction equation: ∂2u ∂2u ∂u ∂t = D ∂x2 + ∂y 2 + λu. a) × (0. Terrestrial Ecosystem Monitoring Sites. but we get equivalent results. conversion of forests into agricultural areas and suburbanization. we assume that u(t. The concept of critical patch size is related to habitat fragmentation in ecological studies. sometime that is more convenient since we do not need to change the domain. the eigenfunction associated with ﬁrst eigenvalue k1 = −π 2 /L2 . we obtain U ′ (t) = (Dk + λ)U (t). y) ∈ R. but can also be caused by natural processes such as ﬁre. we have exclusively considered one dimensional spatial domain. DIFFUSION EQUATIONS If we assume that the growth rate a and D are intrinsic parameters determined by the nature of species and environment. u(0. The number L0 = D/aπ is called the critical patch size. From (2. (x. y) = u (x. http://www.org/gtos/tems/ . y). We say that a bifurcation occurs at L = L0 as the equilibrium solution u = 0 changes stability type from stable for L < L0 to unstable for L > L0 . y) ∈ ∂R. 0 t > 0. From our mathematical result in this section.34 CHAPTER 2. which is considered to be one of several spatial processes in land transformation. Habitat fragmentation1 is the breaking up of a continuous habitat. then the size of the habitat L will play the deciding role here. ∇u · n = 0. (x. It is commonly used in relation to the fragmentation of forests. we can also use a or D as bifurcation parameter instead of L.80) (2. then the species can become extinct. b > 0. y) ∈ R = (0. the equation has other equilibrium solutions u(x) = sin(πx/L). Then similar to one-dimensional case. for the simplicity of the mathematical analysis. b). λ. We will have more discussion on bifurcation problems when the growth rate is nonlinear. 1 (2. Habitat fragmentation is mainly caused by human activities such as logging. We notice that at L = L0 .fao. and ∂R is the boundary of rectangle R. (x. as the population cannot survive if the habitat is too small (the outgoing ﬂux wins over the growth). y).79) where D. (2. we discuss the separation of variables in a two-dimensional rectangle. In this section. 2.81) From TEMS. x. y) = U (t)V (x. but the population will not only survive but thrive to an exponential growth if the habitat is large enough (growth outpaces the emigration). a.7 Separation of variables: rectangles Most of this chapter. From the method of separation of variables.78). if the the fragmentation of habitat limits the spatial movement of the plants and animals. However any interesting application happens in 2-D or 3-D spaces. ecosystem.

84) so both W ′′ (x)/W (x) and Z ′′ (y)/Z(y) must be constants. · · · .0 = −1.86) (2. Therefore W and Z satisfy W ′′ (x) = k1 W (x).85) and (2.0 = 0. Thus we ﬁnd that the solution of (2. · · · . Then we obtain that W ′′ (x) Z ′′ (y) + = k. 1. n = 0.m = − n2 π 2 m2 π 2 nπx mπy − . 2 2 a b a b (2.87) and Both (2. y ∈ (0.82). SEPARATION OF VARIABLES: RECTANGLES and 2 2 ∂ V + ∂ V = kV. Wn (x) = cos 2 a a (2.82) are kn. x.86) are familiar eigenvalue problems in one dimensional space which have been studied in Chapter 2.92) . (2. b).91) where cn. (x. y) = W (x)Z(y). 1. Zm (y) = cos . (2. ∂x2 ∂y 2 ∇V · n = 0. a).2 = −5. y) = ∞ n=0.85) are k1n = − nπx n2 π 2 . m = 0. · · · .1 = −2. k1. k2.79) is u(t.1 = k1.82) For (2. Here we consider a square R = (0. m = 0. a) × (0. the Neumann boundary condition implies that W ′ (0) = W ′ (a) = 0 and Z ′ (0) = Z ′ (b) = 0. y) ∈ R = (0. The spatial patterns of the eigenfunctions for rectangle are much more complex than those of one-dimensional. 2. we need a further separation of variables: V (x. (2. k2. On the other hand. k = k1 + k2 .m e(Dkn. x ∈ (0. π). 2.m=0 cn.2 = −4. W (x) Z(y) (2. then the ﬁrst ﬁve distinctive eigenvalues are (see the contour graph of eigenfunctions below) k0. b). W ′ (0) = W ′ (a) = 0.88) and the eigenvalues and eigenfunctions of (2. (x.7. y) ∈ ∂R. π) × (0.85) Z ′′ (y) = k2 Z(y).0 = k0.m +λ)t cos mπy nπx · cos .1 = k1. Vn. 2 b b (2. 35 t > 0. Z ′ (0) = Z ′ (b) = 0. y) = cos · cos .m (x. k0.90) where n. The eigenvalues and eigenfunctions of (2.86) are k2m = − m2 π 2 mπy . . a b (2. 2.89) Therefore the eigenvalues and eigenfunctions of (2. (2. 1.83) (2.m can be determined by the initial conditions.2.

