You are on page 1of 92

Class notes for MAE 294c:

Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky


May 30, 2005

Table of Contents
1

Mathematical Modeling of Physical Phenomena

1.1

Stages of a modeling process . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

A historic perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Derivation of Equations of Mathematical Physics

2.1

Vibrating strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.1

Physical phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1.2

Construction of a physical model . . . . . . . . . . . . . . . . . . . . .

2.1.3

The choice of a function describing the process . . . . . . . . . . . . . .

2.1.4

Physical laws governing the vibration of a string . . . . . . . . . . . . .

2.1.5

Derivation of governing equations for u . . . . . . . . . . . . . . . . . .

2.1.6

Initial and boundary conditions . . . . . . . . . . . . . . . . . . . . . .

2.2

Vibrating membranes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.3

Generalizations of the wave equation . . . . . . . . . . . . . . . . . . . . . . . .

2.4

Heat conduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.4.1

Physical phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.4.2

Construction of a physical model . . . . . . . . . . . . . . . . . . . . .

2.4.3

The choice of a function describing the process . . . . . . . . . . . . . .

10

2.4.4

Physical laws governing heat conduction in solids . . . . . . . . . . . . .

10

2.4.5

Derivation of governing equations for u . . . . . . . . . . . . . . . . . .

11

2.4.6

Initial and boundary conditions . . . . . . . . . . . . . . . . . . . . . .

12

2.4.7

Stationary temperature fields . . . . . . . . . . . . . . . . . . . . . . . .

13

MAE 294c. Methods in Applied Mechanics III


2.5

Instructor: Daniel M. Tartakovsky

Continuity equations for fluids . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

2.5.1

Physical phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

2.5.2

Construction of a physical model . . . . . . . . . . . . . . . . . . . . .

13

2.5.3

The choice of functions describing the process . . . . . . . . . . . . . .

14

2.5.4

Derivation of governing equations . . . . . . . . . . . . . . . . . . . . .

14

Classification of PDEs

15

3.1

Basic terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

15

3.2

Well-posed problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16

3.3

Classification of second-order PDEs . . . . . . . . . . . . . . . . . . . . . . . .

16

3.3.1

Transformation of second-order PDEs to their canonical form . . . . . .

17

3.3.2

Classification of second-order PDEs . . . . . . . . . . . . . . . . . . . .

19

3.3.3

A curse of dimensionality . . . . . . . . . . . . . . . . . . . . . . . . .

20

3.3.4

Classification of second-order PDEs in n = 2 dimensions . . . . . . . .

20

Method of Characteristics

23

4.1

Linear first-order PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

4.1.1

Constant wave velocity . . . . . . . . . . . . . . . . . . . . . . . . . . .

24

4.1.2

Variable wave velocity . . . . . . . . . . . . . . . . . . . . . . . . . . .

24

4.2

Classification of second-order PDEs in two dimensions (Cntd.) . . . . . . . . . .

25

4.3

Linear advection with a source . . . . . . . . . . . . . . . . . . . . . . . . . . .

26

4.4

Linear advection and reactions . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

4.5

Quasi-linear first-order PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . .

28

4.6

Nonlinear first-order PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

29

4.6.1

Elementary traffic model . . . . . . . . . . . . . . . . . . . . . . . . . .

31

Shock waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

4.7.1

Discontinuous solutions . . . . . . . . . . . . . . . . . . . . . . . . . .

32

4.7.2

Elementary traffic model with shocks . . . . . . . . . . . . . . . . . . .

34

4.7.3

Conditions for shock formation . . . . . . . . . . . . . . . . . . . . . .

36

Nonlinear first-order PDEs with a source . . . . . . . . . . . . . . . . . . . . . .

37

4.7

4.8

- ii -

MAE 294c. Methods in Applied Mechanics III


5

Laplace Transformation

39

5.1

Motivation and Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

5.2

Properties of the Laplace transformation . . . . . . . . . . . . . . . . . . . . . .

40

5.2.1

Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

40

5.2.2

Applicability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

40

Solutions of linear ODEs with constant coefficients . . . . . . . . . . . . . . . .

43

5.3.1

Forced vibrations with damping . . . . . . . . . . . . . . . . . . . . . .

43

5.3.2

A damped absorber . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

43

5.4

Solutions of integral equations with convolution kernels . . . . . . . . . . . . . .

44

5.5

Solutions of linear PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

45

5.5.1

Wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

45

5.5.2

Diffusion equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

47

5.3

Instructor: Daniel M. Tartakovsky

Nonhomogeneous Problems and Greens Functions

49

6.1

Motivation and basic ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

49

6.2

The Dirac delta function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

6.3

Properties of the Greens functions . . . . . . . . . . . . . . . . . . . . . . . . .

52

6.3.1

Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

52

6.3.2

Reciprocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

52

6.4

Calculation of the Greens functions . . . . . . . . . . . . . . . . . . . . . . . .

53

6.5

Free surface flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

Nonlinear Diffusion

59

7.1

Time-dependent diffusion coefficients1 . . . . . . . . . . . . . . . . . . . . . . .

59

7.2

Boltzmanns transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

7.3

Kirchhoffs transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

61

7.3.1

Steady-state diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . .

61

7.3.2

Steady-state diffusion with gravity . . . . . . . . . . . . . . . . . . . . .

61

7.4

Method of spatial moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

7.5

Hopf-Coles transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

- iii -

MAE 294c. Methods in Applied Mechanics III


8

Instructor: Daniel M. Tartakovsky

Approximate Solutions of ODEs: Local Analyses

67

8.1

Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67

8.2

Fuchs theory of ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

68

8.3

Local behavior of linear ODEs near ordinary points . . . . . . . . . . . . . . . .

69

8.4

Local behavior of linear ODEs near regular singular points . . . . . . . . . . . .

70

8.5

Local behavior of linear ODEs near irregular singular points . . . . . . . . . . .

70

8.6

Local behavior of inhomogeneous linear ODEs . . . . . . . . . . . . . . . . . .

71

8.7

Nonlinear ODEs: Spontaneous singularities . . . . . . . . . . . . . . . . . . . .

72

8.8

Early-time behavior of nonlinear ODEs . . . . . . . . . . . . . . . . . . . . . .

73

8.9

Late-time behavior of nonlinear ODEs . . . . . . . . . . . . . . . . . . . . . . .

74

8.10 Local behavior of nonlinear ODEs on bounded domains . . . . . . . . . . . . . .

74

Scaling and Self-Similarity

77

9.1

Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

9.2

Dimensional analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

9.2.1

Transformation to dimensionless parameters . . . . . . . . . . . . . . .

78

9.2.2

Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

79

9.3

Similarity and scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

79

9.4

Self-similar solutions: Linear diffusion . . . . . . . . . . . . . . . . . . . . . . .

80

9.5

Self-similar solutions: Nonlinear diffusion . . . . . . . . . . . . . . . . . . . . .

82

9.6

Anomalous Exponents and Self-Similarity of the Second Kind . . . . . . . . . .

84

- iv -

Chapter 1

Mathematical Modeling of Physical


Phenomena
1.1

Stages of a modeling process

A typical modeling process of a physical phenomenon consists of the stages shown in Figure 1.1.
Physical
Model

Mathematical
Model

Solutions

Interpretation

Figure 1.1: A schematic representation of a typical modeling process.


Construction of a model consists of the following steps
1. Construction of a physical model, i.e., idealization of a physical process,
2. The choice of (i) a function u(x1 , . . . , t) describing the process and (ii) physical laws to
which it obeys,
3. Derivation of governing equations for u,
4. Derivation of complementary conditions
(a) boundary conditions
(b) initial conditions.

1.2

A historic perspective

Some of the earliest seminal contributions to the field of mathematical physics include
Isaac Newton, 1687, Philosophiae Naturalis Principia Mathematica.
1

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Newtons laws of motion


Newtons Law of Gravitation
Brook Taylor, 1715, Methodus Incrementorum. An equation for a vertically vibrating string,
2u
2u
=
D
,
t2
x2

where u is a vertical displacement.

Joseph-Louis Lagrange, 1788, Mecanique analytique. An equation for electric fields,


2u 2u
+ 2 = 0,
x2
y

where u is electric charge.

Jean Baptiste Joseph Fourier, 1822, Theorie analytique de la chaleur. An equation for heat
conduction in a rod,
u
2u
= D 2,
t
x

where u is temperature.

James Clerk Maxwell, 1873, Treatise on Electricity and Magnetism.


Maxwells equations.
Sofia Kovalevskaya, 1874, PhD dissertation at the University of Gottingen.
A prove of the existence and uniqueness of solutions to the Cauchy problem.

-2-

Chapter 2

Derivation of Equations of
Mathematical Physics
Below we use the general framework outlined in Section 1.1 to derive a few mathematical models.

2.1
2.1.1

Vibrating strings
Physical phenomenon

vibration of a string

Figure 2.1: A highly stretched string with firmly fixed ends. It is shown in the state of equilibrium,
i.e., before a force has been applied.

2.1.2 Construction of a physical model


The process is idealized by assuming that
A string is perfectly flexible, i.e., it offers no resistance to bending;
A string is perfectly elastic, i.e., the tension in the string depends only on its local stretching.
All points of the string move in the direction, perpendicular to its equilibrium state;
The vibrations (deviations from the state of equilibrium) are small.
3

MAE 294c. Methods in Applied Mechanics III

2.1.3

Instructor: Daniel M. Tartakovsky

The choice of a function describing the process

The process is described by the vertical displacement of a string u(x, t) from its equilibrium state
(Figure 2.2). Note that
u(x0 , t) provides the law of motion for a point x0 of the string,
u(x, t0 ) describes a profile of the string at t0 = 0

(x, t)

u(x, t)

x=0

x=L

Figure 2.2: A profile of the vibrating string (solid line) at time t away from its equilibrium state
(dashed line). The function u(x, t) designates the vertical displacement of a string at point x and
time t.
The following useful relations hold. Let (x, t) denote the slope of the string, i.e., the angle
between the positive x-axis and the positive direction of the tangent, as shown in Figure 2.2. Then
tan (x, t) =

u
.
x

(2.1)

The velocity v and acceleration a of a point x of the string are given by


u(x, t)
t

v(x) =

and

a(x) =

2 u(x, t)
,
t2

(2.2)

respectively.
Since we assumed that the vibrations u are small, the terms similar to


u
u ,
x

u ,

u
x

2
,

etc

(2.3)

\
can be neglected. This fact has an interesting implications. For example, since the length |M
1 M2 |
\
of a segment M1 M2 of the string in Figure 2.3 is
Z
\
|M
1 M2 | =

x+4x

s
1+

u
x

it does not change with time!


-4-

2

Z
dx

x+4x

dx = 4x,
x

(2.4)

MAE 294c. Methods in Applied Mechanics III

(x, t)

Instructor: Daniel M. Tartakovsky

M1

(x + 4x, t)

M2

T(x + 4x, t)
T(x, t)
x + 4x

Figure 2.3: A profile of the vibrating string (solid line) at time t away from its equilibrium state
(dashed line). The function u(x, t) designates the vertical displacement of a string at a point x and
time t. A tangential force T is called the tension in the string.

2.1.4

Physical laws governing the vibration of a string

Hookes law (R. Hooke, 1676): A force applied to a body is directly proportional to the
resulting deformation of the body;
dAlemberts (J. dAlembert, 1742) principle,
X
Fi = 0,

(2.5)

which states that all forces Fi acting on a body (including the inertial force) must be in
equilibrium.

2.1.5

Derivation of governing equations for u

\
Consider a small segment of the string M
1 M2 (Figure 2.3). Its deformation induces the tension in
the string, which is the tangential force T shown in Figure 2.3. Let T = |T| denote the magnitude
of T. According to (2.4), the length of the string does not change with time. Hence neither does
T, i.e., we showed that T = T(x).
According to dAlemberts principle (2.5), the projections x and u of all forces on the
horizontal and vertical coordinates, respectively, must equilibrate. Since
x {T(x + 4x)} = T (x + 4x) cos (x + 4x, t),

x {T(x)} = T (x) cos (x, t),

u {T(x + 4x)} = T (x + 4x) sin (x + 4x, t),

u {T(x)} = T (x) sin (x, t),


(2.6)

dAlemberts principle (2.5) gives two equations


T (x + 4x) cos (x + 4x, t) T (x) cos (x, t) = 0

(2.7)

and
T (x + 4x) sin (x + 4x) T (x) sin (x, t) + 4xf (x, t) = 4x

2u
.
t2

(2.8)

In (2.8), f denotes the density of an external force, so that 4xf is a total external force acting
\
on the segment M
1 M2 ; and is the density of the string, so that 4x is the mass of the segment
\
M 1 M2 .
-5-

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Next we recall that


1

cos = p

1 + (tan )2

(2.9)

Substituting (2.1) into (2.9), while recalling (2.3), yields


1

cos (x, t) = q

1+


u 2
x

1.

(2.10)

It then follows from (2.7) that


T (x + 4x) T (x) = 0.

(2.11)

In words, under our assumptions, the tension in the string remain constant T (x, t) T0 .
Since
tan

sin = p

1 + (tan )2

tan ,

(2.12)

recalling (2.1) gives


sin

u
.
x

(2.13)

This allows us to rewrite (2.8) as




u
u
2u
T0
(x + 4x)
(x, t) + 4xf (x, t) = 4x 2 .
x
x
t

(2.14)

Dividing by 4x and taking the limit as 4x 0 yields


2u
2u
+
f
(x,
t)
=

x2
t2

(2.15)

2
2u
2 u
=
c
+ g(x, t),
t2
x2

(2.16)

T0
or

where c2 = T0 / and g = f /. Eq. (2.16) is called a wave equation.

2.1.6

Initial and boundary conditions

To complete the description of the problem one needs to specify initial and boundary conditions
for Eq. (2.16).
Initial conditions describe the behavior of the string at some initial time t = t0 . It is common
to set t0 = 0. Since (2.16) contains second time derivative, there must be two initial conditions,
which specify the initial shape (x) and velocity (x) of the string, i.e.,
u(x, 0) = (x)

and

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

for

0 x L.

(2.17)

Boundary conditions depend on the state of the string at the boundaries x = 0 and x = L.
For example,
-6-

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

If both ends of the string are fixed,


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

t 0;

(2.18)

If the ends of the string move according to prescribed laws 1 (t) and 2 (t),
u(0, t) = 1 (t),

u(L, t) = 2 (t),

t 0;

(2.19)

If forces v1 (t) and v2 (t) acting on the ends of the string are known,
u
(0, t) = v1 (t),
x

u
(L, t) = v2 (t);
x

(2.20)

Etc.

2.2

Vibrating membranes

The process is described by u(x, t), the vertical deviation of a point x = (x1 , x2 )T of a perfectly
flexible, perfectly elastic membrane from its equilibrium position.
u

u(x, t)
x2

x = (x1 , x2 )T

x1

Figure 2.4: Small vibrations of a membrane.


Following the procedure described in Section 2.1, one can show that u(x, t) satisfies the
two-dimensional wave equation
2u
= a2 4u + g(x, t),
t2
-7-

(2.21)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

where 4 2 denotes the Laplacian.


The initial conditions for (2.21) are
u(x, 0) = (x),

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

x .

(2.22)

Here = denotes the union of a domain , on which (2.21) is defined, and its boundary
).
If the membranes boundary is fixed, the corresponding boundary condition is
u(x, t) = 0,

2.3

x .

(2.23)

Generalizations of the wave equation

Consider small vibrations of a uniform isotropic body. Let v = (v1 , v2 , v3 )T denote the vector
of the deviation of a point x = (x1 , x2 , x3 )T in the body at time t from its equilibrium state.
Specifically, consider the deviations
v1 (x2 , x3 , t),

v2 (x1 , x3 , t),

v3 (x1 , x2 , t).

Since each slice of the body vibrates as a membrane, all vi (i = 1, 2, 3) satisfy the two-dimensional
wave equation (2.21),

 2
2 v1
2 v1
2 v1
+
+ f1 (x2 , x3 , t)
=a
t2
x22
x23
 2

2 v2
2 v2
2 v2
+
+ f2 (x1 , x3 , t)
(2.24)
=a
t2
x21
x23
 2

2 v3
2 v3
2 v3
=a
+
+ f3 (x1 , x2 , t).
t2
x21
x22
Equations (2.24) can be rewritten in the vector form as
2v
= a2 4v + f (x, t).
t2

(2.25)

Next we assume that the field is potential, i.e., there exist scalar functions g(x, t and u(x, t)
such that
f = g

and

v = u.

(2.26)

Then it follows from (2.25) that




2u
a2 4u g
t2


= 0.

(2.27)

3
X
2
2,
x
i
i=1

(2.28)

Therefore we may take


2u
= a2 4u + g,
t2
-8-

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

which, of course, is the three-dimensional wave equation.


In general, the process of vibration can be described by the general wave equation


n
X
2u

u
pij
qu + f.
2 =
t
xi
xj

(2.29)

i,j=1

Here the state variable u(x, t) is a function of independent variables x Rn , and , pij , q, and f
are known functions. From physical considerations, > 0, pij > 0, and q 0.

2.4

Heat conduction

2.4.1

Physical phenomena

heat conduction (Fouriers law),


electrical conduction (Ohms law),
diffusive transport (Ficks law)
flow in porous media (Darcys law),
etc.