7: (From left to right) (a) cos x + cos y. (x. DIFFUSION EQUATIONS 3 3 3 2.5 2 2.1 = cos y. 3 3 2. (d) V2. b). y) = u (x. k1. y 0.5 1 1.5 1.5 1 1. y).5 0 2 2 x 1. In such case.5 y 0.5 0 0 1 1 3 2 0 3 2 0 3 2 1 2. x. u(0.5 0.5 1.5 x 1.5 2 2 2 x x x 2 3 2 1 2.91) is always an exponential growth if λ > 0.5 y y Figure 2. x.5 2.5 1. from the equation.93) .n = kn. y) ∈ ∂R. (b) V2.0 = cos x. y) ∈ R = (0. the no-ﬂux boundary condition prevents the emigration through the boundary. (b) V0. thus the total population in R has an exponential growth.5 x 1 1 0. (c) V1.5 1.5 1 1. (x.5 0 0.5 0 0 0. y) ∈ R.6: (From left to right) (a) V0.1 = cos x · cos y. We can consider corresponding Dirichlet boundary problem: ∂2u ∂2u ∂u =D + 2 + λu. In particular. and the eigenspace is span{cos x. For example.5 0 0 0. 3 3 3 2.5 1 1. y) = 0.5 y y Figure 2.5 1.5 y 0.1 = cos(2x) · cos y.5 1 1.5 2.5 3 (2.5 1.5 0.m = −(m2 + n2 ). (d) V1.5 2.5 y 0. We notice that many eigenvalues km.5 0.5 0 2 1. and λ > 0 implies a positive growth rate.5 1.5 2.5 1.5 0 0 1 1 3 2 0 3 2 0 3 2 1 2.5 1. (b) cos y − cos x.5 1.5 0.5 x 2. thus k1 cos x + k2 cos y is a spatial pattern for any k1 and k2 .5 0. (x.5 x y 0. the patterns generated by cos x ± cos y are symmetric with respect to one of the diagonal lines.5 2. cos y}.5 1.5 0.5 0 1 3 2 0 3 2 1 2. the eigenspace is of higher dimensional.2 = cos x · cos(2y).5 2.0 = k0.5 1.5 0 x 2 3 2 1 2.36 CHAPTER 2.5 1 1 0. Indeed.5 y Figure 2. ∂t ∂x2 ∂y u(t.5 2. so more possible spatial patterns are generated.1 = −1.5 0 0.n have multiplicity more than 1.5 0. The solution (2.2 = cos(2y). a) × (0.5 2 2. (c) V1. 0 t > 0.0 = 1.5 x 1 0. for example km.5 2.0 = cos(2x).5 2.5: (From left to right) (a) V0.5 3 0 0.

4. Find the eigenvalues and eigenfunctions of the problem y ′′ + 2y ′ + ky = 0. L). 6. x ∈ (0.95) 2. Find the eigenvalues and eigenfunctions of periodic boundary problem y ′′ = ky.99) (2. Show that the solution of ut = Duxx .m e(Dkn. L) = 0.101) mπx dx. lim u(t. Find the eigenvalues and eigenfunctions of Robin boundary problem (b > 0) y ′′ = ky. y(0) = y(π) = 0. L). (2. the solution is u(t.90). π). y ′ (0) = by(0). STEADY STATES IN RADIALLY SYMMETRIC DOMAINS AND TRANSIT TIMES Similar to the method above. 3. u0 (x) cos L . m ≥ 0. y(0) = y(π). Find the eigenvalues and eigenfunctions of homogeneous Neumann boundary problem y ′′ = ky. x) = ∞.8.94) where bn. the persistence/extinction of the population can be determined: when 1 1 λ > 2 + 2 . x u(0. 2. x.m can be determined by the initial conditions.97) (2.100) is u(t. x ∈ (0. lim u(t. L). when t→∞ Dπ 2 a b (2. 2 L L (2. and plot the graphs of the corresponding eigenfunctions.8 Steady states in radially symmetric domains and transit times Chapter 2 Exercises 1.2. 2 t→∞ Dπ a b 1 1 λ < 2 + 2 .m=1 37 bn. use Maple to ﬁnd the numerical values of the ﬁrst three eigenvalues.m +λ)t sin nπx mπy · sin . L). t > 0. x) = 0. x ∈ (0. Similar to last section. a b (2. π). y ′ (0) = y ′ (L) = 0. and kn.96) 5. L). In Example 2. x ∈ (0.m is deﬁned in (2. 0) = ux (t. y) = ∞ n=1.1. x) = where cm c0 + 2 2 = L ∞ m=1 L 0 (2. x ∈ (0. u (t. cm exp −D mπx m2 π 2 t cos .98) (2. x ∈ (0. x) = u0 (x). t > 0. y ′ (0) = y ′ (π). x ∈ (0. y ′ (L) = −by(L).

unless u(t. x ∈ (0. (2. x) = 0. 1). V > 0. x) = f (x). x) is the solution of ∂u = D∆u.5]. Suppose that u(t.106) u(0. 1). the salt solution drifts to the right with velocity V . x) = u0 (x). −1. ∂x ∂x 1 Show that the total population u(t. 1) = 0. x) satisﬁes the convective-diﬀusion equation: ∂c ∂2c ∂c =D 2 −V . x u(0.107) with D. x) satisﬁes the same initial and boundary conditions in (2. x ∈ (0. Suppose that c(t. 9. x ∈ (0. Then c(t. x ∈ Ω. Suppose that in the chemical mixing problem in Section 2. Find the solutions of equilibrium equation u′′ + 3u′ = 0. x ∈ ∂Ω.102).105) t > 0. u(0) = 5. 0 12. x) is the solution of the convective-diﬀusion equation: (D > 0 and V > 0) 2 ∂u = D ∂ u − V ∂u . x ∈ (0. x) = u0 (x). 0) = (t. ∂t ∂x2 ∂x (2. t > 0. (2.5.5.103) x ∈ [0. ∂u ∂u (t.66). Find the solutions of equilibrium equation u′′ = 0. ∂t u(0. x)dx is a constant if u0 (x) = sin x but not a constant if u0 (x) = cos(x).104) (2. ∇u(t. u(0) + 3u′ (0) = 5. 0) = ux (t.4 to simulate the solution of (2. π) = 0. 1). Show that the mean-squared value I(t) = 11. x ∈ Ω. x) ≡ 0. (a) Find the solution of CHAPTER 2. 1]. ∂t ∂x ∂x (2. . x ∈ (0. 0. u (t.102) where f (x) is given by the formula f (x) = (b) Modify the Maple program in Section 2. 8. π). x)]2 dx is strictly decreasing Ω where a > 0. 10. t > 0. u(1) − 5u′ (1) = 7. 1. (2. π). DIFFUSION EQUATIONS ut = 4uxx . x ∈ (0.38 7. [u(t. x) · n(x) + a · u(t. Suppose that u(t. u′ (1) = 4.