2.4.2

Construction of a physical model

The process is idealized by assuming that 1) a materials properties do not change with temperature
and 2) thermal energy is transported by conduction only, i.e., that convection can be ignored.
Conduction:
The dominant method of heat conduction in metal is through the movement of electrons.
This method of conduction does not operate in non-metals because there are no free electrons (other than graphite). When a metal is heated, the electrons closest to the heat source
vibrate more rapidly. These electrons collide with atoms and gain more kinetic energy
(movement energy). This causes the electrons to move around faster and to collide with
other free electrons which, in turn, gain more kinetic energy. Kinetic energy is transferred
between the electrons and through the metal from the point closest to the heat source toward
points further away. Since the electrons travel very short distances at very large velocities,
conduction of heat happens very quickly.
In metals and insulators, there is conduction of heat due to the vibration of atoms. As
atoms closest to the heat source absorb heat/thermal energy, they induce the vibration in the
neighboring atoms, which, in turn, make their neighboring atoms vibrate more, etc.
Examples of conduction:
-9-

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The wire gauzes used on tripods are metal therefore they are good heat conductors.
Gauzes on cookers are also metal so that heat is conducted quickly and food is cooked
fast.
Poor thermal conductors (insulators) are used for saucepan handles so that they dont
heat up and can still be handled.
Metals are used for the containers which heat liquids e.g. pans and kettles.
Air is a poor conductor therefore materials that trap air are used for insulation in lofts
and hot water cylinders.
Convection:
When particles of the cold air are heated by a heat source, they gain kinetic energy and
the air expands. The density of particles in the hot air decreases relative to that of the
surrounding cold air, which causes the hot air to rise and displace the cool air. Since cool
particles are more dense, they fall and move toward the heat source to take the place of the
warm particles. They then heat up and rise while other particles cool down and fall.
Example of convection:
Refrigerators are kept cool by convection.
Land and sea breezes are due to convection.
Atmospheric winds.
Hot water systems.

2.4.3 The choice of a function describing the process


The process is described by the temperature u(x, t) of a solid body.

2.4.4

Physical laws governing heat conduction in solids

The relationship between thermal energy density (the amount of thermal energy per unit
volume) e(x, t) and temperature u(x, t) is given by
e = u,

(2.30)

where (x) is the specific heat (the amount of heat per unit mass required to raise the
temperature by one degree Celsius) of the rod, and (x) its density (mass per unit length).
Fouriers (Joseph Fourier, 1822) law of conduction [see also Newtons (Isaak Newton, 1701)
law of cooling]: Heat energy will flow from the region of high temperature to the region of
low temperature
q = ku,

(2.31)

where q is the heat flux vector and k is thermal conductivity.


Conservation of energy (the first law of thermodynamics) Thales of Miletus (circa 635
BC - 543 BC), . . . , Julius Robert von Mayer (1842):
- 10 -

MAE 294c. Methods in Applied Mechanics III

Heat conducted IN

Instructor: Daniel M. Tartakovsky

Heat conducted OUT

Heat generated within

Change in energy stored within

Figure 2.5: Conservation of energy.

2.4.5

Derivation of governing equations for u

Inside a heated body , we select a small cube (Figure 2.6), whose dimensions are 4x1 , 4x2 and
4x3 , and whose volume is 4V = 4x1 4x2 4x3 . If (x) is the density of this cube, its mass is
M = 4V .
x3
(x1 , x2 , x3 )

(x1 , x
2 , x
3 )
x2

(x1 + 4x1 , x
2 , x
3 )

x1

Figure 2.6: An elementary volume used to derive the heat conduction equation.
Next, we write down the conservation of energy (Figure 2.5) in mathematical terms. Recalling (2.30), the amount of heat energy in the cube at times t is
H(x, t) = e(x, t)4V = (x)(x)u(x, t)4V.

(2.32)

The change in heat energy during the time interval [t, t + 4t] is
4H = H(x, t + 4t) H(x, t) = (x)(x)4V [u(x, t + 4t) u(x, t)] .

(2.33)

The amount of heat entering the cube through the side x1 + 4x1 during the time interval
[t, t + 4t] is Q1 (x1 + 4x1 ) = q(x1 + 4x1 )4x2 4x2 4t. According to Fouriers law (2.31) this
- 11 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

gives
Q1 (x1 + 4x1 ) = k(x1 + 4x1 )

u
(x1 + 4x1 )4x2 4x3 4t.
x1

(2.34)

Likewise, the amount of heat leaving the cube through the side x1 during the time interval [t, t +
4t] is
Q1 (x1 ) = k(x1 )

u
(x1 )4x2 4x3 4t.
x1

(2.35)

The amount of heat 4Q1 = Q1 (x1 + 4x1 ) Q1 (x1 ) conducted in and out of the cube in the x1
direction is


u
u
4Q1 = k(x1 + 4x1 )
(x1 + 4x1 ) k(x1 )
(x1 ) 4x2 4x3 4t.
(2.36)
x1
x1
In a similar manner, one can derive expressions for 4Q2 and 4Q3 , the amounts of heat conducted
in and out of the cube in the x2 and x3 directions, respectively.
Let us assume that inside the cube there are sources of heat, whose density if f (x, t). Then
the total amount of heat generated inside the cube during the time interval [t, t + 4t] is
Qg = f 4x1 4x2 4x3 4t.

(2.37)

The conservation of energy in Figure 2.5 can now be written as 4H = 4Q1 + 4Q2 +
4Q3 + Qg , or


u(t + 4t) u(t)
1
u
u

=
k(x1 + 4x1 )
(x1 + 4x1 ) k(x1 )
(x1 )
4t
4x1
x1
x1


1
u
u
+
k(x2 + 4x2 )
(x2 + 4x2 ) k(x2 )
(x2 )
4x2
x2
x2


1
u
u
+
k(x3 + 4x3 )
(x3 + 4x3 ) k(x3 )
(x3 ) + f. (2.38)
4x3
x3
x3
Taking the limit as 4xi 0 (i = 1, 2, 3) and 4t 0 yields a diffusion equation


3
u X
u

=
k
+ f,
x .
t
xi
xi

(2.39)

i=1

If a heated body is homogeneous, i.e., , , k = const, then (2.39) can be rewritten as


u
= c2 4u + g,
t

x ,

(2.40)

where c2 = k/, g = f /, and 4 = 2 is the Laplacian.

2.4.6

Initial and boundary conditions

Since diffusion equation (2.40) contains only first time derivative, initial conditions consist of only
one equation that specifies the initial temperature distribution (x), i.e,
u(x, 0) = (x),
- 12 -

x .

(2.41)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Boundary conditions for (2.39) or (2.40) are defined on the surface of a heated body. For
example,
If temperature is prescribed on the bounding surface,
u(x, t) = (x, t),

x ,

t 0;

(2.42)

If the surface of a body is insulated, i.e., has a prescribed heat flow v(x, t),
ku(x, t) n(x) = v(x, t),

x ,

t 0,

(2.43)

where n is the unit outward normal vector for the surface ;


If the surface is perfectly insulated, v = 0.
Conditions (2.42) and (2.43) are called the Dirichlet and Neumann boundary conditions, respectively.

2.4.7

Stationary temperature fields

If the boundary conditions and sources of thermal energy do not vary with time, temperature
will eventually reach a steady-state regime, i.e., the regime in which u = u(x). In steady state,
equation (2.39)


3
X

u
k
+ f = 0,
x .
(2.44)
xi
xi
i=1

If k = const, this gives Poissons equation


f
= 0,
x .
k
In the absence of sources of heat (f = 0), we arrive at Laplaces equation
4u +

4u = 0,
which lies at the foundation of the potential theory.

2.5
2.5.1

Continuity equations for fluids


Physical phenomena

flow of fluids

2.5.2

Construction of a physical model

single-phase flow
isothermal conditions,
etc.
- 13 -

x .

(2.45)

(2.46)

MAE 294c. Methods in Applied Mechanics III

2.5.3

Instructor: Daniel M. Tartakovsky

The choice of functions describing the process

The process is described by


the density of a fluid (x, t)
pressure in a fluid p(x, t)
the velocity of a fluid v(x, t)

2.5.4

Derivation of governing equations

Consider a volume occupied by a moving fluid. At time t, the mass of a fluid in this volume is
Z
M (t) =
(x, t)dx.
(2.47)

The change in mass 4M = M (t2 ) M (t1 ) from time t1 to t2 is


Z

Z Z

4M =

t2

[(x, t2 ) (x, t1 )] dx =

t1

dtdx =
t

t2

t1

dxdt.
t

(2.48)

The amount of a fluid leaving the volume through its bounding surface during the time
interval [t1 , t2 ] is
Z

t2

4Q =

t2

(v) ndsdt =
t1

(v)dxdt.
t1

(2.49)

Here n is the unit outward normal vector to the surface .


If sources of a fluid with density f (x, t) are present inside the volume , the amount of fluid
generated during the time interval [t1 , t2 ] is
Z

t2

4Qg =

f (x, t)dxdt.
t1

(2.50)

Since the mass conservation implies that 4M = 4Q + 4Qg , we obtain



Z t2 Z 

+ (v) f dxdt = 0.
t1
t

(2.51)

Since (2.51) holds for an arbitrary t1 , t2 , and , it gives a continuity equation

+ (v) f = 0.
t

(2.52)

For incompressible fluids ( = const), and in the absence of sources (f = 0), (2.52) gives
v = 0.

- 14 -

(2.53)

Chapter 3

Classification of PDEs
3.1

Basic terminology

Consider a function u(x1 , . . . , xn ). An equation


F

u
u k1 +...+kn u
x1 , . . . , xn , u,
,...,
,
x1
xn xk11 . . . xknn

!
=0

(3.1)

is called a partial differential equation (PDE) if n > 1. Otherwise, it is called an ordinary differential equation (ODE).
A function u? (x1 , . . . , xn ) that satisfies (3.1), i.e., turns it into identity, is called a solution.
The order of the PDE (3.1) is determined by the order of the highest derivative.
The dimensionality of the PDE (3.1) is determined by the number n of independent variables
x1 , . . . , xn . When one of these variables is time t, it is common to define the dimensionality as
n 1.
Equation (3.1) is called linear if the corresponding differential operator L(u) is linear, i.e.,
has the following property,
L(c1 u1 + c2 u2 ) = c1 L(u1 ) + c2 L(u2 ),

(3.2)

where u1 (x1 , . . . , xn ) and u2 (x1 , . . . , xn ) are any two functions and c1 and c2 are arbitrary constants. For example, a differential equation
u
2u
=
+ u + f (x, t)
t
x2

(3.3)

can be written as
L(u) = f (x, t),

L(u) =
15

u 2 u
2 u.
t
x

(3.4)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

It is easy to check that L in (3.4) satisfies the condition (3.2). Hence, equation (3.3) is a linear
PDE. If the condition (3.2) does not hold, the corresponding PDE is called nonlinear. An example
of nonlinear PDEs is
2u
u
=
+ u2 + f (x, t).
t
x2

(3.5)

To distinguish between linear and nonlinear differential operators, it is common to denote the
former by L and the latter by N . For example, (3.5) can be written as
N (u) = f (x, t),

N (u) =

u 2 u
2 u2 .
t
x

(3.6)

Equation (3.1) is called a homogeneous PDE if u = 0 is its solution. Otherwise, it is called


a nonhomogeneous PDE. It is easy to check that both (3.4) and (3.6) are homogeneous linear and
nonlinear PDEs, respectively, if and only if f = 0.

3.2

Well-posed problems

Consider a function u = f (z). Let z1 , z2 M and u1 = f (z1 ), u2 = f (z2 ) N , where M and


N are two metric spaces with the corresponding metrics M (z1 , z2 ) and N (u1 , u2 ).
A problem is called stable on the spaces M and N if for any  > 0 there exists = () > 0
such that M (z1 , z2 ) <  implies N (u1 , u2 ) < .
A problem is well-posed [Jacques Solomon Hadamard, 1902] on spaces M and N , if
1. a solution to the problem exists,
2. it is unique, and
3. it is stable.
Otherwise, a problem is ill-posed.

3.3

Classification of second-order PDEs

The significance of the wave (2.16) and diffusion (2.40) equations goes beyond their fundamental
importance for practical applications. In this section we demonstrate that they represent examples
of canonical forms for a general class of second-order PDEs.
As an example, consider an equation
2u
= f.
xy
- 16 -

(3.7)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The change of variables = x y and = x + y, transforms this equation into


2u
2u
=
+ f,
2

(3.8)

which, of course, is the wave equation (2.16).


Let u(x) be a system state that depends on x Rn , where n 2. Consider a differential
equation
n
X

aij (x)

i,j=1

2u
+ F (x, u, u) = 0,
xi xj

x .

(3.9)

It is called a quasi-linear second-order differential equation, since it is linear with respect to the
highest (second) derivative. The coefficients aij (x) are continuous, i.e., aij (x) C(), on
Rn . Without the loss of generality, we further assume that aij = aji .
We wish to simplify (3.9) by transforming it to its canonical form.

3.3.1

Transformation of second-order PDEs to their canonical form

Consider a transformation of coordinates


= (x),

or

l = l (x),

l = 1, . . . , n

(3.10)

whose Jacobian
(1 , . . . , n )
6= 0.
(x1 , . . . , xn )

(3.11)

The latter condition guarantees the existence of the inverse transformation x = x().
Under this transformation, u(x) = u[x()] u(). Hence
n

X u k
u
=
xi
k xi

(3.12)

k=1

and
2u

=
xi xj xj
=

n
X

u
xi





n
n
X
u k X u
k

=
+
xj k xi
k xj xi
k=1

k=1

n
X
k l
u 2 k
+
.
k l xi xj
k xi xj

2u

k,l=1

(3.13)

k=1

Substituting (3.12) and (3.13) into (3.9) yields


n
X
k,l=1

a
kl ()

2u
+ F1 (, u, u) = 0,
k l
- 17 -

(3.14)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

where
a
kl () =

n
X

aij (x)

i,j=1

k l
xi xj

(3.15)

and
F1 (, u, u) = F (, u, u) +

n
X
u 2 k
.
k xi xj

(3.16)

k=1

We wish to find transformations (3.10), for which


a
kl = 0, 1.

(3.17)

In general such coordinate transformations exist not over the whole domain , but only at a fixed
point x0 , i.e., the are local rather than global. This point is transformed into 0 = (x0 ), and
(3.15) gives
a
kl ( 0 ) =

n
X

k 0 l 0
(x )
(x ).
xi
xj

aij (x0 )

i,j=1

(3.18)

Denote
a?ij = aij (x0 ),

ki =

k 0
(x ),
xi

and

lj =

l 0
(x ),
xj

(3.19)

so that (3.18) can be rewritten as a quadratic form


n
X

a
kl ( ) =

a?ij ki lj .

(3.20)

i,j=1

We can now make use of the existing theory1 for quadratic forms. In particular, it is proved
that for any real quadratic form in n variables,
Q(z) =

n
X

bij zi zj ,

(3.21)

i,j=1

there exists an orthogonal point-transformation which reduces it to the diagonal form,


Q(z) =

r
X

zi2

i=1

m
X

zi2 ,

m n,

(3.22)

i=r+1

i.e., that the coefficient a?ij become

a?ij = 0,

i 6= j

(3.23a)

a?ii
a?ii

1ir

(3.23b)

= 1,
= 1,

r < i m.

see, e.g., http://mathworld.wolfram.com/QuadraticForm.html and references therein

- 18 -

(3.23c)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Moreover, according to Sylvesters Inertia Law2 , when a quadratic form (3.21) in n variables is
reduced by a nonsingular linear transformation to the form (3.22), the number r of positive squares
appearing in the reduction is an invariant of the quadratic form (3.21) and does not depend on the
method of reduction.
Since the numbers r and m are solely properties of the quadratic form (3.18), and this
quadratic form derives directly from a quasi-linear PDE (3.9), they can be used to classify this
equation. Moreover, a transformation that recasts (3.20) as (3.23), also transforms the general
quasi-linear PDE (3.9) into its canonical form,
r
X
2u
i=1

3.3.2

i2

m
X
2u
+ F3 (, u, u) = 0,
i2
i=r+1

m n.

(3.24)

Classification of second-order PDEs

Equation (3.9) is called hyperbolic at a point x0 , if in (3.24)


m = n and
one of its terms has the sign opposite to the others, i.e., if r = 1 or r = n 1.
Equation (3.9) is called elliptic at a point x0 , if in (3.24)
m = n and
all its terms have the same sign, i.e., if r = 0 or r = n.
Equation (3.9) is called parabolic at a point x0 , if in (3.24)
m < n.
Equation (3.9) is called ultrahyperbolic at a point x0 , if in (3.24)
m = n and
a few terms have opposite signs, i.e., if 1 r n 1.
Examples:
A wave equation for u(x1 , . . . , xN , t),
N

X 2u
2u
2
=
a
,
t2
x2i

a2 = 1.

i=1

It has n = N + 1 variables, and m = N + 1 = n second derivatives, one of which has the


opposite sign. Hence the wave equation is hyperbolic.
2

http://mathworld.wolfram.com/SylvestersInertiaLaw.html

- 19 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

A diffusion equation for u(x1 , . . . , xN , t),


N

X 2u
u
= a2
,
t
x2i

a2 = 1.

i=1

It has n = N + 1 variables, and m = N < n second derivatives. Hence the diffusion


equation is parabolic.
Poissons equation for u(x1 , . . . , xN ),
a2

N
X
2u
i=1

x2i

a2 = 1.

+ f = 0,

It has n = N variables, and m = N = n second derivatives, all of which have the same
sign. Hence the diffusion equation is elliptic.

3.3.3

A curse of dimensionality

The transformation of coordinates (3.10) consists of n equations for n independent variables l


(l = 1, . . . , n). Let us count the number of conditions in (3.23), which these variables have to
satisfy. Equation (3.23a),
a?ij = 0,

i 6= j,

i, j = 1, . . . , n

contains

n(n 1)
2!
such conditions. Equations (3.23b) and (3.23c) can be combined into one equation
a?ii = i a?11 ,

i = 1,

i = 2, . . . , n

and consist of (n 1) conditions. Thus the total number of conditions contained in (3.23) is
n(n 1)
+ (n 1),
2!

(3.25)

which, for n 3, is always larger than the number n of independent variables l . In other words,
the problem does not have a solution in n 3 dimensions!
It does have a solution in n = 2 dimensions, which we explore below.

3.3.4

Classification of second-order PDEs in n = 2 dimensions

In n = 2 dimensions, equation (3.9) becomes


a11

2u
2u
2u
+
2a
+
a
+ F (x, u, u) = 0,
12
22
x1 x2
x21
x22
- 20 -

x = (x1 , x2 )T ,

(3.26)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

the transformation of variables (3.10) takes the form


= (x),

or

l = l (x),

l = 1, 2

(3.27)

with the Jacobian




(1 , 2 )

=
(x1 , x2 )

1
x1
2
x1

1
x2
2
x2



1 2
1 2

6= 0.
=
x1 x2 x2 x1

(3.28)

Without the loss of generality, we assume that a11 (x) 6= 0 in the region of interest.
Transformation of variables (3.27) recasts (3.26) as (3.14) and (3.15) with n = 2,
a
11 ()

2u
2u
2u
+
2
a
()
+ F1 (, u, u) = 0,
+
a

()
12
22
1 2
12
22

(3.29)

where



1 2
1 1
1 2
= a11
+ 2a12
+ a22
,
x1
x1 x2
x2


1 2
1 2
1 2
1 2
= a11
+ a12
+
+ a22
,
x1 x1
x1 x2 x2 x1
x2 x2




2 2
2 2
2 2
= a11
+ 2a12
+ a22
.
x1
x1 x2
x2


a
11
a
12
a
22

(3.30a)
(3.30b)
(3.30c)

We note that we can make a


11 = a
22 = 0 by defining the transformation 1 (x) and 2 (x) as
solutions of




v 2
v v
v 2
a11
+ 2a12
+ a22
= 0.
(3.31)
x1
x1 x2
x2
The quadratic form in the left hand side of (3.31) factors as
!
!
p
p
a12 a212 a11 a22 v
a12 + a212 a11 a22 v
v
v
= 0.

a11

a11
x2
a11
x2
x1
x1
(3.32)
A solution of this equation and its behavior are determined by the term a212 a11 a22 . The following
cases can occur.
Case 1: a212 a11 a22 > 0. The quasi-linear second-order PDE (3.26) is hyperbolic.
Since a11 6= 0, in this case (3.32) gives a system of first-order PDEs
p
a12 + a212 a11 a22 v
v

=0
x1
a11
x2

(3.33a)

and
a12
v

x1

a212 a11 a22 v


= 0.
a11
x2
- 21 -

(3.33b)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Indeed, if a transformation of coordinates 1 (x) and 2 (x) is defined as solutions of (3.33),


equation (3.29) becomes
2
a12 ()

2u
+ F1 (, u, u) = 0,
1 2

(3.34)

a11 a22 a212 1 2


6= 0,
a11
x2 x2

(3.35)

It is easy to show that


a
12 = 2

provided that 1 /x2 , 2 /x2 6= 0. Hence,


2u
+ F2 = 0,
1 2

F2 =

F1
,
2
a12

(3.36)

which, as we showed earlier, can be transformed into a wave equation.


Case 2: a212 a11 a22 = 0.

The quasi-linear second-order PDE (3.26) is parabolic.

In this case (3.32) becomes




v
a12 v
+
x1 a11 x2

2
= 0,

(3.37)

which gives rise to one first-order PDE,


v
a12 v
+
= 0.
x1 a11 x2

(3.38)

If the transformation 1 = 1 (x) is defined as its solution, it follows from (3.30) that a
11 = 0.
Then one can show that, regardless of the choice of a transformation 2 = 2 (x), the condition
a212 a11 a22 = 0 leads to a
12 = 0. Hence, (3.29) reduces to
2u
+ F2 = 0,
2
Case 3: a212 a11 a22 < 0.

F2 =

F1 (, u, u)
.
a
22

(3.39)

The quasi-linear second-order PDE (3.26) is elliptic.

This case requires complex analyses to bring (3.26) to its canonical form. We only note that
if the coefficients aij in (3.26) are analytical functions, then it is possible to reduce (3.26) to
2u
= G (, u, u) .
1 2

(3.40)

To eliminate the complex variables, we now use the transformation of coordinates = 1 + i2


and = 1 i2 to obtain the canonical form for elliptic equations


2u 2u
u u
+
= G1 , , u,
,
.
(3.41)
2
2

- 22 -

Chapter 4

Method of Characteristics
In the previous Chapter we demonstrated that any second-order quasi-linear PDE with variable
coefficients (3.9) can be transformed into much simpler canonical forms (3.24). In two dimensions (n = 2), we also identified a set of first-order PDEs, either (3.33) or (3.38), which the
corresponding transformations of coordinates must satisfy. In this Chapter, we use the method of
characteristics to solve these and similar first-order PDEs.1

4.1

Linear first-order PDEs

Both (3.33) and (3.38) can be rewritten as a first-order wave equation


v
v
+c
= 0,
t
x

(4.1)

where we relabeled the coordinates, t x1 and x x2 . This equation describes the behavior of
a function v(x, t) in the Eulerian coordinate system (x, t).2
The method of characteristics is based on recasting the process described by (4.1) in the
Lagrangian (material) framework, in which a process v is measured by a moving observer, x =
x(t). While in an Eulerian coordinate system v = v(x, t), in a Lagrangian coordinate system
v = v[x(t), t]. In the Lagrangian framework, the rate of change of v = v[x(t), t] is described by
the total (also known as substantial or convective) derivative,
dv
v dx v
=
+
.
dt
t
dt x

(4.2)

Here dx/dt is the velocity of a moving observer.


The comparison of (4.1) and (4.2) shows that if the observer moves with velocity c, i.e., if
dx
= c,
dt

(4.3)

1
Much of our presentation of the method of characteristics borrows heavily from R. Haberman, Applied Differential
Equations, 4th Ed., Prentice Hall, 2004.
2
The Eulerian (field) description of a process v relies on a fixed coordinate system (frame of reference), in which an
immobile observer takes measurements of v(x, t).

23

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

then the PDE (4.1) can be replaced with an ODE


dv
= 0.
dt

(4.4)

In other words, (4.3) defines a family of curves x(t), along which an original PDE reduces to an
ODE. Such curves are called characteristics.

4.1.1

Constant wave velocity

For a constant c, (4.3) gives a family of linear characteristics


x = ct + ,

(4.5)

where is a constant of integration, such that x = at t = 0. It follows from (4.4) that v(x, t) is
constant along characteristics, i.e., v propagates as a wave with velocity c along the characteristics
(4.5). The shape of this wave is determined by the initial condition.
Consider the initial condition
v(x, 0) = V (x).

(4.6)

Along the characteristics (4.5), x = at t = 0 so that (4.6) gives v(, 0) = V (). Since along the
characteristics (4.5) v is constant,
v(x, t) = v(x, 0) = V ().

(4.7)

At any point (x, t), the parameter can be determined from (4.5),
= x ct.

(4.8)

Substituting (4.8) into (4.7) yields the solution of the PDE (4.1) subject to the initial condition
(4.8)3 ,
v(x, t) = V (x ct).

4.1.2

(4.9)

Variable wave velocity

If the wave velocity is c = c(x, t), it follows from (4.3) that the characteristics are no longer linear.
The procedure outlined above is generalized by using the following recipe:
1. define a characteristic coordinate = (x, t) by
dx
= c(x, t),
dt

x(0) = ;

(4.10)

2. find v(, t) by integrating the ODE (4.4) subject to the initial condition (4.6);
3

Note that any function F (x ct) satisfies the PDE (4.1). Hence v(x, t) = F (x ct) is the general solution of
(4.1).

- 24 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

3. eliminate in favor of x and t.


Consider, for example,
c(x, t) = (x + t)2 .

(4.11)

The ODE (4.3) becomes


dx
= (x + t)2 ,
dt

x(0) = .

(4.12)

The substitution w = x + t gives


dw
= w2 + 1,
dt

(4.13)

whose solution defines a family of characteristics


arctan(x + t) = t + arctan .

(4.14)

Writing the initial condition (4.6) along the characteristics (4.14) gives v(x, 0) = v(, 0)
= V (). Since along the characteristics (4.14) v is constant,
v(x, t) = v(, 0) = V ().

(4.15)

Substituting from (4.14) into (4.15) yields the final solution


v(x, t) = V {tan[arctan(x + t) t]}.

4.2

(4.16)

Classification of second-order PDEs in two dimensions (Cntd.)

We are now ready to revisit the problem of classification of second-order quasi-linear PDEs in two
dimensions. As an example, consider a PDE
x2

2u
2u

x
= 0,
1
x21
x22

x1 > 0,

x2 > 0.

(4.17)

Using the notation of Section 3.3, we have a11 = x2 , a12 = 0, a22 = x1 . Hence a212 a11 a22 =
x1 x2 > 0, which identifies (4.17) as hyperbolic.
To determine a coordinate transformation i = i (x1 , x2 ) (i = 1, 2) that transforms (4.17)
into its canonical form, we have to solve (3.33), which now become
r
r
v
v
x1 v
x1 v

=0
and
+
= 0.
(4.18)
x1
x2 x2
x1
x2 x2
Applying the method of characteristics to the first of these equations gives
r
dx2
x1
=
,
x2 (0) = ,
dx1
x2
- 25 -

(4.19)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

whose solution defines a family of characteristics




3/2
3/2 2/3
= x1 + x2
.

(4.20)

Since along these characteristics v is constant, any function F of solves the first PDE in (4.18),
3/2
3/2
i.e., v = F (x1 + x2 ) is its general solution. Any choice of F will provide an appropriate
transformation of coordinates 1 = 1 (x1 , x2 ). Let us select
3/2

1 (x1 , x2 ) = x1

3/2

+ x2 .

(4.21)

Similarly, applying the method of characteristics to the second equation in (4.18) gives a transformation of coordinates 2 = 2 (x1 , x2 ),
3/2

2 (x1 , x2 ) = x2
The inverse transformation is


1 2 2/3
,
x1 =
2

3/2

x1 .


x1 =

1 + 2
2

(4.22)

2/3
.

(4.23)

Note that the inverse transformation is invariant with respect to the choice of F . Substituting
(4.23) into (4.17) transforms the latter into


2u
1
2 u
2 u
= 3/2
2
1
.
(4.24)
3/2
1 2
1
2
2(1 2 )
As we showed in the beginning of Section 3.3, this equation is readily transformed into a secondorder wave equation under the transformation of coordinates x = 1 2 and t = 1 + 2 .
The importance of first-order PDEs and the method of characteristics for solving them goes
way beyond their use in classification of second-order quasi-linear PDEs. It stems from a plethora
of physical phenomena described by first-order PDEs, which include the immiscible displacement
of one fluid by another in a porous medium, traffic flow, shock waves, and contaminant transport
in rivers. Some of these applications are discussed and analyzed below.

4.3

Linear advection with a source

Consider the following problem. A contaminant has been released into a river, whose flow velocity
is c. Determine the concentration u(x, t) at a location x downstream from the place of release. We
assume that the river flows fast enough to disregard the diffusive effects on the contaminants concentration, i.e., that the contaminant transport is due to advection only. This problem is described
by a first-order PDE
u
u
+c
= f (x, t),
t
x

u(x, 0) = 0.

(4.25)

The initial condition implies that the river was free of the contaminant before the release. The
source function f (x, t) specifies the strength, location, time, and duration of the release.
- 26 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The presence of a source function f necessitates only a slight modification to the method of
characteristics recipe described in Section 4.1. Specifically, in Step 2 of the recipe the ODE (4.4)
should be replaced with an ODE
du
= f [x(, t), t].
dt

(4.26)

Following this modified recipe we obtain an ODE for the characteristics,


dx
=c
dt

x = ct + .

(4.27)

Hence (4.26) becomes


du
= f (ct + , t).
dt

(4.28)

Recalling that is a constant on a characteristic, and using the initial condition (4.25), gives
Z

u=

f (ct0 + , t0 )dt0 .

(4.29)

The final step of the recipe is to eliminate = x ct in favor of x and t,


Z
u(x, t) =

f [x + c(t0 t), t0 ]dt0 .

(4.30)

4.4

Linear advection and reactions

Here we generalize the contaminant transport problem in Section 4.3 in two important ways. First,
we allow the rivers velocity c to vary in space x and time t, i.e., consider c = c(x, t). Then the
transport equation (4.25) takes the form
u cu
+
= f (x, t),
t
x

u(x, 0) = uin (x).

(4.31)

This equation can be rewritten as


u
u
+ c(x, t)
= a(x, t)u + b(x, t),
t
x

u(x, 0) = uin (x).

(4.32)

where a cx and b f .
Second, we allow the contaminant to undergo linear chemical reactions, which adds an extra
term to (4.31),
u cu
+
= (u ueq ) + f (x, t),
t
x

u(x, 0) = uin (x),

(4.33)

where ueq is the equilibrium concentration and is the reaction rate constant. This, of course, is
the same as (4.32) with a = cx and b = f + ueq .
- 27 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Since the right-hand side of the PDE in (4.32) is now a function of u, the method of characteristics recipe described in Section 4.1 once again needs a slight modification. This time, Step 2
of the recipe consists of solving an ODE
du
= a[x(t), t]u + b[x(t), t],
dt

u(0) = uin ().

(4.34)

As an example, we set c(x, t) ex and uin ex in (4.31). This results in (4.32) with
a = ex , b = 0, and c = ex . To solve the resulting PDE, we follow the recipe. Step 1 is to solve
an ODE for the characteristics,
dx
= ex ,
dt

x(0) = .

(4.35)

Integrating this equation gives a family of characteristics


ex e = t

= ln (ex t) .

(4.36)

Step 2 is to solve the ODE (4.34), which now becomes


du
= ex u,
dt

u(0) = uin ().

(4.37)

Substituting an expression for ex from (4.36) and integrating along the characteristics yields
Z

uin ()

du0
=
u0

Z
0

dt0
t0 + e

u(, t) = uin ()

t + e
.
e

(4.38)

Step 3 is to eliminate in favor of x and t. Substituting from (4.36) into (4.38) yields
u(x, t) = uin [ln(ex t)]

ex
.
ex t

(4.39)

For uin () = e , this gives a solution for the contaminant distribution in a river,


u(x, t) = exp ln2 (ex t)

4.5

ex
.
ex t

(4.40)

Quasi-linear first-order PDEs

The approach we used in the previous section can be readily generalized to account for nonlinear
chemical reactions. Such a generalization replaces the reactive term in (4.33) with a reaction term
r(u)
u cu
+
= r(u) + f (x, t),
t
x

u(x, 0) = uin (x).

(4.41)

To be specific, we consider chemical reactions described by r(u) = (u ueq ) and set = 2.


- 28 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Equation (4.41) can be recast in the form of a general quasi-linear first-order PDE,
u
u
+ c(x, t)
= g(x, t, u),
t
x

u(x, 0) = uin (x),

(4.42)

where g = cx (x, t)u(x, t) + r(u) + f (x, t). We use the method of characteristics to solve the
quasi-linear PDE (4.42) with constant velocity c and without sources f = 0.
Step 1 of the recipe is to find the characteristics by solving
dx
= c,
dt

x(0) =

= x ct.

(4.43)

Step 2 of the recipe is modified to solve an ODE


du
= r(u),
dt

u(0) = uin [x(, t)].

Recalling that r(u) = 2(u2 u2eq ) and integrating along the characteristics gives
Z u
u ueq
uin () ueq 4t
du
= 2t
=
=
e
.
2
2
u + ueq
uin () + ueq
uin u ueq

(4.44)

(4.45)

Finally, we use Step 3 to eliminate by substituting (4.43) into (4.45),


u ueq
uin (x ct) ueq 4t
=
e
.
u + ueq
uin (x ct) + ueq

4.6

(4.46)

Nonlinear first-order PDEs

Consider a nonlinear (often referred to as quasi-linear) first-order PDE,


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

u(x, 0) = uin (x),

(4.47)

which describes, for example, traffic flow under an assumption that the car velocity v depends only
on the density u(x, t) of cars on the road. The coefficient c(u) in (4.47) is called the characteristic
velocity of traffic flow.
The presence of nonlinearity c(u) does not affect the fundamental idea of the method of
characteristics, which is to reduce first-order PDEs to first-order ODEs by writing the former in
the Lagrangian framework. Thus, as before, we have
dx
= c(u),
dt

x(0) =

(4.48)

and
du
= 0,
dt

u(0) = uin ().

(4.49)

However, one cannot obtain a family of characteristics from (4.48) alone, since its right-hand side
depends on the unknown u. In other words, now ODEs (4.48) and (4.49) are fully coupled.
- 29 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The fact that the right-hand side of the ODE in (4.49) is zero significantly simplifies the
analysis, since it implies that u is constant along characteristics,
u = uin ().

(4.50)

Hence we can integrate (4.49) along characteristics to obtain


x = c(u)t.

(4.51)

Thus, as in Section 4.1, the characteristics are straight lines. The difference is that now these
straight lines are not parallel to each other.
Finally, using (4.51) to eliminate in favor of x and t in (4.50) we obtain
u(x, t) = uin [x c(u)t].

Fanlike characterisitics. Consider the following initial value problem,


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

(4.52)

(4.53)

Using the general forms of its characteristics (4.51) and its solution (4.52), we obtain
x = 2u(, 0)t

(4.54)

and
(
1
u(x, t) =
2

x 2ut < 0
x 2ut > 0,

respectively. The latter can now be solved analytically,


(
1 x < 2t
u(x, t) =
2 x > 4t,

(4.55)

(4.56)

Since the distance between u = 1 and u = 2 increases, this solution is called an expansion wave.
It remains to determine the behavior of u(x, t) in the region 4t < x < 2t. The reason for this
gap in our solution is the discontinuity of the initial condition in (4.53). We imagine that all values
of u between 1 and 2 are present initially at x = 0. Straight line characteristics corresponding to
each of these values start at the point x = 0 and t = 0. Since at this point = 0, it follows from
(4.54) that the equation for these characteristics is
x = 2ut,

1 < u < 2.

(4.57)

This equation gives u(x, t) for the wedge-shaped region


u=

x
,
2t

2t < x < 4t.


- 30 -

(4.58)

MAE 294c. Methods in Applied Mechanics III

4.6.1

Instructor: Daniel M. Tartakovsky

Elementary traffic model

As an example, we consider an elementary traffic model, which replaces the function c(u) in (4.47)
with its linear counterpart. This is accomplished as follows. We postulate that cars move with the
maximum velocity v = vmax at u = 0 and practically stop moving after the density of cars reaches
some u = umax . A linear relationship that satisfies these conditions is


u
.
(4.59)
v(u) = vmax 1
umax
Since c(u) (uv)u , this gives

c(u) = vmax

2u
1
umax


,

which, when combined with (4.47), gives an elementary traffic equation




u
2u
u
+ vmax 1
= 0.
t
umax x

(4.60)

(4.61)

The initial condition u(x, 0) = uin (x) is determined by the traffic scenario we are interested in.
Consider a situation when the traffic light switches from red to green. Behind a red light (x = 0),
the traffic density is maximum u = umax , while ahead of the light it is u = 0, so that at the
moment t = 0 when the light turns green
(
umax ,
x<0
u(x, 0) = uin (x) =
(4.62)
0,
x > 0.
For c(u) and uin given by (4.60) and (4.62), respectively, the general solution (4.52) becomes




x vmax 1 u2u
t<0
umax ,
max
u(x, t) =
(4.63)



2u
0,
x vmax 1 umax t > 0.
This gives a final expression for the density of cars in traffic,
(
umax ,
x < vmax t
u(x, t) =
0,
x > vmax t.

(4.64)

The comparison of (4.62) and (4.64) shows that there is a delay between the time the light switches
to green and the time cars start moving. It is because the information propagates backward at the
speed umax .

4.7

Shock waves

For nonlinear (quasilinear) first-order wave equation (4.47),


u
u
+ c(u)
= 0,
t
x
- 31 -

u(x, 0) = uin (x),

(4.65)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

t
c(u1 ) < c(u2 )

c(u1 ) > c(u2 )

x
1

Figure 4.1: Linear characteristics (4.51) for nonlinear (quasilinear) first-order wave equation
(4.65).
the method of characteristics will not alway work in a way described in the previous section.
Specifically, it fails when the characteristics intersect. Figure 4.1 shows two possible scenarios
for the behavior of characteristics (4.51). Two characteristicsone starting at x = 1 with u1 =
uin (1 ) and the other starting at x = 2 with u2 = uin (2 )intersect if the faster one catches up
with the slower one, i.e., if c(u1 ) > c(u2 ). In this case, which is know as a compression wave, the
distance between the densities u1 and u2 decreases with time, and eventually they intersect, i.e.,
at some point x two densities coexist. In other words, density becomes a multi-valued function of
space, which in many cases is not physically realizable.

4.7.1

Discontinuous solutions

What is the reason for arriving at this nonphysical behavior (e.g., multi-valued density)? Since our
mathematical treatment of (4.65) did not use any approximations, the source of error must be in
the mathematical formulation of a physical process, i.e., in the formulation of (4.65).
All our analysis so far implicitly assumed that u(x, t) is a continuous function. This assumption will now be overturned by postulating that at some point xs the function u(x, t) has a
jump discontinuity, also called a shock wave, that propagates in time, i.e., xs (t), with the velocity
dxs (t)/dt. Schematic representation of a jump discontinuity of u(x, t) at point xs is shown in
+
Figure 4.2, where x
s and xs indicate that the point xs is being approached from the left and the
right, respectively. At the moment, the point of discontinuity xs , its trajectory xs (t), and velocity
dxs (t)/dt are all unknown.
We start by reformulating our mathematical model. Consider the conservation law for a
segment 0 < x < b of the road, which now includes a jump discontinuity 0 < xs (t) < b,
d
dt

u(x, t)dx = Mass IN Mass OUT

(4.66)

where Mass IN = q(a, t), Mass OUT = q(b, t), q = vu, and v(u) is the velocity of a substance
whose density is u. Accounting for the shock, this equation can be rewritten as
d
dt

"Z

xs (t)

u(x, t)dx = q(a, t) q(b, t)

u(x, t)dx +
a

xs (t)

- 32 -

(4.67)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

u(x, t)

u(x
s , t)

u(x+
s , t)

x
xs (t)

Figure 4.2: Jump discontinuity of u(x, t) at a point xs (t).


Using Leibnitzs rule of differentiation, we get

dxs 
+
u(x
s , t) u(xs , t) +
dt

Z
a

xs (t)

u
dx +
t

xs (t)

u
dx = q(a, t) q(b, t).
t

(4.68)

Since the continuity equation


u q
+
=0
t
x

(4.69)

holds on both sides of the shock x = xs (t), the integrals become


Z
a

xs (t)

u
dx +
t

xs (t)

u
dx =
t

xs (t)

q
dx
x

xs (t)

q
dx
x

+
= q(x
s , t) + q(a, t) q(b, t) + q(xs , t)

(4.70)

and (4.68) yields an expression for the shock velocity


+
q(x
dxs
{q}
s , t) q(xs , t)
=
=
,

+
dt
{u}
u(xs , t) u(xs , t)

(4.71)

where {A} denotes the magnitude of the jump of a function A. In gas dynamics, (4.71) is called
the Rankine-Hugoniot condition.
A useful property of the shock velocity is provided by the entropy condition,
c[u(x
s , t)] >

dxs
> c[u(x+
s , t)].
dt

(4.72)

This general property is tested below.

A shock and a compression wave.

Consider the following initial value problem


(
2 x<0
u
u
+ 2u
= 0,
u(x, 0) =
t
x
1 x > 0.
- 33 -

(4.73)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Following the recipe, we write an equation for characteristics,


dx
= 2u.
dt

(4.74)

Since u(x, t) is constant along characteristics, (4.74) gives a family of linear characteristics
x = 2u(, 0)t + .

(4.75)

Taking into account the initial condition in (4.73), one can see that the characteristics < 0 travel
twice as fast as the characteristics > 0. Hence the two sub-families will intersect, leading to
non-physical multi-valued solutions.
The problem is fixed by introducing a jump discontinuity that travels with the shock velocity
(4.71). It follows from the initial condition in (4.73) that u(x, t) = 2 on the left-hand side of the
shock and u(x, t) = 1 on the right-hand side of the shock. It remains to find the trajectory of the
shock xs (t) from (4.71).
To make use of (4.71), we note that q in the continuity equation (4.69) is related to c(u) in
(4.65), or to c(u) = 2u in (4.73), by
c(u)

q u
u
=
.
x
u x

(4.76)

Hence for u(x, t) in (4.73) to be a conservative quantity,


Z
Z
q
= c(u),
=
q = c(u)du = 2 udu = u2 .
u

(4.77)

Equation (4.71) becomes


2 +
dxs
u2 (x
s , t) u (xs , t)
+
=
= u(x
s , t) + u(xs , t) = 2 + 1 = 3,
+
dt
u(x
s , t) u(xs , t)

(4.78)

i.e., the velocity of the shock is constant. Since the discontinuity in u(x, 0) occurs at x = 0, we
find that the trajectory of the shock is
dxs
= 3,
dt

xs (0) = 0

xs (t) = 3t.

(4.79)

+
Since c(u) = 2u, we find that c[u(x
s )] = 4 and c[u(xs )] = 2. Hence the entropy condition
(4.72) holds.

4.7.2

Elementary traffic model with shocks

Let us return to the task of modeling traffic flow. Following our approach in Section 4.6.1, we
employ the elementary traffic model which results in (4.61),


2u
u
u
+ vmax 1
= 0.
(4.80)
t
umax x
- 34 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

However, now we are interested in the opposite traffic scenario, i.e., a situation when the traffic
light switches from green to red. Before the light turns red, the traffic density is uniform, u = u0 .
The traffic light is located at x = 0 and turns red at time t = 0. At this time, the traffic density
behind the light is uniform,
u(x, 0) = u0 ,

x < 0.

(4.81)

Since the cars stop at the light, the traffic density is maximum at x = 0,
u(0, t) = umax ,

t > 0.

(4.82)

This is the first time we have introduced a boundary condition for the first-order wave equation.
An equation for characteristics is


dx
2u
= vmax 1
,
dt
umax

x(0) = .

(4.83)

Thus characteristics are given by



x = vmax

2u
1
umax

They form two families of parallel curves




2u0
x = vmax 1
t
umax

and


t.

(4.84)

x = vmax t

(4.85)

that propagate the traffic densities u0 and umax . The characteristic velocity c1 of the first family
of characteristics (originating at points x < 0) is


2u0
c1 = vmax 1
,
(4.86)
umax
while that of the second family (originating at the point x = 0) is
c2 = vmax .

(4.87)

The characteristics from x = 0 move backward, since c2 < 0. The characteristics originating at
points x < 0 can move either backward or forward depending on the ratio 2u0 /umax . Regardless,
the latter move faster the the former, since c1 > c2 .
Thus a shock separating u0 and umax will form. Since it follows from (4.77) that the traffic
flow is



Z
Z 
2u
u
q(u) = c(u)du = vmax
1
du = vmax u 1
,
(4.88)
umax
umax
the shock velocity is determined from (4.71) as
+
q(x
q(u0 )
dxs
q(u0 ) q(umax )
s , t) q(xs , t)
=
=
=
.

+
dt
u

u
u
u(xs , t) u(xs , t)
0
max
0 umax

(4.89)

Since u0 < umax and q(u0 ) > 0, the shock velocity is negative, and the shock propagates backward. The corresponding shock trajectory is
xs =

q(u0 )
t.
u0 umax
- 35 -

(4.90)

MAE 294c. Methods in Applied Mechanics III

4.7.3

Instructor: Daniel M. Tartakovsky

Conditions for shock formation

We have established that shocks form when faster waves start behind slower waves, i.e., when the
characteristic velocity c(u) is a decreasing function of x,
c(u)
dc u
=
< 0.
x
du x

(4.91)

Thus for equations with dc/du < 0 shocks form when u(x, t) is an increasing function of x,
so that u must increase with x at a shock. Likewise, for equations with dc/du > 0, shocks
form when u decreases with x at a shock. If these conditions are not met, characteristics do not
intersect, shocks do not form, and discontinuous initial conditions correspond to expansion waves
(see Section 4.6.1).
If dc/du does not change signs, then discontinuous initial conditions result in either a shock
or an expansion wave.
If dc/du does change signs at least ones, then discontinuous initial conditions may result in
both a shock and an expansion wave.

A shock and an expansion wave. Consider the following initial value problem
(
1 x < 0
u
2 u
+u
= 0,
u(x, 0) =
t
x
2
x > 0.

(4.92)

The initial conditions indicate that u will increase with x, i.e., u/x > 0. On the other hand,
dc
= 2u
du

(4.93)

can be negative (leading to a shock) and positive (leading an expansion wave).


An equation for the characteristics is
dx
= u2 ,
dt

x(0) = ,

(4.94)

and an equation for u gives


du
= 0,
dt

(
1
u(, 0) =
2

<0
>0

(
1
u(, t) =
2

<0
> 0.

(4.95)

Since along the characteristics u is constant, (4.94) gives


x = u2 (, 0)t.

(4.96)

Eliminating in favor of x and t gives


(
1
u(x, t) =
2
- 36 -

x<t
x > 4t.

(4.97)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The fan-like characteristics for the region 4t < x < t are obtained as before (Section 4.6). These
characteristics correspond to the discontinuity in the initial condition in (4.92) and originate at the
point x = 0. Hence, (4.96) gives an equation for these characteristics
x = u2 t,

4t < x < t,

which leads to a fan-like family of characteristics


r
x
ufan =
,
1 < ufan < 2.
t
In accordance with (4.91), the part of this solution,
r
x
ufan =
,
0 ufan < 2,
t

(4.98)

(4.99)

(4.100)

combined with the positive part of the solution (4.97) represents the expansion wave.
Consider now the remaining characteristics < 0, along which u = 1. It follows from
(4.94) that their characteristic velocity is 1. They will intersect the slower moving characteristics
in the expansion wave, which correspond to 1 < u < 0. This leads
p to the formation of a shock,
which separates the region with u = 1 and the region with u = x/t.
Since
u3
,
3

(4.101)

dxs
1 1 + (xs /t)3/2
=
.
dt
3 1 + (xs /t)1/2

(4.102)

Z
q(u) =

c(u)du =

the shock velocity is obtained from (4.71) as

This ODE has to be solved numerically.

4.8

Nonlinear first-order PDEs with a source

Here we generalize the nonlinear (quasi-linear) first-order PDE in (4.47) by adding a state-dependent
right-hand side,
u
u
+ c(u)
= 2u,
t
x

u(x, 0) = uin (x),

(4.103)

and set c(u) u. This example loosely corresponds to traffic flow, in which the cars on a
highway are not conserved, but instead leave it at a rate proportional to the car density u.
Along the characteristics
dx
= u,
dt
- 37 -

x(0) = ,

(4.104)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

the density of cars is no longer constant, but is determined instead as a solution of an ODE
du
= 2u,
dt

u(, 0) = uin ().

(4.105)

The solution of (4.105) is


u = uin ()e2t .

(4.106)

Substituting it into (4.104) gives


x =


uin ()  2t
e
1 .
2

(4.107)

In general, the last step of eliminating in favor of x and t has to be done numerically.
Consider a simple initial condition uin (x) = x. Now (4.107) can be solved analytically to
give
=

2x
.
1 + e2t

(4.108)

Eliminating leads to the car density distribution


u=

2x
2x
e2t = 2t
.
1 + e2t
e +1

- 38 -

(4.109)

Chapter 5

Laplace Transformation
In this Chapter we introduce the Laplace transformation as another tool to solve PDEs by reducing
them to ODEs.

5.1

Motivation and Definition

Consider one-dimensional diffusion equation for u(x, t),


u
2u
= D 2,
a < x < b, t > 0.
(5.1)
t
x
It is subject to the initial condition u(x, 0) = uin (x) and some appropriate boundary conditions.
Multiplying this equation by est and integrating from t = 0 to t = gives
Z
Z
u st
2
e dt = D 2
uest dt.
(5.2)
t
x
0
0
The left-hand side of this equation can be integrated by parts to yield, after accounting for the
initial condition,
Z
Z
2
st
uin (x) + s
ue dt = D 2
uest dt.
(5.3)
x 0
0
Denoting
Z
u
(x, s) =
u(x, t)est dt,
(5.4)
0

we obtain an ODE
d2 u

,
dx2
where (real or imaginary) transformation constant s acts a parameter.
uin (x) + s
u=D

(5.5)

The operation on a function u in (5.4) is called the Laplace transformation, the new function u
is called the Laplace transform. Often it is convenient to use a shorthand notation, which
replaces the integral in (5.4), by writing
u
(x, s) = L{u(x, t)}.
39

(5.6)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

5.2 Properties of the Laplace transformation


Some of the basic properties of the Laplace transformation are discussed below.

5.2.1

Linearity

The Laplace transformation L{u} is linear, i.e., for a linear combination of two functions u1 and
u2 ,
L{c1 u1 + c2 u2 } = c1 L{u1 } + c2 L{u2 }.

5.2.2

(5.7)

Applicability

Piecewise continuous functions. A function u(t) is called piecewise continuous on an interval


a t b, if the interval can be subdivided into a finite number of sub-intervals, in each of
which u(t) is continuous and has finite limits at the ends of these sub-intervals. Such functions are
integrable on the interval a t b. As an example of piecewise continuous functions, consider
a unit step function
(
0 0<t<p
Sp (t) =
(5.8)
1 t > p.
Its Laplace transform exists and is given by
Z
Z
Sp (t)est dt =
0

1
est dt = eps ,
s

(5.9)

provided s > 0.
However, not all piecewise continuous functions have Laplace transforms. For instance, a
function
(
et 0 < t < p
fp (t) =
(5.10)
2
et t > p
does not have the Laplace transform because the integral
Z
Z p
Z
st
(1s)t
fp (t)e dt =
e
dt +
0

2 st

et

dt

(5.11)

does not exist.


Functions of exponential order. A function u(t) is of exponential order as t , or O et ,
if there exists such that
|u(t)|et <
- 40 -

(5.12)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

for all t larger than some finite number T . Clearly e2t satisfies this condition, while the function
fp (t) defined above does not.
Any function u(t) that is piecewise continuous and of exponential order as t has the
Laplace transform. This condition can be easily proved by noting that |est u(t)| < M e(s)t ,
where M is some constant. Since the upper bound of |est u(t)| is integrable provided s > , it is
integrable as well.
One has to keep
in mind that this condition is sufficient but not necessary. For example, the
function u(t) = 1/ t has an infinite discontinuity at t = 0, but still has the Laplace transform,
r
  Z
Z
1
1 st
2

x2
e dt =
L =
e dx =
,
s > 0.
(5.13)
s
s 0
t
t
0
Transforms of derivatives.

The passage from (5.2) to (5.3),


 
du
L
= sL{u} u(0),
dt

(5.14)

implicitly assumed that a solution of the diffusion equation u is of exponential order as t .


Other requirements on u, i.e., that u be continuous with a piecewise continuous derivative, are
typically implied by the fact that u is a solution of a differential equation.
For example,
t

Z
L

u(t )dt

The inverse Laplace transform.

1
= L{u}.
s

(5.15)

The inverse Laplace transformation is denoted by L1 . If


L{u(t)} = u
(s),

(5.16)

u(t) = L1 {
u(s)}.

(5.17)

then

The inverse Laplace transformation is a linear operation. Indeed, since


L{c1 u1 + c2 u2 } = c1 L{u1 } + c2 L{u2 }

(5.18)

L1 {c1 u
1 + c2 u
2 } = c1 u1 + c2 u2 = c1 L1 {
u1 } + c2 L1 {
u2 }.

(5.19)

we have

This property will be used again and again to compute the inverse Laplace transforms of solutions
of differential equations. A number of the pairs of direct and inverse transforms are tabulated.

A theorem of substitution. Let u(t) be such that its Laplace transform u


(s) exists when s > .
Then
u
(s a) = L{eat u(t)},
- 41 -

s > + a.

(5.20)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

A theorem of scaling. If u
(s) = L{u(t)} when s > , then
1 s
u

= L{u(at)},
s > a,
a
a

a > 0.

(5.21)

Consider a function u(t), and let ub (t) denote its translation (see

A theorem of translation.
Fig. 5.1),

(
0
0<t<b
ub (t) =
u(t b) t > b.

(5.22)

ebs u
(s) = L{ub (t)}.

(5.23)

Then

u(t)

ub (t)

Figure 5.1: The translation of a function u(t).


This theorem is critical for understanding convolutions.

Convolution.

The convolution u v of the functions u(t) and v(t) is defined as the function
Z t
u(t) v(t) =
u( )v(t )d,
0 t < .
(5.24)
0

A useful property of the convolution is that


u
(s)
v (s) = L{u(t) v(t)}.

(5.25)

To prove this, we multiply the Laplace transform of u(t) with v(s),


Z
u
(s)
v (s) =
u( )es v(s)d

(5.26)

Recalling the theorem of translation, we obtain


Z
Z
u
(s)
v (s) =
u( )L{v (t)}d =

est

u( )v (t)d dt.

(5.27)

Using the definition of v (t) in (5.22), this gives the final result
Z t

Z
Z t
st
u
(s)
v (s) =
e
u( )v(t )d dt L
u( )v(t )d .

(5.28)

- 42 -

MAE 294c. Methods in Applied Mechanics III

5.3

Instructor: Daniel M. Tartakovsky

Solutions of linear ODEs with constant coefficients

The application of the Laplace transformation to the solution of linear ODEs with constant coefficients is relatively straightforward.

5.3.1

Forced vibrations with damping

As an example, we find the solution of the differential equation


d2 u du

6u = 2
dt2
dt

(5.29)

subject to the initial conditions


u(0) = 1,

du
= 0.
dt

(5.30)

This problem describes forced vibrations with viscous damping.


Applying the Laplace transformation to the both sides of this equation gives an algebraic
equation
2
s2 u
s s
u + 1 6
u= .
s

(5.31)

Hence
2 + s2 s
.
s(s2 s 6)

(5.32)

11
8 1
4 1
2 + s2 s
=
+
+
s(s 3)(s + 2)
3 s 15 s 3 5 s + 2

(5.33)

u
=
By writing this as
u
=

and consulting a table of Laplace transforms, we obtain the solution


1
8
4
u(t) = + e3t + e2t .
3 15
5

5.3.2

(5.34)

A damped absorber

If a damped absorber is added to forced vibrations, the process is described by a coupled system
of ODEs, such as
d2 u d2 v dv
2 +
u = et 2,
dt2
dt
dt

(5.35)

du
d2 u d2 v
2 2
+ v = t.
2
dt
dt
dt

(5.36)

- 43 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

We assume the homogeneous initial conditions


u(0) =

du
dv
(0) = v(0) =
(0).
dt
dt

(5.37)

Applying the Laplace transformation to these ODEs gives a system of algebraic equations
s2 u
s2 v + s
vu
=

1
2
,
s1 s

(5.38)

1
.
s2

(5.39)

s2
,
s(s 1)2

(5.40)

1
.
1)

(5.41)

2s2 u
s2 v 2s
u + v =
These equations can be rewritten as
(s + 1)
u s
v=

2s
u (s + 1)
v=

s2 (s

The solutions of this system are


u
=

1
1
1
1
=
+
2
s(s 1)
s s 1 (s 1)2

(5.42)

and
v =

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

s2 (s

(5.43)

Consulting a table of Laplace transforms, we arrive at the solution of the problem,


u(t) = 1 et + tet

5.4

and

v(t) = t + tet .

(5.44)

Solutions of integral equations with convolution kernels

Consider a Fredholm integral equation of the second kind,


Z t
u(t) = at +
u(t) sin(t )d.

(5.45)

Since the integral on the right-hand side is a convolution, taking the Laplace transformation of this
equation gives an algebraic equation
u
(s) =

a
1
+u
(s) 2
,
2
s
s +1

(5.46)

whose solution is

u(s) = a

1
1
+ 4
2
s
s


.

Taking the inverse Laplace transform, we obtain the solution




6
1
u(t) = a
+ 3
t
t
- 44 -

(5.47)

(5.48)

MAE 294c. Methods in Applied Mechanics III

5.5

Instructor: Daniel M. Tartakovsky

Solutions of linear PDEs

In this section we use the Laplace transformation to solve linear PDEs, whose coefficients do not
vary with time.

5.5.1

Wave equation

Consider the wave equation,


2
2u
2 u
=
c
,
t2
x2

x > 0,

t>0

(5.49)

u
(x, 0) = 0
t

(5.50)

u(x , t) = 0.

(5.51)

subject to the initial conditions


u(x, 0) = (x),
and the boundary conditions
u(0, t) = 0,

This problem describes vertical displacements in a long string that has been initially displaced.
Accounting for the initial conditions,
Z 2
Z
u st
u st
e dt = s
e dt = s2 u
s(x),
2
t
t
0
0

(5.52)

so that the Laplace transformation of the wave equation leads to the ODE
s2 u
s(x) = c2

d2 u

,
dx2

x>0

(5.53)

subject to the boundary conditions


u
(x , s) = 0.

u
(0, s) = 0,

(5.54)

We shall solve this ODE by using the Laplace transformation with respect to x. Let
Z
u
(p, s) =
u
(x, s)epx dx.

(5.55)

Then taking the Laplace transformation of the ODE (5.53), while accounting for the boundary
conditions, yields the algebraic equation



s2 u
s(p)
= c2 p2 u
C(s) ,

C(s)

d
u
(0, s),
dx

(5.56)

whose solution is
u
=

c2 C s
.
c2 p2 s2
- 45 -

(5.57)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Taking the inverse Laplace transform with respect to p gives


(
)
sx
s 1

cC
sinh
2L
u
(x, s) =
.
s
c
c
p2 s2 /c2
Using the convolution to invert the last term in this expression gives
Z
sx 1 x
s
cC
sinh

(y) sinh (x y)dy.


u
(x, s) =
s
c
c 0
c
Now we can determine C(s) from the boundary condition at infinity,


Z
sx cC
1
sy/c
0 = exp

(y)e
dy ,
x .
c
s
c 0

(5.58)

(5.59)

(5.60)

Hence,
Z

cC
1
=
s
c

(y)esy/c dy.

(5.61)

Thus the Laplace transform of the solution is


Z
Z x
sx
s
sy/c
c
u(x, s) = sinh
(y)e
dy
(y) sinh (x y)dy.
c 0
c
0

(5.62)

To compute the inverse Laplace transform of this expression, we rewrite it as


Z
Z
Z x
s(x+y)/c
s(xy)/c
2c
u(x, s) =
(y)e
dy +
(y)e
dy +
(y)es(xy)/c dy. (5.63)
0

The change of variables


t=

x+y
,
c

t=

yx
,
c

t=

xy
,
c

in the first, second and third integrals, respectively, gives


Z
Z
Z
2
u(x, s) =
(ct x)est dt +
(ct + x)est dt +
x/c

x/c

(5.64)

(x ct)est dt.

(5.65)

The first and third integrals can be combined by introducing a function


(
(z)
z0
1 (z) =
(z) z 0,
which is called the odd extension of (z). This gives
Z
Z
st
2
u(x, s) =
(x + ct)e dt +
0

1 (x ct)est dt,

(5.66)

(5.67)

whose inverse Laplace transform is


2u(x, t) = (x + ct) + 1 (x ct).

(5.68)

Finally, since in the domain of interest x ct, (x ct) = 1 (x ct) so that


u(x, t) =

(x + ct) + (x ct)
.
2
- 46 -

(5.69)

MAE 294c. Methods in Applied Mechanics III

5.5.2

Instructor: Daniel M. Tartakovsky

Diffusion equation

Consider the diffusion equation,


u
2u
= D 2,
t
x

x > 0,

t>0

(5.70)

subject to the initial condition


u(x, 0) = 0

(5.71)

and the boundary conditions


k

u
(0, t) = q,
x

u(x , t) = 0.

(5.72)

This problem describes the distribution of temperature in a long rod when a constant flux of heat
q is maintained at the boundary x = 0.
The Laplace transform of this problem is
s
u=D

d2 u

,
dx2

x>0

(5.73)

subject to the boundary conditions


k

d
u
q
(0, s) = ,
dx
s

u
(x , s) = 0.

(5.74)

The solution of this problem is

q D xs/D
.
u
(x, s) = e
ks s
It can be inverted by means of a table of Laplace transforms.

- 47 -

(5.75)

MAE 294c. Methods in Applied Mechanics III

- 48 -

Instructor: Daniel M. Tartakovsky

Chapter 6

Nonhomogeneous Problems and


Greens Functions
The Greens functions are an ideal tool for solving linear nonhomogeneous PDEs with nonhomogeneous initial and boundary conditions.

6.1

Motivation and basic ideas

Consider a linear differential equation


c
= (Dc) vc + qc + f
t

t>0

(6.1)

which describes, among many other processes, contaminant transport in porous media. All coefficients in this equation are functions of the N -dimensional space x and time t. The coefficient
D is a positive semi-definite tensor whose components are Dij (x, t) (i, j = 1, . . . , N ), and v is a
vector with components vi (x, t) (i = 1, . . . , N ). Equation (6.1) is subject to the initial condition
c(x, 0) = cin (x)

(6.2)

and the boundary conditions


x D

c(x, t) = (x, t)

n Dc = (x, t)

n (Dc vc) = r(x, t)

x N

t > 0,

t > 0,

x R

t > 0,

(6.3)

(6.4)

(6.5)

where D , N , and R are the Dirichlet, Neumann, and Robin segments of the boundary =
D N R of . The boundary value problem (6.1) (6.5) is nonhomogeneous due to the
presence of the driving forces f , , , and r.
49

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

We start by multiplying (6.1) with as yet unspecified function G(x, t; y, ) and integrating it
in space and time,

Z tZ 
Z tZ
c
(Dc) + vc qc Gdyd =
f Gdyd.
(6.6)
0

0

Since
Z tZ
0

c
Gdyd =

Z tZ

[cG]t =0 dy

Z tZ

c
0

G
dyd,

Z tZ
(Dc)Gdyd =

Z tZ
n [GDc cDG] dyd +

(6.7)

c (DG)dyd,
0

(6.8)
and
Z tZ

Z tZ
G vc =

Z tZ
Gn vc

G vcdyd
0

(6.9)

equation (6.6) can be written as



Z tZ 
Z
G

c
+ (DG) + G v + qG dyd +
[cG]t =0 dy

Z tZ
Z tZ

n [GDc cDG Gvc] dyd =


f Gdyd.
0

Due to the initial and boundary conditions (6.2) (6.5), this gives
Z tZ
Z tZ
c(y, )(x y)(t )dyd =
f Gdyd
0

Z
Z tZ
Z tZ
Z tZ
+ cin G(x, y, t, 0)dy
n DGdyd +
Gdyd +

(6.10)

rGdyd.

(6.11)
where we chose the function G(x, y, t, ) to be the solution of the differential equation

G
= y (Dy G) + v y G + qG + (x y)(t )

(6.12)

subject to the homogeneous initial condition


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

(6.13)

and the homogeneous boundary conditions


G(x, y, t, ) = 0
n [DG vG] = 0

y D ,
y N ,

(6.14)
(6.15)

and
n DG = 0
- 50 -

y R .

(6.16)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Thus defined G(x, t; y, ) is called the Greens function for the boundary value problem (6.1)
(6.5).
You might recognize the function in (6.12) as the Dirac delta function. We will explore its
properties below. At this point, it suffices to say that it allows us to obtain from (6.11) the solution
of (6.1) (6.5),
Z tZ
Z
c(x, t) =
f Gdyd +
cin G(x, y, t, 0)dy
0

Z tZ
Z tZ
Z tZ

n DGdyd +
Gdyd +
rGdyd.
(6.17)
0

Note that the Greens function is the solution of the boundary value problem (6.12) (6.12), which
is similar to the original problem (6.1) (6.5). What have we gained? First, homogeneous problems are usually easier to solve than their nonhomogeneous counterparts. Second, the boundary
value problem for the Greens function is independent from external forces and, thus, describes
the internal structure of a system. Once determined (analytically or numerically), it can be used to
analyze the response of the system to applied external forces in accordance with (6.17).

6.2

The Dirac delta function

The one-dimensional Dirac delta function is defined as


(
0 x 6= x0
(x x0 ) =
x = x0 .

(6.18)

It is not a function in the usual sense of the word. Instead, it is an example of the so-called
generalized functions that are defined in relationship to other functions. Specifically, the definition
of the Dirac delta function (and its main property) is that for any continuous function f (x),
Z
f (x) =
f (x0 )(x x0 )dx0 .
(6.19)

An immediate consequence of this definition is that it has unit area,


Z
(x x0 )dx0 = 1.

(6.20)

Other important consequences are that the Dirac delta function is even,
(x x0 ) = (x0 x),

(6.21)

that its derivative is given by


df
=
dx

f (x0 )

d(x x0 ) 0
dx ,
dx0

(6.22)

that it is the derivative of the Heaviside step function,


dH(x x0 )
(x x ) =
,
dx
0

(
0
H(x x ) =
1
0

- 51 -

x x0
x > x0 ,

(6.23)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

and, finally, that it has the following scaling property:


[c(x x0 )] =

1
(x x0 ).
|c|

(6.24)

The best way to think of the Dirac delta function (x x0 ) is in terms of a source of infinite
strength concentrated at a point.
The properties of the one-dimensional Dirac delta function are easy to generalize to higher
dimensions by using its definition
Z
f (x) =
f (y)(x y)dy.
(6.25)

This gives
(x y) = (x1 y1 ) (xN yN ).

6.3
6.3.1

(6.26)

Properties of the Greens functions


Causality

The analysis of the boundary value problem (6.12) (6.12) shows that the Greens function
G(x, t; y, ) represents the response of the system at (x, t) to a perturbation introduced by the
infinite strength instanteneous point source located at (y, ). It becomes obvious from this interpretation that
G(x, t; y, ) = 0

for > t,

(6.27)

since the system does not know about perturbations that occur in the future. This property is called
a causality principle.

6.3.2

Reciprocity

Equations (6.1) and (6.12) can be written in the operator form as


L(c) = f (y, )

L=

y Dy + y v q

(6.28)

and
= y Dy v y q,
L

L(G)
= (x y)(t )

(6.29)

is called the adjoint of the operator L. The two are related via the
respectively. The operator L
Greens formula (6.10),
Z tZ
Z

[L(c)G L(G)c]dyd = [cG]t =0 dy


0

Z tZ

n [GDc cDG Gvc] dyd.


(6.30)
0

- 52 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Let us replace c with the Greens function G(z, t0 ; y, ), which satisfies


L[G(z, t0 ; y, )] = (z y)(t0 )

(6.31)

subject to the homogeneous initial and boundary conditions (6.2) (6.5). Then (6.30) becomes
Z tZ
Z tZ
0

L[G(z, t ; y, )] G(x, t; y, )dyd =


L[G(x,
t; y, )] G(z, t0 ; y, )dyd. (6.32)
0

Recalling (6.12) and (6.31) gives


Z tZ
Z tZ
0
(z y)(t ) G(x, t; y, )dyd =
(x y)(t ) G(z, t0 ; y, )dyd.
0

(6.33)
Using the property of the Dirac delta function, we obtain
G(x, t; z, t0 ) = G(z, t0 ; x, t),

(6.34)

which demonstrates that the Greens function is symmetric. This is the so-called principle of
reciprocity.

6.4

Calculation of the Greens functions

The Greens functions can be computed analytically for many linear differential equations defined
on simple domains. This is typically accomplished by either the method of separation of variables
or the method of images (see, e.g., Haberman). A large number of the Greens functions are
tabulated in e.g., Butkovsky. Here we use relatively simple examples to demonstrate key features
of the Greens functions.

Steady-state flow.

Consider steady-state flow in a porous column


d2 u
= f (x)
dx2

0 < x < L;

u(0) = ,

u(L) = .

(6.35)

The corresponding Greens function G(x, x0 ) satisfies


d2 G
= (x x0 ),
dx2

G(0, x0 ) = G(L, x0 ) = 0.

Since (x x0 ) = 0 for x 6= x0 , (6.36) leads to


(
a1 + a2 x
0
G(x, x ) =
a3 + a4 x

x < x0
x > x0 .

The boundary conditions in (6.36) determine the two constants of integration,


(
a2 x
x < x0
0
G(x, x ) =
a4 (x L)
x > x0 .
- 53 -

(6.36)

(6.37)

(6.38)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The remaining to constants of integration are determined from the following considerations. Integrating (6.36) over the infinitesimal interval x0 < x < x0+ , we obtain
dG 0
dG 0
(x+ )
(x ) = 1,
dx
dx

(6.39)

which shows that the derivative of the Greens function dG(x, x0 )/dx has a jump discontinuity at
x = x0 . The Greens function itself must be continuous at x = x0 . Otherwise the left-hand side of
(6.36) would be more singular than its right-hand side.
Applying the continuity condition to (6.38) gives
a2 = a4

x0 L
,
x0

(6.40)

so that
(
0
a4 x xL
0 x
0
G(x, x ) =
a4 (x L)

x < x0
x > x0 .

(6.41)

The jump condition (6.39) gives


a4 a4

x0 L
=1
x0

which yields the solution for the Greens function,


( 0
x L
L x
G(x, x0 ) = xL
0
L x

a4 =

x0
,
L

x < x0
x > x0 .

The solution of (6.35) is


Z L
dG
dG
0
u(x ) =
f (x)G(x, x0 )dx +
(L, x0 )
(0, x0 ).
dx
dx
0
Substituting (6.43), we obtain the final expression for u(x),
Z
Z L
yL
xL x
x
xL
u(x) =
f (y)ydy + x
f (y)
dy +
.
L
L
L
L
0
x
Transient flow.

(6.42)

(6.43)

(6.44)

(6.45)

Consider transient flow in a long (mathematically infinite) porous column,

u
2u
= K 2 + f (x, t),
t
x

u(x, 0) = uin (x),

|u(, t)| < .

(6.46)

The corresponding Greens function G(x, t; y, ) satisfies the adjoint differential operator (6.29)

2G
G
= K 2 + (x y)(t ),

(6.47)

subject to the homogeneous initial and boundary conditions. If we introduce T = t , this


equation becomes
G
2G
= K 2 + (x y)(T )
T
y
- 54 -

T > 0,

(6.48)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

that corresponds to the systems response at x to a perturbation induced by the concentrated infinite
strength source at point y at T = 0. In other words, the Greens function for the diffusion equation
is G(x, t; y, ) = G(x, t ; y, 0) or
G(x, t; y, ) = G(x, y, t ),

(6.49)

which is called the translation property. It is not a general property of all Greens functions.
We use Laplaces transformation to solve (6.48). Taking the Laplace transform gives the
ODE
=K
pG

d2 G
+ (x y),
dy 2

, T )| < .
|G(x,

(6.50)

whose general solution is

(
p/Ky
p/Ky
a
e
+
a
e
2
1
y, p) =

G(x,
a3 e p/Ky + a4 e p/Ky

y<x

(6.51)

y > x.

Using the conditions at infinity, we obtain


(
p/Ky
y, p) = a1 e
G(x,
a4 e p/Ky

y<x

(6.52)

y>x

The continuity of G at y = x leads to


(
p/Ky
y, p) = a1 e
G(x,
a1 e p/K(2xy)

y<x

(6.53)

y>x

The jump condition for the derivative takes the form

dG
1
dG
(x+ )
(x ) =
dy
dy
K

(6.54)

and results in
y, p) = 1
G(x,
2K

K p/K|xy|
.
e
p

(6.55)

According to the Table of Laplace transforms in Crank (Entree 7),


s
r
K x2 /4KT
K p/Kx
=
e
.
e
T
p

(6.56)

Since T = t , this gives the Greens function for the one-dimensional diffusion equation
1

G(x, y, t ) = p

4K(t )

e(xy)

The solution of (6.46) can now be written as


Z tZ
Z
u(x, t) =
f (y, )G(x, y, t )dyd +
0

(6.57)

- 55 -

2 /4K(t )

uin (y)G(x, y, t)dy.

(6.58)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

6.5 Free surface flow


Consider the motion of a fluid-fluid interface in a homogeneous porous medium T (bounded by
the surface T ) when gravity, capillary length, and the viscosity of one fluid are zero. In the inviscid fluid (air), the pressure is constant and may be set to zero. The viscous, incompressible fluid
(water), occupies the flow domain ( T ) which is bounded either entirely by a free surface
or by a combination of and some segments of T . Such flow is described by a combination of
Darcys law and mass conservation,
q(x, t) = Kh(x, t),

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

x (t),

(6.59)

subject to boundary conditions


x D ,

h(x, t) = H(x, t),


n(x) q(x, t) = Q(x, t),

x N ,

n(x, t) q(x, t) = Vn (x, t),

h(x, t) = 0,

(6.60a)

(6.60b)
x (t).

(6.60c)

The flux q is down gradients of hydraulic head h subject to constraints imposed by the hydraulic
conductivity K. The boundary = D N consists of Dirichlet segments, D , Neumann
segments, N , and a moving front, . At the free surface , h equals atmospheric pressure which
we set equal to zero without loss of generality; Vn (x, t) is the velocity of the moving boundary
in the normal direction. Mass conservation requires that
V n = ne

d
,
dt

(6.61)

where ne is the medium porosity.


Let us introduce the Greens function G(y, x) satisfying
K2y G(y, x) + (y x) = 0,

y, x T ,

(6.62)

subject to the boundary conditions


G(y, x) = 0,
n(y) y G(y, x) = 0,

y D ,
y N .

(6.63a)

(6.63b)

Note that our Greens function G(y, x) is defined for the entire domain T rather than just for the
flow domain , so that there are no conditions on G along the moving boundary . Specifying G
for just the flow domain, , would require recalculating G at each time as evolves, which is not
computationally efficient.
Rewriting (6.59) in terms of y, multiplying with G, and integrating over the entire domain
recasts (6.59)(6.60), after some mathematical manipulations,
Z
Z
h(x, t) =
f (y)G(y, x)dy + K n [G(y, x)y h(y) h(y)y G(y, x)] dy.
(6.64)

- 56 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Taking into account boundary conditions (6.60a) (6.60c), we obtain the integro-differential equation
Z
Z
h(x, t) = f (y, t)G(y, x)dy
Q(y, t) G(y, x)dy

N
Z
Z
K
H(y, t)n y G(y, x)dy Vn (y, t)G(y, x)dy.
(6.65)
D

This equation serves as a foundation for one of the so-called front tracking numerical methods.

- 57 -

MAE 294c. Methods in Applied Mechanics III

- 58 -

Instructor: Daniel M. Tartakovsky

Chapter 7

Nonlinear Diffusion
The standard diffusion equation
u
= q,
q = Du
(7.1)
t
assumes that materials properties, such as the diffusion coefficient D (or thermal conductivity,
electrical resistivity, hydraulic conductivity, etc.), are constant. This is an approximation, since
one should expect the diffusion coefficient to vary with the concentration u(x, t) of diffusing substance. This approximation works well at low concentrations, but often fails when concentrations
become large (as is the case with the diffusion of vapors in high-polymer substances1 ). Other
examples of varying diffusion coefficients include the dependence of electrical resistivity on temperature and the dependence of the hydraulic conductivity of a porous medium on saturation. In
these and many other applications, D = D(u) and the diffusion equation (7.1) becomes nonlinear.
In this chapter we outline a few mathematical approaches for analytical analysis of the nonlinear diffusion equation (7.1).

7.1

Time-dependent diffusion coefficients1

We start by considering linear diffusion, but one in which the diffusion coefficient varies with
time, D = D(t). Then (7.1) becomes
u
= D(t)2 u.
t

(7.2)

The introduction of a new variable


Z

T (t) =

D(t0 )dt0

(7.3)

transforms (7.2) into the standard diffusion equation


u
= 2 u.
T
The latter can be readily solved by using, for example, Greens functions.
1

Crank, Mathematics of Diffusion.

59

(7.4)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

7.2 Boltzmanns transformation


One-dimensional diffusion in infinite and semi-infinite media is described by the nonlinear diffusion equation


u

u
=
D(u)
,
(7.5)
t
x
x
where t > 0 and either < x < or 0 < x < . We further prescribe the uniform initial
condition u(x, 0) = uin = const. To be specific, we consider the following initial and boundary
conditions
u(x, 0) = uin ,

u(0, t) = u0 ,

u(, t) = uin ,

(7.6)

where u0 > uin .


Following Boltzmann (1894), we introduce a new variable
x
= ,
t

(7.7)

x

=
=
t

2t

2t t

(7.8)

1
=
.
x
t

(7.9)

so that

and

This transforms the nonlinear PDE (7.5) into the nonlinear ODE for u = u()


du
d
du

=
D(u)
2 d
d
d

(7.10)

subject to the boundary conditions


u(0) = u0 ,

u() = uin .

(7.11)

In general, the boundary value problem (7.10) (7.11) has to be solved numerically. There is
a class of functional relationships D = D(u) for which analytical solutions exist. This class can
be obtained by rewriting (7.10) as
Z u

1 d
D(u) =
(u0 )du0 + A .
(7.12)
2 du u0
where A is a constant of integration. Substituting a specific expression = (u) into the righthand side of this expression, we obtain the diffusion coefficient D(u) for which = (u) is an
implicit solution of (7.10). Its explicit form u = u() is obtained by the inversion of = (u).
The choice of = (u) is constrained by the following conditions:
- 60 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

1. It follows from (7.11) that


(u0 ) = 0;
2. For D(u) in (7.12) to exist it is sufficient that both
Z u
d
and
(u0 )du0 exist on uin u u0 ;
du
u0

(7.13a)

(7.13b)

3. Since D has to be positive and = x/ t > 0 on x > 0, it follows from (7.12) that
d
< 0.
du

(7.13c)

Unfortunately, none of the solutions obtained by this approach fit experimental data well.2

7.3
7.3.1

Kirchhoffs transformation
Steady-state diffusion

Steady-state diffusion with the concentration-dependent diffusion coefficient is described by


[D(u)u] + f = 0

(7.14)

subject to appropriate boundary conditions. For an arbitrary D(u) this equation can be transformed
into the Poisson equation,
2 + f = 0,

(7.15)

by means of the Kirchhoffs transformation,


Z

[u(x)] =

D(s)ds.

(7.16)

7.3.2

Steady-state diffusion with gravity

Equation (7.14) can be generalized by adding the gravitational term,


[D(u)u] +

dD u
+ f = 0,
du x3

(7.17)

which gives rise to the Richards equation for flow in partially saturated porous media. The Kirchhoff transform linearizes this equation only if
D(u) = D0 eu .
2

Crank, Mathematics of Diffusion, Sec 7.2.6.

- 61 -

(7.18)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

In this case, (7.17) becomes


f

+
=0
x3 D0
and can be solved analytically for a variety of flow scenario.
2 +

(7.19)

As an example, we consider one-dimensional infiltration into a porous column of length l. In


the dimensionless form, the process is described by the Kirchhoff transformed equation
d2
d
+a
=F
dz 2
dz
subject to the boundary conditions
(0) = H,

0<z<1

d
(1) + a(1) = Q.
dz

The change of variables


(z) = (z)ebz ,

b=

a
2

(7.20)

(7.21)

(7.22)

transforms (7.20) (7.21) into


d2
b2 = F ebz
dz 2

0<z<1

(7.23)

subject to
(0) = H,

d
(1) + b(1) = qeb .
dz

(7.24)

The Greens function g(z, ) for (7.23) (7.24) satisfies


d2 g
b2 g = (z )
dz 2

0<z<1

(7.25)

subject to
g(0, ) = 0,

dg
(1, ) + bg(1, ) = 0.
dz

(7.26)

The solution is
i
1 h b
e
sinh(bz)H( z) + ebz sinh(b)H(z ) .
b
The solution of (7.23) (7.24) can now be obtained as
Z 1
dg
(z) =
F ebz gd Qeb g(1, z) H (0, z).
d
0
g(z, ) =

(7.27)

(7.28)

It follows from (7.27) that


1
g(1, z < 1) = eb sinh(bz),
b

g(1, 1) = 0,

dg
( = 0, z > 0) = ebz .
d

(7.29)

Hence, for all 0 < z < 1,


1

1
F ebz gd + q sinh(bz) + Hebz .
b

(z) =
0

(7.30)

Recalling (7.22) yields


(z) = eaz/2

Z
0


1
F ebz gd + q 1 eaz + Heaz .
a
- 62 -

(7.31)

MAE 294c. Methods in Applied Mechanics III

7.4

Instructor: Daniel M. Tartakovsky

Method of spatial moments

The method of spatial moments (MSM) is a useful tool for solving nonlinear diffusion equations,
which is not restricted to particular functional forms of the diffusion coefficient. Rather than computing the spatio-temporal evolution of concentration u(x, t), the MSM evaluates the dynamics of
its spatial moments
Z
Mn (t) =

u(x, t)xn dx

n = 0, 1, 2, . . .

(7.32)

The moments have the following physical interpretation. The zeroth moment M0 (t) corresponds
to the total amount of the diffusion substance; the first moment M1 (t) yields the evolution of
the center of mass of the plume; and the second moment M2 (t) defines the spread of the plume
(spatial variances).
As an example, consider the one-dimensional nonlinear diffusion equation


u

u
=
D(u)
t
x
x

0 < x < 1,

t>0

(7.33)

subject to the initial and boundary conditions


u
(0, t) = 0,
x

u(x, 0) = 0,

u(1, t) = 1.

(7.34)

We start by multiplying (7.33) with xn and integrating in space,


Z
0

1



u

D(u)
xn dx = 0
t
x
x

n = 0, 1, 2, . . .

(7.35)

The first two moments are computed as follows.3 We start by noting that (7.33) (7.34) describes
the evolution of a diffused front (Figure 7.1). The front separates the region with c(x, t) 6= 0 from
c(x, t)
1

t4

t3

t2

t1

t=0
x

0 xf (t4 )

xf (t3 )

xf (t2 )

xf (t1 )

Figure 7.1: A schematic representation of the evolution of a diffused front. The toe of the front is
denoted by xf (t).
3

Crank, Mathematics of Diffusion, Sec. 7.4.

- 63 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

the region where c(x, t) 0. The outermost point of the front, or its toe, is denoted by xf (t).
We seek a solution c(x, t) in the form of a cubic polynomial,
(
a2 (t) [x xf (t)]2 + a3 (t) [x xf (t)]3 xf (t) x 1
u(x, t) =
(7.36)
0
0 x xf (t)
with three unknowns a2 (t), a3 (t), and xf (t). Note that this particular form of the polynomial is
chosen so that the number of unknowns does not exceed the number of equations. In this example
we have two,


Z 1

u
u

D(u)
dx = 0
(7.37)
t
x
x
0
and
Z
0

1



u

D(u)
xdx = 0,
t
x
x

(7.38)

plus the constraints introduced by the initial and boundary conditions (7.34). The initial condition
in (7.34) gives
xf (0) = 1.

(7.39)

The first boundary condition in (7.34) is satisfied by (7.36) automatically. Applying the second
boundary condition in (7.34) to u(x, t) in (7.36) gives the third equation,
U = a2 (t) [1 xf (t)]2

U (t) + V (t) = 1,

and V = a3 (t) [1 xf (t)]3 .

Substituting (7.36) into (7.37) and integrating gives the first ODE


d
1 xf
1 xf
2U + 3V
U
+V
= D(1)
.
dt
3
4
1 xf

(7.40)

(7.41)

Since
Z

xf



Z 1
D(u)

u
u
D(u)
xdx = D(1) (1) D(1)u(1) +
u
dx
x
x
x
x
xf

and
Z

u
xf

D(u)
dx =
x

u
xf

dD(u) u
dx =
du x

D(1)

u(1)

udD(u) = D(1)u(1)
D(xf )

substituting (7.36) into (7.38) and integrating yields


Z 1
Z 1
u

uxdx = D(1) (1)


D(u)du.
t xf
x
0
Combining (7.36), (7.40), and (7.41) with this equation, we obtain

 Z 1
Z 1
d
1 xf
1 xf
U
+V

uxdx =
D(u)du.
dt
3
4
xf
0
- 64 -

D(u)du,
u(xf )

(7.42)

(7.43)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Evaluating the integral gives the second ODE




Z 1
d
(1 xf )2
(1 xf )2
U
+V
=4
D(u)du.
dt
3
5
0
Using the first equation in (7.40) to eliminate V in (7.41) and (7.44) yields



d
U
3U
(1 xf ) 1 +
= 4D(1)
dt
3
1 xf

(7.44)

(7.45)

and



Z 1
d
2U
2
(1 xf ) 1 +
= 20
D(u)du.
dt
3
0
Integrating (7.46) with the initial condition in (7.34), we obtain


Z 1
2 3

+ U = t,
(t) 1 xf ,
30
D(u)du.
2
0
Using (7.47) to eliminate U in (7.45), gives the ODE for (t),




d t
12D(1) 3
t
+
=

dt 2 3

2 3 2
subject to the initial condition (0) = 0. The solution of this equation is in the form
r
t
=
.

(7.46)

(7.47)

(7.48)

(7.49)

The constant is obtaind by substituting (7.49) into (7.48). This gives as a solution of the
quadratic form
3
24D(1) 2 + [ 108D(1)] + = 0.
(7.50)
2
Inserting (7.49) into (7.47) gives
3
U = .
(7.51)
2
Inserting (7.51) into (7.40) yields
5
V = + .
(7.52)
2
Note that both U and V turned out to be constants! Recalling the definition of in (7.49) yields
r
t
xf (t) = 1
.
(7.53)

Substituting these equations into (7.36) leads to the solution for the concentration of the diffusing
substance
r 2 
r 3

 
  3/2 
3
t
5

t
u(x, t) =
x1+

x1+
, (7.54)
2 t

2
t

p
when 1 t/ x 1. Otherwise, c(x, t) = 0. Note that this solution assumed that the
diffusive front xf (t) did not reach the boundary x = 0. Let t? denote the time at which the front
reaches the boundary. This is the time at which xi = 1 in (7.49). The corresponding concentration
profile is




3
5
2
u(x, t) =
x
x3 .
(7.55)
2
2
- 65 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

7.5 Hopf-Coles transformation


Consider the Burgers equation
2u
u
u
=a 2 u ,
(7.56)
t
x
x
which is often used to model turbulent flows. It can be reduced to the linear diffusion equation
through the following series of integral transformations
The first mapping is defined by
Z x
w(x, t) =
u(y, t)dy

or

u(x, t) =

w(x, t)
.
x

(7.57)

Integrating (7.56) in space gives


w
2w 1
=a 2
t
x
2

w
x

2
.

(7.58)

The second mapping is defined by


v = (w),

(7.59)

v
w
= 0
.
t
t

(7.60)

"

 #
v
2 w 1 w 2
0
= a 2
.
t
x
2 x

(7.61)

so that

Substituting (7.58) gives

It follows from (7.59) that


2v
= 00
x2

w
x

2

+ 0

2w
.
x2

(7.62)

Combining (7.61) and (7.62) yields





v
2v
0
w 2
00
= a 2 a +
.
t
x
2
x

(7.63)

Next, we define the mapping v = (w) in (7.59) as a solution of


2a00 + 0 = 0,

(7.64)

 w
= exp
.
2a

(7.65)

e.g.,

Then (7.63) becomes a linear PDE


v
2v
= a 2.
t
x
Mappings (7.57) and (7.59) are known as the Hopf-Cole transform.
- 66 -

(7.66)

Chapter 8

Approximate Solutions of ODEs: Local


Analyses
The material in this chapter is derived from Bender and Orszag, Advanced Mathematical Methods
for Scientists and Engineers.

8.1

Basic definitions

Analytical functions. Complex analysis is essential for the proper understanding of analytical
functions. For our purpose the following will suffice. A function f (x) is analytic in a neighborhood of x0 if it is locally given by a convergent power series
f (x) =

X
f (n) (x0 )

n!

n=0

(x x0 )n ,

u(n)

dn f
dxn

The radius of convergence. The radius of convergence of a power series


g(x) =

cn (x a)n ,

n=0

is the nonnegative quantity R (which may be a real number or ), such that the series converges
if
|x a| < R
and diverges if
|x a| > R.
The radius of convergence is infinite if the series converges for all x.
Consider an n-th order homogeneous ODE
u(n) (x) + pn1 (x)u(n1) (x) + . . . + p0 (x)u(x) = 0.
It has n linearly independent solutions.
67

(8.1)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Ordinary points. The point x0 (x0 6= ) is an ordinary point of (8.1) if all pi (x) (i =
0, . . . , n 1) are analytic in a neighborhood of x0 .
(Examples: y 000 = ex y excepting x0 = , every point x0 is an ordinary point; x5 y 00 = y
every point x0 , other than x0 = 0 and x0 = , is an ordinary point.)

Regular singular points. The point x0 (x0 6= ) is a regular singular point of (8.1) if not all
pi (x) (i = 0, . . . , n 1) are analytic in a neighborhood of x0 , but if all (x x0 )i pi (x) (i =
0, . . . , n 1) are analytic.
(Examples: x0 = 1 is a regular singular point (x 1)y 000 = ex y.)

Irregular singular points. The point x0 (x0 6= ) is an irregular singular point of (8.1) if its
neither an ordinary point nor a regular singular point.

The point x0 = . To classify the point x0 = , we introduce


y=

1
,
x

d
d
= y 2 ,
dx
dy

...

and then classify the point y = 0. The point x0 = is an ordinary, regular singular, or irregular
singular point if the point y = 0 is correspondingly classified.

8.2

Fuchs theory of ODEs

In 1866 Fuchs proved that

In a neighborhood of the ordinary point x0 , solutions of (8.1) are analytic with the radius of
convergence of the corresponding power series at least as large as the distance from x0 to the
nearest singular point. Thus in a neighborhood of x0 solutions of (8.1) can be represented
by their Taylor series
u(x) =

an (x x0 )n .

n=0

At the regular singular point x0 , a solution of (8.1) may be either


analytic or
singular. The singularity may be either
a pole or
algebraic or logarithmic.
- 68 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

At least one solution must be in the form of a Frobenius series,


u(x) = (x x0 ) A(x) =

an (x x0 )+n ,

n=0

where A(x) is an analytical function and the number is called the indicial exponent. At
worst, the form of the nth solution of (8.1) with n 2 at the regular singular point x0 is
u(x) = (x x0 )

n1
X

ln[(x x0 )]i Ai (x),

i=0

where all Ai (x) are analytical functions.


No comprehensive theory exists for irregular singular points.

8.3

Local behavior of linear ODEs near ordinary points

Consider the Airy equation


u00 = xu.

(8.2)

Find the solution u(x) near x = 0. Since x = 0 is an ordinary point of (8.2), we seek a solution in
the form of the Taylor series

u(x) =

an xn .

(8.3)

n=0

Substituting (8.3) into (8.2) yields

n2

n(n 1)an x

n=0

an xn+1 .

(8.4)

n=0

Equating the coefficients of equal powers of x gives


n(n 1)an = 0,

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

n(n 1)an = an3 ,

n = 3, 4, 5, . . . .

(8.5)

It follows from the first three equations that a1 and a2 are arbitrary constants and a2 = 0. The
remaining coefficients are obtained as solutions of the remaining equations an = an3 /n/(n 1)
as follows
a0
a0 (2/3)
= n
3n(3n 1)(3n 3)(3n 4) 9 8 6 5 3 2
9 n!(n + 2/3)
a1
a1 (4/3)
=
= n
(3n + 1)3n(3n 2)(3n 3) 10 9 7 6 4 3
9 n!(n + 4/3)
= 0.

a3n =
a3n+1
a3n+2

(8.6)
(8.7)
(8.8)

Substituting these coefficients into (8.3) gives the solution


u(x) = c1

X
n=0

X
x3n+1
x3n
+ c2
.
n
n
9 n!(n + 2/3)
9 n!(n + 4/3)
n=0

- 69 -

(8.9)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

8.4 Local behavior of linear ODEs near regular singular points


Find a solution of the ODE
u00 +

u
=0
4x2

(8.10)

near the regular singular point x = 0. It is easy to check that the Taylor series representation of
the solution (8.3) is not adequate. Indeed, substituting (8.3) into (8.10) yields

n2

n(n 1)an x

n=0

1X
=
an xn2 ,
4

(8.11)

n=0

which results in

an

1
n
2

2
=0

an = 0

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

(8.12)

i.e., in the trivial solution.


The general theory of Fuchs suggests that a solution can be found in the form of a Frobenius
series,
u(x) =

an x+n .

(8.13)

n=0

Substituting (8.13) into (8.10) yields

( + n)( + n 1)an x+n2 =

1X
an x+n2
4

(8.14)

n=0

n=0

Assuming that a0 6= 0, equating the leading terms in these expansions determines the indicial
exponent
( 1) +

1
=0
4

1
= .
2

(8.15)

Equating the coefficients of nth powers of x (n 1) gives an = 0 for all n 1. Thus the exact

solution of (8.10) is u(x) = c1 x, where c1 is an arbitrary function.

8.5

Local behavior of linear ODEs near irregular singular points

Find a solution of the ODE


x3 u00 = u

(8.16)

near the irregular singular point x = 0. As expected from the general Fuchs theory, a Frobenius
series does not represent such solutions. Indeed, substituting the Frobenius series (8.13) with
a0 6= 0 into (8.16) gives

( + n)( + n 1)an x+n+1 =

n=0

X
n=0

- 70 -

an x+n .

(8.17)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Equating the coefficients of x yields a0 = 0, which is a contradiction.


The progress can be made by considering the asymptotic behavior of a solution at x = x0 .
When such a solution is given by an infinite series, its first term determines the leading behavior
of a solution. The most rapidly changing component of the leading behavior as x x0 is called
the controlling factor. For example, in the Frobenius series (8.13) the leading behavior is a0 x
and the controlling factor is x . The controlling factors are usually in the form of an exponential.
To derive the leading behavior of a solution of (8.16) near the irregular singular point x = 0,
we use the substitution
u(x) = eS(x)

(8.18)

S 00 + (S 0 )2 = x3 .

(8.19)

which gives

If x = 0 is an irregular singular point, it is usually true that


S 00  (S 0 )2 ,

x 0.

(8.20)

(The validity of this statement can be checked on an example where the leading behavior of u(x)
at x = 0 has the form exp(axb ) with b > 0.) Then the asymptotic differential equation for (8.19)
is
(S 0 )2 x3 ,

x 0,

(8.21)

which has two solutions


S(x) = 2x1/2 ,

x 0.

(8.22)

Thus the controlling factors for two linearly independent solutions at x 0 are
u1 (x) e2x

8.6

1/2

u1 (x) e2x

and

1/2

x 0.

(8.23)

Local behavior of inhomogeneous linear ODEs

Analytical inhomogeneities. Find the local behavior of the general solution of


u0 + xu = x3

(8.24)

near x = 0. Since x = 0 is both an ordinary point of the homogeneous equation y 0 + xy = 0 and


a point where the inhomogeneity x3 + 1 is analytic, we use a Taylor series to represent u(x),
u(x) =

an xn .

(8.25)

an xn+1 = x3 ,

(8.26)

n=0

This gives

nan xn1 +

n=0

X
n=0

which results in an arbitrary a0 , a1 = 0, and


nan + an2

(
0 n 2,
=
1 n = 4.
- 71 -

n 6= 4

(8.27)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Nonanalytical inhomogeneities. Find the local behavior of the general solution of


u0 + xu = x4

(8.28)

near x = 0. The following situations are possible:


1. u0 xu and x4  xu as x 0,
2. xu x4 and u0  x4 as x 0,
3. u0 x4 and xu  x4 as x 0.
The first scenario results in
u(x) aex

2 /2

a,

x 0,

(8.29)

which is not consistent with the condition x4  xu.


The second scenario implies that
u(x) x5 ,

x 0,

(8.30)

which violates u0  x4 .
The third scenario leads to
1
u(x) x3 ,
3

x 0,

(8.31)

which is consistent with the condition that xu  x4 . Hence, this is the only consistent leading
behavior. The corrections to this leading behavior is obtained by setting u(x) = x3 /3 + C(x),
where C(x)  x3 . Substituting this into (8.28) gives the ODE
1
C 0 + xC = x2 .
3

(8.32)

Repeating the above analysis leads to C = 1/(3x), so that


u(x)

1
1

,
3
3x
3x

x 0,

(8.33)

The process can continue leading to more and more terms in this series.

8.7

Nonlinear ODEs: Spontaneous singularities

One of the key features of linear ODEs is that their singularities are determined by their coefficients
and are thus fixed. This means that the points of singularity are independent of the choice of initial
or boundary conditions. For example, the linear ODE
u0 +

u
= 0,
x1
- 72 -

u(0) = u0

(8.34)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

has a regular singular point at x = 1. Its solution,


u=

u0
,
1x

(8.35)

has a pole at x = 1, whose location does not change with the choice of a particular value of u0 .
In contrast, nonlinear ODEs often have singularities that appear spontaneously. Their locations are not fixed and depend on the choice of initial or boundary conditions. Such singularities
are called spontaneous or movable singularities. For example, the nonlinear ODE
u0 u2 = 0,

u(0) = u0

(8.36)

is not singular. Yet its solution,


u(x) =

u0
,
1 u0 x

(8.37)

is singular. Moreover, the location of the singularity varies with the initial condition, e.g., this
solution has a pole at x = 1 if u0 = 1 and at x = 1/2 if u0 = 2.

8.8

Early-time behavior of nonlinear ODEs

The general theory for local analyses of nonlinear ODEs states that the solution u(x) of the firstoder nonlinear ODE
u0 = F (x, u),

u(0) = u0

(8.38)

is analytic in a neighborhood of x = 0 as long as F (x, u) is an analytical function of its two


arguments at x = 0 and u = u0 .
As an example, consider the nonlinear ODE
u0 =

u2
,
1 xu

u(0) = 1.

(8.39)

We wish to determine the early time behavior of u(x), i.e., its behavior near x = 0. Since F (x, u)
at x = 0 is analytic with respect to its two arguments, u(x) is analytic as well, i.e., it can be
represented by the Taylor series
u(x) =

an xn ,

a0 = 1,

(8.40)

n=0

with a nonzero radius of convergence. Substituting (8.40) into (8.39) yields


!
!2

X
X
X
1
an xn+1
nan xn1 =
an xn ,
n=0

n=0

(8.41)

n=0

which can be recast as

X
n=0

n1

nan x

ak an (1 + n)xk+n .

k,n=0

- 73 -

(8.42)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

Equating the coefficients of equal powers of x gives


a1 = 1,

3
a2 = ,
2

8
a3 = ,
3

...,

an =

(n + 1)n1
.
n!

(8.43)

The radius of convergence of the Taylor series (8.40) whose coefficients are given by (8.43) is


an
= lim = 1 .
R = lim
(8.44)
n
n an+1
e
This implies that the solution u(x) of (8.39) has a singularity at x = e1 .

8.9

Late-time behavior of nonlinear ODEs

Find the leading behavior of the Riccati equation


u0 = u2 + x

(8.45)

as x .
We start by introducing the change of variables

u(x) = xv(x)

(8.46)

that transforms (8.45) into

v
v 0 = (1 + v 2 ) x
.
2x

(8.47)

Next we notice that for large x the second term in the right-hand side of (8.47) can be neglected

because 1 + v 2 v and x  1/(2x). Hence

v 0 (1 + v 2 ) x,
x ,
(8.48)
with the solution
2
v arctan x3/2 ,
3

x .

(8.49)

This gives the late-time solution of the Ricatti equation,


2
u x1/2 arctan x3/2 ,
3

8.10

x .

(8.50)

Local behavior of nonlinear ODEs on bounded domains

Examine the leading behavior of the solution of


uu00 = 1,

u(0) = u(1) = 0

near the boundaries, i.e., as x 0 and x 1.


- 74 -

(8.51)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

There exists no general theory for asymptotic analysis of nonlinear higher-order ODEs. Such
analyses have to be carried out by trial and error. For example, the simplest guess of the asymptotic
behavior of the solution to (8.51) near x = 0 is that it approaches 0 algebraically, i.e., that u(x)
aeb as x 0. Substituting this expression into (8.51) gives
a2 b(b 1)x2b2 1,

x 0.

(8.52)

This asymptotic behavior requires that b = 1. Yet this value turns the left-hand side into zero.
Hence our first guess is incorrect.
Next we examine if the logarithmic asymptotic behavior of the solution to (8.51) near x = 0
is possible, i.e., we test the hypothesis that u(x) ax( ln x)b as x 0. Substituting this
expression into (8.51) gives
a2 b( ln x)2b1 + a2 b(b 1)( ln x)2b2 1,

x 0.

(8.53)

Since ( ln x)2b1  ( ln x)2b2 as x 0, the second term can be neglected to give


a2 b( ln x)2b1 1,
This asymptotic behavior dictates that b = 1/2 and a
as x is

u(x) x 2 ln x,
A similar analysis can be carried out for x 1.

- 75 -

x 0.

(8.54)

2. Hence the leading behavior of u(x)

x 0.

(8.55)

MAE 294c. Methods in Applied Mechanics III

- 76 -

Instructor: Daniel M. Tartakovsky

Chapter 9

Scaling and Self-Similarity


The material in this Chapter comes from Barenblat, Scaling, self-similarity, and intermediate
asymptotics.

9.1

Dimensions

Physical quantities can be divided into those that are characterized either by fundamental units
or by derived units. A complete set of fundamental units for a physical phenomenon is called a
system of units. For example, all properties of kinematic phenomena can be characterized by a
system of three fundamental units, those of length (L), mass (M ), and time (T ). A set of systems
of units that differ only in the magnitude but not in physics is call a class of systems of units.
We will use Maxwells notation [u] to denote the dimension of a physical quantity u. In the
LM T system of units, any physical quantity describing a kinematic phenomenon is expressed by
a power-law monomial of these units, [u] = L M T . For example, velocity v has the dimension
[v] = LT 1 , density has the dimension [] = M L3 , etc. It can be proved that this, a power-law
monomial representation, is a rule rather than an exception.
Physical quantities {ui }ki=1 are said have independent dimensions if none of them has a dimension that can be represented as a product of powers of the dimensions of the remaining quantities. For example, density ([] = M L3 ), velocity ([v] = LT 1 ) and force ([f ] = [ma] =
M LT 2 ) have independent dimensions, while density, velocity and pressure ([p] = L1 M T 2 )
do not.

9.2

Dimensional analysis

The ultimate goal of any study of a physical system is to establish relationships between the quantities and parameters that characterize the system. The fact that a quantity of interest u depends on
the known quantities {
pi }N
i=1 is expressed by
u = f (
p1 , . . . , pN ).
77

(9.1)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

To understand the behavior of a physical system is to find the functional form of f .


The quantities {
p i }N
i=1 are called governing parameters. Let us denote the parameters with
independent dimensions by {pi }ki=1 and the parameters with dependent dimensions by {qi }m
i=1 , so
that k + m = N . The definition of parameters with dependent dimensions implies that they can
be expressed as the products of powers rij of their independent counterparts,
[qi ] = [p1 ]ri1 . . . [pk ]rik ,

i = 1, , m.

(9.2)

It then follows from (9.1) that the dimension [u] of the quantity of interest u is given by a powerlaw monomial of the independent dimensions,
[u] = [p1 ]r1 . . . [pk ]rk .

(9.3)

If this were not so, the dimensions [u], [p1 ], . . . , [pk ] would be independent, which is a contradiction.
Of course, the powers ri are as yet unknown. What have we gained? In (9.1) we do not
know the functional form of the dependence of u on the governing parameters, while in (9.3) we
do know the functional form of the dependence of [u] on the dimensions of governing parameters.
In other words, while (9.1) has an infinite number of degrees of freedom, (9.3) has only k.

9.2.1

Transformation to dimensionless parameters

The relationships (9.1) and (9.3) allow one to introduce dimensionless parameters
u
qi
= r1
i = ri1
, i = 1, . . . , m.
rk ,
p1 pk
p1 prkik

(9.4)

Then (9.1) can be rewritten as


f (p1 , . . . , pk , 1 pr111 prk1k , . . . , m pr1m1 prkmk )
=
= F (p1 , . . . , pk , 1 , . . . , m ), (9.5)
pr11 prkk
where F is as yet unknown function.
It is intuitively obvious, and can be proved mathematically, that within the same class of
systems of units (e.g., the LM T class) it is possible to pass from one system (e.g., the SI system
of units) to another (e.g., the cgs system of units) by changing one of the parameters with the
independent dimension by a factor, while keeping the rest of these parameters unchanged. Since
dimensionless quantities remain unchanged under such a transformation, this implies that in
(9.5) is independent of p1 , . . . , pk , i.e., that
= (1 , . . . , m ),

(9.6)

which reduces the dimensionality of the original function f by k. Recalling the definition of the
dimensionless parameters, this gives


q1
qm
rk
r1
f (p1 , . . . , pk , q1 , . . . , qm ) = p1 pk
, . . . , rm1
.
(9.7)
pr111 . . . prk1k
p1 . . . prkmk
This is the celebrated Buckinghams theorem of dimensionless analysis1 .
1

Buckingham, E., On physically similar systems; illustrations of the use of dimensional equations. Phys. Rev. 4,
345-376 (1914)

- 78 -

MAE 294c. Methods in Applied Mechanics III

9.2.2

Instructor: Daniel M. Tartakovsky

Example

By way of example, consider fluid flow in a long cylindrical pipe. We are interested in the pressure
gradient J = dp/dx, i.e., in the pressure drop per unit length of the pipe. It depends on the fluid
velocity averaged over the pipes cross-section U , the diameter of the pipe d, the fluid density ,
and the fluid viscosity . In other words, J is a function of N = 4 governing parameters,
J = f (U, d, , ).

(9.8)

The quantities involved have the following dimensions


[J] = M L2 T 2 ,

[U ] = LT 1 ,

[d] = L,

[] = M L3 ,

[] = M L1 T 1 .

(9.9)

The dimensions [U ], [d], and [] are independent, while the dimension [] can be expressed as
their power-law monomial,
[] = [][U ][d].

(9.10)

Thus the number of independent parameters is k = 3 and the number of dependent parameters is
m = 1. We should expect the dimension [J] to be given by a power-law monomial of [], [U ], and
[d]. This indeed is the case,
[J] = [][U ]2 [d]1 .

(9.11)

According to the theorem (9.7), we can now replace (9.8) with


= (1 ),

J
U 2 d1

1 =

.
U d

(9.12)

Since is a function of one argument only, it is completely feasible to infer the function experimentally. In fact, this what Osborne Reynolds did. You might recognize the 1 = Re as the
Reynolds number.

9.3

Similarity and scaling

Many large objects, such as airplanes and ships, are designed and tested in laboratory settings, such
as wind tunnels. The reason this modeling approach works is that the physical phenomena on these
two vastly different scales are similar. The concept of physically similar phenomena implies that
the quantity of interest u depends on the same governing parameters {
p }N
i=1 in exactly the same
way. The only thing that is different are scales: the scale of the modeldenoted by the superscript
(M )and the scale of the prototypedenoted by the superscript (P ). In other words,
(M )

u(M ) = f (
p1

(M )

, . . . , pN )

and

(P )

(P )

u(P ) = f (
p1 , . . . , pN ).

(9.13)

For such a modeling approach to work it is necessary to know how the quantities of interest scale
from the scale of the model to the scale of the prototype.
The progress is made by realizing that in physically similar phenomena the values of the
corresponding dimensionless governing parameters do not change with the scale, i.e., that
(M )

(P )

= 1

)
(P )
= 1 , . . . , (M
m = m = m .

- 79 -

(9.14)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The quantities 1 , . . . , m are called similarity parameters, and the conditions (9.14) are called
similarity criteria. It follows from (9.6) that on both scales (9.13) can be written in the dimensionless form as
(M )

(M ) = (1

)
, . . . , (M
m )

and

(P )

)
(P ) = (1 , . . . , (P
m ).

(9.15)

Since the function does not change with scale, the similarity criteria (9.14) imply that (M ) =
(P ) . Using (9.4) to return to the dimensional parameters we obtain the scaling law for the results
of the experiment,
#
" (P ) #rk
"
(P ) r1
pk
p
1
(M
.
(9.16)
u(P ) = u(M ) (M
)
)
p1
pk
We also obtain the rules for choosing the governing parameters with dependent dimensions,
"
#
" (P ) #rik
(P ) ri1
pk
(P )
(M ) p1
(M
,
i = 1, . . . , m.
(9.17)
qi = qi
(M )
)
p1
pk
The latter insure the similarity between the prototype and its model.

9.4

Self-similar solutions: Linear diffusion

Consider heat conduction in a long homogeneous rod. We assume that at time t = 0 heat concentrated in a narrow zone centered around x = 0, while the rest of the rod was at zero temperature.
The temperature of the rod u(x, t) at point x and time t is described by the one-dimensional diffusion equation
2u
u
=D 2
t
x

<x<

t>0

(9.18)

subject to the initial condition


x2
u(x, 0) =
exp 2
2l
2l2



,

(9.19)

where l is the width of the initial distribution and Q is the area under u(x, 0). The integral of
motion (or mass conservation) is
Z
Z
M0 (t) =
u(x, t)dx =
u(x, 0)dx = Q = const.
(9.20)

We will attempt to solve this problem by means of the dimensional analysis described in the
previous sections. The analysis starts by identifying the governing parameters affecting temperature u(x, y). These are x, t, heat diffusivity D, the width of the source l and the total amount of
energy Q. In other words, u is a function of N = 5 governing parameters
u = f (x, t, D, l, Q).
- 80 -

(9.21)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

In the LT class of units, where denotes the dimension of temperature, the quantities involved
have the following dimensions
[u] = ,

[x] = L,

[t] = T,

[D] = L2 T 1 ,

[l] = L,

[Q] = L.

(9.22)

The dimensions of D and Q are obtained by making sure that the left- and right-hand sides of
(9.18) and (9.19) have consistent dimensions.
Next, we identify the governing parameters with independent dimensions. The dimensions
[t], [D], and [Q] are independent, while the dimensions of the remaining parameters can be expressed as their power-law monomials
[u] = [Q][t]1/2 [D]1/2 ,

[x] = [Q]0 [t]1/2 [D]1/2 ,

[l] = [Q]0 [t]1/2 [D]1/2 .

(9.23)

This gives rise to the dimensionless parameters


=

u
,
Q(Dt)1/2

1 =

x
,
(Dt)1/2

2 =

l
.
(Dt)1/2

(9.24)

That is, k = 2. According to the theorem (9.7),


= (1 , 2 ),

(9.25)

or
Q
u(x, t) = (1 , 2 ).
Dt

Thus without solving (9.18) we see that u(x, t) varies as t.

(9.26)

The only thing we still do not know is the functional form of . It can be obtained from the
long-time asymptotics, when t . How large t should be is determined by how small l is. At
any rate, in this limit 2 0. Assume that = (1 , 0) is finite and call it fs (1 ). Then it
follows from (9.24) that
Q
u = fs (1 ),
Dt

x
= .
Dt

(9.27)

Substituting (9.26) into (9.18) gives an ODE


1 0 1
00
fs + fs + fs = 0.
(9.28)
2
2
R
The boundary condition fs () = 0 is necessary for udx = Q to exist. The solution is
1
2
fs () = e /4 .
2

(9.29)

Note that time has disappeared in (9.29). This is a characteristic of self-similar solutions.
Of course, the boundary value problem (9.18) (9.19) is simple enough to allow an exact
analytical solution. It could be readily obtained by using the Greens function (6.57). This solution
is




Z
1
Q
(x )2
x2
u(x, t) =
u(, 0) exp
d = p
exp
.
2l2
2(2Dt + l2 )
2 Dt
2(2Dt + l2 )
(9.30)
- 81 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

For large times, i.e., for Dt  l2 ,




x2
Q
.
exp
u(x, t) =
4Dt
4Dt

(9.31)

This solution holds when either t or l 0. Note that an important characteristic of the
initial condition, the width of the initial distribution l, is forgotten.
A formal definition of self-similarity is in order:
A time dependent phenomenon is called self-similar if the spatial distributions of its
properties at different times can be obtained from one another by a similarity transformation.
Comparing the exact analytical solution with its counterpart obtained via dimensional analysis, we
observe that the self-similar solution describes the intermediate asymptotic behavior of the general
solution in the region where the latter becomes independent from the initial conditions but still has
not reached its equilibrium.

9.5

Self-similar solutions: Nonlinear diffusion

Here we extend our dimensional analysis to the problems of nonlinear heat conduction. Consider
the nonlinear diffusion


u

u
=
D(u)
,
< x < , t > 0.
(9.32)
t
x
x
subject to the initial condition
x2
u(x, 0) =
exp 2
2l
2l2
Q

u(x, t)dx = Q.

(9.33)

and conditions at infinity


u(, t) = 0.

(9.34)

Similarity solutions are very sensitive to the choice of D(u). This is because a general D(u) might
introduce additional length scales, e.g., a scale on which D(u) increases and then decreases. We
select
D(u) = D0 un ,

D0 > 0,

n > 0.

(9.35)

which describes flow in porous media, plasma, etc.


Once again, we commence our analysis by identifying the governing parameters affecting
temperature u(x, t). These are x, t, D0 , l, Q, n, so that u is a function of N = 6 governing parameters,
u = f (x, t, D0 , l, Q, n).
- 82 -

(9.36)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

In the LT class of units, the dimensions of the relevant quantities are


[u] = ,

[x] = L,

[t] = T,

[D0 ] = L2 T 1 n ,

[l] = L,

[Q] = L,

[n] = 1.
(9.37)

As before, we identify t, D0 , and Q as governing parameters whose dimensions are independent.


The remaining quantities are expressed as their power-law monomials,

[u] =

[Q]2
[t] [D0 ]

1/(2+n)

[x] = [l] = [t] [D0 ] [Q]n

This gives rise to the following dimensionless parameters




D0 t 1/(2+n)
x
,
=u
,
1 =
n
Q2
(D0 tQ )1/(2+n)

1/(2+n)

2 =

(D0

tQn )1/(2+n)

(9.38)

(9.39)

which are related to each other via the theorem,


= (1 , 2 ).

(9.40)

Without solving the boundary value problem, we can observe that temperature u varies with time
t as a function of (D0 t)1/(2+n) .
The intermediate (long-time) asymptotic solution is obtained by setting 2 0 and denoting
(1 , 0) = fs (1 ). This gives

u(x, t) =

Q2
D0 t

1/(n+2)
fs (),

x
(D0

tQn )1/(n+2)

(9.41)

Since
u
t1
=
t
n+2

Q2
D0 t

1/(n+2) 

dfs
fs +
,
d

d
= (D0 tQn )1/(n+2) ,
x
d

substituting (9.41) into (9.32) gives the ODE




d
dfs
n dfs
(n + 2)
fs
+
+ fs = 0.
d
d
d

(9.42)

(9.43)

This ODE is subject to the boundary conditions2


dfs
=0
d |=0

fs () = 0,

(9.44)

and the normalization condition


Z

fs ()d = 1.

(9.45)

The solution of (9.43) -(9.44) is



fs () =
2

1/n
n
2
2
( )
2(n + 2) 0

The second condition in (9.44) implies the symmetry of the profile.

- 83 -

for

0 ,

(9.46)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

and fs = 0 for > 0 . The constant 0 is determined from (9.45),




n
1=
2(n + 2)

1/n Z

(02

2 1/n

n
d = 2
2(n + 2)

1/n Z

(02 2 )1/n d.

(9.47)

This gives

 


n + 2 1/(2+n)
1
1 n/(2+n)
B
,1 +
,
0 = 2
n
2
n

(9.48)

where B(, ) is the Beta function.


This completes the derivation of the intermediate-asymptotic (self-similar) solution to the
problem.

9.6

Anomalous Exponents and Self-Similarity of the Second Kind

So far, we were able to construct self-similar solutions by using dimensional analysis alone. Such
solutions are called self-similar solutions of the first kind. Here we consider cases where it is not
possible.
Consider heat conduction is a homogeneous rod whose thermal diffusivities (D1 and D in
the cooling and heating regions, respectively) are constant but different. The process is described
by the following modified diffusion equation
(
u
D
u
2u
t > 0
= Di 2 ,
Di =
(9.49)
u
t
x
D1 t < 0,
subject to the initial condition for a concentrated source of heat
Z
Q x
u(x, 0) = u0
,
Q=
u(x, 0)dx
l
l

(9.50)

and the boundary conditions


u(, t) = 0.

(9.51)

A cursory comparison of this problem and the problem we analyzed in Section (9.4) might
suggest that the two are sufficiently similar to allow for dimensional analysis of (9.49) (9.51). In
analogy with (9.24), such an analysis would identify the governing parameters with independent
dimensions t, D, and Q, as well as the dimensionless groups
=

u
Dt,
Q

x
1 = ,
Dt

2 =

l
,
Dt

3 =  =

D1
1,
D

(9.52)

the only difference being the presence of the dimensionless constant parameter . Then, according
to the theorem,
= (1 , 2 , ).
- 84 -

(9.53)

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

The intermediate asymptotic solution (t or l 0) requires that 2 0 so that (1 , 0, 3 )


f (1 , ) exists, is finite, and 6= 0. Then
Q
u(x, t) = f (, ),
Dt

x
= .
Dt

(9.54)

Since u(x, t) is a solution of (9.49), the function f should (i) be continuous, (i) have a continuous
derivative, and (iii) be even f (, ) = f (, ).
Let x0 (t) denote an inflection point of the temperature profile, i.e., the point where u/x
changes the sign. To find it, consider
2u
=0
x2

Q d2 f
= 0.
(9.55)
(Dt)3/2 d 2

Let 0 () correspond to the inflection point, i.e., x0 (t) = 0 () Dt. Substituting (9.54) into (9.49)
gives
(1 + )

or

d2 f
1 df
+
= 0,
2
d
2 d

0 0

(9.56)

and
d2 f
1 df
+
= 0,
2
d
2 d

0 <

(9.57)

Integrating once, we obtain


(1 + )

df
1
+ f = c1 ,
d
2

0 0

(9.58)

and
1
df
+ f = c2 ,
d
2

0 <

(9.59)

By virtue of symmetry, df /d = 0 at = 0. Hence c1 = 0. As , both f and its derivative


must vanish, which implies c2 = 0. Integrating again,


2
f = c3 exp
,
0 0
(9.60)
4(1 + )
and
 2

f = c4 exp
,
4

0 <

(9.61)

The constant c3 and c4 are determined from the continuity of f and f at = 0 . If  = 0, we have
c3 = c4 . If  6= 0, the only solution is trivial, c3 = c4 = 0. It clearly does not satisfy the initial
condition.
To understand what went wrong, we must revisit our assumptions and approximations. In
fact, we have made a single assumption, which is that (1 , 0, 3 ) is finite. We now have to
conclude that it is incorrect, i.e., that (1 , 0, 3 ) might be either zero or infinite. To account for
- 85 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

this possibility, we look for the following asymptotic representation of (1 , 2 , 3 ) as 2 0


(which corresponds to either t or l 0),
(1 , 2 , 3 ) = 2 f (1 , 3 )

as

2 0,

(9.62)

where f is finite, and is called an anomalous exponent. This replaces (9.54) with
u(x, t) =

Ql
f (, ),
(Dt)(1+)/2

x
= .
Dt

(9.63)

Note that that the width of the heat source l, a property of the initial condition that was
absent in our previous self-similar solutions of the first kind, is now present. In fact, (9.63) is an
asymptotic solution in the limit as l 0. If this limit is taken for fixed Q, x, and t, temperature
u(x, t) in (9.63) becomes either 0 and , depending on the sign of . This is not adequate,
since it follows from the initial condition (9.50) that such a limit corresponds to the point source
of strength Q, which should result in a finite temperature distribution u. This suggests that the
product Ql must remain finite as l 0, i.e., that
u(x, t) =

A
(Dt)(1+)/2

A = lim Ql ,

f (, ),

l0

x
= ,
Dt

(9.64)

where is a dimensionless constant that depends on the normalization of the function f (, ). For
example, we can normalize f (, ) by the condition
f (0, ) = 1.

(9.65)

To find the value of the anomalous exponent and the functional form of f , we substitute (9.64)
into (9.49), which gives
(1 + )

d2 f
df
1+
+
+
f = 0,
2
d
2 d
2

0 0

(9.66)

and
df
1+
d2 f
+
+
f = 0,
2
d
2 d
2

0 < .

(9.67)

As before, 0 is an inflection point at which df /d = 0, i.e.,

df
+ (1 + )f = 0
d

at

= 0 .

(9.68)

Due to the natural symmetry of the solution, f (x, ) must be even and satisfy the symmetry condition
df
(0, ) = 0.
d

(9.69)

Also, since temperature and its derivative must be continuous at every point including the inflection point x0 , the function f and its derivative df /d must be continuous at = 0 . With these
conditions, the boundary value problem (9.65) (9.69) can be solved exactly in terms of the confluent hypergeometric functions (see equations 3.42 3.45 of Barenblatt). This solution contains
an eigenvalue problem for , which in general has to be solved numerically.
- 86 -

MAE 294c. Methods in Applied Mechanics III

Instructor: Daniel M. Tartakovsky

When   1, a perturbation solution for = () is possible. This solution is similar to that
given in the book by Goldenfeld (p. 305). Consider the change of total mass in (9.49)
d
dM

dt
dt

Z
u(x, t)dx = 2D1

x0 (t)

2u
dx + 2D
x2

x0 (t)

2u
u
dx = 2D ,
2
x
x

(9.70)

where now u = u[x0 (t), t]. It shows that mass is not conserved! To compute the first-order
(in ) approximation of M (t) in (9.70), it is enough to obtain the zeroth-order approximation of
u[x0 (t), t]. The latter is derived from the late-time temperature distribution



x2
M (t)

exp
,
Dt  l
(9.71)
u(x, t) =
4Dt
2 Dt

Since x0 (t) = 2Dt, this gives


u
M (t)

[x0 (t), t] =
x
2Dt 2l

(9.72)

dM
M
.
=
dt
t 2l

(9.73)

and (9.70) becomes

Thus,

M (t) = Const t/

2l

(9.74)

On the other hand,


Z

M (t)

u(x, t)dx 2 Q

l
Dt


.

(9.75)

Hence
2
=
+ O(2 ).
2l

- 87 -

(9.76)

You might also like