y) = u(a. (a) Find the equilibrium solution of (2. y) ∈ R. b). x ∈ (0. 39 (b) Show that the equilibrium solution is positive. y) ∈ R = (0. Suppose that c(t.111) . x) satisﬁes the same initial and boundary conditions in (2. x) satisﬁes a linear diﬀusion equation: ∂2c ∂c = D 2 − kc. ∂t ∂x2 ∂x u(t. x. L) of the equilibrium solution with that of (2. and explain the eﬀect of the reaction to the geometric properties of the equilibrium solution. 0) = bu(t. Then c(t. b).107).67). L). (x. ∂t ∂x (2. Consider the doubly periodic boundary value problem: 2 2 ∂u = D ∂ u + ∂ u + λu. y). and the series representation of the solution. L).2. ∂t ∂x2 ∂y 2 u(0. a) × (0. the salt is consumed in a chemical reaction at a rate k. y). Determine the critical patch for a diﬀusive Malthus equation with Robin boundary conditions: 2 ∂P = D ∂ P + aP. (2.108) with same parameters in Maple. k > 0. x) = u0 (x). (c) Compare the exiting ﬂux −Dcx (t. 0) = uy (x. u(0. y). y) = u0 (x. L) = −bu(t.66). decreasing and concave up. 15. Consider a convective-diﬀusion equation 2 ∂u = ∂ u + 4 ∂u + 8u. L) of the equilibrium solution with that of (2. L). L). t > 0. x ∈ (0.5. u(0. x ∈ (0. STEADY STATES IN RADIALLY SYMMETRIC DOMAINS AND TRANSIT TIMES (a) Find the equilibrium solution of (2. 14. u(t.67). 13. L) = 0. u(x. ux (t. t > 0. x ∈ (0. and explain the eﬀect of the drifting to the geometric properties of the equilibrium solution. (x. Find the solution of the equation with the following steps: (2.) (b) Show that the equilibrium solution is positive. y) = ux (a. ux (0. (2. u(0. 0). decreasing and concave down. L).108) with D. ∂t ∂x2 (2. Suppose that in the chemical mixing problem in Section 2. x) = u0 (x). 0) = u(x. b). t > 0.108).110) Find the eigenvalues and eigenfunctions of the problem. (c) Compare the exiting ﬂux −Dcx (t.8.67). (d) Plot the graphs of the equilibrium solutions of (2. uy (x.109) ux (t.107) and (2. 0) = 0. 16. (You can use Maple to ﬁnd the general formula.

V (0) = V (L) = 0. (f) Suppose that the species lives in a river with length L. DIFFUSION EQUATIONS (a) Use separation of variables method to show that if u(t. x) = U (t)V (x) is a solution. then for some constant k. V ′′ (x) + 4V ′ (x) + 8V (x) = kV (x). (e) Describe the population distribution qualitatively when L > L0 .) e u0 (x) sin L 0 L (d) Determine the critical patch L0 of the problem. U and V satisfy U ′ (t) = kU (t). and the convection is due to the drifting of the river. Explain your mathematical results in this context. .) (c) Find the solution of the equation in a series form. V (0) = V (L) = 0. 2 L 2x nπx (Hint: cn = dx. (b) Find the eigenvalues and eigenfunctions of V ′′ (x) + 4V ′ (x) + 8V (x) = kV (x). k = 4 and k < 4 separately. (Hint: treat the cases of k > 4.40 CHAPTER 2.

- Solutions Tute 2
- Must See Befor Final
- MScdiagquiz
- final_22b_08_solutions.pdf
- end_sem_new
- Lagrangiano Dos Masas Con Resorte
- first oder diff eqs
- MATH 2930 - Worksheet 4 Solutions
- Everyman Diff Eq's
- Midterm A Solutions S10
- 2003_Paper I
- [Herbert Kreyszig, Erwin Kreyszig] Student Solutio(BookZZ.org)
- 9781461436171_sm
- Mathematical Methods Coursework!
- Linear Differential Equations
- Algebraic Methods to Compute Mathieu Functions - Frenkel
- Ecuaciones diferenciales parciales de Logan - Chapter 3 Solutions
- Differential Equations - Solved Assignments - Semester Spring 2007
- Multivariable 17 Differential Equations
- diffeq1
- Forced Vibration
- Assignment
- ODE_Main
- 2nd Order Homogeneous Differential Equations
- ODE LargeFont
- 19_3
- LectnotesCh3
- Calculus one and several variables 10E Salas solutions manual ch09
- AMATH 350 Notes
- MTH101
- Diffusion Equation

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd